US semiconductors and semiconductor equipment: UBS sees a durable AI-led semiconductor upcycle, but says elevated expectations are limiting the market reward for earnings beats.

Overview

This UBS semiconductor update combines takeaways from its Taiwan Summit, an assessment of the MATCH Act, MLPerf v6.1 benchmark results, and recent earnings-season stock action. The report argues that AI infrastructure demand continues to strengthen the underlying cycle, while share prices face a higher bar because strong results are increasingly expected.

Core views

UBS’s Taiwan Summit takeaways point to a widening AI-compute bottleneck beyond accelerators. Compute performance is scaling at roughly 3x every two years, compared with about 1.6x for memory bandwidth and 1.4x for I/O. UBS therefore sees memory and data movement as increasingly important to extracting accelerator performance. As clusters exceed 100,000 GPUs or XPUs, networking becomes part of the compute architecture: processing is moving from CPUs toward NICs, SmartNICs and DPUs. UBS views this as favorable for Broadcom, Marvell and Astera Labs through larger “XPU attach” opportunities and greater emphasis on switching features. Advanced packaging remains central to AI hardware roadmaps, although customer approaches are diverging. TSMC is targeting CoPoS process and equipment selection for 310x310mm by mid-calendar 2027 and mass production in 2028. Intel’s EMIB-T roadmap is also becoming more relevant, with silicon-bridge yields targeted to exceed 50% in the first half of 2027 and support substrates above 2m for MediaTek TPU v9 demand. Supply is improving: industry CoWoS capacity is expected to reach 260 thousand wafers per month exiting 2027, versus 160 thousand exiting 2026, including 180 thousand at TSMC, 60 thousand at ASE and 20 thousand at Amkor. UBS believes AMD and NVIDIA are well placed to capture the incremental capacity, supporting continued accelerator ramps into 2027. For memory, UBS argues that AI demand is creating a multi-year DRAM and NAND shortage that is constraining server shipments, raising hardware bills of materials and accelerating capacity investment. Its industry checks indicate DRAM fulfillment near 60% of unconstrained demand and enterprise SSD fulfillment near 30%. In UBS’s view, availability rather than end-customer demand is becoming the main constraint on server and storage shipments through 2027. Together with advanced-node and packaging expansion, this supports a broader semiconductor-capital-equipment upcycle; UBS forecasts global wafer-fab equipment spending of about $226 billion in 2027 and $275 billion in 2028, with demand spanning lithography, wet processing, metrology, probing and test. On policy, UBS considers the MATCH Act more “bark” than “bite” for US semiconductor-equipment suppliers. Based on discussions with Washington experts, it considers inclusion and passage under the current Congress unlikely, though the topic may return in the next Congress. The Act aims to align restrictions on US suppliers with foreign competitors such as ASML and TEL, but UBS’s industry discussions suggest those companies have largely complied with existing US-Dutch-Japanese arrangements. Additional statutory restrictions could also make allied negotiations harder. UBS notes that China is already seeking to source 50% of tools domestically, while questioning whether local suppliers such as SMEE can achieve adequate yields. MLPerf Inference v6.1, released September 16, had a record 30 participants and produced the first peer-reviewed results for NVIDIA Vera Rubin and AMD Instinct MI350P. NVIDIA reported about 2.0 million server tokens per second on DeepSeek-R1 using a 288-GPU GB300 NVL72 system, more than 1.1 million on GPT-OSS-120B with a 72-GPU GB300 configuration, and preliminary Vera Rubin performance of roughly 1.2 million server tokens per second on DeepSeek-R1 using 72 GPUs. UBS says these results demonstrate NVIDIA’s system and software advantage and scaling across Blackwell and Rubin. AMD’s 512-GPU MI355X cluster with Crusoe delivered roughly 2.4 million server tokens per second on DeepSeek-R1 and 5.4 million on GPT-OSS-120B, the highest aggregate MLPerf throughput reported, while a 72-GPU MI355X exceeded 950,000 on GPT-OSS-120B. UBS views AMD’s results as encouraging for inference deployments, though current-generation scaling requires customization; it expects MI450 Helios to improve scaling into larger clusters. Intel remained focused on Xeon 6 CPUs and Arc Pro B70 GPUs and, in UBS’s assessment, lacks a competitive data-center AI accelerator roadmap versus NVIDIA and AMD. Finally, UBS highlights a disconnect between operating results and equity-market reactions. During the recent earnings season, 84% of semiconductor companies guided above consensus, down from 97% three months earlier, yet guide-above companies received little reward and guide-below companies were penalized less than in the prior quarter. UBS interprets this as evidence that upside surprises have become a baseline expectation rather than a catalyst. It remains bullish on fundamentals, but believes the market may be in a consolidation phase until companies demonstrate that the elevated pre-July expectation level is achievable.

Analysis framework

UBS combines industry discussions from its Taiwan Summit and policy experts with supply-chain checks, AI benchmark results from MLPerf, earnings guidance versus consensus, and post-earnings price reactions. It then applies company-specific valuation approaches and assesses risks to the covered semiconductor universe.

Methodology notes

• Industry AnalysisSupply-demand framework

  AI-driven memory supply-demand analysis

  UBS compares unconstrained demand with DRAM and enterprise SSD fulfillment to argue that memory availability is limiting server and storage shipments.

• Industry AnalysisUpstream-Midstream-Downstream Transmission

  AI infrastructure supply-chain transmission

  The report traces rising accelerator and cluster demand through networking, memory, advanced packaging and semiconductor-equipment demand.

• Valuation methodsP/E and PEG Valuation

  Next-twelve-month P/E valuation

  UBS states that it values AMD and Astera Labs on NTM P/E, and uses a P/E multiple for NVIDIA and Marvell.

• Valuation methodsEV/EBITDA valuation

  EV/FCF valuation

  UBS values Broadcom using enterprise value to free cash flow, focusing on free cash flow and dividend payout factors.

• Other

  Guidance-versus-consensus and post-earnings price-reaction analysis

  UBS uses the share of companies guiding above consensus and their subsequent stock reactions to assess whether investor expectations are already elevated.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• AMD (AMD)

  Potential beneficiary of additional advanced-packaging capacity and AI accelerator ramps; MLPerf results support an expanding inference role.

  Strengths

  MI355X posted strong aggregate inference throughput, and UBS expects MI450 Helios to improve scaling into larger clusters.

  Weaknesses

  Current-generation GPU scaling requires customized efforts.

  Comparison

  UBS sees AMD as an inference option, but NVIDIA retains a systems and software advantage.

  Risks

  Aggressive Intel price competition, margin pressure, delayed production shift to TSMC, or weaker cloud/data-center and AI traction.

• Astera Labs (ALAB)

  Potential networking beneficiary as AI clusters scale and data movement becomes more important.

  Strengths

  Networking architecture expansion can increase XPU-attach opportunities.

  Comparison

  UBS groups ALAB with AVGO and MRVL as beneficiaries of increasing networking importance.

  Risks

  Slower Scorpio X ramp or retimer growth, or a faster-than-expected transition to CPO.

• Broadcom (AVGO)

  Potential beneficiary of AI-cluster networking and compute-architecture changes.

  Strengths

  UBS sees networking as increasingly embedded in AI compute architecture.

  Comparison

  Grouped with Marvell and Astera Labs as a networking beneficiary.

  Risks

  Further M&A could affect free cash flow and dividend payout ratios; tighter market conditions could pressure dividend yield and valuation.

• Marvell Technology Group (MRVL)

  Potential beneficiary of the networking buildout in large AI clusters.

  Strengths

  Cloud customers seek another source beyond Broadcom and Cisco.

  Weaknesses

  The opportunity may take time to open; UBS may be underestimating internal controller efforts at SSD suppliers.

  Comparison

  Potential alternative source in cloud networking beyond AVGO/CSCO.

  Risks

  Timing of cloud-customer adoption and stronger internal SSD-controller capabilities.

• NVIDIA (NVDA)

  Primary AI-accelerator and advanced-packaging beneficiary in the report’s industry discussion.

  Strengths

  MLPerf results demonstrate strong scaling and, in UBS’s view, a systems and software advantage across Blackwell and Rubin.

  Comparison

  UBS views NVIDIA ahead of AMD and Intel in data-center AI acceleration; AMD’s results are improving for inference.

  Risks

  Competition from AMD in GPUs, ARM-based processors, Intel’s processor initiatives, and broader semiconductor demand sensitivity to corporate profitability.

Key data

• AI compute performance scaling~3x every two yearsVersus ~1.6x for memory bandwidth and ~1.4x for I/O.
• DRAM fulfillment~60% of unconstrained demandUBS industry check; memory availability is becoming a shipment constraint.
• Enterprise SSD fulfillment~30% of unconstrained demandUBS sees server and storage supply as constrained into 2027.
• CoWoS capacity260kwpm exiting 2027Versus 160kwpm exiting 2026; includes 180kwpm at TSMC, 60kwpm at ASE and 20kwpm at Amkor.
• Global WFE spending forecast~$226B in 2027; ~$275B in 2028UBS projection for the semiconductor-equipment cycle.
• Companies guiding above consensus84%Recent earnings season, down from 97% three months earlier.
• AMD MI355X benchmark throughput~2.4MM DeepSeek-R1 tokens/s; ~5.4MM GPT-OSS-120B tokens/s512-GPU Crusoe cluster; UBS notes the latter was the highest aggregate throughput reported.

Impact & implications

UBS believes AI demand is broadening beyond accelerators into networking, memory, advanced packaging and equipment. It sees the underlying semiconductor cycle as strong, but argues that elevated valuation and earnings expectations mean incremental beats may no longer lift stocks unless they validate the sector’s ambitious demand and capacity assumptions.

Risks

• Macroeconomic weakness, geopolitical developments, trade disruption, currency movements and tax-policy changes could affect the semiconductor universe.
• Technology disruption, changing industry business models, M&A and market-share shifts could alter unit sales, average selling prices, revenue and valuation assumptions.
• AI infrastructure demand may be constrained by powered-shell capacity, while memory shortages may restrict server and storage shipments.
• The report identifies company-specific execution and competitive risks for AMD, Astera Labs, Broadcom, Marvell and NVIDIA.

What to watch

• DRAM and enterprise SSD fulfillment rates and their effect on server and storage shipments.
• CoWoS capacity expansion, including TSMC, ASE and Amkor supply additions, and accelerator ramps through 2027.
• Progress in TSMC CoPoS and Intel EMIB-T roadmaps and associated yield targets.
• Whether the MATCH Act advances in future US legislation and whether existing allied export-control arrangements change.
• MLPerf progress for NVIDIA Vera Rubin, AMD MI450 Helios and competing AI accelerator platforms.
• Whether earnings beats and guidance above consensus begin receiving stronger stock-market rewards.