U.S. Semiconductors and Semiconductor Capital Equipment: Silicon Valley meetings reinforce a durable AI-infrastructure buildout, with supply and deployment constraints replacing demand as the main bottlenecks.

Overview

The report summarizes meetings with 11 semiconductor and infrastructure companies during a Silicon Valley investor tour. Its central message is that AI infrastructure demand remains robust through 2027-28, but physical deployment capacity, semiconductor supply, memory, packaging and power availability increasingly determine how much demand converts into revenue.

Core views

Broadcom’s management argued that custom AI accelerators are becoming a broader opportunity than the market assumes. The company described Anthropic’s 5GW opportunity as separate from Google’s internal TPU demand, while OpenAI is seeking to accelerate deployment and Meta remains primarily a 2028 opportunity, with possible late-2027 revenue. Broadcom’s current outlook is $115 billion of AI revenue in 2027 and $230 billion in 2028; management indicated revenue could be higher if customers deploy their full planned gigawatt capacity. It sees the principal limitations shifting from demand to power, wafers and memory, and said it has detailed visibility into physical sites supporting its 2027-28 outlook. Broadcom also maintained that custom silicon will coexist with, rather than displace, NVIDIA. It believes large predictable workloads favor custom hardware, especially where Broadcom can design the XPU, memory subsystem, networking and rack rather than only supplying a chip. Its networking position lets customers retain lower-cost, lower-latency copper in AI racks for longer; management said roughly 80% of Google scale-up connectivity remains copper. Optics should move closer to compute as bandwidth rises, but Broadcom does not expect CPO to become a major scale-up technology until roughly 2028-29. Supply remains strategically important: TSMC is its key foundry partner, Broadcom has secured memory through LTAs through 2029, and it is building proprietary Malaysian advanced-packaging capacity to support increasingly complex multi-die designs. NVIDIA emphasized that its platform’s value rests on performance, flexibility and the CUDA ecosystem rather than GPUs alone. Management estimated its share of hyperscaler capex rose from roughly 5% before ChatGPT to about 20% and then the mid-20% range despite TPU and custom-ASIC deployment. It described data-center content approaching roughly $40 billion per GW as CPUs, LPUs, networking and system components expand its addressable content; the approximately $20 billion of current-quarter Vera Rubin revenue largely excludes separate CPU-rack and LPU opportunities. NVIDIA expects FY27 growth to be roughly 70% supply-constrained, with wafers, CoWoS, memory, optics, transceivers, power and site availability all constrained. It stated that capacity allocation will increasingly depend on whether customers’ physical infrastructure is sufficiently ready to absorb delivered hardware. NVIDIA also described a broader customer base and a continued role for programmable systems as workloads evolve. Hyperscalers are expected to remain close to 50% of revenue over time, while neoclouds, industrial customers, enterprises, sovereign buyers and AI labs could grow faster; frontier AI labs represent roughly 20% of revenue, including purchases routed through clouds and neoclouds. Compute-rental backstops are intended to bridge long-lived infrastructure commitments with enterprise customers seeking shorter contracts, rather than finance idle demand; management said no GPUs are currently idle and expects these arrangements to contribute a few billion dollars of revenue next year. It also expects efficiency gains to reduce memory consumed per token over time, while rising performance requirements still drive absolute memory and infrastructure demand higher. Intel’s discussion centered on strong server demand and a strategic pivot toward a larger addressable market, but with substantial execution work remaining. Management expects to sell essentially everything it can make in 2027, with tightness potentially continuing into the first half of 2028. It expects to more than double 18A wafer capacity next year, with Nova Lake as the main volume driver, while advanced packaging and ASICs should add revenue but dilute gross margin initially. Intel is prioritizing a “rule-of-45” framework—revenue growth plus operating margin—over maximizing gross margin. Its stated TAM is about 15 times its historical opportunity set, supporting near-term margin trade-offs if new businesses can generate attractive returns. The report highlights Intel’s product-cost gap as a critical issue. Management used an illustrative integrated AMD/TSMC comparison of roughly $15-20 of wafer cost per $100 of revenue versus about $60 at Intel, with only part of the gap attributable to Foundry. It is targeting die-size reduction, tile reuse and better cost competitiveness through Panther Lake, Nova Lake and potentially Coral Rapids, but recognizes this as a multi-year process. Intel’s ASIC revenue has risen from a few hundred million dollars roughly a year ago to an approximately $2 billion annualized run rate, with an approximately $4 billion run rate expected relatively quickly. A compute-ASIC win in 2027 would be an important KPI. Key risks include PC memory inflation, distorted demand signals from supply tightness, roadmap slippage, margin headwinds and the late-2027/2028 test of normalized demand as new capacity arrives. NXP’s tone was constructive on automotive and industrial recovery. Automotive channel inventory has returned to about 11 weeks, its first pre-COVID level, though management does not see broad restocking. The company believes auto demand troughed in 3Q24 and reiterates an 8-12% long-term automotive-growth algorithm despite broadly flat industry production, driven by software-defined vehicles, processing, connectivity and vehicle electronics. China represents 17% of revenue under NXP’s reporting methodology; management sees local competitive pressure most clearly in power discretes but cites ongoing capability gaps in microcontrollers and platform solutions. Industrial is tracking roughly a year ahead of internal expectations, and the acquired Kinara platform has a roughly $1.5 billion pipeline, much of it still at proof-of-concept stage. NXP targets gross margin toward 60% by end-2028 through manufacturing restructuring, foundry partnerships and ventures including VSMC. Applied Materials conveyed stronger conviction in the wafer-fab-equipment outlook. Its Semi Systems growth expectations progressed from above 20% in February to above 30% in May and roughly 40% by August, with about 80% of incremental WFE growth expected from leading-edge logic, DRAM and advanced packaging. AMAT is preparing to double quarterly output capacity by 2028 and said its planned capacity could support a roughly $300 billion WFE market at its current share, versus Street expectations of roughly $250-260 billion in 2028. Customer cleanroom availability, rather than demand, is the near-term constraint. The company sees AI driving more complex DRAM architectures, backside power, nanosheets, stacked transistors and integrated process systems, raising materials intensity and favoring co-optimized equipment solutions. Semi Systems gross margin is already above 55%; management expects higher-value integrated systems, value-based pricing and service pull-through to support longer-term margin expansion. The non-covered companies reinforced the breadth of the AI-infrastructure cycle. Monolithic Power Systems described enterprise data center as its fastest-growing business and communications as another major contributor through power for switches, DPUs and optical networking; 800V data-center power is viewed as a late-2027 or 2028 opportunity. Astera Labs sees expanding connectivity content in CXL memory expansion, UALink, NVLink Fusion and optical networking, with UALink deployment in 2027 and more meaningful revenue in 2028. Lattice expects AI-server FPGA content to rise materially, with blended server attach rates increasing from about 1.5x in 2024 to 2.5x in 2025 and 3.5-3.6x in 2026; it also expects PQC-related revenue around mid-to-late 2027 and sees its AMI acquisition expanding TAM from about $6 billion to $12 billion by 2030. GlobalFoundries highlighted silicon photonics as a major growth vector, targeting photonics revenue to double in 2025 and again in 2026 to a $1 billion annualized run rate by 2028. Capacity rather than demand is its primary bottleneck, with 300,000-350,000 photonics-capable wafers available in New York and total tooled capacity of about 2.8 million wafers annually. Ambarella remains positive on edge and physical AI, lifting its SAM to $23 billion by CY31/FY32, but identifies memory availability and inflation as the largest near-term uncertainty. Lightmatter expects NPO trials in 2027 and a larger volume ramp in 2028, arguing that near-packaged optics can capture much of CPO’s bandwidth and power benefit while retaining existing pluggable-optics manufacturing and test infrastructure.

Analysis framework

Bernstein uses management meetings to assess demand, supply bottlenecks, technology roadmaps, competitive positioning, capacity expansion, margins and capital allocation. It then relates those observations to the rated companies’ operating outlooks and valuation frameworks.

Methodology notes

• Industry AnalysisSupply-demand framework

  AI-infrastructure supply-demand analysis

  The report assesses whether AI demand can translate into revenue by examining wafer, memory, packaging, power, cleanroom and data-center-site constraints.

• Industry AnalysisUpstream-Midstream-Downstream Transmission

  AI semiconductor supply-chain transmission

  The analysis links accelerator demand to foundries, memory, packaging, semiconductor equipment, power delivery and optical-connectivity providers.

• Valuation methodsP/E and PEG Valuation

  Forward EPS multiple valuation

  Bernstein derives price targets for AMAT, AVGO, NVDA and NXPI by applying valuation multiples to average forward non-GAAP or pro-forma EPS estimates.

Asset mapping & comparison

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

• Applied Materials (AMAT)

  Direct beneficiary of AI-led WFE, DRAM and advanced-packaging intensity.

  Strengths

  Strong exposure to key technology inflections, integrated systems and attractive valuation versus peers.

  Weaknesses

  Near-term output is constrained by customer cleanroom availability.

  Comparison

  Its planned capacity could support a roughly $300B WFE market versus Street expectations of roughly $250-260B in 2028.

  Risks

  Industry downturn, detrimental end-market mix, share loss and geopolitical risk.

• Broadcom (AVGO)

  Beneficiary of custom AI accelerators, networking and AI infrastructure.

  Strengths

  Broadening customer base, end-to-end custom-chip architecture capabilities and supply visibility.

  Weaknesses

  Revenue realization depends on customer deployment of physical infrastructure.

  Comparison

  Management sees OpenAI and Anthropic potentially becoming at least as large as Google in custom silicon by 2028.

  Risks

  Unexpected AI weakness, customer socket losses, merger-synergy execution, weaker cash return or management change.

• NVIDIA (NVDA)

  Core beneficiary of accelerated-computing demand and expanding system-level data-center content.

  Strengths

  Platform fungibility, CUDA ecosystem, supply commitments and growing content beyond GPUs.

  Weaknesses

  Growth is constrained by supply and physical data-center deployment capacity.

  Comparison

  Management estimates hyperscaler-capex share rose from roughly 5% pre-ChatGPT to the mid-20% range.

  Risks

  Near-term business lumpiness, slower key-end-market growth, competition, internal customer silicon and export regulation.

• Intel (INTC)

  Server recovery, foundry, advanced packaging and ASIC expansion are potential growth drivers.

  Strengths

  Very strong server demand, expanding 18A capacity and growing ASIC revenue.

  Weaknesses

  Product-cost competitiveness, margin dilution and roadmap execution remain multi-year issues.

  Comparison

  Management illustrated roughly $60 wafer cost per $100 of revenue at Intel versus roughly $15-20 in a synthetic AMD/TSMC model.

  Risks

  Macro headwinds, roadmap slippage, margin pressure and further share losses.

• NXP Semiconductors (NXPI)

  Automotive content, industrial edge processing and infrastructure-related data-center products support the recovery thesis.

  Strengths

  Solid execution, automotive content expansion and industrial business tracking ahead of expectations.

  Weaknesses

  The pace and makeup of the cyclical recovery remain uncertain.

  Comparison

  Data center is roughly 5% of revenue but growing more than 20% annually.

  Risks

  Macro deterioration, inventory flush, key-customer socket losses and failure to meet long-term growth or margin targets.

Key data

• Broadcom AI revenue outlook$115B in 2027; $230B in 2028Management indicated revenue could be higher if customers deploy their full planned GW capacity.
• NVIDIA data-center content opportunity~$40B/GWReflects CPUs, LPUs, networking and broader system content.
• NVIDIA supply-constrained FY27 growth~70%Constraints span wafers, CoWoS, memory, optics, transceivers, power and site capacity.
• Intel ASIC revenue run rate~$2B annualized, targeting ~${4}B relatively quicklyCurrently almost entirely networking ASIC-related revenue.
• NXP automotive channel inventory~11 weeksReturned to pre-COVID levels; management does not see broad restocking.
• Applied Materials Semi Systems growth expectation~40%Expectation increased from above 20% in February and above 30% in May.
• GlobalFoundries photonics target$1B annualized run rate by 2028Management expects photonics revenue to double in 2025 and again in 2026.

Impact & implications

Bernstein’s meetings support the view that AI spending is broadening across compute, custom silicon, memory, packaging, equipment, power and connectivity. The report nevertheless stresses that physical supply and deployment readiness, rather than headline demand alone, will determine near-term revenue realization and company-specific outcomes.

Risks

• AI demand may weaken or fail to convert into revenue if power, site readiness, wafers, memory, packaging or cleanroom capacity remain constrained.
• Intel faces roadmap, margin, product-cost and market-share risks despite improving server demand.
• NXP remains exposed to macro conditions, inventory normalization, customer socket losses and execution against its margin targets.
• AMAT faces semiconductor-cycle, end-market-mix, share-loss and geopolitical risks.
• Broadcom and NVIDIA face customer-concentration, competitive, supply-chain and export-regulation risks.

What to watch

• Whether AI customers can complete power, land, shell and data-center infrastructure in time to absorb planned hardware deployments.
• Broadcom’s customer deployment pace at OpenAI, Anthropic, Meta and Google, and conversion of GW roadmaps into revenue.
• NVIDIA’s supply expansion across wafers, CoWoS, memory, optics and transceivers.
• Intel’s 18A ramp, server wafer availability, cost improvements and progress toward additional ASIC wins.
• The durability of WFE growth, DRAM technology pull and customer cleanroom availability for Applied Materials.
• The timing of NPO, CPO, 800V power and broader optical-connectivity adoption.