Our fact base and framework for comparing fundamental upside and identifying winners across the AI value chain.

ARTIFICIAL INTELLIGENCE: THE AI INFRASTRUCTURE VALUE CHAIN

The AI boom versus bubble debate has raged on since 2023. The range of outcomes could scarcely be wider, from advanced superintelligence to a crash of epic scale. While it is clearly difficult to pinpoint the exact trajectory, we see AI spend remaining healthy for the foreseeable future.

However, we primarily focus on the more analyzable question: Where is AI infrastructure spend going, and who could benefit?

We break down the GB200 NVL72 data center bill of materials and estimate an all-in AI data center capex at $36Bn per GW. This spend is notably dominated by the GPU and by Nvidia gross profit dollars. Networking is the other big-ticket item, and there is clear upside for foundry, HBM, WFE, and mechanical and electrical equipment.

We further estimate incremental AI profit dollars as GWs * TAM/GW * Market Share * Incremental Margins. Based on this framework, we find that Ibiden, Unimicron, and other PCB and substrate names could have further upside, while Intel, Cisco, and server OEMs have a lower upside potential relative to their prominence in the debate.

The range of outcomes in AI could scarcely be wider: at one extreme, we could see a crash of epic proportions, wiping out potentially trillions of dollars of shareholder value. At the other, we could achieve advanced superintelligence and the obsolescence of the human race.

We hope for a golden mean in which AI boosts productivity, driving the next leg of economic growth and helping to develop technological solutions for many of the world's problems. However, we lack the power to meaningfully influence the outcome, or even to advance a high-conviction prediction. Instead, we focus on what we can do: help investors make money.

While we lack conviction in the long-term outcome, we see AI spend remaining healthy for the foreseeable future. The medium- to long-term outcome depends heavily on the scaling laws, which are not knowable, alongside building the next generation of frontier models and seeing how capable they are. However, we observe that AI believers dominate top-level decision-making at most key technology firms. Moreover, given the continued progress in model capability, we lack visibility on any downside catalyst that would cause decision-makers to change their views — suggesting AI spend will remain healthy.

We model an estimated AI data center capex at $36Bn per GW ($6Mn per rack). Data center capex is dominated by the graphics processing unit (GPU), which we estimate at 38% of total costs. Networking is the other big-ticket item at ~12% of spend. Storage is relatively small, while spending on mechanical and electrical equipment is significant but less concentrated.

Based on this analysis, we construct a framework for estimating AI upside across companies and sectors. At a high level, our framework is extremely simple: first, GWs of capacity coming online *TAM/GW* company market share = incremental revenue by company. Furthermore, incremental revenue * incremental margins = incremental profit dollars.

While these estimates of AI upside are imprecise, we find that PCB, substrate, GPU, ASIC, and electricals stocks could still have further upside. Beyond industry favorites such as NVIDIA and Broadcom, we find that Ibiden, Unimicron, GPU and ASIC names such as Advanced Micro Devices (AMD) and Mediatek (covered), as well as electrical names such as Eaton, could all see very large upside opportunities relative to current profit footprints.

While the US-China AI race remains a hot topic, the US is clearly adding more compute, and China is not close. However, this adds context to the US bans across AI and semicap.

SIGNIFICANT RESEARCH CONCLUSIONS 5 SURVEYING REVISIONS AND STOCK MOVES 17 A quick scan of how the GenAI boom has translated into fundamental revisions and stock moves COMPARING FRONTIER MODELS 35 Who is leading – and, more importantly, how industry-wide model progress is trending THE GPU DEPRECIATION DEBATE 49 Addressing one bear case: Why is it reasonable to depreciate GPUs over a six- to seven-year lifespan? THE DATA CENTER BILL OF MATERIALS 55 What actually goes into a GW of data center capacity? OUR AI UPSIDE FRAMEWORK 69 GPU, GPU components, and electrical names appear to have the most upside leverage to the theme US VERSUS CHINA COMPUTE CAPACITY ADDITIONS 83 The US-China AI race adds a geopolitical dimension, but the US is likely to remain ahead in the near-to-medium term

O - Outperform, M - Market-Perform, U - Underperform, NR - Not Rated, CS - Coverage Suspended

NVDA, AVGO, AMD, INTC, HUBB, ETN, IFX.GR, 6723.JP estimate is Adjusted EPS; NVDA, AVGO, AMD, INTC, HUBB, ETN, 6723.JP valuation is Adjusted P/E (x); NVDA base year is 2026; 4062.JP, 6146.JP, 6857.JP, 8035.JP, 6525.JP, 7735.JP base year is 2024; Source: Bloomberg, Bernstein estimates and analysis.

AI has been the biggest theme driving markets. It is hardly controversial to say that, since the November 2022 launch of ChatGPT and NVIDIA's subsequent beat and guide-up in May 2023, AI has been the single biggest theme driving public markets. Hyperscale capex guidance pointed to 400Bn+ in data center capex in 2025, amounting to ~20% of expected global GDP growth in 2025, and likely 600Bn+ by 2027E (see our report 4Q25 AI Server Pulse: joining the OpenAI club to keep the party going?). As a result, from May 2023 to February 2026, our rough basket of AI stocks $ ^{1} $ saw an aggregate ~9% revision to 2025 revenue, ~26% to EBIT and net income, and ~22% to FCF (Exhibit 1).

At the peak of every cycle, bulls argue that “this time will be different.” Long-tenured investors may have a sense of déjà vu. The tech industry has a rich history of overexuberance. Moreover, the tech industry is defined by massive cost deflation: Moore’s Law implies that semis and hardware computing costs should decline by a staggering 30% CAGR over time. AI is clearly not an exception: the GPUs that underpin the AI computing build-out have actually seen NVIDIA deliver a 1,000x performance increase from the K20X in 2012 to the H100 in 2022 (roughly 100% CAGR over 10 years). Progress is also continuing, with Rubin boasting a 5x performance increase in inference and 3.5x in training over Blackwell. The flip side of cost deflation is that it creates a strong incentive to match the pace of capacity build-outs to that of end-user adoption, often driving technology digestion cycles following an initial overbuild. In this case, there seems to be little doubt that the capacity build has outpaced end-user demand thus far. We previously estimated that 2024 enterprise end-user demand could be served with ~50k H100s (for more details see our report The Intelligence Revolution: How big is it? And what signposts are we looking for along the way?), against an estimated four million GPU shipments in that year (and 9.5 million expected for 2026). Accordingly, bubble fears have been around since the beginning (for our July 2023 take, see: AI Infrastructure: The build-out is huge. Bonanza or bubble?). It is not hard to imagine a scenario where the build-out outpaces end-user adoption, causing operators to pull back on investment until demand grows into their capacity.

But maybe this time is different. Although it might be true that AI capacity exceeds end-user demand, the capabilities of leading models continue to improve and are reaching human levels even on practical benchmarks such as OSWorld, which evaluates the ability of LLMs to use computers (Exhibit 2). Tech industry participants are increasingly debating over accelerationist scenarios, in which each generation of AI models is more and more capable of helping to build the next generation of models, allowing AI leaders pull further and further ahead of rivals, in turn driving a winner-takes-most outcome, justifying aggressive investment until AI adoption reaches critical mass levels. Accordingly, there is still ample demand to deploy additional capacity toward training, and there appears to be very limited idle capacity.

EXHIBIT 1: Aggregate revisions from May 23, 2023, to February 23, 2026, for affected stocks

Source: Bloomberg (as of February 23, 2026), Bernstein analysis

EXHIBIT 2: OSWorld Agentic Task Benchmark Performance by leading-edge models and select historical examples

Note: Models are a mix of agentic frameworks, specialized models, and generalist models. While agentic frameworks generally show best performance, and on-device agents will likely ultimately use such a framework, there is more risk that human-level performance on OSWorld does not translate to other activities for agentic frameworks and specialized models compared with general models. Salesforce, Alibaba, and Google are covered; others are privately held.

Source: OSWorld Benchmark, Bernstein analysis

The range of outcomes in AI is clearly wide. We model the enterprise AI TAM and arrive at a range of 600Bn in the bear case to 11Tn in the bull case. At the highest level, AI demand will primarily depend on two related questions, which both appear impossible to answer: first, how big is the productivity uplift from generative AI, and second, how will model capabilities scale with incremental computing power. There are clearly smart people on both sides of the debate, and as far as we can tell, there is no way to answer this question besides building the next generation of AI models and seeing how good they are. But the consequences are stark. In one extreme, if the next generation of models offers limited incremental capabilities, that could limit the appetite to invest in training new models, and the precedent set by DeepSeek $ ^{2} $ suggests that it is possible to drastically shrink inferencing workloads. In that scenario, supply could continue to ramp up even as demand declines, leading to a significant digestion cycle that wipes out massive amounts of shareholder capital and eventually settles into a steady state below current levels (when spending is needed to support the training of new models). In the opposite extreme, one could envision a scenario where AI models simply keep getting bigger and better until they achieve advanced superintelligence and replace humanity — in which case the AI industry will scale as fast as supply permits until reaching that point. As for what happens beyond that point, that is beyond the forecasting capability of mere humans like ourselves. At any rate, we anchor our estimates to Anthropic Economic Index's estimate that 30% of knowledge worker tasks are addressable by AI and attempt to model both scenarios, and arrive at a range of 600Bn in the bear case (Exhibit 3) to 11Tn in the bull case (Exhibit 4).

While it is difficult to pinpoint the exact trajectory, we see AI spend remaining healthy for the foreseeable future. We observe that AI believers dominate decision-making at most key tech firms. Moreover, even if the bear case materializes, we believe continued frontier model progress will likely support continued conviction in AI investments, and we lack visibility over downside catalysts that could cause tech decision-makers to change their views. On net, we see AI spend remaining healthy for the foreseeable future.

If we can't make a high-conviction call on the cycle, what can we do? We dedicate the rest of this Blackbook to the questions we can actually feasibly answer: where is AI infrastructure spending going and who could benefit. In doing so, we develop a framework for identifying which names have the most AI upside, although we also observe that if a digestion cycle happens, the same framework could be applied to identifying downside in a digestion cycle, or to pair trade names that have received relatively more or less credit for their fundamental exposure.

EXHIBIT 3: Bernstein AI infrastructure model – bear case

Note: Growth numbers are YoY, all other numbers are in $Bn unless otherwise specified. Red numbers are input assumptions; while green and black are Bernstein estimates. See online version for colors.

Source: Anthropic Economic Index, International Monetary Fund, Bernstein analysis and estimates

EXHIBIT 4: Bernstein AI infrastructure model – bull case

Note: Growth numbers are YoY, all other numbers are in $Bn unless otherwise specified. Red numbers are input assumptions; while green and black are Bernstein estimates. See online version for colors.

Source: Anthropic Economic Index, International Monetary Fund, Bernstein analysis and estimates

We estimate that a typical GB200 NVL72 rack $ ^{3} $ costs ~3.5Mn per rack. Coupled with ~2.5Mn in physical infrastructure costs per rack, this points to all-in AI data center capex of 6Mn per rack or 36Mn per GW. This estimate is notably lower than the 50-60Bn number given by NVIDIA on its Q2 FY26 earnings call — we believe NVIDIA is looking ahead to future product cycles (Exhibit 5 and Exhibit 6).

Data center capex is notably dominated by GPUs, which we estimate at 38% of total costs, and by NVIDIA gross profit dollars (32% of total costs). NVIDIA's ~70% gross margin implies that its gross profit dollars account for ~30% of total AI data center spending. Even with lower application-specific integrated circuit (ASIC) margins, AI ASICs would likely remain the largest cost item by a significant margin: assuming COGS is the same for GPUs versus ASICs but ASIC margins are 50%, this would take down the price of accelerated compute from ~$2.3Mn to ~$0.8Mn, a ~25% saving on total capex and still ~18% of ASIC data center capex.

Networking is the other big-ticket item at ~12% of spend, although networking spend is also more dispersed across different types of equipment. Storage is relatively small at <2% of spend.

How much foundry, high-bandwidth memory (HBM), and wafer fabrication equipment (WFE) suppliers capture varies significantly between GPU and ASICs. With GB200 NVL72, a foundry gets 2.5-3% of data center capex ($1.1Bn/GW) and ~1% more if CPU ($0.3Bn/GW) too. HBM gets 3-3.5% ($1.1Bn/GW). WFE gets 3-4% ($1.2Bn/GW). With ASIC, they get much more, as the same data center capex can buy more chips.

The spending on mechanical and electrical equipment is less concentrated, but major items include diesel and gas generators and turbines (~6% of spend), uninterruptible power supplies (~4%), and transformers (~5%). Thermal management is a relatively small part of spend (~4%) and remains split between air cooling and liquid cooling, although we expect spend to continue to shift toward liquid cooling. We model in-rack power components content will increase to 7-8x in Rubin Ultra in 2027 with the 800V HVDC design.

Given the shorter depreciation lifespan of IT hardware such as servers and networking compared with mechanical and electrical equipment or land and buildings, and given that operating costs are relatively low, the true economic costs are likely even more heavily weighted toward servers and networking compared with what cash capex would imply. Even at an elevated cost of $0.15/kWh, it costs ~$1.3Bn in electricity to run a GW of data center capacity for a year. Personnel costs are also negligible, with 20GW data centers reportedly operating with 8 to 10 people, costing $30-$80k per year each.

EXHIBIT 5: GB200 NVL72 estimated rack capex breakdown

Note: GPU ASP ($2,304k) is reported price (listed in blue); other numbers listed in black or green are Bernstein estimates triangulated based on third-party data and conversations with experts. See online version for colors.

Source: DRAMeXchange, expert conversations, company reports, literature search, Bernstein analysis and estimates

EXHIBIT 6: Data center infrastructure estimated capex breakdown

Note: Power usage per rack (in blue) is from SMCI technical user manual; other numbers (in green or black) are Bernstein estimates triangulated based on third-party data and conversations with experts. See online version for colors.

Source: Omdia, DRAMeXchange, expert conversations, company reports, literature search, Bernstein analysis and estimates

Based on this analysis, we construct a top-down framework for estimating 2025-27 AI upside across sectors. As a reminder, this is not how we are forecasting the companies in our coverage — the purpose is to serve as a top-down framework to simplify comparisons across names in very different sectors and to extend the comparison to companies not under coverage.

We extend the analysis to estimate market share and revenue. In addition to our estimated incremental TAM for each vertical, BloombergNEF's forecast of 16 GW in capacity in 2027E and our own estimates of market share give per company incremental revenue estimates as: 16 GW * TAM/GW * market share.

We further regress QoQ incremental revenue versus EBIT, and use the slope as an estimate of margins on incremental AI revenue, allowing us to estimate incremental profit dollars as: incremental revenue * incremental margins (Exhibit 7).

On a forward-looking basis, we find that Ibiden, Unimicron, and other PCB and substrate names could have very high torque to the upside, with Unimicron in particular likely to benefit from several large opportunities in ABF substrate and HDI. Ibiden could benefit the most from NVIDIA's AI substrate upgrade as content grows 2x generation to generation. Our analysis further finds that (beyond industry favorites such as NVIDIA and Broadcom), GPU and ASIC names such as AMD, Mediatek, and Marvell (not covered), as well as electrical names such as Eaton, could all see very large upside opportunities relative to current profit footprints.

Conversely, Intel and Cisco (not covered), as well as server OEMs (Dell and Hewlett Packard Enterprise, both covered) have relatively lower exposure relative to their prominence in the AI debate.

While these estimates of AI upside are imprecise, and valuations factor in a myriad of non-AI factors, a first-pass comparison of estimated AI upside to multiples would point to Unimicron having further room to run. Conversely, Intel and, to a lesser extent, Arista and Amphenol (both not covered), screen as expensive despite more moderate AI opportunities.

We acknowledge that the analysis is built on a static view of the GB200 cycle, and that some of the dislocations could be because of forward-looking trends that do not yet reflect in our estimates. In particular, we see significant upside to Delta on the in-rack power side, NVIDIA and Broadcom on content and backlog growth, MediaTek on the Tensor Processing Unit (TPU) ramp-up, as well as memory and storage players such as SanDisk, Samsung, Micron, SK Hynix, and Kioxia (all covered) due to rapid memory price surge and content growth in the Rubin cycle.

EXHIBIT 7: Two-year EBIT growth opportunity from AI data centers

Note: Values are in $Bn unless otherwise stated. Red numbers are input estimates from Bernstein coverage analysis; green numbers are estimates from published Bernstein industry models; all other numbers are Bernstein estimates derived from methodology discussed above. See online version for colors. NVIDIA, Broadcom, AMD, Intel, TSMC, Mediatek, Hynix, Micron, Samsung, SanDisk, Kioxia, Seagate, Western Digital, Quanta Computer, Supermicro, Dell, HPE, Unimicron, Ibiden, Delta, Eaton, ABB, Siemens, Caterpillar, Legrand, Cummins, and Quanta Services are covered by Bernstein; the rest are not covered.

Source: Bloomberg, Bernstein analysis and estimates (all)

The burgeoning US-China AI race adds a geopolitical dimension to the AI debate. In that comparison, it's often observed that the US has chips but no power, while China has power but no chips, raising the question of which country is actually bringing more compute online.

However, on closer scrutiny, the US is clearly adding more compute, and China is not particularly close. We estimate that the US and its allies added $ >25 \times 10^{21} $ floating point operations per second (ZFLOPS) in AI accelerated compute capacity (FP16 sparse) in 2025 versus China at $ <1 $ ZFLOPS (Exhibit 8 and Exhibit 9). The simple math here is that we estimate China shipped $ \sim1.5 $ million local AI chips in 2025. Using the Huawei Ascend 910B as a benchmark, the chip delivers $ 0.4 \times 10^{15} $ FP16 floating point operations per second (PFLOPS),

implying that China added 0.6 ZFLOPS in incremental compute capacity in 2025. In addition, some low-end chips from NVIDIA and AMD were shipped to China in 2025 (we estimate that to be another 0.2-0.3 ZFLOPS), but it's safe to say that China in total added <1 ZFLOPS in 2025. By contrast, 4 million NVIDIA Blackwell chips at 4.5 PFLOPS each would've added 18 ZFLOPS in compute capacity. Factoring in TPU and AI ASICs, the total should be at least 26 ZFLOPS.

Zooming out, despite the fact that China has more total power capacity added, the US added more data center capacity in 2025, both in terms of overall capacity and for AI specifically. China added far more total power than the US — >500GW equivalent in total power capacity in 2025 versus the US at ~30GW. For data centers, however, China added 3.9GW in 2024 versus the US at 5.3GW, and for AI specifically we believe the gap is even wider. Similarly, we observe that Chinese cloud service providers (CSPs) have been much more conservative in AI investments than the US — we estimate China AI capex at 20% of the US.

In a supply-constrained build-out, something will always be the bottleneck, and China's lack of chip capacity bottlenecks it at a much lower level of capacity addition than the US' power-constrained build-out. While China has a surprising amount of total chip wafer capacity at 30% of the global total, most of this is lagging edge, and the country lacks logic or especially advanced logic (7% of 7nm or less), with actual output share on advanced logic much lower than that due to lower yield and less advanced nodes.

China is adding capacity quite rapidly, but even if you expect 7nm equivalent Al chip shipments to grow at a 100% CAGR over 2025-30, that still points to China reaching 19 ZFLOPS in 2030, below the US today. That said, the further away from the end goal of advanced general intelligence (AGI) and the longer the AI race lasts, the more opportunities there will be for China to close the gap versus the US and its allies.

The big caveat is that, strictly speaking, compared with China the US actually has (mostly) neither chips nor power. It is US allies, especially South Korea and Taiwan, which have the chips at this point. This puts the US bans (across both AI and semicap) in the correct light (China has power to spare, so limiting its ability to buy or build AI seems appropriate), as well as justifying US onshoring efforts. Over the longer term, it also seems possible that the US pulling ahead in AI could further cause tensions, as China potentially feels pressured to take more aggressive actions (Taiwan?) or risk losing the race.

EXHIBIT 8: China versus NVIDIA AI compute additions in 2025: we believe China added <1 ZFLOP of compute, far below US competitors...

Note: Z, P, and T are standard metric prefixes (indicating $ 10^{21} $, $ 10^{15} $, and $ 10^{12} $, respectively). Performance data is from specifications listed on company websites, power usage is from third-party sources; all other data points are estimates.

Source: Company websites, channel checks, Bernstein analysis and estimates

EXHIBIT 9: ... who we believe have easily added >25 ZFLOPS in 2025, orders of magnitude more than what has been added in China

Note: When specs were only available for FP16 dense, assumed sparse is 2x dense. Specifications are either from company website product listings (in black) or estimated (in red/gray); shipments are Bernstein estimates (see online version for colors). Amazon, Meta, and Microsoft are covered.

Source: Company websites, channel checks, Bernstein Asia Semiconductors Team, Bernstein analysis and estimates

INVESTMENT IMPLICATIONS Broadcom (Outperform, price target $525): A strong 2025 AI trajectory seems set to accelerate into 2026, bolstered by software, cash deployment, and superb margins and FCF.

NVIDIA (Outperform, price target $300): The data center opportunity is enormous, and is still early, with material upside still possible.

AMD (Market-Perform, price target $235): AI expectations remain high, but a new deal with OpenAI has the prospect to drive further (possibly substantial) growth.

Intel (Market-Perform, price target $36): Intel's problems have broken through to the forefront, though President Trump's support is helpful.

Delta Electronics (Outperform, price target NT$1,830): We are positive on Delta's clearer AI product roadmap.

Chroma ATE (Outperform, price target NT$1,660): We view Chroma as a key structural beneficiary of rising power/thermal requirements in the AI era. We are also encouraged by its potential in CPO-related testers.

Unimicron Technology (Outperform, price target NT$610): We are encouraged by Unimicron's involvement in long-term projects such as EMIB-T and CPO, and by the visibility on substrate price hikes.

Quanta Computer (Underperform, price target NT$250): The buy-sell model will likely remain dominant through 2026. However, some customers might shift toward consignment to

reduce financing costs. Rising GB300 mix and potential Rubin shipments in 4Q26 will add to margin pressure.

Eaton Corporation (Outperform, price target $428): Eaton benefits from attractive multi-year end market growth, particularly in the data center space, where bottlenecks persist, demand for equipment is high, and backlog/orders remain healthy.

Hubbell Incorporated (Outperform, price target $553): Hubbell benefits from attractive multi-year end market growth, particularly in the data center space, where bottlenecks persist, demand for equipment is high, and backlog/orders remain healthy.

ASML (Outperform, price target €1,700 and $1,971): ASML benefits from the global capacity expansion for leading edge logic and memory to enable AI capacity build out.

Ibiden (Outperform, price target ¥9,200): Ibiden dominates GPU substrate supply and benefits from substrate area increase and the migration to EMIB-T.

Infineon (Outperform, price target €52): As a major supplier of AI power semi, Infineon sees its AI revenue doubling this year before growing another 67% next year to drive earnings growth.

Renesas (Outperform, price target ¥3,300): As a major supplier of AI semis across power, memory-interface, and clock/timing ICs, with exposure to both AI GPUs and ASICs, Renesas expects its AI revenue to double this year.

DISCO (Outperform, price target ¥70,800): DISCO is the dominant supplier of grinders and dicers, which are used for all kinds of advanced packaging used in AI.

Advantest (Market-Perform, price target ¥26,000): Advantest is the dominant supplier of high-end SoC testing equipment for HPC and AI.

Besi (Outperform, price target €195): Besi is the leading supplier of hybrid bounders, which enable advanced logic processors for HPC and AI.

Tokyo Electron (Outperform, price target ¥56,800): As the #4 WFE player, TEL provides many kinds of equipment, including track, etching and deposition. It also provides wafer probers for testing, as well as temporary bonders for HBM.

Kokusai (Outperform, price target ¥8,620): Focusing on a niche area of batch atomic layer deposition (ALD), Kokusai has the biggest memory exposure in our coverage, especially NAND.

Lasertec (Outperform, price target ¥41,000): We expect its actinic extreme ultraviolet (EUV) inspection penetration to enable advance logic progression.

Screen (Market-Perform, price target ¥21,600): With an exposure tilt toward advanced logic, we expect Screen to benefit from capacity expansion from TSMC (covered).

A quick scan of how the GenAI boom has translated into fundamental revisions and stock moves

ChatGPT's November 30, 2022, launch did not go unnoticed by investors (for initial Bernstein takes, see our reports: NVIDIA (NVDA): A bottoms-up approach to sizing the ChatGPT opportunity, NVIDIA (NVDA): Getting Smarter...Real opportunities for AI - ChatGPT (Part 2) - A bottoms-up approach to sizing LLM training, and US Internet Weekender: Will AI kill Google Search?, as well as Bob Brackett's unforgettable joint note with ChatGPT: Bernstein Energy & Power: Why I (and not ChatGPT) should be your trusted analyst). However, it wasn't until NVIDIA's massive beat and guide-up on its FQ1 24 earnings call on May 24, 2023 (see our report NVIDIA (NVDA): FQ124 Recap - The Big Bang), that it became clear that Generative AI would be one of the defining themes of the year. Two years on, enterprise CIOs are describing the focus on AI as "all-encompassing", influential voices are openly discussing the possibility of artificial superintelligence by 2027, $ ^{1} $ and AI capex is driving 45% of US GDP growth, according to Dell. $ ^{2} $ Against that backdrop, we zoom out and take stock of how estimates for AI-related stocks have shifted since May 2023 — and what it has meant for AI-exposed stocks.

Our analysis is based on consensus expectations for 2024-26 revenue, EBIT, net income, and FCF as of May 23, 2023 (before NVIDIA's FQ1 24), versus the latest data as of February 23, 2026, which we compare with the changes in market cap and enterprise value (EV). We run the analysis across a long list of US-listed stocks with potential AI exposure — which includes all hardware and semiconductor stocks, as well as a select list of industrial, energy, and utility Global Industry Classification Standard (GICS) industries and subindustries. See the Appendix at the end of this chapter for more details on the methodology.

Revisions across the AI value chain have been significant, and investors appear most willing to pay for revenue revisions. In aggregate, companies in our list of affected stocks have seen significant revisions, with an 8.9% revision to 2025 revenue, 25.6% to EBIT, 25.5% to net income, and 21.5% to FCF compared with what was expected in May 2023 — driving a 111% increase in EV from May 2023 to February 2026. We also observe that subsector-level price reactions appear most correlated for 25 revenue revisions — suggesting that this is what investors are most willing to pay for.

• Revisions across our universe of affected stocks have been significant. In aggregate, companies in our list of 137 relevant stocks have seen material revisions, with total 2025 numbers across all companies seeing an 8.9% increase in revenue, 25.6% in EBIT, 25.5% in net income, and 21.5% in FCF compared with expectations in May 2023 (Exhibit 1).

• Revisions have driven significant increases in market value. The revisions have helped drive a significant re-rating in market value, with total value across companies on the list seeing a 111% increase in EV terms and a 133% increase in market cap terms over May 23, 2023, to February 23, 2026 (Exhibit 2). This is in part driven by market multiple inflation, with the S&P P/E ratio up 39% from 19.0 to 26.6 and EV/revenue up 41%. However, even accounting for overall market multiple inflation, in-scope companies saw an 85% increase in aggregate P/E multiple, representing a ~46% increase in the relative multiple over this period.

• Investors appear to see much of the revisions as structural rather than cyclical, which is helping drive multiple re-rating. Consistent with intuition, there is a meaningful correlation between the magnitude of subsector-level revisions $ ^{3} $ and the shifts in value (Exhibit 3). We also observe that there is a positive correlation between sector-level revisions and the corresponding multiple (Exhibit 4 to Exhibit 6), which suggests that investors have seen much of the positive revisions as a structural acceleration in growth, as opposed to more cyclical sectors where positive revisions are often seen as more driven by cycle timing and are negatively correlated with the multiple (see Oil & Gas Storage and Transport, for example, Exhibit 7). This is especially notable for semiconductors, which have historically traded as a more cyclical sector (for more details, see our report U.S. Semiconductors & Semiconductor Capital Equipment - What are multiples and revisions doing as we enter the 2H?). This, in turn, suggests that despite continued debate around the risk of a digestion cycle, consensus remains willing to underwrite those risks.

• Investors appear most willing to pay for revenue revisions. Across subsectors, we observe that EV/revenue and revenue revisions seem to have a more consistent relationship compared with other multiples (Exhibit 8), while EV/EBIT multiples also appear to have strong correlations. This suggests that investors are most willing to pay for revenue acceleration, and are only starting to look ahead more to 2026 numbers. EV/FCF multiples appear to have a strong correlation for 2025, but not for 2026, which may suggest that some of the movement in 2026 FCF estimates is seen as noise, compared with trailing FCF being more real.

At the sector level, semiconductor stocks have been the biggest winners — which appears justified by more significant revisions. Semiconductor stocks have been the biggest sector-level winners, with aggregate EV up 277%, compared to the S&P at 62%, although electronic components are trailing closely behind at 274% from May 23, 2023, to February 23, 2026 (Exhibit 9). However, this appears justified by greater revisions, with semiconductor 2025 revenue estimates up 84% compared with the next highest sector, electronic components, up 44%. We also observe that electric component stocks (e.g., Amphenol and Corning, both not covered) have seen comparable multiple expansion despite more limited revisions. Hardware stocks have shown wide dispersion, which is consistent with the idiosyncratic and high-dispersion nature of product portfolios in the sector.

We re-run the analysis on a sector level in order to get better visibility to individual stock level moves (Exhibit 10 to Exhibit 20; screens of analyzed companies can be found in the Appendix at the end of this chapter). At the individual stock level, NVIDIA remains the standout winner. Supermicro (covered) has also received limited credit for outlier performance, but potentially justified by low margins and accounting concerns. Vertiv (not covered) also appears as an outlier within the industrials group: its revenue revisions stand out compared with peers, and while they appear modest compared with tech comps, revisions look healthier when factoring in margin expansion.

• NVIDIA appears to be the standout winner at a single-stock level — and has actually had multiple compression on an EBIT basis. NVIDIA has actually seen multiple compression on an EBIT basis (Exhibit 10), with the ~6x move in the stock driven by 4.7x revenue revisions (Exhibit 11) and higher-than-expected margins from May 23, 2023, to February 23, 2026; the EV/2025 EBIT multiple actually contracted 6%. Accordingly, despite the fact that the company was one of the highest-flying stocks during this period, the move actually looks relatively modest compared with the magnitude of revisions.

• Supermicro also has received little credit for outlier performance, but potentially justified by low margins and accounting concerns. Supermicro has been another outlier in terms of standout revenue performance, with revenue estimates up over 200% from May 23, 2023, to February 23, 2026 (Exhibit 13). However, its revenue growth has been margin-dilutive (Exhibit 12) and, coupled with concerns about accounting quality and the sustainability of growth, this potentially justifies its comparatively modest multiple re-rating. For more details, see the Bernstein IT Hardware initiation report IT Hardware: The Intelligence Revolution - Sector Initiation (Outperform - AAPL, DELL, STX, SNDK).

• Vertiv has been another standout, albeit with more contribution from margin expansion than tech comps. Vertiv has also seen outlier performance relative to its peer set, as the company has established itself as a critical player in the liquid cooling ecosystem. That said, while Vertiv's revenue revisions are more material than those of its peers (Exhibit 15), they appear more modest than tech comps. However, Vertiv has also seen significant margin expansion, making its re-rating appear more justifiable on an EV/EBIT basis than an EV/revenue basis.

Conversely, companies such as Broadcom and Pure Storage (not covered), as well as utilities names such as Vistra and AES, have seen significant multiple expansion despite more modest revisions. Likewise, despite being frequently featured in AI discussions, the data center real estate investment trusts (REITs) have seen surprisingly limited impact to numbers.

• Broadcom has seen a comparable increase in market cap to NVIDIA despite more limited revisions — which we see as justified. Within the semiconductor universe, Broadcom has seen a similar increase in market cap to NVIDIA (468% versus 476%), despite much smaller revisions (69% increase in 2025 FCF, versus 725% for NVIDIA). However, we see the move as justified by the fact that: (1) there is a timing difference in that much of Broadcom's inflection occurs in 2026-27 (especially given the Broadcom/OpenAI partnership; see our report Broadcom (AVGO): Joining the OpenAI 10GW club...), so revisions

in 2025 numbers do not reflect the incremental revenues; (2) Broadcom is a beneficiary of the ASIC build-out, which is a perceived downside risk for NVIDIA; and (3) there's an argument to be made that NVIDIA has seen too little re-rating, as opposed to Broadcom seeing too much.

• More broadly, semiconductors have seen a stark bifurcation into winners and losers. While semiconductor sector-level performance is very strong in terms of both revisions and re-rating, this is primarily driven by a very small number of winners. NVIDIA alone accounts for 97% of revenue revisions for 2025. Once we include a broader set of clear AI winners (for our purposes: NVIDIA, Broadcom, AMD, Micron, Marvell, Lam Research, AMAT, and KLA Corp$^{4}$), the winners have seen +101% revenue revision and have received a commensurate reward in the form of a 392% increase in EV. By contrast, the remaining names in the space have seen significant negative revisions (-19% for 2025 revenue), with >100% of the EV increase coming from multiple expansion (Exhibit 21). This picture generally holds when looking at 2026 revisions or at EBIT instead of 2025 revenue (Exhibit 22). Negative revisions across the losers were relatively broad-based, with 19 of the 23 companies in the set seeing negative 2025 revenue revisions.$^{5}$ Although Intel was the biggest loser in terms of absolute dollar revisions by a pretty significant margin, Microchip, Enphase, and Solaredge (all not covered) saw worse than -50% revisions and were the worst hit in percentage terms. That said, we note that the winners were generally larger and more important companies even in May 2023 (median market cap of $91Bn in May 2023 versus $16Bn for the losers); winners accounted for 66% of market cap in 2023 and 88% as of February 23, 2026, resulting in sector performance being more driven by the winners than the losers.

• Pure Storage has also seen a major re-rating despite limited revisions. Pure Storage also appears notable, given that the stock has tripled almost entirely on multiple re-ratings (Exhibit 12). However, bulls will likely argue that upside from Pure Storage's software licensing deals (notably with Meta) are not fully in the numbers, which is especially notable, given that these deals are likely to be much higher margin that Pure Storage's traditional business.

• Investors don't see oil tanker and LNG stocks as real AI plays. There is occasional discussion about LNG pipelines and oil tankers as AI plays (for an example, see our report US Gas: Trying not to get run over by the AI train…upgrading EQT from Underperform to Market-Perform), leading us to include these stocks in the analysis. However, investor consensus does not appear to see these companies as real AI plays, and their revisions show negative correlation with shifts in the multiple (Exhibit 7 and Exhibit 18), which is more characteristic of cycle-driven revisions than thematic re-rating.

• Data center REITs have seen surprisingly limited impact to numbers. Despite being frequently featured in AI discussions, the data center REITs (Digital Realty and Equinix, both covered) have seen surprisingly limited impact to numbers (Exhibit 19 and Exhibit 20). That said, Digital Realty has reported a record development pipeline, so this could reflect longer build-to-revenue cycles, with impacts coming further in the future, as opposed to truly not seeing material impact.

• Utilities has seen four companies with outsized multiple re-ratings (Vistra, Constellation, NRG Energy, and AES; all not covered), despite the fact that AES has had negative revisions and NRG is more middle of the pack. Much of this appears to be due to the fact that these players have significant nuclear power capacity — with bulls arguing that nuclear power will be critical to provide stable energy generation capacity for the AI buildout, they may conclude that these companies will benefit longer term, despite seeing limited impact to near-term numbers thus far.

EXHIBIT 1: Aggregate revisions from May 23, 2023, to February 23, 2026, for affected stocks

Source: Bloomberg (as of February 23, 2026), Bernstein analysis

EXHIBIT 2: Aggregate change in value from May 23, 2023, to February 23, 2026, across affected stocks

Source: Bloomberg (as of February 23, 2026), Bernstein analysis

EXHIBIT 3: Consensus EV change versus 2026 revenue revisions since May 23, 2023

Source: Bloomberg (as of February 23, 2026), Bernstein analysis

EXHIBIT 4: Consensus 2026 revenue revisions versus EV/2026 revenue multiple change since May 23, 2023

Source: Bloomberg (as of February 23, 2026), Bernstein analysis

EXHIBIT 5: Consensus 2025 revenue revisions versus EV/2025 revenue multiple change since May 23, 2023

Source: Bloomberg (as of February 23, 2026), Bernstein analysis

EXHIBIT 6: Consensus 2026 EBIT revisions versus EV/2026 EBIT multiple change since May 23, 2023

Source: Bloomberg (as of February 23, 2026), Bernstein analysis

EXHIBIT 7: Oil & Gas Storage & Transport: EV/2025 consensus revenue revisions versus multiple re-rating since May 23, 2023

Note: Oil & Gas Storage & Transport defined as 10102040 GICS subindustry. Cheniere is covered by Bernstein; all others are not covered. Source: Bloomberg (as of February 23, 2026), Bernstein analysis

EXHIBIT 8: Correlation between subsector level multiple re-ratings and revisions in the respective denominator since May 23, 2023

Note: Aggregates done based on GICS subindustry.

Source: Bloomberg (as of February 23, 2026), Bernstein analysis

EXHIBIT 9: EV change versus change in revenue and EBIT estimates since May 23, 2023

Source: Bloomberg (as of February 23, 2026), Bernstein analysis

EXHIBIT 10: Semiconductors: EV/2025 consensus EBIT revisions versus multiple re-rating since May 23, 2023

Note: 4530 GICS Industry Group used as definition of semiconductors. NVIDIA, Broadcom, Intel, AMD, Qualcomm, AMAT, Micron, Lam Research, and KLA are covered by Bernstein; Marvell, SiTime, ACM Research, Teradyne, and ON Semiconductor are not covered. Regetti (not covered) excluded as an outlier. Source: Bloomberg (as of February 23, 2026), Bernstein analysis

EXHIBIT 11: Semiconductors: EV/2025 consensus revenue revisions versus multiple re-rating since May 23, 2023

Note: 4530 GICS Industry Group used as definition of semiconductors. NVIDIA, Broadcom, Intel, AMD, Qualcomm, Lam Research, AMAT, Micron, and KLA are covered by Bernstein; Marvell, SiTime, ACM Research, Teradyne, and ON Semiconductor are not covered. Regetti (not covered) excluded as an outlier.

Source: Bloomberg (as of February 23, 2026), Bernstein analysis

EXHIBIT 12: Hardware: EV/2025 consensus EBIT revisions versus multiple re-rating since May 23, 2023

Note: 4520 GICS Industry Group used as definition of hardware; Western Digital excluded from analysis due to impact of SanDisk spin. Apple, Dell, and Supermicro are covered; IonQ, Arista, Amphenol, Viasat, and Pure Storage are not covered.

Source: Bloomberg (as of February 23, 2026), Bernstein analysis

EXHIBIT 13: Hardware: EV/2025 consensus revenue revisions versus multiple re-rating since May 23, 2023

Note: 4520 GICS Industry Group used as definition of hardware. Western Digital excluded from analysis due to impact of SanDisk spin. Source: Bloomberg (as of February 23, 2026), Bernstein analysis

EXHIBIT 14: Electrical Equipment and Engineering: EV/2025 consensus EBIT revisions versus multiple re-rating since May 23, 2023

Note: Engineering defined as 201030 GICS industry and Electrical Equipment defined as 201040 GICS. Quanta, Eaton, and Hubbell are covered by Bernstein. Vertiv, Emerson, Rockwell, and Fluence are not covered. Bloom Energy, NuScale, Plug Power, and Enovix (all not covered) excluded as outliers. Source: Bloomberg (as of February 23, 2026), Bernstein analysis

EXHIBIT 15: Electrical Equipment and Engineering: EV/2025 consensus revenue revisions versus multiple re-rating since May 23, 2023

Note: Engineering defined as 201030 GICS industry and Electrical Equipment defined as 201040 GICS. Quanta, Eaton, and Hubbell are covered by Bernstein. Vertiv, Emerson, Rockwell, and Fluence are not covered. Bloom Energy, NuScale, Plug Power, and Enovix (all not covered) excluded as outliers. Source: Bloomberg (as of February 23, 2026), Bernstein analysis

Note: Utilities defined as 60108050 and 60108030 GICS codes. All companies are not covered.

Source: Bloomberg (as of February 23, 2026), Bernstein analysis

EXHIBIT 17: Utilities: EV/2025 consensus revenue revisions versus multiple re-rating since May 23, 2023

Note: Utilities defined as 60108050 and 60108030 GICS codes. All companies are not covered.

Source: Bloomberg (as of February 23, 2026), Bernstein analysis

EXHIBIT 18: Oil & Gas Storage & Transport: EV/2025 consensus EBIT revisions versus multiple re-rating since May 23, 2023

Note: Oil & Gas Storage & Transport defined as 10102040 GICS subindustry. Cheniere is covered by Bernstein; all others are not covered. Source: Bloomberg (as of February 23, 2026), Bernstein analysis

EXHIBIT 19: Data Center REITs and Crypto Miner: EV/2025 consensus EBIT revisions versus multiple re-rating since May 23, 2023

Note: 60108030 and 60108050 GICS subindustries used as REITS; Riot Platforms was the only crypto miner considered. All companies are covered by Bernstein. Source: Bloomberg (as of February 23, 2026), Bernstein analysis

EXHIBIT 20: Data Center REITs and Crypto Miner: EV/2025 consensus revenue revisions versus multiple re-rating since May 23, 2023

Note: 60108030 and 60108050 GICS subindustries used as REITS; Riot Platforms was the only crypto miner considered. All companies are covered by Bernstein. Source: Bloomberg (as of February 23, 2026), Bernstein analysis

EXHIBIT 21: Semis winners and losers: EV/2025 consensus revenue revisions versus re-rating since May 23, 2023

Note: Winners defined as NVIDIA, Broadcom, AMD, Micron, Lam Research, Advanced Materials, and KLA Corp (all covered), as well as Marvell (not covered). Losers defined as other companies in 4530 GICS code. Aggregate re-rating calculated as change in EV minus aggregate revisions.

Source: Bloomberg (as of February 23, 2026), Bernstein analysis

EXHIBIT 22: Semis winners and losers: 2025 and 2026 revenue and EBIT revisions since May 23, 2023

Note: Winners defined as NVIDIA, Broadcom, AMD, Micron, Lam Research, Advanced Materials, and KLA Corp (all covered), as well as Marvell (not covered). Losers defined as other companies in 4530 GICS code. Aggregate re-rating calculated as change in EV minus aggregate revisions.

Source: Bloomberg (as of February 23, 2026), Bernstein analysis

• For each in-scope name, we collected:

• Revenue, EBIT, net income, and FCF for

• As of May 23, 2023, and February 23, 2026

• Revisions calculated as percentage of change in value for the same metric in the same time span between May 23, 2023, and February 23, 2026

• Multiple change computed from change in value and from revisions

• Analysis was conducted on US-listed stocks with at least >$2Bn in market cap that fall within the following GICS industry groups/industries/subindustries (selected because high-profile AI names fall within these categorizations):

• Hardware – GICS 4520 industry group

• Semiconductors – GICS 4530 industry group

• Construction & Engineering Firms — GICS 201030 Industry

• Electrical Equipment — GICS 201040 Industry

• Electric Utilities — GICS 55101010 Subindustry

• Independent Power Producers — GICS 55105010 Subindustry

• Renewable Energy — GICS 55105020 Subindustry

• Oil & Gas Storage and Transport – GICS 10102040 Subindustry

• Data Center/Tower REITs and Crypto Miners

• Data Center REITSs — GICS 60108050 Subindustry

• Telecom Tower REITs — GICS 60108030 Subindustry

• Crypto Miners: IREN Ltd, Core Scientific, Riot Platforms, MARA Holdings, CleanSpark (all covered), Applied Digital, Terawulf, Cipher Digital, and Bitfarms (all not covered), although ultimately only Riot is included in the analysis.

• Excluded all companies that didn't have 2026 revenue estimates on May 23, 2023

• Analysis is highly dependent on ordinary least squares (OLS) regression, which is very sensitive to outliers for a dataset with many outliers. While least absolute deviations may have been a better fit, it is not well-supported by Excel.

• The number of outliers also made it difficult to run analysis on the full list of stocks, which is why the analysis was either done on a sector basis or by looking at sector aggregates.

• 2027 was excluded because Bloomberg lacks 2027 estimates for some critical companies as of May 23, 2023.

• Several key companies were excluded from the data due to lack of estimates as of May 23, 2023, notably GE Vernova (not covered), which was not a standalone company at the time.

Who is leading – and, more importantly, how industry-wide model progress is trending

With Grok 4.1 on November 17, Gemini 3 Pro on November 18, Claude Opus 4.5 on November 24, and DeepSeek v3.2 on December 1, $ ^{1} $ we've had four models come out at the end of 2025 that each claim to be the leading frontier model. In this chapter, we scrutinize these claims and their implications for stocks in our coverage.

We see Gemini 3 Pro and Claude Opus 4.5 as neck and neck, and view DeepSeek's claim to leadership with more skepticism. All three model builders have published benchmarks showing themselves as the leader, in accordance with the long history of model builders cherry-picking favorable-looking benchmarks. While third-party benchmarking is still in progress, we believe early results show Gemini and Claude as neck and neck, and Grok 4.1 ahead of GPT-5.

• All four model builders have published benchmarks showing themselves as the leader (Exhibit 1 to Exhibit 4). Viewed uncritically, it would suggest that as the last of the four models to release, DeepSeek is the leader.

• However, model builders have a long history of cherry-picking benchmarks that look favorable, with DeepSeek's benchmarking appearing particularly questionable. We observe that model builders have a long history of cherry-picking benchmarks that look favorable for their models. DeepSeek's chosen benchmarks are particularly questionable, since Claude Opus 4.5 is not included in its benchmarking — and even then, DeepSeek acknowledges lagging behind in many of the agentic use cases even as it claims leadership on reasoning (Exhibit 4). To be fair, given that DeepSeek V3.2 is an open model, end users can easily run their own tests, making external benchmarking less important, and potentially explaining DeepSeek's questionable benchmarking as carelessness rather than data manipulation.

While third-party benchmarking is still in progress, we believe early results show Gemini and Claude as neck and neck – and Grok 4.1 ahead of GPT-5. Frustratingly, one of our preferred third-party benchmarks, OSWorld, omits several models. While OSWorld does show Claude Opus 4.5 as the leading general model (Exhibit 5), Gemini 3 Pro and DeepSeek are not included in its benchmark data (and neither is the base version of GPT-5 – only agentic frameworks built on GPT-5). That said, we note that LLMArena shows Gemini 3 Pro as the leading model, both overall and for text (Exhibit 6), with Claude Opus 4.5 close behind across a range of use cases (Exhibit 7). We also note that Grok 4.1 ranked ahead of Claude on the text

rankings and on the overall rankings (which appear to assign very high weight to text), giving it a stronger claim to leadership than OpenAI (private).

Zooming out, the bigger and more important takeaway is that the scaling laws are clearly not dead, for both pre- and post-training. This, in turn, could renew confidence for the AI labs and their financial backers to build out AI infrastructure as fast as they can.

• Model builders are reportedly still seeing strong results for pre-training. Despite a great deal of hand-wringing about the end of pre-training, catalyzed in part by OpenAI's struggles with Orion and with its subsequent training runs, competing model builders appear to continue to see strong results from pre-training. Commentary from our industry contacts suggests that Gemini 3 in particular had a pre-training run that was so successful that the team was leaking its impending success in late-September 2025, long before post-training had even finished. As a result, while model builders have generally moved away from disclosing parameter counts (other than DeepSeek, which has disclosed that v3.2 has 685 billion total parameters with ~37 billion parameters activated per token), we believe that model sizes continue to increase, with Gemini repeatedly described by industry participants as a "trillion-scale" model.

• Post-training is also still scaling. While reasoning model step counts appear to be stabilizing at around 100 steps, model builders have also increasingly emphasized the importance of further scaling post-training, with synthetic data and reinforcement learning (RL) in particular seeing increasing adoption.

• As model improvement continues, why shouldn’t model training spend continue as well? We have argued in the past that the pace of model improvement is the most important fundamental leading indicator in the space. Current models may not drive sufficient productivity to justify the pace of spend. However, we observe that technology leaders justify spending anyway in large part because they believe model improvement will continue — or even accelerate — in turn expanding the addressable use cases to cover a much wider opportunity than current models are capable of addressing. As models continue to improve — and addressable use cases continue to expand — we believe technology leaders will see this as validation for their belief in the scaling laws, which in turn could renew confidence for the AI labs to build out AI infrastructure as fast as possible. For more details, see the Bernstein IT Hardware team’s report The Intelligence Revolution: How big is it? And what signposts are we looking for along the way?

That said, DeepSeek's (and other open source models') continued ability to fast follow could renew concerns that if and when model improvement decelerates, fast followers building smaller models could quickly erode frontier model economics. However, given model improvement still appears strong, we believe it is too early to worry about what might happen when model improvement decelerates. While we are skeptical of DeepSeek's claim to be a true leader in the space, we acknowledge it is at least a fast follower, with results that are at least competitive with leading models despite having more constrained compute (more details in the chapter titled "US Versus China Compute Capacity Additions" of this Blackbook) and reportedly using a meaningfully smaller model. We wonder if these results may reinforce the concerns that, if model improvement does decelerate, fast followers may be able to close the gap.

versus leading labs, eroding model builder economics and driving the ecosystem into a race to the bottom (for thoughts about what a digestion might look like, see our report $ \underline{\text{AI Infrastructure: The build-out is huge. Bonanza or bubble?}} $). That said, we note that first, given second-tier models often depend on distilling frontier models, frontier model improvement deceleration may also drive deceleration in second-tier models as well, meaning that the capability gap between frontier and second-tier models may not actually shrink. Perhaps more importantly, we believe there is a real long-term risk to model builder economics from smaller distilled models. However, given model improvement still appears strong and, accordingly, spending is likely to continue, we believe it is too early to worry about what might happen when model improvement decelerates.

While industry-wide scaling laws appear healthy, OpenAI has now fallen from being the clear leader in the space to arguably not even being in third place. It is hard for external observers to have conviction on the exact issues, but leadership is clearly worried. Given the importance of the scaling laws to tech stocks (and to the overall economy), we believe investors should monitor incremental data points here.

• OpenAI has gone from being the clear leader in the space to arguably not even being third amid a series of failed pre-training runs and high-profile departures. OpenAI's launch of ChatGPT is what launched the whole LLM race (and propelled the company to become a household name). However, OpenAI's benchmark performance now lags all three of Gemini, Claude, and Grok (Exhibit 7). Our industry contacts have reported that OpenAI has failed three pre-training runs in a row, starting with Project Orion, which could be contributing to the company falling behind competitors that are still seeing significant pre-training scaling, and contributing to its greater emphasis on post-training versus competitors. With the bull case on the space depending heavily on continued model improvement, OpenAI lagging behind appears to be a serious risk.

What's wrong with OpenAI? If not data, could it just be poor execution amid an exodus of key talent? On one hand, the fact that Anthropic (and Grok) appear to be doing well suggests that the issues may be less of a data gap and could be more idiosyncratic, given the other two startups face similar data constraints. We also observe that OpenAI has seen a series of high-profile departures over the past few years. Most obviously, much of the core team on GPT-2 and GPT-3 left to found Anthropic, including GPT-2 and GPT-3 development lead, and Anthropic CEO Dario Amodei. However, even beyond the Anthropic team, former CTO and development lead on GPT-4/4o, Mira Murati, left to found Thinking Machines (private), OpenAI co-founder and Chief Scientist Ilya Sutskever has left to found Safe Superintelligence Inc (private), and the majority of the OpenAI o1 reasoning model team has been poached by Meta. One potential explanation is that OpenAI's brain drain has hobbled its ability to execute against its model roadmap — leading to the company falling behind in terms of performance.

• Maybe it's a data gap after all? In that case, wouldn't Google be the front-runner? However, Anthropic has also been more focused on coding and tool use, and its performance has shined on most related benchmarks. That suggests it is still possible that the issue is due to a data gap after all, and Anthropic has seen better performance due to specialization rather than superior engineering or execution — which could eventually suggest that Anthropic could eventually hit similar issues. For investors who land on this explanation, the natural conclusion is that Google — which has both a leading model and more access to data — should have a longer runway to scaling than any of the startup competitors, and should be considered the

front-runner in the space. However, we also note there is historical precedent for initial leaders in technology transitions ultimately falling behind.

• OpenAI still has a potential moat in the form of its large and apparently sticky consumer user base. OpenAI remains the most widely adopted AI platform, with the company disclosing that ChatGPT had 800 million monthly average users (MAUs) as of November 2025 (and some OpenAI investors disclosing 900 million in the same month) versus sources reporting 650 million MAUs for Gemini in October 2025. We note that consumer behavior is often quite sticky in technology, and OpenAI's large (and growing) user base potentially provides a moat and some breathing room as the company tries to right the ship in terms of product capability.

• OpenAI's leadership seems to be sounding the alarm... could it be the next Mistral? Concerns are intense enough that OpenAI CEO Sam Altman has reportedly declared a "code red" in an internal memo, de-prioritizing work on a range of initiatives, including advertising, AI agents for health and shopping, and a personal assistant, in order to focus on improving ChatGPT — both model performance and personal context memory. We note that while OpenAI is a leader in the AI race, industry participants generally see the industry as remaining quite early-stage. If OpenAI doesn't turn things around, it could go the way of other early players such as Mistral (not covered), which had a high-profile launch with prominent backers, but which has since descended into relative obscurity amid struggles to compete on core capabilities (and in fact now uses DeepSeek's models to inform its own, effectively ceding any claim to be at the true innovation frontier).

• Or perhaps these issues are simply transitory? With all that said, given the rapid pace of progress in the space — typically newer models outperform older ones, and it is worth acknowledging that part of the reason why GPT-5 underperforms competitor models is that the GPT-5 launch in August 2025 and even the GPT-5.1 launch on November 12, 2025, simply pre-date those models. ChatGPT 5.2 and 5.4 both saw subsequent improvements in performance — so it is also possible that all the hand-wringing around OpenAI was simply an overreaction to short-term issues that the company has already worked through.

None of the four models were trained on Blackwell, and only GPT-5.1 and Grok 4.1 were trained on GPUs, which has driven renewed concerns about the threat from ASIC and TPUs to NVIDIA's moat. However, we believe the defining theme is compute scarcity, and that both GPU and ASIC should thrive.

• Is NVIDIA's reign over? We note that Gemini 3 was trained on a combination of TPU v5 and v6, making it the first frontier model to completely bypass NVIDIA. Likewise, while Anthropic uses a combination of chips, our industry contacts believe Opus 4.5 is primarily trained on Trainium2 and TPUs, and Anthropic has taken pride in building out a relatively hardware-agnostic tech stack. While GPT-5.1 and Grok were trained on GPUs, our industry contacts believe both were trained on H100s rather than on Blackwells, further raising concerns about the erosion of NVIDIA's moat.

• However, we believe the GPU versus TPU versus ASIC debate misses the forest for the trees. The defining theme is compute scarcity, which is a positive for both GPUs and

ASIC competitors, and Blackwell-trained models should be released soon. We continue to believe that right now the real question should really be "is the opportunity in front of us still big, or is it not?" as (hopefully!) we are not yet in a mature, saturated market for AI hardware.

In other words, it's still the size of the pie that matters — if it's big, both GPUs and ASICs should thrive (and if it's not, they are both in trouble). For more details, see our report NVIDIA, Broadcom - What could you do with a TPU?.

• Google is the only one of the major model builders which is publicly listed, and Gemini's success is clearly directionally positive, although the focus is shifting from pure model performance to product adoption and monetization. While Gemini remains a small part of Google's financials relative to core search, Google's increasing strength in AI has helped dispel existential fears that it might be disrupted by model builders, and there's certainly some potential upside to numbers from monetizing Gemini and also TPUs. However, we believe there is a paradigm shift away from pure model performance toward product adoption and monetization. While Gemini 3 is one of the two most perforant models today, Gemini was still unable to consistently displace ChatGPT from the top of the app stores.

• For NVIDIA and Broadcom, as mentioned, we continue to see a stronger outlook on AI spending as positive for both, and believe that debating GPU versus ASICs is missing the point to an extent. We continue to be positive on the AI build-out, and accelerated compute continues to be the most mission-critical aspect of the build. Accordingly, as the two leading accelerated compute vendors, NVIDIA and Broadcom both stand to benefit from a stronger AI outlook. We would continue to own both (and NVIDIA in particular appears very attractive at current valuations, with the recent stagnation overblown in our opinion).

• AMD is probably the most dependent on OpenAI's success, given its limited traction with other AI labs. However, the impact of OpenAI's struggles on AMD are somewhat mixed, as OpenAI may feel pressured to spend its way out, which would be positive for all its suppliers, though we suspect it would also be more likely to pull back from AMD first should it ever decide to dial back its aspirations.

• More broadly, we continue to be fundamentally positive on the AI infrastructure theme. While sentiment has clearly turned, especially amid weakening macro, and it is hard to fight shifts in sentiment, we believe the fundamental outlook continues to look strong, and would see recent weakness as a buying opportunity for higher conviction longs. We believe announcements of incremental data center spending and capacity are likely, and could serve as a positive catalyst for exposed names. Exhibit 8 and Exhibit 9 summarize our framework for estimating the degree of exposure of different verticals to the theme, which is laid out in full detail in the chapter titled “The Data Center Bill of Materials” of this Blackbook. For more details on memory and upstream wafer fab equipment (WFE) suppliers in particular, see Bernstein’s Asia Semiconductor team’s report Asian Semis & Global Memory: Two findings, one simulation & one model for AI.

EXHIBIT 1: Grok 4.1: x.ai disclosed benchmarks

Source: x.ai

For details on our evaluation methodology please see deepmind.google/models/evals-methodology/gemini-3-pro

Source: Google DeepMind

EXHIBIT 3: Claude Opus 4.5: Anthropic disclosed benchmarks

EXHIBIT 4: DeepSeek v3.2: DeepSeek disclosed benchmarks

EXHIBIT 5: OSWorld Agentic Task Benchmark Performance by leading-edge models and select historical examples

Note: Models are a mix of agentic frameworks, specialized models, and generalist models. While agentic frameworks generally show best performance, and on-device agents will likely ultimately use such a framework, there is more risk that human-level performance on OSWorld does not translate to other activities for agentic frameworks and specialized models compared with general models. Salesforce, Alibaba, and Alphabet (Google) are covered; others are privately held. Source: OSWorld Benchmark (as of December 8, 2025), Bernstein analysis

EXHIBIT 6: AI model performance versus size

Note: Alphabet (Google), Meta, Amazon, Alibaba, and Tencent are covered; all others are privately held.

Source: LLM Arena (as of December 8, 2025), Bernstein analysis and estimates

Source: LLM Arena (as of December 8, 2025)

EXHIBIT 8: GB200 NVL72 estimated rack capex breakdown

Note: GPU ASP ($2,304k) is reported price (listed in blue); other numbers listed in black or green are Bernstein estimates triangulated based on third-party data and conversations with experts. See online version for colors.

Source: DRAMeXchange, expert conversations, company reports, literature search, Bernstein analysis and estimates

EXHIBIT 9: Data center infrastructure estimated capex breakdown

Note: Power usage per rack (in blue) is from SMCI technical user manual; other numbers (in green or black) are Bernstein estimates triangulated based on third-party data and conversations with experts. See online version for colors.

Source: Omdia, DRAMeXchange, expert conversations, company reports, literature search, Bernstein analysis and estimates

Addressing one bear case: Why is it reasonable to depreciate GPUs over a six- to seven-year lifespan?

Investor sentiment is growing increasingly cautious around the AI trade (as evidenced by surging credit-default-swap spreads on CoreWeave and Oracle $ ^{1} $ debt). Against this backdrop, investors have resurfaced the debate around whether GPUs can really run for six to seven years and, by extension, whether depreciation accounting (Exhibit 1) accurately reflects the true economics of owning and operating a GPU.

In short: Yes. GPUs can profitably run for ~6 years, and the depreciation accounting of most major hyperscalers is reasonable. We observe that cash costs of operating a GPU are very low compared with market prices for GPU rental which, in turn, means that the contribution margins of running old GPUs for longer are quite high. Even with meaningful improvements in price and performance with each GPU generation, vendors can make comfortable margins on five-year-old A100s; it's only when we look at seven-year old Volta GPUs that we start to reach cash breakeven. This, in turn, implies that a five- to six-year depreciation lifespan is reasonable. We suspect that the economics may become less favorable if compute demand softens, potentially leading to older GPUs being switched off even if they are still functional. However, we also believe that if compute demand softens materially, GPU depreciation accounting would be the least of investors' concerns.

• Old GPUs generally still function beyond ~6 years. Through our conversations with industry participants, we have usually received consistent feedback that GPUs generally still function at six to seven years or more. While there are some high-profile stories of GPUs burning out after ~6 months, most of these are more attributable to burn-in: as GPU power density increases with each generation, the power and thermal configurations needed to reliably run those GPUs evolve over time. Data center operators sometimes make configuration mistakes early on when working with new GPUs, resulting in some GPUs burning out. However, once GPUs are properly configured, they can last a surprisingly long time — our conversations with Lambda (private) highlighted six to seven years (for more details, see our report The AI Cloud Market: Making sense of the compute backdrop - a conversation with a Neocloud industry expert).

• Cash costs of operating GPUs for ~6 years are very low compared with the market prices of GPU rental, implying contribution margins are quite high. It is hard to be too precise on estimates, given significant variance on pricing between service providers (Exhibit 2) and sometimes between configurations within the same service provider. The cost side is even less transparent, with service providers offering little transparency into their cost structure. However, if we zoom out, this does not really matter to this debate — GPU rental

prices appear to be an order of magnitude higher than cash operating costs such as power and co-location costs (Exhibit 3), which remains true even if flexing assumptions on price and cost are within a reasonable range. This implies that the contribution margins of running old GPUs for longer are quite high, making it economically rational to run GPUs for longer.

• In a compute-constrained world, there is still ample demand for running A100s. Comments from industry participants suggest that many of the leading AI labs firmly believe that higher intelligence is an emergent property from training models on more data and with more compute, as evidenced by the huge jump in capabilities from GPT-2 to GPT-3 and the broader history of machine learning. With that context in mind, AI labs don't want to lose the AI race because they were outbid on compute, and there remains ample demand for any compute available, even older machines, with industry analysts highlighting that A100 capacity remains close to fully booked out. This relationship only appears to break down with Volta and older machines, in part because Volta's older architecture does not support the Brain Floating Point 16 (BF16) format and the price-performance ratio is too low to justify re-architecting software to still function on machines older than Ampere.

• On net, the fact that it is both possible and economically viable to run six-year-old GPUs makes six-year-old GPU depreciation appear reasonable.

While most vendors are power-constrained, and newer GPUs offer both better price-performance ratio and power-performance ratios, differing power densities and architectures mean it does not make sense to replace old GPUs with new ones to better utilize limited power, with new GPUs generally coming online in new data center builds. The exception that proves the rule seems to be crypto mining ASICs, which generate worse economics than AI data centers, and where we see examples of Bitcoin miners unplugging mining rigs and replacing them with GPUs as part of a broader pivot to AI (for an overview of the theme, see our report AI investment case for Bitcoin miners - IREN top pick, slide deck & key takeaways). Other than crypto miners, data center operators appear to favor keeping Ampere machines running as long as there is still demand, while trying to build new data center footprints to house newer machines — perhaps helping to fuel the demand for powered shells that can be brought online quickly.

One nuance is that GPUs probably lose more value after the first year than a linear six-year depreciation would imply, but they appear to retain value fairly well beyond that point. We observe that data center operators often lose a significant number of GPUs to burn-in, as configurations based on the previous generation of GPU may not be quite right for newer hardware and operators take time to figure out the configuration. Likewise, users often prefer to run demanding workloads such as AI training on the latest generation hardware, with older machines being relegated to running less performance-sensitive workloads. This is consistent with the observation that GPUs on the resale market often lose 20-30% of their list price after the first year, but value seems to be better retained after that point.

We also observe that, given the prevalence of long-term contracts, even if GPUs depreciate faster than companies are modeling, the cost may be borne by end users in the form of artificially high prices. For instance, if OpenAI signs a five-year contract for CoreWeave H100 capacity, even if the H100s are worth less than a five-year depreciation cycle would imply, it would be OpenAI that bears the economic cost. CoreWeave has stressed that the majority

of its capacity is locked up in long-term contracts, and we suspect the same is true of many other cloud service providers, which suggests that even if the depreciation lifespan is overstated (understating the real economic costs), it may be end customers that ultimately bear the cost.

Another observation is that, in contrast to memory and storage, accelerated compute does not appear to price as a commodity, with older GPUs commanding higher prices than price-performance parity would suggest. This suggests that end users are still running legacy workloads that they are unwilling or unable to run on more modern hardware, and that cloud vendors are upcharging for these captive workloads.

• Accelerated compute does not appear to price as a commodity... While it is often observed that memory and storage are commodity products, this is not true of accelerated compute. In a commodity product, you would expect the price of older GPUs to fall until price and performance are at parity with more modern products. We observe that this is in fact not the case, with older GPUs offering meaningfully lower price-performance parity than the modern B200 even at market prices (with the notable exception of the H200, Exhibit 4), and this appears to remain true if we look at single-vendor pricing for better comparability (Exhibit 5 and Exhibit 6).

• ... Likely because older GPUs are still needed to run legacy workloads, supporting pricing for old GPUs. We suspect that this is because much of the capacity of older GPUs is being occupied to run legacy workloads that aren't expensive or important enough to justify re-architecting workloads to run on more modern machines. This, in turn, helps support pricing for old GPUs.

While in theory this would be negative for GPU vendors, given it implies lower demand for GPU replacements, in practice much of the dynamic is driven by the fact that compute demand is so overwhelming that it still makes sense to keep running older, lower-efficiency hardware. Our main takeaway instead is that despite concerns that GPU lifespans are overstated, understating the real cost of GPU depreciation and flattering hyperscaler margin, the assumptions (and by extension, hyperscaler depreciation) appear more fair than bears fear.

EXHIBIT 1: Hyperscaler disclosures on depreciation lifespans for relevant categories

Note: Google, Amazon, Meta, Microsoft, Oracle, and Coreweave are covered; Nebius is not covered. Some entries are excluded because the company's disclosure category labels don't clearly align with our categorization. *Effective January 2025, the useful lives of certain servers and network assets are extended to 5.5 years. Source: Company reports, Bernstein analysis

EXHIBIT 2: GPU pricing differs significantly by vendor and by configuration, complicating much of the analysis

On-demand GPU pricing by service provider

Note: Alphabet (Google, GCP), Amazon (AWS), Microsoft (Azure), and CoreWeave are covered; Nebius is not covered; the rest are privately held.

Source: Product listings on company websites (as of November 14, 2025), Bernstein analysis

EXHIBIT 3: Cash costs are very low relative to pricing, implying contribution margins from operating GPUs for longer are high

Note: Assumed power at $0.1/KWh with GPU as 45% of total data center power consumption, as well as 60% utilization, $80 per month in colocation cost, and $1,000 per year in maintainance. Numbers in red/gray are assumptions, numbers in black are Bernstein estimates; see online version for colors.

Source: Product listings on company websites, GPU specs from Nvidia website, Bernstein analysis and estimates

EXHIBIT 4: GPU Rental Price Performance is significantly lower for older GPUs

Source: Product listings from company website, GPU specs from Nvidia website, Bernstein analysis and estimates

Source: Product listings on company website (as of November 14, 2025)

Source: Product listings on company website (as of November 14, 2025)

Based on a series of conversations with industry experts, supply chain checks, and third-party data on out-of-rack costs of land and physical infrastructure, we construct an estimate breaking down rack-level economics for an GB200 NVL72 AI data center. This chapter includes our analysis and key takeaways.

We estimate that a typical GB200 NVL72 rack costs ~$3.5Mn per rack. Coupled with physical infrastructure costs of ~$2.5Mn per rack, this points to an all-in AI data center capex of $6Mn per rack or $36Bn per GW.

• 3.5Mn per rack for GB200 NVL72. Based on commentary from industry experts, updated to reflect increasing prices for DRAM and NAND, $ ^{1} $ we estimate the cost of a GB200 NVL72 rack at ~3.5Mn for a "typical" rack (Exhibit 1).

• $36Bn per GW in total capex. We estimate 132kW in rack power consumption and 169kW in total power consumption per rack (i.e., the rack is 78% of data center power). This points to 5.9k racks per GW, in turn implying $21Bn in rack cost per GW. Industry contacts further pointed us to ~$15Bn in out-of-rack physical infrastructure costs (translating to $2.5Mn per rack), for an all-in data center capex of $6Mn per rack or $36Bn per GW (Exhibit 2).

• What about NVIDIA's comments? Is that just “Jensen optimism”? While the per GW opportunity for semis in general and NVIDIA specifically is clearly huge, this estimate is notably lower than the $50-60Bn/GW number given by NVIDIA on its Q2 2026 earnings call, and more consistent with comments from Broadcom and AMD of ~$15-20Bn in addressable opportunity per GW. Our industry contacts were explicit in calling out that they believed NVIDIA's number was too high, but we believe it is looking ahead to future GPU cycles when per rack and per GW costs will be higher.

Data center capex is notably dominated by the GPU, which we estimate at 38% of total costs, and by NVIDIA gross profit dollars (~32% of total costs).

• NVIDIA gross profit dollars are the single largest line item in the AI data center BOM. Given NVIDIA's ~75% gross margins, this implies that its gross profit dollars account for ~32% of total AI data center spending (Exhibit 3), about as much as the entirety of the mechanical and electrical equipment (~31%). Despite capturing a relatively small share of GPU economics, the

sheer size of GPU spend means that HBM, GPU silicon, and other GPU costs remain relatively sizable line items in the overall BOM.

• Even at lower ASIC margins, AI ASICs remain the largest cost item. Assuming COGS is the same for GPUs and ASIC but ASIC margins are 50%, this would reduce the price of accelerated compute from ~$2.3Mn to ~$0.8Mn, a ~25% saving on total capex and still ~18% of ASIC data center capex.

• Are CPUs still relevant? More or less... At face value, CPUs actually remain a fairly significant cost item (at 4% of total capex, it's largely in line with components such as switches that are more often argued to be AI winners). That said, given CPUs are a small share of spend relative to GPUs and CPUs, and GPUs are often bundled as superchips, CPUs are often seen as an add-on in the broader bundle.

Networking is the other big-ticket item at ~12% of spend, although networking spend is also more dispersed across different types of equipment. Switches are the largest piece of networking equipment (~3% total spend), but other scale-up networking such as copper cabling (~2%) and backplane connectors such as those supplied by Amphenol (~2%) as well as the scale-out networking infrastructure (e.g., network accelerators) are also significant.

• Switches are ~3% of spending, with relatively high margins for both networking vendors and switch silicon designers. While switches are not a massive cost item (likely 40-50% of scale-up networking, which in turn is 10-15% of rack costs), it was still fairly significant at ~3% of total spend. We also note that given networking vendors such as Arista earn relatively high margins (60%+) on the switch hardware, and switch silicon designers are also earning relatively high margins on the silicon (which is the biggest component of COGS), switching is potentially more significant as a share of the profit pool than it is as a share of end-user spend.

• DPUs (3%), copper cabling (2%), backplane connectors (2%), network interface controllers (NICs) (2%), and transceivers (1%) are all also significant portions of spend. Moreover, many of these categories have potential shifts in per rack content, especially as co-packaged optics gain adoption.

Even at elevated prices, storage is a fairly small share of spend at ~2%. One of our industry contacts observed that a typical rack had ~2PB of storage — at $0.01/GB for hard disk drives (HDDs), this would point even lower at ~$20k per rack, or ~0.6% of rack costs. Notably, this recontextualizes why data center operators may prioritize other factors such as power efficiency over storage price per GB, and why data center operators have been willing to swallow increasing memory storage prices even as price increases drove demand destruction in other sectors such as smartphones.

• 2PB of storage translates to ~0.6% of rack costs. One of our industry contacts observed that a typical rack had ~2PB of storage — at $0.01/GB for HDDs, this translates to ~$20k per rack, or ~0.6% of rack costs — which is actually lower than the industry experts' feedback, likely because some storage is on higher-cost NAND.

• Have we been thinking about the quad-level cell (QLC) flash debate the wrong way? Does the HDD versus NAND cost difference even matter? This also recontextualizes the widely debated Meta white paper $ ^{2} $ in which Meta argued for more widespread use of QLC NAND flash in data centers instead of HDDs. Even with a significantly higher per GB cost, the incremental cost is not very relevant compared with overall costs, and is perhaps justifiable, given (as Meta argued) QLC SSD is more power efficient than either HDD or triple-level cell (TLC) flash, which may be more important than the incremental cost savings in a power-constrained environment.

• What about indirect spend? That said, we observe that while direct AI data center spending on storage is relatively small, storage may also experience indirect benefits from AI. For instance, Dell and others have argued that enterprise AI adoption will require modernization of storage footprints which, in turn, could drive a stronger refresh cycle in enterprise storage.

• Further reading. For the prior US IT Hardware team's view on why storage is a relatively small part of spend, see our report Dell, HPE: Dimensioning the AI storage opportunity — Bonanza or wishful thinking?. For the current Bernstein US IT Hardware team's primer on the HDD industry following the initiation on IT Hardware in September 2025, see: Global Hard Disk Drives: It's HAMR time! A primer on the HDD industry and why STX is poised to reap outsized rewards

How much value foundries, HBM and semicap, and other upstream suppliers can capture varies significantly between GPU and ASIC.

• We estimate with GB200 the foundry opportunity at ~2.5-3% for GPU, the networking and the high-bandwidth memory (HBM) opportunity at 3-3.5%, and wafer fab equipment (WFE) at 3-4%. For the data centers deployed with GB200 NVL72, we estimate foundries receive 2.5-3% of the data center capex through GPU, networking chips, etc., and ~1% more if CPU is also fabbed by foundries. Memory suppliers receive 3-3.5% through HBM (regular DRAM is not included). WFE suppliers receive 3-4%. This is equivalent to foundries receiving $1.1Bn/GW, and another $0.3Bn/GW when CPU is also made by foundries. Memory companies will have $1.1Bn/GW through HBM and WFE suppliers will enjoy US $1.2Bn/GW (Exhibit 4).

• Lower price and margins and, thus, higher volumes imply that ASIC offers more favorable economics to foundries and other upstream suppliers. If a data center is deployed with servers based on ASIC, low margins demanded by ASIC vendors imply that more servers and hence more chips can be bought with the same amount of data center capex. Using the example mentioned earlier in this chapter where ASIC vendors demand 50% gross margin versus 70% enjoyed by NVIDIA, the data center capex can be lowered by ~16%, or equivalently ~19% more racks, servers, and chips can be purchased if the capex amount is unchanged. For the same reason, if ASIC is provided by Asian suppliers such as MediaTek or Alchip (not covered) which ask for gross margin below 50%, or GPU is from less demanding suppliers (AMD, and even Intel?), upstream suppliers such as foundry, memory,

and WFE suppliers will receive even more. The percentage numbers (as a percentage of the data center capex) above hence should be scaled accordingly based on the savings from adopting these suppliers.

• Power efficiency adds complexity. The per GW revenue estimates above, however, won't change if we simplistically assume GPU and ASIC differ in margins of suppliers but are equal in power efficiency. The reality is certainly more complex, and GPU and ASIC likely may optimize power efficiency for different use cases.

• This is a snapshot in time... but how might it evolve? We also believe the estimate here is subject to technology evolution and may change over time, as shown in the trend from A100 to B300 in Exhibit 5 for example. Our exercise here is based on a data center with GB200 NVL72, and should be adjusted to predict the future.

What about Lam's comment that it had an $8Bn opportunity per $100Bn in data center investment? Some may find 3-4% above of data center capex going to WFE low, as Lam noted $8Bn in WFE spending per $100Bn data center investment. We attribute the difference to the fact that Lam's number likely assumes data center deployments with a mix of chips from NVIDIA, AMD, and ASIC suppliers. Lam's numbers may also embed some future changes that our analysis based on GB200 NVL72 doesn't capture. We also observe that Lam's estimates are about right assuming that WFE spend is 20% of GPU spend (given GPUs are 40% of total spend). While the 20% number appears about right for many of the ASIC suppliers (Exhibit 5), it is high for NVIDIA, given its higher capital efficiency (and conversely, ASICs are typically lower than 40% of total spend).

The spending on mechanical and electrical equipment was less concentrated, but major items include diesel and gas generators and turbines (~6% of spend), uninterruptible power supplies (~4%), and transformers (~5%). Thermal management was a relatively small part of spend (~4%) and remained split between air cooling and liquid cooling, although we expect spend to continue to shift toward liquid cooling (Exhibit 2).

• Standby diesel and natural gas generators and turbines used for redundancy are the single largest piece of mechanical and electrical equipment at ~6% of spend. These products include generators produced by Caterpillar (covered), MTU (owned by Rolls-Royce, covered), Kohler (not covered), and now Generac (not covered), which entered the data center power generation market in April 2025 and started international deliveries in 2H25. While utility-scale power generation is out of scope of the analysis (since it does not sit within the data center and is instead reflected in power costs), they are increasingly in focus for the debate.

• Power distribution was collectively ~10% of spend across a number of verticals and vendors. Power distribution equipment makers include Eaton (covered), Schneider Electric (covered), Vertiv (not covered), ABB (covered), nVent (not covered), and Legrand (covered) to name a few. A larger repository of power distribution players can be found in our report Grid investment - who to play? A compendium of the 20 key electrical players. Increased electrical content is a key catalyst for players in this space. Eaton generates ~20% of

its revenues from data centers and distributed IT. In a traditional data center, Eaton can generate $1.2-$1.5Mn per MW in potential sales, representing 6-10% of total compute and infrastructure capex. In an AI data center, this opportunity can expand by 50% to $1.2-$2.9Mn in potential sales, up to 8% of total compute and infrastructure capex. For this reason, Eaton is currently investing $1.5Bn to add capacity in three-phase UPS, low- and medium-voltage assemblies, switchboards, and panelboards to satisfy incremental data center demand. It has also prioritized investing in modular construction via M&A (Fibrebond). Electrical component maker Hubbell (covered) has also entered the market through its acquisition of PCX.

• Thermal management was relatively smaller at 4%. Thermal management appears to be a relatively small part of spend (~4%) and remains split between air and liquid cooling. That said, we expect spend to continue to shift toward liquid cooling. Moreover, players in the space such as Vertiv appear more bullish on the opportunity than our analysis would suggest, with Vertiv citing a $2.5-3Mn per MW opportunity for traditional compute applications that rises by ~20% to $3-3.5Mn per MW for high-density compute applications.

Given the shorter depreciation lifespan of IT hardware such as servers and networking compared with mechanical and electrical equipment or land and buildings, and given that operating costs are relatively low, the true economic costs are likely even more heavily weighted toward servers and networking than what cash capex would imply. Even at an elevated cost of 0.15/kWh, it costs ~1.3Bn in electricity to run a GW of data center capacity for a year. Personnel costs are also negligible, with multi-GW data centers reportedly operating with 8-10 people costing 30-80k per year each. By contrast, even on a six-year depreciation cycle, $ ^{3} $ compute plus networking costs ~3.4Bn in annual depreciation. Given the relatively high cost of capex compared with ongoing operating costs, and the fact that hardware has a shorter depreciation lifespan than physical infrastructure (Exhibit 6), the true total cost of ownership (TCO) economics may actually be even more heavily weighted toward servers and networking than what cash capex would imply.

Going forward, incremental economics are likely to accrue to companies that pose bottlenecks or which have potential for content increases. Beyond our present day snapshot of data center economics, we observe that companies which serve as a bottleneck are likely to capture an increasing share of economic value, as undersupply drives increased pricing and margins. Content increases also serve as an opportunity for companies to capture disproportionate economics. Lastly, we believe companies that do not serve as bottlenecks but which hold crucial technology or capabilities can manage supply in response to demand.

• Utility-scale power is increasingly becoming a bottleneck to the space. Industry participants increasingly cite the availability of powered land with commitments from grid utilities as a bottleneck to the continued build-out. Utility-scale combined cycle gas turbine OEMs include GE Vernova (not covered), Mitsubishi Heavy Industries (not covered), and Siemens Energy (covered), which collectively own 70-80% of global production capacity.

Bloomberg consensus expects their collective capacity to grow from ~40GW in 2025 to ~70GW by 2027 to satisfy increased demand, particularly from data centers.

• Hyperscalers are looking for ways to circumvent this bottleneck. Due to ongoing capacity constraints, hyperscalers are increasingly seeking alternatives via smaller industrial gas turbines from players such as Caterpillar, which owns Solar Turbines, the world's largest manufacturer of industrial gas turbines. Caterpillar's power generation capabilities also extend into standby diesel and natural gas generators used for redundancy. Other players include Cummins (covered), Kohler (private), MTU, and now Generac (both not covered), which entered the market in April and started international deliveries in 2H25. Alternative energy sources such as small modular nuclear reactors (SMRs) are also being widely explored as potential alternatives.

• Could vendors with crucial capabilities and market cap take advantage? Possibly...

We observe that industry participants generally seem confident in 1-2 year demand visibility, likely because the pace of the AI build-out is being dictated by power constraints and so AI labs are matching their build-outs to the pace at which power becomes availability. While on the surface this would lend increased bargaining power to the power ecosystem, we wonder if vendors who own critical technologies, capabilities, or market power can take advantage by then matching the pace of supply additions to the growth in demand. In doing so, vendors such as TSMC or the memory vendors may be able to maintain a favorable demand-supply balance, helping to manage cyclicality in historically cyclical industries.

• There is a significant opportunity for content increases in in-rack power. We model in-rack power content for Vera Rubin (VR200) will be 2-3x that of the GB200 NVL72, and will increase to 7-8x in Rubin Ultra in 2027. The power content increase in GB300 versus GB200 NVL72 is limited due to similar PSU and battery backup unit (BBU) specs, while a higher adoption rate of BBU could increase the power TAM by mid-teens to $46k (Exhibit 7). While the power design of VR200 is not finalized yet, we expect power component content in terms of $ per rack to at least double to $80-$120k with higher power rate per rack and higher requirements for power stability (Exhibit 8), increasing the demand for BBU and super capacitors. The deployment of first-generation 800V HVDC (high-voltage direct current) is likely to start in 2H26. Applying that to Bernstein volume estimates for NVIDIA chips, we model power component TAM to increase by ~70% in 2026 and 30% in 2027 (Exhibit 9). More details can be found in our report Delta Electronics: sizing the upside from power rack; PT raised to NT$1130.

For reference: Exhibit 10 includes a map of an AI server supply chain. Exhibit 11 includes companies in the AI server hardware supply chain. Detailed discussions of competition landscape in each subsegment can be found in the following reports: Server ODM/OEM primer (2025) - Players, Positioning and Profitability, Delta Electronics: sizing the upside from power rack; PT raised to NT$1130, Liquid Cooling Primer - Delta Electronics poised to gain from the cooling revolution, and China Next Winners: Tech Hardware - Luxshare's AI play and China's dominance in optical transceiver & HDI.

EXHIBIT 1: GB200 NVL72 estimated rack capex breakdown

Note: GPU ASP ($2,304k) is reported price (listed in blue); other numbers listed in black or green are Bernstein estimates triangulated based on third-party data and conversations with experts. See online version for colors.

Source: DRAMeXchange, expert conversations, company reports, literature search, Bernstein analysis and estimates

EXHIBIT 2: Data center infrastructure estimated capex breakdown

Note: Power usage per rack (in blue) is from SMCI technical user manual; other numbers (in green or black) are Bernstein estimates triangulated based on third-party data and conversations with experts. See online version for colors.

Source: Omdia, DRAMeXchange, expert conversations, company reports, literature search, Bernstein analysis and estimates

EXHIBIT 3: We estimate that GBL NVL72 capex is dominated by GPU and, to a lesser extent, networking

Source: Omdia, expert conversations, company reports, literature search, Bernstein analysis and estimates

EXHIBIT 4: How much value foundries, HBM and semicap, and other upstream suppliers can capture varies significantly between GPU and ASIC

Estimated Value Share in Data Center Capex

Source: Bernstein analysis and estimates

EXHIBIT 5: We believe the WFE required to generate XPU value may change over time

Note: NVIDIA, AMD, Alphabet (Google), Amazon, Broadcom, and Mediatek are covered by Bernstein.

Source: Company websites and reports, literature search, Bernstein analysis and estimates (all)

EXHIBIT 6: Hyperscaler disclosures on depreciation lifespans for relevant categories

Note: Google, Amazon, Meta, Microsoft, Oracle, and Coreweave are covered; Nebius is not covered. Some entries are excluded because the company's disclosure category labels don't clearly align with our categorization. *Effective January 2025, the useful lives of certain servers and network assets are extended to 5.5 years. Source: Company reports, Bernstein analysis

EXHIBIT 7: We expect VR Ultra's power content to grow by 7.6x versus GB200

Note: BBU's content factored in adoption rate.

Source: Company reports, channel checks, literature search, Bernstein analysis and estimates (all)

EXHIBIT 8: Power component per rack in NVIDIA's AI server

Power content in Nvidia AI server racks

Source: Delta company reports, channel checks, literature search, Bernstein analysis and estimates (all)

EXHIBIT 9: NVIDIA's AI server power component TAM

Note: Bloomberg consensus in purple, Bernstein industry model inputs in green, and other Bernstein estimates in black; see online verison for colors.

Source: Bloomberg, company websites and reports, literature search, Bernstein analysis and estimates

EXHIBIT 10: AI server supply chain

Note: Bernstein covers Quanta, Delta Electronics, Unimicron, Chroma ATE, TSMC, Samsung, SK Hynix, Micron, NVIDIA, AMD, Broadcom, DISCO, Ibiden, and BE. Semi; the rest are not covered.

Source: NVIDIA, SK Hynix, Bernstein analysis

EXHIBIT 11: Summary of AI supply chain companies

Note: Other than Bernstein estimates for Quanta, all other data are Bloomberg consensus estimates. Bernstein covers Dell, Supermicro, HPE, Quanta, Luxshare, and Delta; the rest are not covered.

Source: Bloomberg (data as October 25, 2025), company websites, Bernstein analysis and estimates

GPU, GPU components, and electrical names appear to have the most upside leverage to the theme

In this chapter, we construct a top-down framework for estimating 2025-27 AI upside across sectors. As a reminder, this is not how we are forecasting the companies in our coverage — the purpose is to serve as a top-down framework to simplify comparisons across names in very different sectors and to extend the comparison to companies not under coverage.

Accordingly, while we sense check estimates against our forecasts for companies under coverage, the approximations in this analysis will inherently be more high level, and will fail to capture nuances that we address in our company-specific forecasts, including issues such as companies' ability to play across different verticals and content increases or decreases as AI data center form factors evolve. Moreover, there may be even more margin for error for companies not under coverage. With that disclaimer out of the way, our framework is available on request, and we encourage investors who disagree on specific inputs to enter their own assumptions in order to develop their own estimates.

In the chapter titled “The Data Center Bill of Materials” of this Blackbook, we broke down GB200 NVL72 data center capex. We estimate an all-in capex at $6Mn per rack or $36Bn per GW (Exhibit 1 and Exhibit 2).

For nine key hardware and semiconductor verticals, we extend the analysis using market share estimates to estimate incremental revenue per GW of capacity. In general, we leveraged past industry analysis where available. Where unavailable, we used overall revenue shares of the companies in the group as a proxy, and made new assumptions based on our covering analysts' expertise where those estimates appear inadequate. More specifically:

• GPU, ASIC and server, and rack. Given our estimate of $ <40\% $ share for ASICs, we allocated ASIC share as 20% to Broadcom, 5% to MediaTek, 5% to Marvell, and the rest to Alchip (not covered) and others. We acknowledge that this is a very rough approximation, as the revenue of GPU and ASIC suppliers is evolving, e.g., to extend from chip to rack, but sometimes to exclude HBM from chip revenue. For server and rack market share, we leveraged estimates from the Bernstein Asia Tech Hardware team's AI server market model (link to model here, $ ^{1} $ for more details see our report 4Q25 AI Server Pulse: joining the OpenAI club to keep the party going?), except for Dell and HPE, where we leveraged the AI server revenues they disclosed.

• CPU. The majority of CPUs are sold bundled with GPUs, e.g., the GB200 super chip, which includes two Blackwell GPUs and one Grace CPU. However, we assumed Intel might be able

to capture 15% of CPU share (ASICs at ~30% and Intel CPU at ~50% share among ASIC tech stacks).

• Memory. For memory market share, we leveraged estimates from the Bernstein Asia Semiconductor team's CoWoS/HBM (link $ \underline{\text{here}} $,² for more details see our report $ \underline{\text{Global}} $ Memory: Any upside left after the recent rally?) for HBM market share. We added it to their respective market share in conventional DRAM on a weighted average basis to reach their blended share in memory TAM.

• Storage. For storage, we estimated a 70% NAND to HDD revenue mix, which translates to a 13% NAND bit mix. We then leveraged estimates of NAND from the Bernstein Asia Semicondutors team's NAND industry model and HDD from the Bernstein US IT Hardware team's HDD industry model.

• Power and thermal. For the power and thermal names, we estimate Delta at 60%, followed by Flex at 20% and Lite-on at 10% (both not covered).

• Switches and backplanes. We assume that AI data center switch market shares are consistent with industry-wide revenue shares in data center switches, pointing to 26% for Arista and 22% for Cisco. While we lack good data on backplanes, given Amphenol competes with TE Connectivity, $ ^{3} $ as well as in-house builds from original design manufacturers (ODMs, e.g., Luxshare), we assumed it has at most a ~50% share. We also note that NVIDIA and Broadcom have significant AI networking businesses, and have adjusted their addressable markets to factor in the networking opportunity on top of their GPU and ASIC opportunity.

• PCB, substrate, and transceivers. For the remaining verticals, given our lack of industry data to serve as a basis, we used company-specific estimates where available and otherwise assumed that public companies had an AI market share proportional to their overall revenue share in order to build our high-level view.

• Foundry. Given the dominant position of TSMC and the dependence of key GPU and ASIC players on TSMC's CoWoS packaging, we assume TSMC had a 100% share of foundry revenue.

• Mechanical and electrical. We estimated that Eaton could capture 40% of the $3.5Bn power and distribution opportunity. We further estimated that Caterpillar and Cummins would capture 40% and 25%, respectively, of the diesel and gas generator and turbine opportunity.

We further regress QoQ incremental revenue versus EBIT, and use the slope as an estimate of margins on incremental AI revenue. For each company, we looked at QoQ incremental revenue and incremental EBIT of the past 32 quarters (i.e., CQ1 2018-CQ4 2025) for each. We then regressed incremental EBIT on incremental revenue, and used the slope of the regression as an estimate for incremental margins (i.e., for every incremental dollar of revenue, how many cents flow through to EBIT, Exhibit 3). Assuming that AI revenues come at comparable incremental margins to the existing business and that incremental margins are relatively consistent going forward, this suggests we can use historical incremental margins

as an estimate of incremental margins on go-forward AI revenue. We then adjusted estimates where appropriate based on our experience covering these names (Exhibit 4). Notably: (1) the GPU and ASIC names have disclosed that margins would be below historical averages, and we have accordingly adjusted down incremental margins compared with what the analysis would imply; (2) Unimicron's margins have benefited from cyclical recovery, and so we have estimated through cycle margins; (3) MediaTek margins should be impacted by the MediaTek TPU ramp-up starting in 2H26; (4) for memory and storage, the recent supply shortage and price surge could make the revenue and EBIT outlook very different from the 32 quarters of history analyzed by the regression; and (5) within mechanical and electrical, Eaton, Caterpillar, Cummins, and Quanta all have margin profiles that we needed to adjust significantly compared with what the analysis suggested.

Putting it all together: BloombergNEF's estimate of 16GW of incremental capacity in 2027E, multiplied by our estimate of incremental TAM per GW, gives an estimate of incremental TAM per vertical. Multiplying by our estimate of market share gives a simple estimate of incremental revenue, and furthermore multiplying by estimated incremental margins gives a rough estimate of incremental profit dollars each company could capture by 2027 (Exhibit 5).

• BloombergNEF estimates 16GW in incremental data center capacity additions. We note that assuming a 15% CAGR in global new data center capacity would point to 16GW in incremental capacity additions in 2027 (Exhibit 6). For more details on the implications for the onsite and standby power generation TAM, see our report Data Center Power Gen: How big is the onsite and standby TAM and what's different about CAT and CMI?.

• The (too) simple math: Incremental GWs * TAM/GW * Market Share * Incremental Margin = Profit. Coupling the estimate of 16GW of incremental capacity with our estimates of incremental TAM/GW gives an estimate of incremental TAM per vertical. Multiplying the TAM for each vertical with estimated market share for each company within each vertical gives a simple estimate of incremental revenue for each company. Furthermore, multiplying incremental revenue by estimated incremental margins gives a rough estimate of incremental profit dollars by company (Exhibit 5).

• The caveat: This is a high-level estimate that fails to capture nuances, and even then there is room to disagree on key input assumptions. As a reminder, this is not how we build our company-level forecasts — the top-down estimate fails to capture nuances that we attempt to capture in our company-specific forecasts, including issues such as companies' ability to play across different verticals and content increases or decreases as AI data center form factors evolve. However, we hope this framework is still helpful for investors in screening for potential AI winners that may be worth spending more time on — and we encourage investors who disagree on specific inputs to enter their own assumptions in order to develop their own estimates.

On a forward-looking basis, we find that Ibiden, Unimicron, and other PCB and substrate names could have very high torque to the upside, with Unimicron in particular benefiting from several large opportunities in ABF substrate and HDI. Our analysis further finds that in addition to the industry favorites (NVIDIA and Broadcom), GPU and ASIC names such as AMD, MediaTek, and Marvell, as well as electrical names such as Eaton, could all see very large upside opportunities.

• PCB and substrate players such as Unimicron could be beneficiaries. The PCB and substrate names do not fit as cleanly into our framework, given they have upside exposure across multiple parts of the BOM. However, our past work finds that Unimicron could see as much as NT$70Bn or $2.2Bn across its ABF substrate, HDI, and PCB opportunities. Coupled with our estimate of 25% incremental margins, this would imply that Unimicron's has an opportunity 4x that of its existing profit base (Exhibit 7) (for more details, see our report Asia Tech Hardware 2026 Outlook: AI drives growth, but risks demand selectivity).

• Within substrates, Ibiden is the dominant supplier for NVIDIA GPU substrate and captures substantial value from the upgrades. NVIDIA's transition from the Blackwell to the Rubin architecture meaningfully increases substrate intensity, with Rubin requiring double the substrate value content. Ibiden has regained its position as the sole supplier of ABF substrates for the Rubin generation, and we expect the company to maintain an 80%+ share over 2026-27 at least. Given the complexity and extremely tight pitch requirements of the Rubin substrate, we see limited risk of price erosion, as peers continue to face material yield disadvantages. We forecast Ibiden's revenue contribution from NVIDIA to reach JPY115Bn or $1.6Bn (27% of total revenue) in FY26/3E. Looking ahead, we expect revenue from NVIDIA to grow at a 44% CAGR over FY26/3E-FY28/3E, ultimately representing ~40% of Ibiden's total revenue in FY28/3E. Driven by NVIDIA demand and broader AI-related substrate applications, we forecast Ibiden to deliver an 18% revenue CAGR and a 37% operating profit CAGR over FY26/3E-FY28/3E.

• Going forward, the smaller GPU and ASIC players could see fairly high torque to the upside. GPUs and ASICs are the largest part of the TAM, and while NVIDIA and Broadcom have dominated the market thus far, we see increasing opportunities for other players to capture some of the incremental upside opportunity, especially given increasing adoption of TPUs (please see our deep-dives on TPU — MediaTek & ASIC: TPU Economy and Long View: TPU economy (part 2) & the long-term prospects of ASIC). This share gain opportunity, coupled with an opportunity for operating leverage, means that (in addition to industry favorites such as NVIDIA and Broadcom), smaller GPU and ASIC players such as AMD, MediaTek, and Marvell could see significant torque to the upside (Exhibit 5). This is because while substantially smaller on an absolute basis than their larger peers, they are also coming from a smaller baseline and hence with more room for incremental growth off it. In our coverage, we like the favorites (NVIDIA and Broadcom alike), but also see MediaTek as an attractive way to invest in the theme for investors who are bullish on incremental spend. This is especially true as we believe that recent demand increases for TPUs have not been reflected in consensus numbers and should drive upward revisions for MediaTek (for more details, see our report MediaTek: Near-term headwind but faster growth starting late 2026).

• The opportunity for electricals is real. Some members of the team had the initial hypothesis that the mechanical and electricals names have gotten more credit despite having less exposure than hardware and semiconductor names. While the electricals names do not have the highest exposure among the names in our analysis, the opportunity still appears quite real, and this analysis also does not account for potential upside on the grid power side. Accordingly, we believe investors should explore upside opportunities in the space, and we are particularly bullish on Eaton, given its strength in low- and medium-voltage power distribution. Increased electrical content is a key catalyst. Eaton generates ~20% of its revenues from data centers and distributed IT. In a traditional data center, Eaton can generate 1.2-1.5Mn per MW in potential sales, representing 6-10% of total compute and infrastructure capex. In an AI data center, this opportunity can expand by 50% to 1.2-2.9Mn in potential sales, up to 8% of total compute and infrastructure capex. After Eaton's recent entry into liquid cooling through its acquisition of Boyd Thermal, a 40%+ market share player (to better compete with Schneider's acquisition of Motivair and Vertiv $ ^{4} $), this is set to add an incremental 500k per MW in its total data center opportunity, raising its current outlook for AI data centers by 25% to 2.5Mn per MW at the midpoint. Eaton is currently investing 1.5Bn to add capacity in three-phase UPS, low- and medium-voltage assemblies, switchboards, and panelboards to fulfill incremental data center demand. It has also prioritized investing in modular construction via M&A (Fibrebond). Electrical component maker Hubbell has also entered the market through its acquisition of PCX.

Conversely, Intel and Cisco (not covered), as well as server OEMs (Dell and HPE, both covered) have lower exposure relative to their prominence in the AI debate.

• Low incremental margins for server OEMs limit their exposure. While the AI server opportunity is huge, and server OEMs (especially Dell and Supermicro) have managed to capture meaningful revenue, incremental margins have been much lower than in the traditional enterprise server business (though Dell has fared much better than SMCI), and are likely to remain low as NVIDIA continues to vertically integrate and take on a greater share of design (for more details, see our report The Intelligence Revolution: Can NVIDIA's vertical integration disintermediate OEMs? What's the impact to OEMs and ODMs?). Accordingly, the profit dollar upside has been more limited, though still driving a substantial portion of both Dell's and Supermicro's profit growth. While both names could see further upside from selling to enterprises (particularly Dell), our CIO survey suggests that enterprises have limited appetite to own their own AI servers thus far (Exhibit 8) and continue to expect AI workloads to be primarily hosted in the cloud (Exhibit 9) (for more details, see our report IT Hardware: December 2025 CIO survey results — Perspectives on spending levels, AI, PCs, and vendors). As a result, the margin upside story continues to be largely theoretical at this stage as opposed to having real visibility to a margin inflection. Regardless, due to the sheer size and rapid growth of the AI server TAM, Dell's absolute profit dollar growth from AI servers has been strong and consistent.

• Intel has struggled to capture share. Intel has struggled to capture a meaningful market share in the space and, accordingly, its revenue opportunity appears small relative to its historical core business in CPUs. While Intel could be a beneficiary of tighter CPU supply,

especially if CPU intensity in AI workloads increases, the direct opportunity in AI data centers appears relatively modest versus what peers are likely to experience.

While these estimates of AI upside are imprecise, and valuations factor in a myriad of non-AI factors, a first-pass comparison of estimated AI upside to multiples would point to Unimicron having further room to run. Conversely, Intel and, to a lesser extent, Arista and Amphenol (both not covered), screen as expensive despite more moderate AI opportunities (Exhibit 10). Gigabyte and Wiwynn (both not covered) screen as being relatively inexpensive ways to invest in potential AI upside, although we acknowledge that, like Dell and Supermicro, they could face margin pressure from NVIDIA's form factor standardization. That said, we highlight that many of these valuations reflect other issues rather than being true valuation dislocations. For instance, given Intel's narrow expected margins, its multiple could move significantly if it is able to realize margin upside from either improving CPU cyclicality or from an idiosyncratic turnaround.

We acknowledge that the analysis is built on a static view of the GB200 NVL72 cycle, and that some of the dislocations could be because of forward-looking trends that do not yet reflect in our estimates. In particular, we see significant potential upside for: (1) Delta on the in-rack power side for content increases in the Vera Rubin cycle (for more details, see our report Delta Electronics: sizing the upside from power rack; PT raised to NT$1130), (2) MediaTek on the TPU ramp, (3) memory and storage players such as SanDisk, SK Hynix, Micron, and Kioxia due to rapid memory price surge (see our report Global Memory: Any upside left after the recent rally?), and content increases in the Vera Rubin cycle (for more details, see our report Global Technology: What happened in Vegas? Takeaways from management meetings at CES and US IT Hardware 2026 Outlook: Winners and losers in the AI driven data explosion - Top picks SNDK (TP $580) and STX (TP $370)), and (4) likely increases to order and backlog dynamics in upcoming product cycles for both NVIDIA (see our report NVIDIA (NVDA): It's back to work for us, starting with a Jensen CES keynote and analyst Q&A) and Broadcom (see our report Broadcom (AVGO): Vegas baby...Takeaways from a CES investor meeting with the Semiconductor Solutions Group President).

EXHIBIT 1: GB200 NVL72 estimated rack capex breakdown

Note: GPU ASP ($2,304k) is reported price (listed in blue); other numbers listed in black or green are Bernstein estimates triangulated based on third-party data and conversations with experts. See online version for colors.

Source: DRAMeXchange, expert conversations, company reports, literature search, Bernstein analysis and estimates

EXHIBIT 2: Data center infrastructure estimated capex breakdown

Note: Power usage per rack (in blue) is from SMCI technical user manual; other numbers (in green or black) are Bernstein estimates triangulated based on third-party data and conversations with experts. See online version for colors.

Source: Omdia, DRAMeXchange, expert conversations, company reports, literature search, Bernstein analysis and estimates

EXHIBIT 3: Micron QoQ incremental revenue versus EBIT

Source: Bloomberg, Bernstein analysis

EXHIBIT 4: Incremental margins implied by regression slopes

Note: All slopes and correlations are regression estimates; for incremental margin assumptions, estimates in black are taking regression slope values directly, while estimates in red are Bernstein coverage analyst estimates. See online version for colors. NVIDIA, Broadcom, AMD, Intel, TSMC, Mediatek, Hynix, Micron, Samsung, SanDisk, Kioxia, Seagate, Western Digital, Quanta Computer, Supermicro, Dell, HPE, Unimicron, Ibiden, Delta, Eaton, ABB, Siemens, Caterpillar, Legrand, Cummins, Quanta Services, Trimble, Hubbell, KLA, AMAT, and Lam Research are covered by Bernstein; the rest are not covered.

Source: Bloomberg, Bernstein analysis and estimates (all)

EXHIBIT 5: Two-year EBIT growth opportunity from AI data centers

Note: Values are in $Bn unless otherwise stated. Numbers in red are input estimates from Bernstein coverage analysis, numbers in green are estimates from published Bernstein industry models, all other numbers are Bernstein estimates derived from methodology discussed above; see online version for colors. NVIDIA, Broadcom, AMD, Intel, TSMC, Mediatek, Hynix, Micron, Samsung, SanDisk, Kioxia, Seagate, Western Digital, Quanta Computer, Supermicro, Dell, HPE, Unimicron, Ibiden, Delta, Eaton, ABB, Siemens, Caterpillar, Legrand, Cummins, and Quanta Services are covered by Bernstein; the rest are not covered.

Source: Bloomberg, Bernstein analysis and estimates (all)

EXHIBIT 6: Incremental global data center capacity additions

Source: BloombergNEF (all data), Bernstein analysis

EXHIBIT 7: We project Unimicron's AI revenue mix to increase to 40% by 2027

Unimicron AI chip Related Revenue (NT$ B)

Unimicron ABF Revenue for Nvidia Blackwell & Rubin GPU

Source: Company reports, Bernstein analysis and estimates

EXHIBIT 8: Weighted percentage of CIOs planning increase or decrease in IT spending by category

Source: Bernstein CIO Surveys (published September 18, 2025)

Source: Bernstein CIO Surveys (published September 18, 2025)

EXHIBIT 10: Forward P/E multiple versus estimated two-year EBIT growth from AI

Note: NVIDIA, Broadcom, AMD, Intel, TSMC, Mediatek, Hynix, Micron, Samsung, SanDisk, Kioxia, Seagate, Western Digital, Quanta Computer, Supermicro, Dell, HPE, Unimicron, Ibiden, Delta, Eaton, ABB, Siemens, Caterpillar, Legrand, Cummins, and Quanta Services are covered by Bernstein; the rest are not covered.

Source: Bloomberg (as of January 26, 2026), Bernstein analysis and estimates

The US-China AI race adds a geopolitical dimension, but the US is likely to remain ahead in the near-to-medium term

When comparing the US-China AI race, it's often observed that the US has chips but no power, while China has power but no chips, raising the question of which country is actually bringing more compute online.

However, on closer scrutiny, the US is clearly adding more compute, and China is not particularly close. We estimate that the US and its allies added >25 x 1021 ZFLOPS in AI accelerated compute capacity (FP16 sparse) in 2025 versus China at <1 ZFLOPS. The simple math here is that we estimate that China shipped ~1.5 million local AI chips in 2025. Using the Huawei Ascend 910B as a benchmark, the chip delivers 0.4 x 10^{15} FP16 FLOPS, implying that China added 0.6 ZFLOPS in incremental compute capacity in 2025. In addition, some low-end chips from NVIDIA and AMD were shipped to China in 2025 (we estimate that to be another 0.2-0.3 ZFLOPS), but it's safe to say that China in total added <1 ZFLOPS in 2025. By contrast, four million NVIDIA Blackwell chips at 4.5 PFLOPS each would've added 18 ZFLOPS in compute capacity (Exhibit 1). Factoring in TPU and AI ASICs, the total should be at least 26 ZFLOPS (Exhibit 2).

Zooming out, despite the fact that China has more total power capacity added, the US added more data center capacity, both in terms of overall capacity and for AI specifically in 2025.

• China may have more total power addition... but it has not translated to higher data center capacity addition. China added far more total power than the US: >500 GW equivalent in total power capacity in 2025 versus the US at ~30GW (Exhibit 3) (for more details, see our report Electric China: If power is the bottleneck to AI, is China winning?). However, for data centers, China added only 3.9GW in 2024 versus the US at 5.3GW (Exhibit 4). Coupled with the fact that leading Chinese chips lag leading US chips on power and performance (Exhibit 1), this translates to a pretty large gap in terms of real capacity.

• Chinese CSPs have been more conservative in spending than US peers. Similarly, we observe that Chinese CSPs have been much more conservative in AI investments than the US — we estimate China AI capex at ~$90Bn in 2025 (Exhibit 5), or ~20% of the US and Europe hyperscaler and neocloud capex of ~$400Bn (Exhibit 6) (for more details, see the report 4Q25

AI Server Pulse: joining the OpenAI club to keep the party going?, and our analysis on China China AI: Sizing the AI chip supply and demand in China).

In a supply-constrained build-out, something will always be the bottleneck, and China's lack of leading-edge chip capacity bottlenecks it at a much lower level of capacity addition than the US' power-constrained build-out.

• While China has a surprising amount of total chip wafer capacity, it lacks leading edge or logic capacity. While China accounts for a surprisingly high share of total semiconductor wafer fab capacity at ~30% of the global total (Exhibit 7), this is largely driven by lagging edge, analog, and discretes. Looking at logic alone, its share is much smaller at ~20% (Exhibit 8). Even then, China has very little of the capacity that matters most. Data center accelerators are generally leading edge minus one. For instance, NVIDIA Blackwell is made on 4NP (4nm class), behind the leading edge 3nm node. However, looking at 7nm or less capacity, China has only 7% (and >50% sits in Taiwan, Exhibit 9).

• The picture is likely even more challenging for China than the data suggests. The data fails to account for capacity and yield issues resulting from the fact that most of China's 7nm or less capacity is believed to be 7nm capacity made using multi-patterning on 14nm machines due to the impact of current export controls on semiconductor equipment. In addition, most of the current 7nm capacity has been allocated to mobile chips rather than Al chips. Using wafer share is also not a fair comparison, as TSMC is already producing 2nm chips, while China is stuck at 7nm production, with far less transistor density, and lower computing performance per transistor.

• Accordingly, we believe that the gap in terms of AI chip output is even wider than total logic foundry capacity would suggest, to the point that despite China having more total power available than the US, it is actually adding less data center power load because it does not have sufficient chips to utilize that power.

China is adding capacity quite rapidly, but even if you expect 7nm-equivalent Al chip shipments to grow at a 100% CAGR over 2025-30, that would still point to China reaching 19 ZFLOPS in 2030, well below the US today. That said, the further away from the end goal of AGI and the longer the AI race lasts, the more opportunities there will be for China to close the gap versus the US and its allies. Anecdotally, Chinese decision-makers appear a lot more cautious about the prospect of AGI in the next few years compared with the US AI labs. Besides supply constraints, we also wonder if the more measured pace of the compute build-out in China may reflect that view. Chinese technology firms may be more willing to accept some lag in terms of frontier model performance in order to leverage dramatic cost deflation and build "good enough" models at a fraction of the compute and dollar cost (we all remember the DeepSeek meltdown in January 2025).

The big caveat is that, strictly speaking, compared with China, the US actually has (mostly) neither chips nor power. It is US allies, especially South Korea and Taiwan, which have the chips at this point, which puts the US bans (across both AI and semicap) in the correct light (China has power to spare, so limiting its ability to buy AI or build it seems appropriate), as well as justifies the recent uptick in onshoring efforts in the US (including NVIDIA starting to

manufacture some Blackwell chips in Arizona). Over the longer term, it also seems possible that the US pulling ahead of China in AI could further cause tensions, as China potentially feels more pressure to take more aggressive actions (Taiwan?) or risk losing the AI race.

EXHIBIT 1: China versus NVIDIA AI compute additions: We believe China added <1 ZFLOPS of compute in 2025, far below US competitors...

Note: Z, P, and T are standard metric prefixes (indicating $ 10^{21} $, $ 10^{15} $, and $ 10^{12} $, respectively). Performance data is from specifications listed on company website, power usage is from third-party sources; all other data points are estimates.

Source: Company websites, channel checks, Bernstein analysis and estimates

EXHIBIT 2: ... who we believe have easily added >25 ZFLOPS in 2025, orders of magnitude more than what has been added in China

Note: When specs were only available for FP16 dense, assumed sparse is 2x dense. Specifications are either from company website product listings (in black) or estimated (in red/gray); shipments are Bernstein estimates. See online version for colors.

Source: Company websites, channel checks, Bernstein Asia Semiconductors Team, Bernstein analysis and estimates

EXHIBIT 3: Annual power capacity additions in major countries: With China on track to add over 500GW of solar and wind capacity in 2025, no other country can add as much power supply

Source: Government data, Bernstein analysis and estimates (2025E)

EXHIBIT 4: US versus China: Data center power load additions as a percentage of global total

Source: BloombergNEF (all data), Bernstein analysis

EXHIBIT 5: We estimate total AI capex in China grew by 53% in 2025, reaching $91Bn...

Note: Alibaba and Tencent are covered by Bernstein; others are either private or not covered.

Source: Company reports, Bernstein analysis and estimates

EXHIBIT 6: ... but still only a fraction of US and Europe hyperscalar and neocloud capex

Source: Bloomberg (estimates), company reports, Bernstein Asia Hardware Team

EXHIBIT 7: China has ~30% of installed semi fab capacity versus the combined total of Taiwan, South Korea, Japan, and the US at ~60%

Source: SEMI World Fab Forecast, Bernstein analysis

EXHIBIT 8: For logic only, China has <20% of installed semi fab capacity versus the combined total of Taiwan, South Korea, Japan, and the US at >70%

Source: SEMI World Fab Forecast, Bernstein analysis

EXHIBIT 9: Most of the China capacity footprint is lagging (>14nm); for 7nm or smaller, >50% sits in Taiwan, and China has only 7% (and that capacity is itself likely impaired, given the state of current export controls)

Installed Semiconductor Capacity by Market and Node (Million 200mm eq WSPM) 2024

Source: SEMI World Fab Forecast, Bernstein analysis

References to "Bernstein" or the "Firm" in these disclosures relate to the following entities: Bernstein Institutional Services LLC (April 1, 2024 onwards), Sanford C. Bernstein & Co., LLC (pre April 1, 2024), Bernstein Autonomous LLP, BSG France S.A. (April 1, 2024 onwards), Sanford C. Bernstein (Hong Kong) Limited 盛博香港有限公司, Sanford C. Bernstein (Canada) Limited, Sanford C. Bernstein (India) Private Limited (SEBI registration no. INH000006378), Sanford C. Bernstein (Singapore) Private Limited, Sanford C. Bernstein Japan KK (サンフォード・C・バーンスタイン株式会社) and analysts employed by Société Générale Africa Technologies & Services to produce Bernstein research under a Global Services Agreement in place between Bernstein and Société Générale.

Bernstein is part of a joint venture between Société Générale (SG) and AllianceBernstein, L.P. (AB). Unless specifically noted otherwise, for purposes of these disclosures, references to Bernstein's "affiliates" relate to both SG and AB and their respective affiliates.

This research publication covers six or more companies. For valuation methodology and other company disclosures: Please visit: https://bernstein-autonomous.bluematrix.com/sellside/Disclosures.action.

Or, you can also write to the Director of Compliance, Bernstein Institutional Services LLC, 245 Park Avenue, New York, NY 10167.

This research publication covers six or more companies. For risks and other company disclosures:

Please visit: https://bernstein-autonomous.bluematrix.com/sellside/Disclosures.action.

Or, you can also write to the Director of Compliance, Bernstein Institutional Services LLC, 245 Park Avenue, New York, NY 10167.

The Bernstein brand rates stocks based on forecasts of relative performance for the next 12 months versus the S&P 500 for stocks listed on the U.S. and Canadian exchanges, versus the Bloomberg Europe Developed Markets Large and Mid Cap Price Return Index EUR (EDME) for stocks listed on the European exchanges and emerging markets exchanges outside of the Asia Pacific region, versus the Bloomberg Japan Large and Mid Cap Price Return Index USD (JPL) for stocks listed on the Japanese exchanges, and versus the Bloomberg Asia ex-Japan Large and Mid Cap Price Return Index (ASIAX) for stocks listed on the Asian (ex-Japan) exchanges - unless otherwise specified.

The Bernstein brand has three categories of ratings:

• Outperform: Stock will outpace the market index by more than 15 pp

• Market-Perform: Stock will perform in line with the market index to within $ \pm 15 $ pp

• Underperform: Stock will trail the performance of the market index by more than 15 pp

Coverage Suspended: Coverage of a company under the Bernstein research brand has been suspended. Ratings and price targets are suspended temporarily, are no longer current, and should therefore not be relied upon.

Not Rated: A rating assigned when the stock cannot be accurately valued, or the performance of the company accurately predicted, at the present time. The covering analyst may continue to publish research reports on the company to update investors on events and developments.

Not Covered (NC) denotes companies that are not under coverage.

Bernstein brand stock ratings are based on a 12-month time horizon.

The Autonomous brand rates common stocks as indicated below. As our benchmarks we use the Bloomberg Europe 500 Banks And Financial Services Index (BEBANKS) and Bloomberg Europe Dev Mkt Financials Large and Mid Cap Price Ret Index EUR (EDMFI) index for developed European banks and Payments, the Bloomberg Europe 500 Insurance Index (BEINSUR) for European insurers, the S&P 500 and S&P Financials for US banks and Payments coverage, S5LIFE for US Insurance, the S&P Insurance Select Industry (SPSIINS) for US Non-Life Insurers coverage, and the Bloomberg Emerging Markets Financials Large, Mid and Small Cap Price Return Index (EMLSF) for emerging market banks and insurers and Payments. Ratings are stated relative to the sector (not the market).

The Autonomous brand has three categories of common stock ratings:

• Outperform (OP): Stock will outpace the relevant index by more than 10 pp

• Neutral (N): Stock will perform in line with the market index to within $ \pm 10 $ pp

• Underperform (UP): Stock will trail the performance of the relevant index by more than 10 pp

Coverage Suspended: Coverage of a company under the Autonomous research brand has been suspended. Ratings and price targets are suspended temporarily, are no longer current, and should therefore not be relied upon.

Not Rated: A rating assigned when the stock cannot be accurately valued, or the performance of the company accurately predicted, at the present time. The covering analyst may continue to publish research reports on the company to update investors on events and developments.

Those denoted as 'Feature' (e.g., Feature Outperform FOP, Feature Under Outperform FUP) are our core ideas.

Not Covered (NC) denotes companies that are not under coverage.

Autonomous brand common stock ratings are based on a 12-month time horizon.

The Autonomous brand has three categories of preferred stock ratings:

• Outperform (OP): The total return of the preferred instrument is expected to outperform preferred securities of other issuers operating in similar sectors or rating categories over the next six months.

• Neutral (N): The total return of the preferred instrument is expected to perform in line with preferred securities of other issuers operating in similar sectors or rating categories over the next six months.

• Underperform (UP): The total return of the preferred instrument is expected to underperform preferred securities of other issuers operating in similar sectors or rating categories over the next six months.

Autonomous preferred stock ratings are based on a 6-month time horizon.

Where this report contains investment recommendations for credit instruments, as defined in article 3(1)(35) of the Market Abuse Regulation, the information below is presented to comply with its disclosure requirements.

The report may also include reference(s) to published opinions by other Autonomous or Bernstein analysts covering the equity securities of the issuer(s) referenced herein. Please note an investment recommendation for credit instruments published by the author(s) of this report may differ from the published view of the analyst covering equity securities for the issuer(s) contained in this report and vice versa.

The Autonomous brand has three categories of credit ratings:

• Credit Outperform (C-OP): The total return of the Reference Credit Instrument is expected to outperform the credit spread of bonds of other issuers operating in similar sectors or rating categories over the next six months.

• Credit Neutral (C-N): The total return of the Reference Credit Instrument is expected to perform in line with the credit spread of bonds of other issuers operating in similar sectors or rating categories over the next six months.

• Credit Underperform (C-UP): The total return of the Reference Credit Instrument is expected to underperform the credit spread of bonds of other issuers operating in similar sectors or rating categories over the next six months.

Autonomous credit ratings are based on a 6-month time horizon.

A list of all investment recommendations produced by the author(s) of this report alongside credit ratings history are available upon request.

It is at the sole discretion of the Firm as to when to initiate, update and cease research coverage. The Firm has established, maintains and relies on information barriers to control the flow of information contained in one or more areas (i.e. the private side) within the Firm, and into other areas, units, groups or affiliates (i.e. public side) of the Firm

• These figures represent the percentage of companies within each equity rating category for which affiliates of Bernstein have provided investment banking services within the previous 12 months.

As of December 31, 2025. All figures are updated quarterly.

Prior to April 1, 2024, Sanford C. Bernstein & Co., LLC. issued the ratings and price target information in the graph(s) below for the following companies: NVIDIA Corp, Broadcom Inc, Advanced Micro Devices Inc and Intel Corp.

This research publication covers six or more companies. For price chart and other company disclosures, please visit https://bernstein-autonomous.bluematrix.com/sellside/Disclosures.action or you can write to the Director of Compliance, Bernstein Institutional Services LLC, 245 Park Avenue, New York, NY 10167.

All statements in this report attributable to Gartner represent Bernstein's interpretation of data, research opinion or viewpoints published as part of a syndicated subscription service by Gartner, Inc., and have not been reviewed by Gartner. Each Gartner publication speaks as of its original publication date (and not as of the date of this report). The opinions expressed in Gartner publications are not representations of fact, and are subject to change without notice.

Stacy A. Rasgon maintains long positions in various crypto currencies.

Bernstein Autonomous LLP or BSG France SA, beneficially, has either a net long or short position of 0.5% or more of the total issued share capital of a class of common equity securities of the following MiFID eligible securities: Broadcom Inc and Eaton Corp PLC.

SG and/or its affiliates beneficially own 1% or more of a class of common equity securities of the following company: Ibiden Co Ltd.

Bernstein has received compensation for non-investment banking securities-related products or services in the previous twelve months from the following clients: NVIDIA Corp.

Bernstein and/or affiliates have received compensation for non-investment banking securities-related products or services in the previous twelve months from the following clients: Intel Corp and Screen Holdings Co Ltd.

Certain affiliates of Bernstein act as market maker or liquidity provider in the equities securities of: NVIDIA Corp, Broadcom Inc, Advanced Micro Devices Inc, Intel Corp, Hubbell Incorporated, BE Semiconductor Industries NV, Eaton Corp PLC, Infineon Technologies AG and ASML Holding NV.

Certain affiliates of Bernstein act as market maker or liquidity provider in the debt securities of: Broadcom Inc, Intel Corp, Screen Holdings Co Ltd, Infineon Technologies AG and ASML Holding NV.

The legal entity(ies) employing the analyst(s) listed in this report, and their location, can be determined by the country code of their phone number, as follows:

+1 Bernstein Institutional Services LLC; New York, New York, USA

+44 Bernstein Autonomous LLP; London UK

+212 Société Générale Africa Technologies & Services; Casablanca, Morocco

+41 Bernstein Autonomous LLP; Geneva, Switzerland

+49 BSG France S.A.; Frankfurt, Germany

+91 Sanford C. Bernstein (India) Private Limited; Mumbai, India

+852 Sanford C. Bernstein (Hong Kong) Limited 盛博香港有限公司; Hong Kong, China

+65 Sanford C. Bernstein (Singapore) Private Limited; Singapore

+81 Sanford C. Bernstein Japan KK; Tokyo, Japan

It is at the sole discretion of the Firm as to when to initiate, update and cease research coverage. The Firm has established, maintains and relies on information barriers to control the flow of information contained in one or more areas (i.e., the private side) within the Firm, and into other areas, units, groups or affiliates (i.e., public side) of the Firm.

Separate branding is maintained for “Bernstein” and “Autonomous” research products.

• Bernstein produces a number of different types of research products including, among others, fundamental analysis and quantitative analysis under both the “Autonomous” and “Bernstein” brands. Recommendations contained within one type of research product may differ from recommendations contained within other types of research products, whether as a result of differing time horizons, methodologies or otherwise. Furthermore, views or recommendations within a research product issued under one brand may differ from views or recommendations under the same type of research product issued under the other brand. The Research Ratings System for the two brands and other information related to those Rating Systems are included in the previous section.

• Autonomous operates as a separate business unit within the following entities: Bernstein Institutional Services LLC, Bernstein Autonomous LLP, Sanford C. Bernstein (Hong Kong) Limited 盛博香港有限公司 and Sanford C. Bernstein (India) Private Limited. For information relating to “Autonomous” branded products (including certain Sales materials) please visit: www.autonomous.com. For information relating to Bernstein branded products please visit: www.bernsteinresearch.com.

Analysts are compensated based on aggregate contributions to the research franchise as measured by account penetration, productivity and proactivity of investment ideas. No analysts are compensated based on performance in, or contributions to, generating investment banking revenues.

This report has been produced by an independent analyst as defined in Article 3 (1)(34)(i) of EU 596/2014 Market Abuse Regulation ("MAR") and the same article of MAR as it forms part of United Kingdom domestic law by virtue of the European Union (Withdrawal) Act 2018.

To our readers in the United States: Bernstein Institutional Services LLC, a broker-dealer registered with the U.S. Securities and Exchange Commission ("SEC") and a member of the U.S. Financial Industry Regulatory Authority, Inc. ("FINRA") is distributing this publication in the United States and accepts responsibility for its contents. Where this material contains an analysis of debt product(s), such material is intended only for institutional investors and is not subject to the US independence and disclosure standards applicable to debt research prepared for retail investors.

Bernstein Institutional Services LLC may act as principal for its own account or as agent for another person (including an affiliate) in sales or purchases of any security which is a subject of this report. This report does not purport to meet the objectives or needs of any specific individuals, entities or accounts.

To our readers in Canada: If this publication pertains to a Canadian domiciled company, it is being distributed in Canada by Sanford C. Bernstein (Canada) Limited, which is licensed and regulated by the Canadian Investment Regulatory Organization. If the publication pertains to a non-Canadian domiciled company, it is being distributed by Bernstein Institutional Services LLC, which is licensed and regulated by both the SEC and FINRA, into Canada under the International Dealers Exemption.

This document may not be passed onto any person in Canada unless that person qualifies as "permitted client" as defined in Section 1.1 of NI 31-103.

To our readers in Brazil: This report has been prepared by Bernstein Institutional Services LLC, and Banco BTG Pactual S.A. ("BTG") is responsible for the distribution of this report in Brazil.

This document is for distribution only to persons who (i) have professional experience in matters relating to investments falling within Article 19(5) of the Financial Services and Markets Act 2000 (Financial Promotion) Order 2005 (as amended, the “Financial Promotion Order”), (ii) are persons falling within Article 49(2)(a) to (d) (“high net worth companies, unincorporated associations, etc.”) of the Financial Promotion Order, (iii) are outside the United Kingdom, or (iv) are persons to whom an invitation or inducement to engage in investment activity (within the meaning of section 21 of the FSMA) in connection with the issue or sale of any securities may otherwise lawfully be communicated or caused to be communicated (all such persons together being referred to as “relevant

(persons"). This document is directed only at relevant persons and must not be acted on or relied on by persons who are not relevant persons. Any investment or investment activity to which this document relates is available only to relevant persons and will be engaged in only with relevant persons.

To our readers in the member states of the EEA: This publication is being distributed by BSG France SA, which is authorised and regulated by the Autorité de Contrôle Prudentiel et de Résolution (ACPR) and Autorité des Marchés Financiers (AMF).

To our readers in Hong Kong: This publication is being distributed in Hong Kong by Sanford C. Bernstein (Hong Kong) Limited 盛博香港有限公司, which is licensed and regulated by the Hong Kong Securities and Futures Commission (Central Entity No. AXC846) to carry out Type 4 (Advising on Securities) regulated activities and subject to the licensing conditions mentioned in the SFC Public Register (https://www.sfc.hk/publicregWeb/corp/AXC846/details). This publication is solely for professional investors, as defined in the Securities and Futures Ordinance (Cap. 571). The purpose of this report is solely to provide an analysis of the issuers referred to in this report and is not intended for any purpose contrary to the laws of Hong Kong.

To our readers in the People's Republic of China: The securities referred to in this document are not being offered or sold and may not be offered or sold, directly or indirectly, in the People's Republic of China (for such purposes, not including the Hong Kong and Macau Special Administrative Regions or Taiwan, the "PRC") in contravention of any applicable laws of the PRC.

This document does not constitute an offer to sell or the solicitation of an offer to buy any securities in the PRC to any person to whom it is unlawful to make the offer or solicitation in the PRC.

We do not represent that this document may be lawfully distributed, or that any securities may be lawfully offered, in compliance with any applicable registration or other requirements in the PRC, or pursuant to an exemption available thereunder, or assume any responsibility for facilitating any such distribution or offering. In particular, no action has been taken by us which would permit a public offering of any securities or distribution of this document in the PRC. Accordingly, the securities are not being offered or sold within the PRC by means of this document or any other document. Neither this document nor any advertisement or other offering material may be distributed or published in the PRC, except under circumstances that will result in compliance with any applicable laws and regulations.

To our readers in Japan: This publication is being distributed in Japan by Sanford C. Bernstein Japan KK (サンフォード・C・バーンスタイン株式会社), which is registered in Japan as a Financial Instruments Business Operator with the Kanto Local Finance Bureau (registration number: The Director-General of Kanto Local Finance Bureau (FIBO) No.3387) and regulated by the Financial Services Agency. It is also a member of Japan Investment Advisers Association. This publication is solely for qualified institutional investors in Japan only, as defined in Article 2, paragraph (3), items (i) of the Financial Instruments and Exchange Act.

For the institutional client readers in Japan who have been granted access to the Bernstein website by Daiwa Securities Group Inc. ("Daiwa"), your access to this document should not be construed as meaning that Bernstein is providing you with investment advice for any purposes. Whilst Bernstein has prepared this document, your relationship is, and will remain with, Daiwa, and Bernstein has neither any contractual relationship with you nor any obligations towards you.

To our readers in Australia: Sanford C. Bernstein (Hong Kong) Limited 盛博香港有限公司 is responsible for distributing research in Australia. It is regulated by the Securities and Exchange Commission under U.S. laws, by the Financial Conduct Authority under U.K. laws, which differs from Australian laws. Sanford C. Bernstein (Hong Kong) Limited 盛博香港有限公司 is exempt from the requirement to hold an Australian financial services license under the Corporations Act 2001 in respect of the provision of the following financial services to wholesale clients:

• providing financial product advice;

• dealing in a financial product;

• making a market for a financial product; and

• providing a custodial or depository service.

• SCB India does not have any disciplinary history as on the date of this report.

• do not have any financial interest in the subject company

• do not have actual/beneficial ownership of one percent or more in securities of the subject company;

• is not engaged in any investment banking activities for Indian companies, as such;

• have not managed or co-managed a public offering in the past twelve months for any Indian companies;

• have not received any compensation for investment banking services or merchant banking services from the subject company in the past 12 months;

• have not received compensation for brokerage services from the subject company in the past twelve months;

• have not received any compensation or other benefits from the subject company or third party related to the specific recommendations or views in this report; and

• do not currently, but may in the future, act as a market maker in the financial instruments of the companies covered in the report.

• do not have any conflict of interest in the subject company as of the date of this report.

• Except as noted above, the subject company has not been a client of SCB India during twelve months preceding the date of distribution of this research report. Neither SCB India nor its Associates (i.e., affiliates/group companies) have received compensation for products or services other than investment banking, merchant banking or brokerage services from the subject company in the past twelve months.

• The Research investor charter and Terms & Conditions of SCB India are available on its website and may be accessed at Sanford C. Bernstein (India) Private Limited (https://bernsteinresearch.in/) for your reference.