GPU compute scarcity remains the core theme; May SIA is weak but does not change a constructive semiconductor view

Overview

This is a Morgan Stanley North America semiconductor weekly report focused on the Meta GPU environment, cloud demand, GPU supply-demand assessment, and May SIA semiconductor sales data. Its core conclusion is that the market reaction to the possibility of Meta leasing GPUs is too negative; the more reasonable interpretation is that there is still significant GPU compute scarcity. Although May SIA data was below forecast, the trajectory of improving semiconductor cycles, leading AI demand, memory supply tightness, and a gradual rebound in traditional end markets remains unchanged.

Core views

First, GPU compute remains scarce; a hyperscaler renting out GPU capacity externally does not necessarily indicate underutilization of internal demand, and may instead indicate that external demand and leasing economics are strong. Second, NVIDIA is the best positioned in the secondary compute market relative to ASICs and AMD due to its GPU ecosystem, de facto standard status, and compute versatility. Third, Meta capex is no longer the single most prominent top-down driver for GPU demand as in prior years, but overall GPU demand is now more diversified. Fourth, May SIA data were softer than expected, especially in broad semiconductor and memory, yet the industry cycle is still improving and AI remains the strongest area of structural outperformance. Fifth, memory remains a key constraint on AI buildout; elevated DRAM pricing and durable NAND recovery support MU and SNDK as primary exposures.

Analysis framework

The report combines company news, channel checks, hyperscaler GPU utilization, SIA monthly semiconductor sales data, regional and product-category m/m and y/y trends, inventory days, valuation, and stock rating tables to form an industry view. For GPU, it uses supply-demand framing and a compute-versatility comparison to position NVIDIA, AMD, and ASIC solutions. For SIA, it compares May actuals against Morgan Stanley forecasts, historical seasonal averages, and prior-month trend.

Methodology notes

• Industry supply-demand analysisGPU compute supply-demand and secondary-market versatility framework

  When GPUs can be leased to external customers and economics remain attractive, this may indicate external compute demand exceeding available supply rather than underutilized internal capacity.

  The report believes NVIDIA GPUs, due to de facto standard status and stronger versatility, are more easily reallocated in a mismatch of compute supply-demand or in secondary leasing scenarios. ASICs are constrained by customization and scale, making secondary-market transferability weaker.

• Cycle monitoringSIA monthly semiconductor sales and seasonal comparison framework

  Use monthly m/m, y/y, three-month average y/y, and historical seasonal averages to assess semiconductor cycle strength.

  The report compares May SIA data with forecasts and long-term averages to gauge whether broad market, analog, discretes, MCU, MPU, DRAM, NAND, and other categories were above or below expectations.

• Memory cycle analysisAI-driven structural supply constraint framework

  The current memory upturn is less a traditional inventory-restocking cycle and more a structural supply constraint from AI buildout.

  The report sees DRAM as an AI deployment bottleneck, with growing evidence that customers are locking in supply early, while NAND also benefits from tighter mix discipline and supply discipline.

Asset mapping & comparison

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

• NVIDIA Corp. (NVDA)

  Core beneficiary in GPU compute scarcity and the secondary compute leasing market

  Strengths

  De facto industry standard status, compute versatility, higher tokens per gigawatt, faster deployment, longer average uptime, and longer asset life.

  Weaknesses

  Meta capex is no longer the single most prominent upside driver as it has been in past years.

  Comparison

  Compared with ASIC, NVIDIA GPUs are easier to sublease and reallocate; compared with AMD, NVIDIA has stronger market share and ecosystem coverage.

  Risks

  If cloud demand slows or client internal compute utilization remains underused, the market may again fear GPU oversupply.

• Advanced Micro Devices Inc. (AMD)

  Meta could be an early adopter of AMD Helios

  Strengths

  Higher versatility than ASIC; as MI300/350 remains an important customer for Meta, warrant structures provide stronger procurement incentives.

  Weaknesses

  Overall market share is below NVIDIA and secondary-market potential may be smaller.

  Comparison

  More transferable than low-volume custom ASICs, but weaker than NVIDIA in de facto standard status and ecosystem scale.

  Risks

  Even if Meta adopts AMD Helios, the report says there is no evidence yet of an impact on NVIDIA procurement.

• Micron Technology Inc. (MU) / SanDisk Corporation (SNDK)

  Primary exposure to memory constraints and pricing durability in AI buildout

  Strengths

  DRAM remains an AI deployment bottleneck, NAND recovery shows more persistence, and customer supply prebooking could extend the profit window.

  Weaknesses

  Both May DRAM and NAND data were below Morgan Stanley forecasts, reflecting higher forecast uncertainty due to monthly volatility and supply-driven moves.

  Comparison

  The report clearly maintains preference for MU and SNDK as memory tightness exposure names.

  Risks

  If supply is released faster than expected or if price strength is not durable, the profit window may contract.

• Broadcom (AVGO), Cerebras (CBRS), LRCX, KLAC, MKSI, ON, ADI, NXP

  Beneficiary themes for leading-edge logic, semiconductor equipment supply chain, and analog/MCU recovery

  Strengths

  The report sees traditional end-market demand improving, and recovery is no longer limited only to AI-related demand.

  Weaknesses

  In May broad market, discretes and analog were below forecast on m/m.

  Comparison

  These names cover different links in the recovery chain, including leading-edge logic, capital equipment, supply chain, and high-end analog/MCU.

  Risks

  If industrial and traditional-end market recoveries disappoint, non-AI semiconductor exposure could come under pressure.

Key data

• Total semiconductor sales in May+16.1% m/mBelow Morgan Stanley's +22.0% forecast, but above the 10-year average of +6.9%.
• May three-month average y/y growth+104.2%Accelerated from April's +93.9%.
• May single-month y/y growth+118.8%Shows y/y growth remains at a high level.
• Regional y/y growthAmericas +150.7%；Asia Pacific +126.3%；China +101.9%；Europe +83.7%；Japan +39.4%Americas had the highest y/y growth.
• DRAM May m/m+27.7% m/mBelow the forecast of +43.0% and the 5-year average of +45.9%, but three-month y/y sales growth reached 304.8%.
• NAND May m/m+40.7% m/mSlightly below the forecast of +43.8%, but above the 5-year average of +25.6%; three-month average y/y revenue reached 364.6%.
• Annual semiconductor forecast revisionreduced from +103% to +99%Primarily due to some deceleration in memory shipments.
• CY26 memory forecast$847bnBelow the prior forecast, but still significantly above last year's $222bn.
• CY27 growth assumption+24% y/y to $1.95 trillionPrimarily driven by memory pricing pass-through, and broadly in line with the prior forecast.
• Semiconductor company inventory days114 daysUp 2 days from the prior month, above the historical median of 23 days.
• Semiconductor customer inventory days60 daysUp 9 days m/m, slightly above a seasonal increase of 8 days.
• Distributor inventory days61 daysDown 2 days m/m, still above the historical median of 7 days.

Impact & implications

For investors, the report reinforces a medium-term logic centered on AI compute chains and memory tightness. If GPU compute scarcity persists, NVIDIA's ecosystem and versatility advantages may be further amplified; AMD could benefit from customers like Meta adopting Helios and warrant incentives, but the secondary compute market may be smaller than NVIDIA's. On memory, continuing DRAM and NAND supply constraints are expected to prolong the profit window, supporting MU and SNDK. Even though May SIA was below expectation, the report still expects Q2 to be strong overall, and recovery is spreading beyond AI into industrial, analog, MCU, and equipment supply-chain areas.

Risks

• The market may continue to interpret GPU leasing or cloud-service expansion as hyperscaler compute oversupply, weighing on AI semiconductor valuations.
• May SIA data were below forecast; if subsequent months remain weak, it could weaken the outlook of a strong Q2 and improvement in H2.
• Monthly trends in DRAM and NAND are heavily supply-constrained, increasing forecast uncertainty.
• If memory prices fall due to faster supply release or demand normalization, profit sensitivity for exposures like MU and SNDK may decline.
• Morgan Stanley disclosed investment bank, market-making, or other service relationships with several covered companies; investors should monitor potential conflicts of interest.

What to watch

• Cloud providers' earnings disclosures on AI compute, GPU leasing, capex, and internal utilization.
• Whether Meta advances a cloud-services segment and whether it becomes an early adopter of AMD Helios.
• Whether subsequent monthly SIA data continue to validate strong Q2 and H2 improvement in 2026.
• Whether DRAM and NAND prices, shipments, and customer supply-commitment behavior continue to support the structural-tightness narrative.
• Subsequent developments and earnings for ASML, TSMC, TXN, AMD Advancing AI, INTC, NXP, KLAC, QCOM, and others.