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

**Generated on:** 2026-09-23 08:07:22.127890  
**Industry:** semiconductor market for AI data center chips  
**Geography:** United States  
**Details:** Focus on AI chip demand. Reference dataset: http://dev-kube-vespa-bdset-large-query-service.service-dir.us-central1-a.c.dev-index-svc-prpg.internal:8080/search/?hits=1

---

# U.S. AI Data-Center Chips: Demand, Bottlenecks, and Winners

## Executive Summary

- **Capex Sets the Demand Floor**: J.P. Morgan estimates that the five largest U.S. hyperscalers will spend **$697B in 2026**, up **$173B** from its estimate at the beginning of the year [21] -> suppliers should plan for strong near-term orders but tie capacity additions to customer commitments, energization dates, and utilization.
- **Forecasts Bound, But Do Not Size, the U.S. Market**: Omdia projects a global data-center AI-chip market of **$286B in 2030**, while Bloomberg Intelligence projects a broader global accelerator market of **$604B in 2033** [20][22] -> use the forecasts as definition-sensitive scenarios, not as a precise U.S. TAM.
- **NVIDIA Remains the Merchant Benchmark**: NVIDIA recorded **$193.737B** of FY2026 Data Center revenue, up from **$115.186B** in FY2025 [13] -> buyers need a deliberate portability plan before assuming that alternative silicon will deliver lower total cost.
- **Custom ASICs Are the Structural Share Challenger**: Bloomberg Intelligence expects custom ASICs to grow at **27% CAGR** and rise from **8% of accelerator revenue in 2024 to 19% in 2033** [22] -> Broadcom, Marvell, hyperscaler design teams, HBM vendors, and packaging suppliers can grow even if merchant-GPU share declines.
- **Demand Is Broadening From Training to Inference**: Reasoning models increase token generation, but smaller specialized models and generation-to-generation efficiency gains restrain compute per task [20] -> build portfolios across GPUs, ASICs, CPUs, memory, and networking rather than extrapolating training clusters indefinitely.
- **The Rack, Not the Processor, Is the Scarce Product**: TSMC CoWoS integrates multiple logic devices and HBM through advanced 2.5D packaging [19] -> procurement must synchronize accelerators, HBM, packaging, switches, optics, cooling, and power.
- **Power Becomes a Deployment Ceiling**: LBNL's 2030 U.S. scenarios range from **521 to 843 TWh**, or **9.5% to 15.3%** of national electricity use [23] -> rank projects by time-to-power and interconnection certainty, not merely land cost or announced GPU count.
- **The Core Downside Is Monetization Lag**: J.P. Morgan identifies overbuilding, uncertain end demand, technology obsolescence, utilization volatility, and customer concentration as financing risks [21] -> stage deployments against contracted demand, cloud growth, revenue per watt, and achieved utilization.

**Bottom line.** The U.S. AI data-center chip market remains in a powerful build cycle as of September 23, 2026. Demand is real, but it is changing shape: from individual accelerators toward integrated racks; from training-only clusters toward mixed training and inference; and from a merchant-GPU near-monopoly toward a coexistence of GPUs and custom ASICs. The best-positioned suppliers own a control point in that full system. The most exposed participants depend on one customer, one package technology, speculative tenants, or power that has not yet been secured.

**Reference-dataset note.** The supplied internal URL was tested exactly with `hits=1`. Static extraction returned no records, and the browser returned `502 Bad Gateway: Upstream proxy error: 454`. No unseen fields or values from that endpoint are represented in this report.

## Market Scope and Demand: From Chips to $697B of Infrastructure

This report defines the market as semiconductors deployed in U.S. AI data centers: merchant GPUs and CPUs; custom accelerators; HBM and related memory; switch, NIC, DPU, and optical-connectivity silicon; and the advanced packaging needed to integrate them. It excludes edge AI, client AI PCs and phones, and non-semiconductor construction spending. Because U.S.-designed chips are fabricated and packaged globally and deployed by clouds across regions, a clean U.S.-only chip-revenue series is not publicly standardized. The most decision-useful U.S. measures are therefore hyperscaler capex, supplier data-center revenue, cloud growth, committed power, and supply-chain capacity.

| Demand layer | What buyers procure | Demand mechanism | Best leading indicator |
|---|---|---|---|
| Training compute | GPU or custom-ASIC clusters | Larger frontier and multimodal models | Cluster orders, HBM allocation, training utilization |
| Inference compute | GPUs, ASICs, CPUs, memory | Tokens served, agents, search and recommendation | Token volume, cost per token, inference utilization |
| Scale-up and scale-out | Switches, NICs, DPUs, retimers, optics | More chips must behave as one system | Network revenue, optical shipments, port speeds |
| Memory and packaging | HBM, interposers, advanced substrates | Bandwidth and package area rise with accelerator complexity | HBM qualification, package yield, lead time |
| Physical capacity | Power, cooling, buildings | Chips cannot generate revenue until energized | Utility interconnection and commissioning dates |

Two credible forecasts illustrate why a single TAM number is misleading. Omdia covers GPUs, AI ASICs and ASSPs, FPGAs, and CPUs with integrated AI acceleration. It estimated **$123B** shipped in 2024, **$207B** in 2025, and **$286B** in 2030; it also expected AI infrastructure spending to peak as a share of data-center spending in 2026 before tapering [20]. Bloomberg Intelligence uses a longer accelerator horizon and projects **$116B in 2024 to $604B in 2033**, a **16% CAGR** [22].

| Forecast | Scope and horizon | Central numbers | Decision-ready interpretation |
|---|---|---|---|
| Omdia | Global data-center AI processors through 2030 | $207B in 2025; $286B in 2030 [20] | Rapid deceleration after the initial build; custom chips gain traction |
| Bloomberg Intelligence | Global GPUs plus custom ASICs through 2033 | $604B in 2033; GPUs $486B, ASICs $118B [22] | Longer inference and sovereign/enterprise adoption sustain growth |
| U.S. demand proxy | Five largest U.S. hyperscalers, all capex | $697B estimated for 2026 [21] | A broad infrastructure envelope, not chip revenue |

### Case study: Project Stargate makes financing part of chip demand

Project Stargate plans up to **$500B** of U.S. data-center and energy infrastructure investment over four years, while J.P. Morgan originated **$9.6B** of construction loans for its Abilene, Texas campus [21]. This is not equivalent to $500B of semiconductor orders. It demonstrates the mechanism that converts an AI thesis into chip demand: financing, long-term contracts, site construction, utility capacity, and only then server deployment.

The implication is that semiconductor forecasts should discount announced projects by financeability and time-to-power. Corporate debt can fund broad compute and networking programs; project debt relies more heavily on visible project cash flow; equity absorbs obsolescence, ramp, utilization, and customer-concentration risk [21]. Chip vendors should therefore monitor financed and energized capacity, not headline project values alone.

## Buyer Economics: Hyperscalers Set the U.S. Demand Curve

Hyperscalers dominate near-term U.S. demand because they can fund full-stack systems, aggregate workloads, and amortize proprietary silicon over enormous internal volumes. The **$697B** 2026 capex estimate is therefore the most important demand signal, but it includes buildings, power, networking, and other infrastructure in addition to chips [21]. Cloud monetization is a useful cross-check: Microsoft's reported Azure revenue-growth KPI was **40% in FY2026 Q4** [13].

The buyer decision is no longer simply NVIDIA versus AMD. It is merchant flexibility versus workload-specific economics. Merchant GPUs offer broad programmability, a mature software environment, and rapid access to new model architectures. Custom ASICs can reduce cost and power for stable, repetitive workloads, but require design scale, software investment, and enough volume to absorb non-recurring engineering expense.

| Demand-side platform | Silicon posture | Public evidence | Market implication |
|---|---|---|---|
| Google | TPU-based custom acceleration plus merchant options | Ironwood can scale to **9,216 chips**, linked at **9.6 Tb/s**, with **1.77 PB** of shared HBM [10] | Custom silicon is now a system-level alternative, not a niche coprocessor |
| Microsoft | Maia custom accelerator plus merchant fleets | Maia 200 is deployed near Des Moines, with a Phoenix-region deployment next [8] | Captive silicon moves from design to geographically distributed deployment |
| Meta | MTIA custom inference silicon plus merchant accelerators | Meta frames low-cost, global model serving as a core infrastructure challenge [12] | Inference economics, not only peak training performance, drives design choices |
| AWS | Mixed merchant and captive-silicon strategy | Public sources reviewed here do not provide a comparable current unit or revenue disclosure | Track instance adoption and customer migration rather than chip announcements |
| Oracle and neoclouds | Primarily merchant accelerated compute | Project finance can be exposed to tenant concentration and utilization volatility [21] | Contract quality matters more as demand moves beyond balance-sheet-funded hyperscalers |

### Case study: Google's Ironwood turns custom silicon into a full system

Ironwood's significance is not simply that Google designed another accelerator. A 9,216-chip superpod, 9.6 Tb/s inter-chip fabric, and 1.77 PB of shared HBM show that competitive advantage has moved into topology, memory, and software-managed scale [10]. The mechanism is vertical integration: Google can optimize chips, interconnect, compiler, cloud service, and workload together.

This does not make merchant GPUs obsolete. Custom silicon works best when the buyer controls large, predictable workloads and can support its own software stack. The recommendation for enterprise buyers is not to imitate a hyperscaler blindly; it is to use custom or alternative accelerators only where workload stability, compiler support, and utilization produce measurable total-cost gains. For merchant suppliers, the response is to sell an integrated platform and make portability less costly.

## Competitive Map: NVIDIA, AMD, Broadcom, Marvell, and Captive ASICs

The competitive structure has three layers. NVIDIA and AMD sell broadly programmable merchant compute. Broadcom and Marvell enable hyperscaler custom accelerators and connectivity. Google, Microsoft, Meta, and AWS internalize part of design and capture value through cloud services rather than chip sales.

| Player or group | Position | Latest demand evidence | Principal risk |
|---|---|---|---|
| NVIDIA | Merchant GPU, networking, and full-stack systems | FY2026 Data Center revenue was **$193.737B**, versus **$115.186B** in FY2025 [13] | Custom ASIC substitution, export restrictions, and buyer concentration |
| AMD | EPYC CPUs and Instinct accelerators | Q2 2026 Data Center revenue was **$6.7B**, up **107% YoY** [11] | Software and deployment friction can offset attractive chip economics |
| Broadcom | Custom AI accelerators and networking | Q3 FY2026 AI semiconductor revenue was **$16.7B**, up **221% YoY**; Q4 outlook was **$21.7B**, up **236% YoY** [2] | A small number of large custom customers and program timing |
| Marvell | Custom compute, electro-optics, networking, storage | FY2026 Data Center revenue was **$6.1003B**, versus **$4.1642B** a year earlier [13] | Design-win timing, concentration, and execution against larger rivals |
| Hyperscaler ASICs | TPU, Trainium/Inferentia, Maia, MTIA | Bloomberg projects ASIC share to rise from **8% in 2024 to 19% in 2033** [22] | High fixed design cost, software burden, and limited external scale |
| HBM suppliers | SK hynix, Micron, Samsung | HBM is physically integrated alongside logic in advanced AI packages [19] | Qualification, yields, capital intensity, and memory cyclicality |
| Foundry and packaging | TSMC-led leading-edge manufacturing and CoWoS | CoWoS has been in production since 2012 and supports interposers up to **3.3x reticle** [19] | Geographic concentration, yields, and synchronized capacity ramps |

Bloomberg Intelligence expects GPUs still to represent **81%** of accelerator revenue in 2033. It projects NVIDIA at **70%-75%** share through 2030 and AMD reaching at least **10%** by decade-end; these are forecasts, not current measured shares [22]. The likely structure is therefore coexistence, not winner-take-all replacement.

### Case study: Broadcom demonstrates that custom ASIC growth can coexist with NVIDIA

Broadcom's Q3 FY2026 AI semiconductor revenue growth of **221%** and its Q4 expectation of **236%** growth show that custom acceleration and networking are expanding while NVIDIA's Data Center business remains enormous [2][13]. The observation matters because market-share narratives can hide a larger economic reality: the overall pool is growing fast enough for merchant GPUs and custom programs to expand simultaneously.

The mechanism differs. NVIDIA monetizes a reusable platform across customers; Broadcom monetizes design expertise, IP, and connectivity in a few very large programs. NVIDIA offers breadth and software reuse, while custom silicon offers workload-specific performance per dollar and per watt. Investors should value these businesses with different leading indicators: platform adoption and ecosystem for NVIDIA; design wins, customer ramps, and networking attach for Broadcom and Marvell.

NVIDIA's exposure to policy is already visible in guidance: its Q2 FY2027 outlook called for **$108B** of total revenue, plus or minus 2%, while assuming no Data Center compute revenue from China [3]. U.S. demand can remain strong even as export policy changes global mix, product design, and supply allocation.

## The Rack Is the Product: HBM, CoWoS, Networking, and Optics

Accelerator demand is increasingly constrained by system completion rather than a single die. A sellable AI rack requires logic wafers, HBM, advanced packaging, substrates, power delivery, liquid cooling, scale-up links, scale-out networking, optics, and software. A shortage or qualification delay in any one component can strand the others.

TSMC's CoWoS is the clearest example. It combines multiple SoCs and HBM stacks in an advanced 2.5D package, using high-density interconnects and a large silicon or redistribution-layer interposer [19]. CoWoS-S supports an interposer of roughly **2,700 mm2**, or **3.3x reticle**; larger designs are directed toward CoWoS-L or CoWoS-R [19]. Bigger packages increase bandwidth and integration, but also raise yield, thermal, substrate, and capacity challenges.

| Constraint | Why it gates demand | Leading signal | Exposed participants |
|---|---|---|---|
| HBM | Accelerators need bandwidth close to compute | Qualification status, contracted supply, stack generation | GPU/ASIC vendors, memory suppliers |
| Advanced packaging | Joins logic chiplets and HBM into a usable module | CoWoS capacity, package yield, interposer size | TSMC, OSATs, substrates, chip designers |
| Scale-up fabric | Connects accelerators inside a system | Link bandwidth, topology, switch availability | NVIDIA, Broadcom, Marvell, hyperscaler ASIC teams |
| Scale-out networking | Links racks into training and inference clusters | Ethernet/InfiniBand ports and optics | Switch, NIC, DSP, laser, and transceiver suppliers |
| Cooling and power delivery | Determines rack density and uptime | Commissioning date, PUE, liquid-cooling readiness | Data-center operators and equipment vendors |
| Software | Converts installed silicon into productive compute | Time-to-train, cost per token, utilization | Cloud operators, model developers, chip vendors |

### Case study: CoWoS shifts bargaining power downstream

The traditional semiconductor model emphasized transistor density and wafer capacity. CoWoS shows why that is insufficient for AI. The package must hold multiple logic elements and HBM at a scale beyond a normal reticle, so packaging becomes part of architecture rather than a back-end commodity [19]. This shifts value toward interposers, substrates, HBM, bonding, test, and thermal engineering.

The strategic implication is portfolio allocation by completed-system output. Ordering more accelerator wafers without matching HBM and package slots does not create more deployable compute. Suppliers should contract and qualify the whole bill of materials; buyers should track delivered racks and energized capacity rather than nominal chip allocations. Investors should separate wafer starts from package-complete shipments.

U.S. industrial policy is beginning to target these system bottlenecks. Announced potential support includes up to **$300M** for GlobalFoundries co-packaged-optics R&D, **$245M** for Kepler's AI memory, **$140M** for Multibeam's advanced packaging, and smaller programs in dielectrics, component identification, and photonic substrates [7]. These initiatives may improve domestic capability, but potential incentives are not the same as qualified high-volume supply.

## Power, Policy, and Geography: 521-843 TWh Changes Site Selection

Power is both a demand indicator and a hard cap. U.S. data centers used **176 TWh in 2023**, about **4.4%** of national electricity. The earlier LBNL outlook projected **325-580 TWh** and **6.7%-12%** by 2028 [16]. The June 2026 update raised the 2030 reference case to **649 TWh**, or **11.8%**, with a **521-843 TWh** uncertainty range [23].

| U.S. power metric | Estimate | Meaning for chips |
|---|---:|---|
| 2023 actual | 176 TWh; 4.4% of U.S. electricity [16] | Starting point before the current build wave matures |
| 2028 range | 325-580 TWh; 6.7%-12% [16] | Near-term grid and interconnection pressure |
| 2030 reference | 649 TWh; 11.8% [23] | Large sustained deployment if modeled assumptions hold |
| 2030 uncertainty | 521-843 TWh; 9.5%-15.3% [23] | Demand can diverge sharply with utilization, shipments, and technology |

The model is useful but should not be mistaken for a semiconductor forecast. It assumes a five-year accelerated-server life, **20%** idle power, **80%** training utilization, and inference utilization reaching **20% by 2030** [23]. It also does not model major departures caused by breakthroughs, market changes, regulation, supply constraints, or grid availability [23].

### Case study: Power availability can delay otherwise valid chip demand

Data-center power demand more than doubled from 2017 to 2023, largely due to AI servers, whose chips and cooling systems are power intensive [16]. A chip order therefore becomes revenue-producing infrastructure only after interconnection, substation, backup generation, cooling, and commissioning. The highest-value geographic metric is time-to-energization.

The mechanism creates regional divergence inside the United States. A market with attractive land but a long interconnection queue can lose deployments to a more expensive region with firm power. Operators should combine onsite generation, storage, flexible load, and utility rate design where feasible; those are among the response options identified by DOE [16]. Semiconductor suppliers should map demand to energized megawatts and commissioning schedules, not public site announcements.

Policy is two-sided. Domestic incentives can deepen U.S. packaging, memory, photonics, and materials capabilities [7]. Export controls and geopolitical tension can reduce addressable sales or force product variants; Bloomberg identifies U.S. export controls, tariffs, and Taiwan concentration as supply-chain vulnerabilities [22]. The practical response is to maintain policy-specific revenue scenarios and qualify geographically diverse critical inputs where technically and economically viable.

## Scenarios Through 2030: Growth, Bottlenecks, or Correction

The following are strategic scenarios, not point forecasts. They translate the evidence into observable triggers and avoid assigning unsupported probabilities.

| Scenario | Mechanism | Chip-demand outcome | Winners | Trigger metrics |
|---|---|---|---|---|
| **Reference: Growth normalizes** | Hyperscaler capex remains high; inference expands; efficiency offsets part of volume growth | Continued expansion, but below the 2022-2025 surge reflected in Omdia's trajectory [20] | Integrated GPU platforms, custom ASIC partners, HBM, networking | Capex, cloud growth, package output, energized MW |
| **Bull: Inference and agents compound** | Reasoning and agentic systems generate more tokens; enterprise and sovereign adoption broadens | Bloomberg-like long-duration growth; both GPUs and ASICs expand [22] | GPUs for flexibility; ASICs for high-volume inference; optics and memory | Tokens served, inference utilization, revenue per watt |
| **Supply-capped bull** | Orders exceed HBM, packaging, optics, or grid capacity | Revenue shifts to scarce enabling components; deployments slip | HBM, CoWoS, power equipment, networking | Lead times, yields, interconnection delays |
| **Correction** | Monetization lags capex; smaller models and efficiency reduce compute per task; speculative tenants retrench | Order pushouts, lower utilization, pricing pressure, customer concentration losses | Flexible suppliers with strong balance sheets and diversified customers | Cancellations, idle capacity, tenant credit, declining cloud growth |
| **Policy fragmentation** | Export controls, tariffs, and regional supply duplication intensify | U.S. demand stays strong but global mix and product roadmaps fragment | Domestic R&D and qualified alternative suppliers | License approvals, redesign cost, regional revenue mix |

The central tension is efficiency. Omdia identifies smaller specialized models, better data curation, model design, and inference optimization as forces that slow AI-compute demand [20]. Yet lower cost per task can expand usage, so efficiency may reduce demand per inference while increasing the number of inferences. Investors should track both sides: compute per token and total tokens served.

The correction risk is credible even without an AI-demand collapse. J.P. Morgan warns that capex can outpace monetization and cites uncertain end demand, technology cycles, shifting customer needs, and debt concentration [21]. A project can fail financially while aggregate chip demand grows. The most vulnerable assets are those financed against weak tenants, dependent on one model customer, or awaiting uncontracted power.

## Strategic Playbooks and KPI Dashboard

| Stakeholder | Priority action | KPI set | Decision rule |
|---|---|---|---|
| Hyperscaler or large buyer | Optimize total delivered cost, not chip price | Cost per token, utilization, revenue per watt, deployment time | Add custom silicon only where stable volume repays software and design cost |
| Enterprise buyer | Preserve workload portability | Benchmark performance, migration effort, cloud egress, availability | Dual-source where operational savings exceed porting and support costs |
| Merchant chip supplier | Sell the system and software outcome | Platform adoption, networking attach, completed-rack delivery | Do not overbuild wafers without matched HBM and packaging |
| Custom-silicon provider | Diversify programs and de-risk ramps | Design wins, customer concentration, tape-out and qualification milestones | Stage capacity against customer-funded or contracted programs |
| Memory/packaging supplier | Lock in multi-year demand without sacrificing flexibility | Contract coverage, yield, node mix, package throughput | Expand against qualified demand, not headline accelerator forecasts |
| Data-center developer/lender | Underwrite power and tenant quality first | Firm MW, energization date, lease term, tenant credit, utilization | Finance contracted cash flow; use equity for technology and ramp risk |
| Investor | Separate shipped chips from productive infrastructure | Supplier revenue, cloud growth, HBM/package output, energized MW | Require confirmation across at least two layers of the demand stack |

A compact monitoring dashboard should contain: (1) hyperscaler capex and guidance revisions; (2) Azure, AWS, Google Cloud, and Oracle cloud growth; (3) NVIDIA, AMD, Broadcom, and Marvell data-center or AI revenue; (4) HBM qualification, pricing, and contracted capacity; (5) advanced-package output and yield; (6) networking and optical revenue; (7) energized megawatts and interconnection delays; (8) utilization and cost per token; and (9) cancellations, customer concentration, and tenant credit.

The key is cross-validation. Capex without cloud growth may signal overbuild. Chip revenue without package-complete systems may signal inventory imbalance. Announced megawatts without firm interconnection may never become demand. Conversely, rising cloud usage, high utilization, booked HBM, package output, and energized capacity together constitute a much stronger demand signal than any single forecast.

## Synthesis

| Strategy or layer | Value mechanism | Scope | Main trade-off | Evidence horizon |
|---|---|---|---|---|
| Merchant GPU platform | Reusable hardware, software, and networking across workloads | Broadest training and inference coverage | Premium economics and vendor dependence versus flexibility | Current revenue plus multi-year platform adoption |
| Custom ASIC | Tailored performance per dollar and watt | Large, stable workloads at hyperscaler scale | Better unit economics versus design cost and software burden | Multi-year design and deployment cycles |
| HBM and advanced packaging | Converts logic into bandwidth-rich modules | Required by both GPUs and high-end ASICs | Scarcity value versus yield and capital intensity | Capacity and qualification cycles |
| Networking and optics | Makes thousands of accelerators act as one system | Training fabrics and distributed inference | More system performance versus complexity and power | Tracks cluster scale and topology |
| Power and data centers | Turns delivered racks into usable compute | Site and regional infrastructure | Long-lived assets versus fast chip obsolescence | Multi-year construction and utility timelines |

The non-obvious conclusion is that GPU versus ASIC is not the most important binary. Both can grow while the binding constraint moves among HBM, packaging, networking, and power. NVIDIA's scale, Broadcom's custom-silicon growth, and Google's system-level TPU deployment are not contradictory; together they show a market expanding and segmenting by workload [13][2][10].

The second tension is between efficiency and aggregate demand. Better models reduce compute per task, but cheaper inference can stimulate more tasks. Omdia explicitly recognizes both reasoning-driven token growth and efficiency-driven restraint [20]. The correct strategy is therefore not to bet solely on ever-larger training runs. It is to build exposure to heterogeneous inference, memory bandwidth, connectivity, and power efficiency.

The third tension is time horizon. Chips turn rapidly; packages and fabs turn more slowly; data centers and grids turn slowest. A synchronized plan must match these clocks. Buyers should secure power and system components before accelerator delivery. Suppliers should align wafer, HBM, and package commitments. Lenders should underwrite contracted utilization. Investors should favor control points that remain valuable across GPU and ASIC architectures.

**Decision-ready conclusion:** near-term U.S. AI-chip demand is strong and supported by extraordinary buyer capex and supplier revenue. The durable opportunity, however, is not simply "more GPUs." It is completed, powered, utilized AI systems. The winning portfolio combines merchant compute, custom silicon, HBM and packaging, networking and optics, and infrastructure exposure, while controlling concentration, policy, utilization, and power risks.

## References

1. *Micron Technology, Inc. - Micron Technology, Inc. Reports Record Results for the Third Quarter of Fiscal 2026*. https://investors.micron.com/news/press-release/2026/Micron-Technology-Inc--Reports-Record-Results-for-the-Third-Quarter-of-Fiscal-2026/default.aspx
2. *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
3. *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
4. *Marvell Technology, Inc. Reports Second Quarter of Fiscal Year 2027 Financial Results*. https://www.businesswire.com/news/home/20260827816134/en/Marvell-Technology-Inc.-Reports-Second-Quarter-of-Fiscal-Year-2027-Financial-Results
5. *SK hynix Announces 2Q26 Financial Results – SK hynix Newsroom*. https://news.skhynix.com/en/q2-2026-business-results
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. *Department of Commerce Announces Letters of Intent With 7 Companies for $874 Million to Accelerate Semiconductor R&D for the Compute Supply Chain | NIST*. https://www.nist.gov/news-events/news/2026/07/department-commerce-announces-letters-intent-7-companies-874-million
8. *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
9. *United States Data Center Energy Usage Report: 2025 Update | Energy Technologies Area*. https://eta.lbl.gov/publications/united-states-data-center-energy-2025
10. *3 things to know about Ironwood, Google's latest TPU*. https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/ironwood-google-tpu-things-to-know
11. *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
12. *Four MTIA Chips in Two Years: Scaling AI Experiences for Billions*. https://ai.meta.com/blog/meta-mtia-scale-ai-chips-for-billions
13. *Fiscal.ai /v2/company/segments-and-kpis*. https://fiscal.ai
14. *The Jevons Paradox in AI Infrastructure*. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6775299
15. *Global AI Investment Is Forecast to Exceed $1 Trillion in 2026*. https://www.goldmansachs.com/insights/articles/global-investment-is-forecast-to-exceed-1-trillion-in-2026
16. *Berkeley Lab Report Evaluates Increase in Electricity Demand from Data Centers - Berkeley Lab – Berkeley Lab News Center*. https://newscenter.lbl.gov/2025/01/15/berkeley-lab-report-evaluates-increase-in-electricity-demand-from-data-centers
17. *Companies Are Racing to Buy GPUs. Many Sit Idle - AIwire*. https://www.hpcwire.com/aiwire/2026/04/21/companies-are-racing-to-buy-gpus-many-sit-idle/
18. *CHIPS Incentives Funding Opportunities | NIST*. https://www.nist.gov/chips/chips-incentives-funding-opportunities
19. *CoWoS® - Taiwan Semiconductor Manufacturing Company Limited*. https://3dfabric.tsmc.com/english/dedicatedFoundry/technology/cowos.htm
20. *New Omdia forecast: AI data center chip market to hit $286bn, growth likely peaking as custom ASICs gain ground*. https://omdia.tech.informa.com/pr/2025/aug/ai-data-center-chip-market-to-hit-286bn-growth-likely-peaking-as-custom-asics-gain-ground
21. *Financing AI infrastructure and U.S. data centers*. https://www.jpmorgan.com/insights/banking/capital-markets/financing-ai-infrastructure-data-centers
22. *AI Accelerator Market Looks Set to Exceed $600 Billion by 2033, Driven by Hyperscale Spending and ASIC Adoption, According to Bloomberg Intelligence | Press | Bloomberg LP*. https://www.bloomberg.com/company/press/ai-accelerator-market-looks-set-to-exceed-600-billion-by-2033-driven-by-hyperscale-spending-and-asic-adoption-according-to-bloomberg-intelligence/
23. *Lawrence Berkeley National Laboratory*. https://www.energy.gov/documents/united-states-data-center-energy-usage-report-2025-update