# semiconductor market for AI data center chips Market Research Report - United States

**Generated on:** 2026-09-23 08:08:43.836506  
**Industry:** semiconductor market for AI data center chips  
**Geography:** United States  
**Details:** Focus on AI chip demand. Reference dataset: http://internal.parallel.ai/

---

# US AI Data-Center Chips: Demand, Bottlenecks, and Winners

## Executive Summary

- **Demand Is Expanding Faster Than The Broader Chip Cycle**: Gartner projects more than **$1.3T** of worldwide semiconductor revenue in 2026, with AI semiconductors at approximately **30%** of the total; AMD separately projects data-center AI accelerator TAM above **$500B by 2028** [25][24] -> Treat the United States as the leading demand center, but use global forecasts as context rather than mislabeling them as US TAM.
- **Hyperscalers Set The Near-Term Order Book**: Five major hyperscalers have announced about **$800B** of 2026 capex and **$1T** for 2027, with the vast majority directed to AI infrastructure [19] -> Track Amazon, Alphabet, Meta, Microsoft, and Oracle procurement and power availability as the primary demand indicators.
- **Merchant Accelerators Remain The Scale Leader**: NVIDIA reported **$89.0B** of quarterly data-center revenue, up **117%** year-over-year, while AMD reported **$6.718B**, up **107%** [27][4] -> Buyers should preserve NVIDIA compatibility while qualifying AMD as a second source where software and workload fit permit.
- **Custom Silicon Is Growing Alongside GPUs**: Broadcom reported **$16.7B** of quarterly AI semiconductor revenue, up **221%**, and Amazon said both its AI business and chips business exceeded **$25B annual revenue run rates** [3][26] -> Expect a mixed market in which GPUs serve broad workloads and ASICs take repetitive, high-volume training and inference.
- **Inference Changes The Winning Specification**: AMD forecasts inference demand above **80% CAGR**, while Microsoft's Maia 200 emphasizes HBM bandwidth, low-precision compute, and performance per dollar [24][14] -> Prioritize tokens per dollar, tokens per watt, memory bandwidth, latency, and software portability instead of peak FLOPS alone.
- **HBM, Packaging, And Power Gate Realized Demand**: TSMC says advanced-packaging capacity remains very tight, while DOE projects US data centers could consume **325-580 TWh**, or **6.7%-12.0%** of US electricity, by 2028 [18][6] -> Secure packaging, memory, networking, cooling, and energization as one system; a chip allocation without power is not deployable demand.
- **Policy Creates Resilience And Friction At The Same Time**: The CHIPS Act appropriated **$39B** for incentives, while current export rules subject specified advanced chips for China and Macau to licensing and conditions -> Separate domestic capacity strategy from global revenue exposure and maintain export-compliant product roadmaps.
- **The Main Risk Is Capital Digestion, Not Immediate End-Demand Collapse**: A Goldman Sachs scenario estimates **$7.6T** of cumulative AI capex from 2026 through 2031, but explicitly says the result is a sensitivity framework rather than a forecast [16] -> Stage commitments against utilization, contracted workloads, electricity milestones, and falling inference cost rather than extrapolating headline capex indefinitely.

**Bottom line:** US AI-chip demand is exceptionally strong and broadening from training GPUs into inference ASICs, HBM, networking, and rack-scale systems. The strongest positions combine a software ecosystem, access to leading-edge foundry and packaging, and a credible path to power-efficient deployment. The key downside is not lack of use cases; it is a mismatch among capital spending, monetization, supply-chain timing, and available power.

## 2026 Demand Baseline: $1.3T Semiconductors, 30% AI Share

There is no reliable public figure for a standalone US AI data-center semiconductor market. Supplier revenues are global, chip billing locations do not consistently identify deployment geography, and hyperscaler capex includes buildings, servers, networking, power, and other equipment. This report therefore triangulates US demand through three indicators: worldwide AI-chip forecasts, spending by US-headquartered hyperscalers, and current supplier segment revenue.

| Demand indicator | Current evidence | Scope and interpretation |
|---|---:|---|
| Worldwide semiconductor revenue | **$1,320.2B in 2026**, up **64%** from $805.3B in 2025 | Global context; AI semiconductors are approximately **30%** of 2026 revenue [25] |
| Server and data-center semiconductors | **$330B in 2025 -> $1.2T in 2030**, a projected **29% CAGR** | Global forecast covering more than accelerators; it is not a US TAM [19] |
| Data-center AI accelerator TAM | **$45B -> more than $500B by 2028**, above **60% CAGR** | AMD vendor forecast; useful as an order-of-magnitude signal, not an independent consensus [24] |
| Five-hyperscaler capex | About **$800B planned for 2026** and **$1T for 2027** | Largely AI infrastructure, but not all semiconductor content and not entirely US deployment [19] |
| AI buildout sensitivity | **$765B annual AI capex in 2026 -> $1.6T in 2031** | Goldman scenario; explicitly not a forecast [16] |

The table shows why a single market-size number would create false precision. The actionable conclusion is the convergence of multiple demand signals: chip-industry revenue, cloud capex, and reported supplier growth all point upward, but their scopes overlap and must not be added together.

For this report, the market includes merchant GPUs and accelerators, custom ASICs, server CPUs, DPUs and NICs, Ethernet and proprietary scale-up fabrics, optical and electrical interconnects, HBM, and advanced packaging. It excludes data-center real estate and power-generation revenue, although both constrain when semiconductor demand can be recognized. McKinsey's base case places the total semiconductor market at **$2.3T in 2030**, within a **$1.7T-$2.6T** range, and says the outcome requires high data-center investment, a stable economy, and sustained pricing for leading-edge chips and memory [19].

The supplied reference endpoint did not expose market records; it returned an API-key resolution error [20]. No figure in this report is attributed to that dataset. Decision-makers should replace public proxies with the dataset's US shipment, deployment, and customer-level records if authorized access becomes available.

## Inference Pulls Demand Toward Bandwidth, Efficiency, and ASICs

Training remains compute intensive, but inference is becoming the larger recurring workload. AMD's forecast shows inference demand growing at more than **80% CAGR**, faster than its overall accelerator TAM projection [24]. This changes procurement from a race for maximum training throughput to a multi-variable optimization problem: token cost, latency, power, memory capacity, memory bandwidth, utilization, networking, and developer productivity.

| Chip layer | Demand mechanism | Representative players | Purchase criterion |
|---|---|---|---|
| Merchant accelerators | Flexible training and inference across many models | NVIDIA, AMD, Intel | Software ecosystem, time to solution, rack-scale throughput |
| Custom ASICs | Lower cost and power for stable, high-volume workloads | Google TPU, Amazon Trainium/Inferentia, Microsoft Maia, Meta MTIA; Broadcom and Marvell as design partners | Workload predictability, volume, compiler maturity |
| Server CPUs and DPUs | Host orchestration, data preprocessing, security, storage and network offload | AMD, Intel, NVIDIA | Memory, I/O, core efficiency, platform integration |
| HBM | Feeds large models and reduces accelerator idle time | SK hynix, Micron, Samsung | Bandwidth, capacity, yield, thermal behavior |
| Network and optics | Connects accelerators within racks and across clusters | NVIDIA, Broadcom, Marvell, Astera Labs, Credo, Coherent | Scale-up latency, scale-out bandwidth, power per bit |
| Foundry and packaging | Produces and integrates logic chiplets with HBM | TSMC, Samsung, Intel Foundry and OSAT partners | Leading-edge yield, CoWoS-class capacity, cycle time |

The market is best understood as a system rather than an accelerator category. More accelerator shipments increase demand for HBM, switches, retimers, optics, CPUs, DPUs, substrates, and packaging; bottlenecks in any one layer can delay the whole cluster.

**Case study - Microsoft's Maia 200.** Microsoft positions Maia 200 as an inference accelerator built on TSMC's 3 nm process with **216 GB of HBM3e at 7 TB/s**, low-precision FP4/FP8 compute, and a **750 W** SoC envelope. Microsoft claims **30% better performance per dollar** than the latest-generation hardware then in its fleet, but this is a vendor claim rather than an independent benchmark [14]. Deployment began in Microsoft's US Central region near Des Moines, with US West 3 near Phoenix next [14].

Maia illustrates the custom-silicon mechanism: a hyperscaler owns the workload, compiler, fleet scheduler, and data center, so it can optimize the chip for utilization and token economics. Yet Microsoft still describes a heterogeneous infrastructure, not a single-chip replacement strategy [14]. The implication is coexistence: merchant GPUs dominate broad and fast-changing workloads, while first-party silicon gains share where workload volume repays design cost. Suppliers should therefore offer open compilers, chiplets, HBM integration, Ethernet, and rack designs that make them useful in both merchant and custom platforms.

## Four Supplier Models Capture Distinct Profit Pools

The latest reported quarters show extraordinary growth across different layers. These figures are global, cover different fiscal periods, and use different segment definitions, so they should not be summed into a market size or used to calculate market share.

| Supplier | Latest relevant metric | Strategic model | Main advantage | Main risk |
|---|---:|---|---|---|
| NVIDIA | **$89.0B** data-center revenue, up **117% YoY**; total revenue **$96.2B** [27] | Full-stack merchant platform: GPU, CPU, networking, systems and software | CUDA ecosystem, rack-scale integration, rapid product cadence | Customer concentration, export limits, third-party foundry/packaging dependence |
| AMD | **$6.718B** data-center revenue, up **107% YoY** and **58%** of total revenue [4] | Challenger platform: Instinct GPU, EPYC CPU, Pensando DPU, ROCm and Helios racks | CPU franchise, open software direction, second-source value | Software maturity, HBM/component access, export and execution risk |
| Broadcom | **$16.7B** AI semiconductor revenue, up **221% YoY** and **54% QoQ** [3] | Custom accelerators plus Ethernet, optics, SerDes and switching | Deep hyperscaler co-design and connectivity portfolio | Large-customer concentration and outsourced manufacturing |
| Marvell | **$1.651B** data-center revenue, **74%** of Q4 sales, up **21% YoY** and **9% QoQ** [1] | Custom compute, electro-optics, switching and data infrastructure | Design wins across custom silicon and connectivity | Smaller scale, customer concentration and vertical integration by customers |

**Case study - NVIDIA scale versus AMD acceleration.** NVIDIA's quarterly data-center revenue is more than an order of magnitude above AMD's, and NVIDIA reported **75.0%** GAAP and non-GAAP gross margins [27]. Its next-quarter outlook was **$108B**, plus or minus 2%, without assuming any data-center compute revenue from China [27]. This demonstrates the economic power of a full stack: software increases switching costs, while networking and complete racks capture more content per deployment.

AMD's smaller base is growing rapidly. Its Q2 2026 total revenue reached **$11.536B**, up **50%**, and management expected approximately **$13B**, plus or minus $300M, in Q3 as second-half data-center sales accelerated [4]. AMD's opportunity is not simply a lower-priced GPU. It can combine EPYC CPUs, Instinct accelerators, Pensando DPUs, ROCm, and rack-scale systems, but must convert design wins into broad software support and repeat deployments.

**Case study - Broadcom and Marvell monetize customization without owning the cloud.** Broadcom expects AI semiconductor revenue to rise to **$21.7B** in Q4, up **236%** year-over-year [3]. Marvell reported record FY2026 revenue of **$8.195B**, up **42%**, with AI demand as the driver, while warning that customers may develop their own solutions or vertically integrate [1]. The divergence is important: custom silicon does not eliminate semiconductor vendors, but shifts value toward firms that supply design IP, packaging interfaces, Ethernet, SerDes, and optics. The recommended portfolio approach is exposure to both merchant platforms and the enabling layers that benefit regardless of which accelerator wins.

## Four Hyperscalers Spend Heavily While Internalizing Silicon

US-headquartered hyperscalers are simultaneously the largest customers, chip architects, cloud distributors, and AI application owners. Their spending is the clearest near-term demand signal, but capex periods and definitions differ and are not directly comparable.

| Hyperscaler | Current demand metric | Internal silicon | Strategic implication |
|---|---:|---|---|
| Amazon | Q2 property and equipment purchases **$54.208B**; TTM **$173.028B**; AWS sales **$42.2B**, up **37%** [26] | Trainium, Inferentia, Graviton | Uses custom chips to improve cloud economics while continuing to offer merchant accelerators |
| Alphabet | Q2 capex **$44.9B**; 2026 guidance **$195B-$205B**; cloud backlog **$514B** | TPU | Deepest vertically integrated accelerator history; still uses third-party capacity as a bridge |
| Microsoft | FY2026 property and equipment additions **$115.948B**; Azure growth **43%** [23] | Maia, Cobalt | Builds first-party silicon inside a heterogeneous Azure fleet |
| Meta | Q2 capex **$31.08B**; 2026 guidance **$130B-$145B** [22] | MTIA | Optimizes ranking, recommendation and GenAI workloads while retaining external silicon |

Alphabet's spending mix makes the semiconductor pull-through visible: approximately **60%** of Q2 technical-infrastructure investment went to servers and **40%** to data centers and networking. It nevertheless described the environment as supply constrained and planned to use third-party capacity while adding internal capacity [21]. This means announced capex can translate into merchant GPU demand, internal TPU deployment, and networking purchases at the same time.

Amazon offers the clearest custom-chip commercialization case. It said both its AI business and chips business exceeded **$25B annual revenue run rates**, each growing at triple-digit percentages, and reported multi-year, multi-gigawatt Trainium commitments from Anthropic and OpenAI [26]. The mechanism is economic rather than ideological: owned silicon can lower price-performance for predictable workloads, while AWS still needs broad accelerator choice to attract customers.

Meta exposes the capital-return tension. Its operating margin fell to **31%** in Q2 2026 from **43%** a year earlier while it raised the lower end of its capex range [22]. Meanwhile, MTIA expanded across four successive generations beyond ranking and recommendation inference into training and GenAI workloads [17]. This is not evidence that AI spending has failed, but it is an early warning that depreciation and infrastructure costs can reach the income statement before new revenue matures.

The practical recommendation is to segment demand by workload. Merchant GPUs should remain strongest in frontier training, rapidly changing models, and external cloud choice. Custom ASICs should gain in recommendation, ads, internal inference, and stable high-volume models. Networking, HBM, and advanced packaging benefit in both cases.

## HBM, CoWoS, And Power Cap Deployable Chip Demand

Demand is no longer constrained by logic dies alone. The critical path runs from leading-edge wafers through HBM and advanced packaging to network fabrics, cooling, and utility interconnection.

| Constraint | Current metric or signal | Demand effect | Recommended response |
|---|---|---|---|
| Leading-edge foundry | TSMC expects 2026 capex near the high end of **$52B-$56B** [18] | Capacity additions lag orders because a new fab takes years | Use multi-year capacity agreements and multiple product tape-outs |
| Advanced packaging | TSMC says capacity remains very tight and it is working with OSAT partners | A packaged GPU with HBM, not a bare die, is the saleable unit | Reserve packaging with wafers and memory, not afterward |
| HBM and DRAM | Gartner projects 2026 memory revenue of **$633.3B** and estimated DRAM prices up **125%** [25] | Memory inflation raises system cost and can crowd out non-AI demand | Qualify multiple HBM vendors and optimize memory per workload |
| US electricity | **176 TWh and 4.4%** of US use in 2023; **325-580 TWh and 6.7%-12.0%** projected for 2028 [6] | Power availability can delay cluster commissioning and chip acceptance | Tie purchase schedules to utility, generation, cooling, and interconnect milestones |
| Networking and optics | NVIDIA, Broadcom, and Marvell all report AI-driven networking demand [3][27] | Larger clusters require more bandwidth and lower power per bit | Treat scale-up and scale-out fabrics as core silicon budgets |

**Case study - TSMC turns AI demand into a US capacity decision.** TSMC says AI demand from cloud providers remains extremely robust and that large CoWoS remains the main advanced-packaging supply approach [18]. Its second Arizona fab will use 3 nm technology and is scheduled to begin volume production in the second half of 2027, with further Arizona expansion intended to serve leading-edge US customers.

The case shows both opportunity and timing risk. Domestic fabrication can reduce geographic concentration, but a fab takes **two to three years** to build and another **one to two years** to ramp [18]. Near-term US demand will therefore continue to depend on Asian wafer, HBM, and packaging ecosystems even as domestic capacity grows.

Power is the harder local constraint. DOE's 2028 range is wide because model efficiency, deployment pace, and facility utilization remain uncertain [6]. A later LBNL update estimates data centers could reach **11.8%** of US electricity use by 2030, with a **9.5%-15.3%** scenario range [9]. Chip suppliers should treat power-efficient architecture as a market-access feature, while data-center buyers should measure deployable accelerator demand in energized megawatts and contracted tokens, not purchase orders alone.

## $39B CHIPS Support Meets Export And Construction Friction

US industrial policy improves long-run resilience but does not remove execution risk. The CHIPS Act appropriated **$39B** through FY2026 for incentives; the associated investment tax credit began at **25%** and was later increased to **35%**. By January 31, 2025, 19 companies had received up to **$30.7B** for 40 commercial fab projects. Approximately **59%** of award funding targeted leading-edge logic and **20%** leading-edge memory.

That allocation aligns with AI-chip demand, but project schedules reveal the failure modes. TSMC's second Phoenix fab moved toward a 2027 mass-production schedule amid AI demand, while Intel delayed its New Albany, Ohio opening to 2030 and Micron delayed its first Clay, New York fab to the third quarter of 2030. Subsidies improve project economics, but customer commitments, permitting, labor, equipment, and ramp yields still determine delivery.

Export policy creates a separate demand segmentation. Effective January 15, 2026, BIS changed specified advanced-computing chips such as NVIDIA H200 equivalents from a presumption of denial to case-by-case review for China and Macau if supply, foundry-capacity, security, and US testing conditions are satisfied. This is not unrestricted access. NVIDIA's own outlook assumed no China data-center compute revenue, while AMD identifies export controls and prior MI308-related charges as material risks [27][4].

| Policy channel | Opportunity | Risk | Decision implication |
|---|---|---|---|
| Fab incentives and tax credits | More US leading-edge logic and memory capacity | Multi-year delays and uncertain ramps | Underwrite on production milestones, not announced investment |
| Export licensing | Conditional access for some products | Revenue volatility, compliance cost, redesign and diversion risk | Maintain country-specific products and conservative forecasts |
| Domestic sourcing | Better resilience for government and strategic buyers | US cost premium and continued foreign packaging/HBM dependence | Measure resilience at the complete system level |
| Energy and permitting policy | Can unlock new data-center regions | Grid queues and local opposition can strand chip commitments | Coordinate semiconductor, utility, and site policy |

The core implication is temporal: export controls affect revenue now, while domestic fab incentives change supply over several years. Vendors should not treat future US capacity as a solution to current packaging or HBM shortages, and investors should separate policy announcements from qualified, yielding output.

## Five Failure Modes Define The Scenario Range

The base case remains strong, but semiconductor cycles often turn when supply, customer budgets, and end-market monetization lose synchronization. The following risk register converts that general warning into observable triggers.

| Failure mode | Mechanism | Leading indicator | Mitigation |
|---|---|---|---|
| AI monetization lags capex | Depreciation and energy costs rise before revenue and productivity gains | Cloud backlog conversion, token revenue, enterprise renewal, margins | Stage clusters; require utilization and revenue gates |
| Accelerators last longer | Replacement demand falls when useful life extends | Secondary-market prices, fleet age, utilization by generation | Use modular racks and redeploy older chips to inference |
| Power and construction slip | Chips arrive before energized capacity | Utility interconnection dates, transformer and turbine lead times | Align semiconductor deliveries with site milestones |
| Customers vertically integrate | Custom chips displace part of merchant demand and pressure margins | TPU, Trainium, Maia and MTIA deployment share | Sell enabling IP, networking, HBM interfaces and software |
| Supply catches demand abruptly | HBM, packaging or leading-edge pricing normalizes | Lead times, inventory, prepayments, memory pricing | Avoid double ordering and negotiate flexible commitments |

Goldman estimates AI accelerators typically have a **four-to-six-year** useful life and shows that small changes in life can move cumulative investment by hundreds of billions of dollars [16]. It also notes that new AI facilities increasingly cost approximately **$15M-$20M per MW**, versus roughly $10M per MW for traditional hyperscale facilities. The mechanism is powerful: longer chip life lowers replacement demand, while higher facility cost raises the return threshold for every deployed accelerator.

McKinsey's **$2.3T** 2030 semiconductor base case requires **$2.7T** of cumulative data-center capex, tangible enterprise returns, and inference-cost reductions of three to four orders of magnitude. It also assumes tight leading-edge and HBM supply helps sustain pricing [19]. Memory supply could begin catching demand around 2027, reducing prices and HBM margins even if unit demand remains healthy.

| US AI-chip demand scenario | Conditions | Expected market behavior | Action |
|---|---|---|---|
| Downside / digestion | Capex pauses, utilization disappoints, power slips, chip life extends | Merchant accelerator growth decelerates; memory and packaging pricing normalize first | Reduce speculative inventory; favor networking upgrades and efficiency |
| Base / constrained expansion | Hyperscaler plans proceed, custom ASICs coexist with GPUs, packaging and power remain tight | Broad growth across accelerators, HBM, networking and advanced packaging | Secure capacity, diversify suppliers, gate deployments by power |
| Upside / elastic inference | Agentic workloads expand tokens, lower cost stimulates more use, new power comes online | ASIC and GPU volumes both rise; networking and memory become larger value pools | Invest ahead in bandwidth, optics, cooling and software portability |

These are directional US demand scenarios, not market-size forecasts. The decision rule is to watch physical and economic throughput: energized megawatts, HBM and packaging lead times, accelerator utilization, tokens served, and cloud backlog conversion.

## Strategic Actions For Vendors, Buyers, And Policymakers

| Participant | Priority action | Why it matters now | 12-24 month metric |
|---|---|---|---|
| Merchant accelerator vendor | Build complete rack economics and portable software | Buyers compare tokens per dollar and watt, not isolated chip FLOPS | Production utilization, software adoption, network attach rate |
| Custom-silicon supplier | Target repeatable workloads and sell enabling IP | Hyperscalers need ASIC economics without rebuilding every subsystem | Tape-out conversion, multi-generation design wins, HBM yield |
| Hyperscaler or AI cloud | Run a merchant-plus-custom portfolio | Workloads differ and no single architecture optimizes every model | Cost per million tokens, job completion time, fleet utilization |
| Memory, packaging or networking supplier | Contract against actual accelerator ramps | These layers can capture scarcity value but face inventory corrections | Lead time, prepayment quality, customer concentration |
| Data-center developer or utility | Co-plan compute, power, cooling and network | A delayed substation can strand depreciating chips | Energized MW, PUE, time from delivery to revenue service |
| Investor or lender | Underwrite cash flow by layer, not an undifferentiated AI theme | Merchant chips, ASIC design, memory and power have different cycles | Backlog quality, gross margin, capex intensity, customer exposure |
| Policymaker | Link incentives to milestones and complete-system resilience | Fab announcements alone do not supply packaged accelerators | Qualified output, yield, packaging/HBM capacity, workforce readiness |

For vendors, the highest-value strategic position is not necessarily the largest logic die. It is control of a bottleneck plus integration into the customer's software and system architecture. NVIDIA demonstrates this through CUDA and rack-scale systems; Broadcom and Marvell do it through custom design and connectivity; TSMC and HBM suppliers do it through scarce manufacturing capabilities [3][27].

For buyers, dual sourcing should be workload specific. Qualifying AMD or an internal ASIC without porting models, compilers, observability, and scheduling does not create real leverage. Buyers should benchmark end-to-end service-level cost under realistic memory, network, and utilization conditions, then allocate frontier, general-purpose, and repetitive inference workloads separately.

For capital providers, the main diligence question is who bears utilization risk. Chip vendors recognize demand at shipment; clouds recover it over years; utilities and developers invest before full occupancy; enterprises monetize it through productivity or product revenue. Contract structures should match that timing, with staged drawdowns and covenants tied to power, deployment, and revenue milestones.

## Synthesis: Three Architectures, One Constrained System

| Strategy | Primary mechanism | Scope | Trade-off | Evidence base | Time horizon |
|---|---|---|---|---|---|
| NVIDIA full-stack merchant platform | General-purpose acceleration plus proprietary software and networking | Broadest training and inference coverage | Highest ecosystem leverage, but concentration and export exposure | **$89.0B** quarterly data-center revenue and **75%** gross margin [27] | Immediate and multi-generation |
| AMD open challenger platform | Combine GPU, CPU, DPU and rack systems with ROCm | Broad workloads and second-source deployments | Faster growth from a smaller base; software execution remains critical | **$6.718B** quarterly data-center revenue, up **107%** [4] | Immediate, with share gains over 1-3 years |
| Hyperscaler custom ASICs with Broadcom/Marvell | Co-design silicon for owned, repeatable workloads | High-volume internal and cloud services | Better workload economics, less flexibility and more design risk | Broadcom AI revenue growth plus Trainium, TPU, Maia and MTIA deployments [3][26][14][17] | Multi-year, workload by workload |
| TSMC, HBM and connectivity enablers | Supply scarce manufacturing, memory bandwidth and data movement | Benefits across both merchant and custom chips | Capital intensive and cyclical; exposed to customer concentration | Tight advanced packaging and strong AI signals | Structural through the buildout, cyclical at the margin |

The non-obvious conclusion is that custom silicon is both a threat to merchant GPUs and an accelerator of the total semiconductor opportunity. If ASICs merely replace GPUs one for one, they pressure merchant revenue. If lower cost expands inference volume, they stimulate a Jevons-like rebound in compute and increase demand for HBM, networking, packaging, and power. Goldman similarly notes that ASICs reduce infrastructure spending only when compute demand is inelastic; with elastic demand, cheaper compute enables broader use.

The second tension is between national capacity and global interdependence. CHIPS incentives can add US leading-edge production, but current AI systems still rely on globally concentrated HBM, packaging, substrates, and equipment. TSMC's Arizona expansion improves resilience, yet its ramp schedule and tight packaging commentary show why the transition is gradual.

The third tension is financial. Supplier growth and hyperscaler commitments support a powerful base case, but Meta's margin compression, multi-year fab delays, and the sensitivity of buildout economics to chip life show how quickly accounting returns can diverge from physical demand [22][16]. The best strategy is therefore balanced rather than maximal: combine merchant and custom compute, invest in the enabling bottlenecks, contract power before chips, and scale only when utilization and token economics validate the next tranche.

For the US market, the likely outcome is not one winning chip. It is a layered ecosystem in which NVIDIA retains broad platform leadership, AMD gains where second sourcing and heterogeneous compute matter, Broadcom and Marvell monetize hyperscaler customization and connectivity, and TSMC plus HBM suppliers govern how fast demand can become deployed capacity. The winners will be those that turn silicon into reliable, powered, software-ready tokens at the lowest full-system cost.

## References

1. *Marvell Technology, Inc. Reports Fourth Quarter and Fiscal Year 2026 Financial Results | Marvell Technology, Inc. (MRVL)*. https://investor.marvell.com/news-events/press-releases/detail/1011/marvell-technology-inc-reports-fourth-quarter-and-fiscal-year-2026-financial-results
2. *Fiscal.ai /v2/company/segments-and-kpis*. https://fiscal.ai
3. *Broadcom Inc. Announces Third Quarter Fiscal Year 2026 Financial Results and Quarterly Dividend | Broadcom Inc.*. https://investors.broadcom.com/news-releases/news-release-details/broadcom-inc-announces-third-quarter-fiscal-year-2026-financial
4. *AMD Reports Second Quarter 2026 Financial Results :: Advanced Micro Devices, Inc. (AMD)*. https://ir.amd.com/news-events/press-releases/detail/1295/amd-reports-second-quarter-2026-financial-results
5. *Global AI Investment Is Forecast to Exceed $1 Trillion in 2026 | Goldman Sachs*. https://www.goldmansachs.com/insights/articles/global-investment-is-forecast-to-exceed-1-trillion-in-2026
6. *DOE Releases New Report Evaluating Increase in Electricity Demand from Data Centers | Department of Energy*. https://www.energy.gov/articles/doe-releases-new-report-evaluating-increase-electricity-demand-data-centers
7. *2026 Market Outlook – “Focus on the HBM-Led Memory Supercycle” – SK hynix Newsroom*. https://news.skhynix.com/en/2026-market-outlook-focus-on-the-hbm-led-memory-supercycle/
8. *Semiconductor Fabrication Facilities Funded by the CHIPS Act: Project Status and Considerations for Congress | Congress.gov | Library of Congress*. https://www.congress.gov/crs-product/R49031
9. *Sector-Wide Modeling & Forecasting | Center of Expertise for Data Center Energy*. https://datacenters.lbl.gov/modeling-forecasting
10. *More compute for AI, not less | Deloitte Insights*. https://www.deloitte.com/us/en/insights/industry/technology/technology-media-and-telecom-predictions/2026/compute-power-ai.html
11. [
      Federal Register
       :: 
      Revision to License Review Policy for Advanced Computing Commodities
    ](https://www.federalregister.gov/documents/2026/01/15/2026-00789/revision-to-license-review-policy-for-advanced-computing-commodities)
12. *Guidance Regarding Enforcement of License Requirements for Advanced Computing Items for Entities Headquartered in Country Group D:5 and Macau [May*. https://www.bis.gov/media/documents/bis-guidance-may-31-2026.pdf
13. *CoWoS® - Taiwan Semiconductor Manufacturing Company Limited*. https://3dfabric.tsmc.com/english/dedicatedFoundry/technology/cowos.htm
14. *Maia 200: The AI accelerator built for inference - The Official Microsoft Blog*. https://blogs.microsoft.com/blog/2026/01/26/maia-200-the-ai-accelerator-built-for-inference
15. *Ironwood: The first Google TPU for the age of inference*. https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/ironwood-tpu-age-of-inference
16. *Tracking Trillions: The Assumptions Shaping the Scale of the AI Build-Out | Goldman Sachs*. https://www.goldmansachs.com/insights/articles/tracking-trillions-the-assumptions-shaping-scale-of-the-ai-build-out
17. *Four MTIA Chips in Two Years: Scaling AI Experiences for Billions*. http://ai.meta.com/blog/meta-mtia-scale-ai-chips-for-billions
18. *Q1 2026 Taiwan Semiconductor Manufacturing Co Ltd Earnings Call (Chinese, English) on April 16, 2026 / 6:00AM*. https://investor.tsmc.com/chinese/encrypt/files/encrypt_file/reports/2026-04/3cef85204275f94fd111485cfdf4adb3c0263c45/TSMC%201Q26%20Transcript.pdf
19. *The 2 Point 3 Trillion Dollar Horizon How Ai Is Rewriting The Semiconductor Story*. https://www.mckinsey.com/industries/semiconductors/our-insights/the-2-point-3-trillion-dollar-horizon-how-ai-is-rewriting-the-semiconductor-story
20. *internal.parallel.ai*. http://internal.parallel.ai/
21. [
	Alphabet Investor Relations - 2026 Q2 Earnings Call
](http://abc.xyz/investor/events/event-details/2026/2026-Q2-Earnings-Call-2026-GgTAq7Is0z/default.aspx)
22. [
	Meta - Meta Reports Second Quarter 2026 Results
](https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Reports-Second-Quarter-2026-Results/default.aspx)
23. *FY26 Q4 - Press Releases - Investor Relations - Microsoft*. https://www.microsoft.com/en-us/investor/earnings/fy-2026-q4/press-release-webcast
24. *CAUTIONARY STATEMENT*. https://www.amd.com/content/dam/amd/en/documents/corporate/events/advancing-ai-2025-distribution-deck.pdf
25. *Gartner Forecasts Worldwide Semiconductor Revenue to Exceed $1.3 Trillion in 2026*. https://www.gartner.com/en/newsroom/press-releases/2026-04-08-gartner-forecasts-worldwide-semiconductor-revenue-to-exceed-us-dollars-one-point-3-trillion-in-2026
26. [
	Amazon.com, Inc. - Amazon.com Announces Second Quarter Results
](https://ir.aboutamazon.com/news-release/news-release-details/2026/Amazon-com-Announces-Second-Quarter-Results)
27. *NVIDIA Announces Financial Results for Second Quarter Fiscal 2027 | NVIDIA Newsroom*. https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-second-quarter-fiscal-2027