Artificial intelligence (AI) compute demand remains robust, and Wall Street brokerage Bernstein made its position clear in its latest research report: even as U.S. semiconductor stocks have pulled back in recent days, the global semiconductor and memory chip industry "still warrants a firmly bullish stance from investors." The report identifies TSMC's foundry price increases, the memory chip upcycle, and the expansion of custom AI chips (ASICs) as the primary investment theses going forward.
The report, titled Asia Semiconductors & Global Memory: 3Q26 Earnings Preview — Is the AI Theme Still Worth Being Long?, was published on the eve of the U.S. third-quarter earnings season. Bernstein's analyst team believes that while the surge in 10-year and 30-year Treasury yields has pushed up global credit financing costs, and environmental constraints around AI safety and large-model iteration cycles could slow compute infrastructure investment, the recent share-price correction has already priced in most of the risk.
Analysts noted that large customers urgently accelerating AI compute infrastructure buildout continue to show strong demand even after cost increases, and cost pass-through pressure is more likely to fall on non-AI demand.
On the operational front, TSMC reported preliminary third-quarter revenue of approximately NT$1.494 trillion (approximately $46.8 billion), up 51% year over year; Samsung Electronics posted preliminary operating profit of ₩107.4 trillion (approximately $80.1 billion), up 782.5% year over year — evidence that foundry and memory demand is translating into actual earnings. Based on this, Bernstein concluded that TSMC's capacity expansion and pricing power amid the AI boom, the long-term sustainability of memory sector profitability, and the shift of TSMC capacity toward CPUs plus the Google ASIC ramp led by MediaTek will constitute the primary drivers of the semiconductor rally.
Earnings Take the Baton: From Multiple Expansion to Profit-Driven Gains
The global AI bull market is searching for an "earnings handoff window." With Treasury yields at more than two-decade highs and valuations compressing, stocks can rise without relying on P/E multiples returning to elevated levels — as long as earnings per share keep growing. Bernstein believes global semiconductor stocks fit this condition precisely.
JPMorgan's strategy team likewise favors semiconductors, citing easing positioning congestion, lower valuations, and earnings resilience as collectively creating a re-entry opportunity. The firm's data shows that since June, forward 12-month EPS estimates for semiconductors have been revised up by roughly 30% to 40%, while global software sector earnings expectations have shown no comparable improvement. JPMorgan cites hyperscaler capex outlook projections of approximately $950 billion for 2026, roughly $1.4 trillion for 2027, and at least approximately $3 trillion by 2030.
The structural shift in AI inference workloads is the core argument underpinning compute demand. Meta's Muse AI agent and OpenAI's Astra model enable a single user command to trigger sustained, multi-stage computational work. A single research or programming task may sequentially involve planning, retrieval, reading files, calling tools, executing code, checking results, and correcting errors — with multiple steps each requiring the model to be re-invoked. Anthropic has observed in its research systems that multi-agent tasks consume roughly 15 times the tokens of ordinary chat interactions.
This means AI compute demand will spread across the entire stack: GPU computation, HBM and DRAM capacity and bandwidth, KV cache and SSD storage, high-performance CPU tool execution, and high-speed data transmission through optical interconnect chips. Bernstein therefore argues that GPUs and TPUs handle model computation, high-performance data center CPUs handle task orchestration, and HBM, server DRAM, storage, and optical interconnect components jointly support data transport and state management — ultimately requiring an increasingly massive cluster of AI compute servers to deliver services.
TSMC Price Increases and Memory Upswing: Two Pillars of Earnings Revisions
Bernstein projects TSMC revenue growth of 41% in 2026, with fourth-quarter revenue rising approximately 10% quarter over quarter or more, and sees room for further upward revisions. Advanced-node price increases in 2027 are expected to range from high single digits to low double digits, with mid-single-digit increases for mature nodes, beginning to show in the first half of next year.
The report notes that TSMC's exclusive 3nm-and-below advanced process capacity and 2.5D/3D advanced packaging capacity remain tight, enabling the company to convert any semiconductor supply-chain demand intensity related to AI compute into revenue growth through sustained foundry price increases and process-plus-packaging mix upgrades. However, reports of Apple cutting component production for certain models make the fourth-quarter guidance a more meaningful test.
On CPUs and ASICs, Bernstein expects TSMC's CPU wafer revenue to approach $40 billion in 2027, roughly on par with custom ASIC/XPU wafer revenue, each accounting for about 15% of total revenue; XPU also contributes the strongest packaging revenue. Incremental capacity is tilting toward CPUs and ASIC/XPU, benefiting Google and Amazon's in-house CPUs. Most CPUs use traditional packaging, and with cleanroom resources tight, TSMC is prioritizing front-end manufacturing expansion, with ASE Technology, Amkor, and Intel's EMIB positioned to capture more packaging incremental volume. The report forecasts TSMC capital expenditure of $64 billion, $75 billion, and $82 billion for 2026–2028.
In the memory segment, Bernstein cites third-party forecasts indicating fourth-quarter DRAM and NAND prices will both rise 15% to 20% quarter over quarter, slightly above the firm's own expectations. However, HBM4 schedule delays, potentially lower-than-expected HBM price increases in 2027, and higher employee bonuses limit the magnitude of earnings upside. Analysts therefore place greater emphasis on long-term supply agreements improving earnings visibility for memory stocks, and shareholder returns driving valuation re-rating.
Among memory names, Bernstein explicitly prefers Samsung and CXMT. Samsung is seen as establishing HBM4 leadership with a relatively low valuation; CXMT is supported by domestic substitution, execution capability, and relative valuation advantages. Chinese memory supply expansion constitutes a long-term competitive pressure on the three major global memory makers.
Top Picks: MediaTek Newly Added, Samsung Offers Largest Upside
Bernstein unveiled its favored semiconductor stock list as the U.S. earnings season kicked off. MediaTek is a newly added top pick, joining Asian semiconductor giants Samsung, TSMC, Kioxia, and CXMT, with Micron (MU) as the only U.S. semiconductor company on the list.
Note: Implied upside figures are 12-month potential gains as stated in Bernstein's research report.
The core bullish logic for MediaTek centers on its partnership with Google to build TPU or broader ASIC custom chip business. Bernstein sharply raised its 2028 global data center XPU ASIC market forecast from $140–160 billion to $250–300 billion; 2026 and 2027 forecasts stand at $50–70 billion and $130–150 billion, respectively.
The report notes that TPU v8 has begun volume shipments, and yield improvements on critical substrates for v9 create conditions for 2027 delivery and earnings revisions. Even factoring in intensified competition for v10 in 2029 and a lower valuation multiple, Bernstein still sees upside. The firm projects MediaTek ASIC revenue of $16 billion and $44 billion in 2027 and 2028, respectively, with revenue contribution rising from 11% in 2026 to 65% in 2028, and profit contribution potentially rising from 15% to 68%. Smartphone SoC market share gains also buffer the impact of memory price increases.
On mature nodes, Bernstein maintains an Underperform rating on UMC, raising its target price from NT$88 to NT$103 (approximately $3.2), implying 30.2% downside. UMC benefits from connectivity and power management demand and Samsung's transfer of display driver orders, with average selling prices expected to rise 7% and 5% in 2027–2028, but new business contributions and earnings forecasts trail market expectations.
Vanguard International Semiconductor (5347.TW) benefits from interposer and power management demand at its new VSMC fab, with the potential to run near full capacity from ramp-up; product upgrades partially offset depreciation pressure, but valuation is already full. Bernstein maintains a Market Perform rating, raising the target price from NT$146 to NT$194 (approximately $6.1), implying 4.0% upside. Novatek Microelectronics (3034.TW) benefits from edge vision and customer inventory restocking, but the low-cost memory inventory tailwind will gradually fade. Bernstein maintains Market Perform, raising the target price from NT$480 to NT$490 (approximately $15), implying 7.7% downside.
Samsung Posts Record Profit, But Market Skepticism Lingers
Samsung Electronics' preliminary third-quarter results released on October 8 showed operating profit expected to reach ₩107.4 trillion (approximately $80.1 billion), up 782.5% year over year, marking the first time quarterly profit has surpassed the ₩100 trillion threshold; revenue came in at approximately ₩195 trillion (approximately $145.4 billion), up 126.6% year over year. This marks Samsung's fourth consecutive quarter of record operating profit, with a single quarter's earnings exceeding the full-year 2025 total of ₩43.6 trillion (approximately $32.5 billion).
Yet on the day the preliminary results were disclosed, Samsung's stock fell more than 1%, and has now declined over 25% from its June all-time high. Market observers note that the divergence between earnings and share price reflects both the market having already priced in the profit upside and underlying anxiety about the sustainability of the AI memory supercycle.
Counterpoint Research data shows third-quarter DRAM price quarter-over-quarter growth expectations were revised up from 5%–10% to 10%–20%, with NAND flash memory prices expected to rise 15%–20% quarter over quarter. JPMorgan forecasts Samsung's HBM market share could rise from 20% to 34% this year, while SK Hynix's share slips from 60% to 46%.
But concerns remain. Rising memory chip prices are pushing up component costs for end products such as smartphones, and Samsung's Mobile Experience division has notified suppliers of plans to cut fourth-quarter smartphone production by 20% to 30%. Additionally, the Korean won appreciated 14.3% against the U.S. dollar in the third quarter — the largest quarterly gain since 1998 — shrinking overseas earnings when converted back to local currency.
TrendForce forecasts fourth-quarter general-purpose DRAM contract prices will rise 10% to 15% quarter over quarter, a sharp deceleration from the roughly 60% quarterly surge in the second quarter. Avril Wu, senior research vice president at TrendForce, said suppliers are cautious about continuing aggressive price increases, concerned about damaging consumer electronics end demand; the rising share of long-term agreements with built-in price cap mechanisms has also slowed the pace of increases.
Strong Demand Evidence in the AI Compute Supply Chain
Compared with the valuation narrative on the AI application side, the AI compute supply chain has more concrete evidence of strong demand. Media reports disclosed Anthropic's infrastructure arrangements of approximately $518 billion over roughly the next decade, of which about 80% is non-cancellable or subject to committed payment terms. Anthropic's 2025 revenue grew to roughly 12 times the prior year, approaching $4.6 billion, with operating losses exceeding $8 billion. Compute and infrastructure spending reached $7.33 billion, roughly triple 2024 levels and accounting for about 58% of total operating expenses of $12.65 billion.
Nvidia's latest quarterly data center revenue reached $89 billion, up 117% year over year. South Korea's September semiconductor exports hit $60.3 billion, up 262.8% year over year, with officials noting that memory export volumes and contract prices rose in tandem.
OpenAI is in talks to raise capital at a pre-money valuation of approximately $1.4 trillion; potential IPO investors have assigned Anthropic a valuation range of $1.8 trillion to $2 trillion, with expectations that the offering could match or exceed SpaceX's issuance size. These financing negotiations and Silicon Valley venture capital-driven IPO expectations confirm capital markets' intense focus on the AI compute supply chain.
However, demand-side penetration is only just beginning. Andreessen Horowitz's seventh edition Consumer AI Applications report shows that only 4.5% of U.S. consumers have ever paid for personal subscriptions to ChatGPT, Gemini, or Claude. Spending among paying users is highly concentrated: the top 1% of payers spend an average of $903 per month — more than the bottom 50% of all users combined. Only about 30% of S&P 500 companies report quantifiable AI impact, and only about 2% report trackable metrics.
Andreessen Horowitz's second edition "State of Markets" report notes that GPU rental prices are rising rather than falling, with A100 instances even more expensive than in early 2026. Compute progress and model progress are not a zero-sum game: as AI intelligence gets cheaper, demand for compute actually rises. Rents for the latest chips are climbing, and rents for older chips are holding steady.
This gap — a compute side sprinting ahead while demand-side penetration is only just beginning — is both the greatest uncertainty in the current AI market and the largest room for future growth. Bernstein's bullish stance rests on the premise that the AI compute supply chain already has concrete evidence of strong demand; market skepticism stems from low consumer paid-penetration rates and long-term questions about capital expenditure sustainability. The tug-of-war between the two will be tested in the coming earnings season.