Goldman Sachs Raises Global Server TAM; AI Racks, ASIC Penetration, and Memory Costs Are the Key Incremental Drivers

Overview

This report updates Goldman Sachs' TAM forecasts for the global server market, with a focus on AI server racks, 8-GPU-equivalent AI servers, general servers, the GPU and ASIC mix in AI chips, and capex from leading CSPs in the U.S. and China. The report believes the AI infrastructure buildout cycle will continue through 2028E, and that a higher mix of AI racks, rising ASIC adoption, higher memory costs, and upgrades to general server specifications will together lift server value.

Core views

The core view is that global server demand comes not only from AI-dedicated servers, but also from general servers' control-plane, data-movement, and ongoing inference needs under Agentic AI workloads. For AI racks, NVIDIA-driven rack forecasts were raised to 92k/148k in 2027E/2028E, while AMD-driven racks were broadly kept at 5k/13k/15k for 2026E-2028E. The global value TAM for AI server racks is expected to grow at a 2025E-2028E CAGR of 118%, reaching $561.4 billion in 2028E and accounting for 51% of the global server value TAM. The 8-GPU-equivalent AI server value TAM is expected to reach $295.5 billion in 2028E, accounting for 27%. The general server value TAM is expected to reach $245.2 billion in 2028E, accounting for 22%.

Analysis framework

The report uses a bottom-up TAM framework that breaks server revenue into general servers, AI server racks (NVL72-equivalent), and 8-GPU-equivalent AI servers, and projects them by combining chip platforms, ODM/OEM suppliers, ASPs, shipment volumes, and CSP capex structure. General server revenue is estimated as shipments of major server brands, ODM direct, and other suppliers multiplied by their respective ASPs; on the AI server side, the model tracks GPU and ASIC implied AI chips, rack-level shipments, and 8-GPU-equivalent shipments.

Methodology notes

• TAM forecastingBottom-up server TAM model

  Bottom-up server TAM calculation

  Based on shipment volumes, ASPs, product types, supplier structure, and CSP capex, the model separately estimates revenue for general servers, AI server racks, and 8-GPU-equivalent AI servers.

• Product equivalenceNVL72-equivalent AI server racks

  NVL72-equivalent AI server racks

  A unified rack-equivalent measure is used to capture high-end AI server rack shipments, reflecting more complex designs, stronger GPU interconnects, and higher dollar content per rack.

• Chip mix analysisGPU vs. ASIC implied AI chips

  GPU and ASIC mix in AI chips

  AI server shipments are used to infer GPU and ASIC chip demand, and to assess how higher ASIC adoption affects server TAM and the supply chain.

• Capex analysisCSP capex structure

  CSP capex structure

  CSP capex is broken down into chipsets, HBM memory, networking and cooling, power, chassis, rails, and other components to gauge the intensity of AI infrastructure spending.

Asset mapping & comparison

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

• ADVANCED MICRO DEVICES INC

  AMD is included in the AI server rack chip platform forecast, and the report expects AMD-driven racks of roughly 5k, 13k, and 15k in 2026E-2028E.

  Strengths

  Growth in AI server racks, expanding overall AI chip demand across GPUs and ASICs, and higher CSP capex are favorable for data center semiconductor demand.

  Weaknesses

  The report shows a much more pronounced upward revision for NVIDIA-driven rack forecasts, while AMD rack forecasts are basically unchanged, indicating limited incremental evidence.

  Comparison

  Compared with the upward revisions to NVIDIA-related rack forecasts for 2027E/2028E, AMD-related rack forecasts are broadly unchanged; CSP chip spending in China may be more distributed across GPUs and ASICs.

  Risks

  ASIC substitution, customer in-house chips, a slowdown in CSP capex, memory and network bottlenecks, and AI server shipments coming in below expectations.

• AI server supply chain

  The report lists Wiwynn, Wistron, Hon Hai/FII, LandMark, VPEC, AVC, Fositek, Auras, King Slide, Chenbro, EMC, GCE, Eoptolink, TSMC, MPI, WinWay, Aspeed, and Hon Precision as Buy names related to AI servers.

  Strengths

  Higher AI server rack value content, rising demand for liquid cooling and high-speed interconnects, and expanded spending on HBM and networking components support multiple supply-chain segments.

  Weaknesses

  Different supply-chain segments are affected differently by customer concentration, product-spec transitions, and pricing pressure.

  Comparison

  ODM direct is expected to be the main supplier, with Hon Hai/FII's market share in AI server racks projected to rise from 55% in 2026E to 69% in 2028E.

  Risks

  CSP order concentration, design changes, yields, capacity, intensifying competition, and capex-cycle volatility.

• General servers

  The report raises general server revenue growth forecasts and argues that Agentic AI is pushing CPU servers to take on control-plane, data-movement, and ongoing inference roles.

  Strengths

  High memory costs and specification upgrades lift ASPs, while spillover from AI workloads supports demand.

  Weaknesses

  Shipment growth is relatively modest versus AI racks, and value growth depends more on ASPs and memory costs.

  Comparison

  General server value TAM in 2028E is $245.2 billion, accounting for 22% of the global server TAM, below the 51% share of AI server racks.

  Risks

  A decline in memory prices, a contraction in enterprise IT budgets, changes in AI inference architecture, or slower-than-expected CPU server upgrade cycles.

Key data

• Global server revenue TAM2026E/2027E/2028E are $604.1 billion, $849.9 billion, and $1.102 trillionUnder the new forecasts, YoY growth is 74%, 41%, and 30%.
• AI server racks (NVL72) revenue2026E/2027E/2028E are $166.7 billion, $336.1 billion, and $561.4 billionThe 2025E-2028E CAGR is expected to be 118%, and 2028E will account for 51% of the global server value TAM.
• 8-GPU-equivalent AI server revenue2026E/2027E/2028E are $248.3 billion, $292.1 billion, and $295.5 billionThe 2025E-2028E CAGR is expected to be 26%, and 2028E will account for 27% of the global server value TAM.
• General server revenue2026E/2027E/2028E are $189.1 billion, $221.6 billion, and $245.2 billionRevenue growth was raised to 28%, 17%, and 11%, driven mainly by memory costs and specification upgrades.
• AI server racks shipments2026E/2027E/2028E are 55k, 105k, and 163k2027E and 2028E were raised by 16% and 20% versus the previous forecast.
• 8-GPU-equivalent AI server shipments2026E/2027E/2028E are 1.858m, 2.403m, and 2.558mRaised by 18%, 24%, and 11% versus the previous forecast.
• Implied AI chips2026E/2027E/2028E are 18.847m, 26.750m, and 32.214mASIC share is expected to be 50%, 52%, and 55%, above the prior 2026E/2027E figures of 43% and 50%.
• U.S. CSP capexYoY growth of 76%, 35%, and 8% in 2026E-2028ECovers Microsoft, Amazon, Meta, Alphabet, and Oracle, and the 2028E forecast was raised to an implied cumulative level of $114.5 billion.
• China CSP capexYoY growth of 80%, 20%, and 18% in 2026E-2028ECovers ByteDance, Tencent, Alibaba, and Baidu, with forecast upgrades of 37%, 44%, and 55%.

Impact & implications

For investment implications, the report leans toward the AI server supply chain and the upward CSP capex chain. Beneficiaries include ODM direct, silicon photonics, liquid cooling, slide rails, chassis, CCL, PCB, optical interconnects, foundry, probe cards, sockets, BMC/fabless chips, FT handlers, and related names. GPU platforms still benefit from high-end rack expansion, while ASIC platforms gain higher penetration from CSP in-house design and customization demand. General servers are no longer just a low-growth category; they are supported by memory costs, CPU platform upgrades, and spillover from AI workloads.

Risks

• AI infrastructure capex could come in below expectations, causing server TAM and supply-chain revenue to miss forecasts.
• Changes in ASIC penetration could reallocate value across GPUs, ASICs, and related supply chains.
• Price and supply fluctuations in key components such as memory, HBM, networking, liquid cooling, and power may affect server ASPs and delivery timelines.
• High concentration among major CSP customers means order changes can amplify supply-chain volatility.
• A slower-than-expected rollout of new networking architectures and server specification upgrades could weaken growth assumptions beyond 2027E.
• PC and smartphone shipment forecasts remain weak, which may affect sentiment toward broader technology hardware demand.

What to watch

• Whether actual AI server rack shipments in 2026E-2028E come close to 55k, 105k, and 163k.
• Whether ASIC share of AI chips rises along the path of 50%, 52%, and 55%.
• Whether capex budgets for U.S. and Chinese CSPs deliver the upgraded high-growth forecasts.
• Changes in the share of ODM direct players such as Hon Hai/FII in AI server racks.
• How HBM, memory, and networking component costs affect server ASPs and TAM.
• Differences in rack shipments between NVIDIA and AMD platforms, and the pace of CSP in-house ASIC rollouts.
• Whether Agentic AI continues to drive general server specification upgrades and CPU server demand.