AI Server Supply Chain Report Interpretation

Overview

Mizuho's industry call examines the AI-server supply chain, with a constructive view on Broadcom's custom-ASIC opportunity and differentiated implications for AI-chip, server and interconnect companies. The report expects strong longer-term AI infrastructure demand but flags near-term Nvidia earnings variability, China-related headwinds and rising competition for several peers.

Core views

The report's central conclusion is that Broadcom remains the principal beneficiary of expanding custom-AI-silicon demand. Mizuho expects strong ASIC ramps from TPUv7p and MTIA2, followed by potential OpenAI Strawberry and Apple Baltra programs in 2026E-27E. It estimates AVGO's 2026 CoWoS capacity at about 95,000 wafers, with upside to more than 100,000 wafers if those customer ramps become material. AVGO's custom-silicon addressable market is projected to expand from about $17.5 billion currently, where the firm has roughly 70% share, to $60-90 billion by 2027E. The report retains Outperform and a $250 target, based on 31.2x F26E P/E. Mizuho sees the wider accelerator market continuing to grow despite China restrictions. It forecasts Nvidia GPU shipments rising at a 17% CAGR from 4.5 million units in 2024 to 7.2 million in 2027E, while AMD GPU units rise at a 38% CAGR to 915,000. Nvidia's near-term AprQ and JulQ outlook is described as mixed: H20 sales may provide some upside, but the report expects an approximately $3 billion JulQ headwind from H20 wind-downs and notes a $5.5 billion H20 write-down. It expects a stronger OctQ and JanQ as GB200 NVL72 and GB200/300 HGX shipments ramp, with GB300 NVL72 more of a 1H26E event. Nvidia remains Outperform with a $168 target based on 38.4x F26E P/E. Middle East agreements are presented as a potential partial offset to lost China AI-accelerator shipments. Mizuho estimates a possible 2026 China H20 headwind of roughly 1 million units, or about $15 billion. It highlights Saudi HUMAIN commitments involving 500MW each for AMD and Nvidia over five years, which it translates into roughly 280,000 Nvidia GPUs and approximately 350,000 AMD units over that period, as well as a UAE G42 import quota for 500,000 Nvidia GB200s, valued at about $15 billion annually. SMCI's $20 billion DataVolt partnership and AMD's $10 billion HUMAIN partnership are additional supply-chain supports. The report differentiates peer outcomes. AMD's target rises to $135 from $117, while its Outperform rating and estimates are retained; the thesis rests on MI355X ramping in 2H25E, AI demand into 2026E and the HUMAIN partnership partially offsetting an $800 million JunQ MI308 China-ban headwind. Dell remains Outperform and its target rises to $145 from $140, supported by a roughly $30 billion AI-server pipeline, $2.1 billion of F4Q25 AI-server sales, a $9 billion backlog exiting February, and potential corporate/AI-PC refresh demand. SMCI remains Neutral despite a higher $40 target, as DataVolt and liquid-cooling deployment strengths are offset by competition, limited margin expansion, customer concentration and internal-control overhangs. Credo remains Outperform with a $70 target based on about 19x F26E P/S. Mizuho cites its SerDes IP, active electrical cable and optical product portfolio, engagement with seven major hyperscalers, approximately 64% gross margin, and a data-center interconnect TAM potentially about 15 times its roughly $400 million annualized AEC revenue run rate. Intel remains Neutral with a $22 target: the report recognizes possible upside from a new CEO, foundry expansion and an FPGA spin-off, but believes weak foundry and AI-accelerator execution, PC share loss and lagging traction versus AMD and Nvidia remain material constraints. The report frames capacity and competition as important transmission mechanisms. Total industry CoWoS capacity is forecast to grow from 788,001 wafers in 2025E to 1,162,751 in 2026E and 1,328,250 in 2027E, but the growth rate slows sharply. This supports AVGO's potential customer-driven demand upside while creating possible 2026 pressure for Marvell and MediaTek: Mizuho expects AlChip share gains on AWS Trainium3 and flags yield issues that could soften MediaTek's TPUv7e-related CoWoS outlook. China domestic AI demand is also rising, with Ascend 910 orders estimated above 700,000 in 2025E, although SMIC yields are estimated near 30%.

Analysis framework

Mizuho combines supply-chain channel observations with estimates of CoWoS capacity, accelerator shipment volumes, hyperscaler capital expenditure, announced customer programs and company-specific operating data. It then applies forward valuation multiples to forecast earnings or sales to establish target prices and presents bull and bear scenarios for the covered companies.

Methodology notes

• Industry AnalysisSupply-demand framework

  AI accelerator, custom-ASIC and CoWoS capacity supply-demand analysis

  The report compares expected chip demand, customer ramps and hyperscaler spending with CoWoS capacity to identify likely beneficiaries and capacity-related constraints.

• Industry AnalysisUpstream-Midstream-Downstream Transmission

  AI-server supply-chain transmission

  It traces how hyperscaler and sovereign AI demand flows through accelerators, custom silicon, packaging, server OEMs and interconnect suppliers.

• Valuation methodsP/E and PEG Valuation

  Forward P/E target-price valuation

  Mizuho values several covered companies by applying stated F26E or F27E P/E multiples to forecast EPS.

• Valuation methodsPS valuation

  Forward price-to-sales valuation for Credo

  Credo's target price is based on a stated multiple of F26E sales rather than earnings.

Asset mapping & comparison

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

• Broadcom (AVGO)

  Primary beneficiary of expanding custom-ASIC, networking and CoWoS demand.

  Strengths

  ~70-80% 2024E ASIC share; potential AI revenue above $50 billion by F27E; strong profitability and recurring software bookings.

  Comparison

  Expected to retain ASIC leadership while AlChip gains AWS share and Marvell faces potential 2026 CoWoS pressure.

  Risks

  Competition, concentrated customers, execution, contract-manufacturing dependence, pricing pressure and regulatory or tariff changes.

• NVIDIA (NVDA)

  Leading merchant AI-silicon supplier benefiting from GB200 and Middle East demand.

  Strengths

  >95% estimated AI training/inference share today; GB200 ramp and broad data-center, gaming and automotive exposure.

  Weaknesses

  Near-term H20 wind-down and slower Blackwell ramp may create mixed earnings.

  Comparison

  Dominates AI accelerators while AMD remains a distant second supplier.

  Risks

  China export restrictions, geopolitical tensions, AI-server spending pullback, CoWoS constraints, litigation and competition.

• Advanced Micro Devices (AMD)

  AI-accelerator competitor positioned to benefit from MI355X and HUMAIN demand.

  Strengths

  Server-share gains, MI325X/MI355X ramp, AI roadmap and $10 billion HUMAIN partnership.

  Weaknesses

  China MI308 restrictions caused an $800 million JunQ headwind; market share remains modest versus Nvidia.

  Comparison

  The report characterizes AMD as the second-place AI accelerator competitor behind Nvidia.

  Risks

  AI slowdown, failed PC recovery, Intel competitive improvement and limited CoWoS allocation.

• Dell Technologies (DELL)

  AI-server OEM expected to gain share from backlog, pipeline and potential PC refresh.

  Strengths

  $2.1 billion F4Q25 AI-server sales, $9 billion backlog, ~$30 billion pipeline, storage margins and negative working capital.

  Weaknesses

  Limited AI-server operating margins.

  Comparison

  Competes with other AI-server builders and primarily serves Tier-2 CSPs and enterprise customers.

  Risks

  Slower AI-server or AI-PC ramps, delayed PC refresh, customer concentration, margin compression, cooling-cost issues and geopolitical risks.

• Super Micro Computer (SMCI)

  AI-server and liquid-cooling supplier supported by DataVolt but constrained by competition and margins.

  Strengths

  >75% ex-CSP share in 2024, fast deployment, early GB200 liquid-cooling ramp and DataVolt partnership.

  Weaknesses

  Expected share decline, limited margin expansion and internal-control overhang.

  Comparison

  Faces expanding capacity and deployment competition from Dell and other server integrators.

  Risks

  AI spending slowdown, customer and supplier concentration, elevated cash-conversion cycle, margin pressure, slow liquid-cooling adoption and accounting issues.

• Credo Technology (CRDO)

  Data-center interconnect supplier benefiting from AI infrastructure investment.

  Strengths

  SerDes IP, AEC and optical portfolio; seven-hyperscaler engagement; ~64% gross margin.

  Weaknesses

  Sensitive to a limited number of large data-center customers.

  Comparison

  Competes mainly with Broadcom and Marvell in relevant interconnect markets.

  Risks

  Hyperscaler spending pullback, competitive pressure, China exposure and sole-foundry dependence on TSMC.

• Intel (INTC)

  Potential turnaround and foundry/FPGA optionality, but with execution challenges.

  Strengths

  Server-CPU position, possible leadership change, IFS expansion and FPGA spin-off potential.

  Weaknesses

  Weak foundry execution, lagging AI roadmap and PC share loss.

  Comparison

  Risks further share loss to AMD in PCs and servers and falling further behind Nvidia in AI.

  Risks

  Process-node or product-ramp setbacks, softer data-center/AI demand and foundry delays.

Key data

• AVGO 2026E CoWoS capacity~95,000 wafers; potentially >100,000 wafersPotential upside if OpenAI and Apple custom-ASIC ramps materialize.
• Nvidia GPU shipments7.2 million units by 2027E17% 2024-27E CAGR from 4.5 million units in 2024.
• AMD GPU shipments915,000 units by 2027E38% 2024-27E CAGR.
• 2025E hyperscaler capex growth~35% year-on-yearCited as support for AI accelerator demand.
• Potential China H20 headwind~1 million units / ~$15 billion in 2026EPotential shipment and revenue impact cited by the report.
• Total industry CoWoS capacity1,162,751 wafers in 2026EUp from 788,001 in 2025E; forecast to reach 1,328,250 in 2027E.

Impact & implications

Mizuho argues that AI infrastructure spending and new customer programs can sustain demand through the semiconductor supply chain, particularly for Broadcom's ASIC and networking exposure. However, China restrictions, slower packaging-capacity growth, margin pressure, customer concentration and competitive share shifts make the effects uneven across chip and server vendors.

Risks

• China export restrictions and geopolitical tensions could weaken AI-accelerator demand or shipments.
• A significant pullback in AI-server spending would pressure chip, server and interconnect suppliers.
• CoWoS capacity constraints and slowing capacity growth may limit shipment ramps or shift share.
• Competitive share gains, customer concentration and margin pressure are material company-specific risks.
• Execution risk remains prominent for Intel's foundry and AI roadmap and SMCI's controls and deployment economics.

What to watch

• Whether AVGO's OpenAI Strawberry and Apple Baltra programs lift 2026E CoWoS demand above 100,000 wafers.
• The pace of Nvidia GB200 NVL72, GB200/300 HGX and later GB300 NVL72 deployments.
• Implementation and scale of HUMAIN, G42 and DataVolt AI infrastructure commitments.
• The extent to which China H20 and MI308 restrictions are offset by non-China demand.
• CoWoS allocation changes, including AlChip's AWS Trainium3 share and MediaTek TPUv7e yield progress.
• AI-server backlog conversion, liquid-cooling adoption and margin trends at Dell and SMCI.