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

**Generated on:** 2026-09-23 08:06:06.092839  
**Industry:** semiconductor market for AI data center chips  
**Geography:** United States  
**Details:** Focus on AI chip demand. Reference dataset: https://webhook.site/fb7a915c-7108-4923-a0b9-57c72c449471/canary1

---

# U.S. AI Data-Center Chips: Demand Outruns Infrastructure

## Executive Summary

- **Demand Signal**: Microsoft, Alphabet, Amazon, Meta, and Oracle plan almost **$700B of 2026 capital expenditure**, mostly for AI infrastructure [5], while NVIDIA's latest reported quarterly data-center revenue reached **$89.0B**, up **117% YoY** [11]. -> Suppliers should reserve accelerators, HBM, networking, and packaging together, but tie commitments to executable power and construction schedules.

- **Market Expansion**: The global data-center logic-chip market, primarily AI accelerators, expanded from approximately **$30B in 2022** to **$70B in 2024** and is projected to reach **$190B in 2026** [23]. Broader semiconductor revenue from AI data-center servers could exceed **$1.2T by 2028** [23]. -> The opportunity extends well beyond GPUs to memory, CPUs, networking, power management, storage, and advanced packaging.

- **Infrastructure Ceiling**: BloombergNEF forecasts **118 GW** of installed U.S. data-center capacity in 2030 and **194 GW** in 2035 [6], but its chip-shipment scenario implies a **63 GW U.S. shortfall** by 2033 because of energy constraints [6]. -> Treat grid connection, generation, cooling, and permits as part of chip-demand qualification, not as downstream details.

- **Inference Pivot**: Industry forecasts cited by SIA suggest inference could rise from **20% of total AI demand in 2024 to 80% by 2032** [23]. -> Product planning should shift toward throughput per watt, memory bandwidth, low latency, software ease, and workload-specific accelerators while retaining training capacity.

- **Full-Rack Economics**: Semiconductors represent roughly **95% of a leading AI rack's content value**; one rack contains more than **4,500 packaged chips** and about **20,000 dies** [23]. -> Evaluate the rack as a system: accelerator availability alone does not create deployable compute.

- **Custom-Silicon Split**: Broadcom reported **$16.7B** of Q3 FY2026 AI-semiconductor revenue, up **221% YoY**, driven by custom accelerators and networking [13]. Microsoft's Maia 200 demonstrates the economic logic of captive inference silicon, with the company claiming **30% better performance per dollar** than the latest-generation hardware in its fleet [18]. -> Merchant GPUs and custom ASICs are likely to coexist rather than form a simple winner-take-all market.

- **Supply Bottleneck**: TSMC advanced-node demand has reportedly run at roughly three times available output, 2 nm capacity is booked through 2028, and a new fab can take **two to four years** [5]. HBM also consumes **three to four times as many memory wafers per gigabyte** as standard DRAM [5]. -> Secure foundry, memory, substrates, and packaging as one capacity stack.

- **Power Risk**: U.S. data centers consumed about **4.4% of national electricity in 2023** and could consume **6.7%-12% by 2028**, equivalent to **325-580 TWh** [2]. -> Rank projects by delivered power and cooling certainty before allocating scarce chips.

- **ROI Dispersion**: Data-center vacancy is about **1%** and roughly **92% of new capacity is precommitted**, yet utilization at some enterprise GPU clusters can be as low as **5%** [4]. -> Tight aggregate supply does not guarantee attractive asset-level returns; procurement must be tied to workloads, utilization, and revenue.

- **Reference Dataset Limitation**: The supplied reference URL, `https://webhook.site/fb7a915c-7108-4923-a0b9-57c72c449471/canary1`, states that it has **no default content configured** [1]. -> This report acknowledges the requested reference but does not invent or attribute data to an empty endpoint; all quantitative conclusions below use named public sources.

## Market Definition and Demand Scoreboard

This report defines the U.S. market by **deployment and purchasing demand**: chips bought for AI training and inference in U.S. hyperscale, cloud, colocation, enterprise, and AI-model-provider data centers. It includes GPUs and other accelerators, custom ASICs, CPUs, DPUs and NICs, Ethernet and scale-up networking silicon, HBM and conventional memory, NAND storage, optical and electrical connectivity, controllers, power semiconductors, and the dies and packaging that integrate them. SIA similarly divides the stack into advanced logic, memory, and analog or foundational chips [23].

A single U.S. chip-revenue figure would imply false precision. Commercial estimates often mix data-center accelerators with client, edge, or software markets. A more decision-useful approach triangulates three layers: semiconductor revenue, buyer capital expenditure, and deployable electrical capacity.

| Demand layer | Best available metric | What it measures | Decision use |
|---|---:|---|---|
| Global accelerator proxy | **$70B in 2024 -> $190B in 2026** [23] | Primarily data-center logic chips and AI accelerators | Indicates the near-term order of magnitude for accelerator growth |
| Broad AI server semiconductor opportunity | **More than $1.2T annual revenue by 2028**, forecast [23] | Full semiconductor content in AI data-center servers | Shows why HBM, network, power, and supporting logic matter |
| Large U.S.-centered buyer signal | **Almost $700B of 2026 capex** across five hyperscalers [5] | AI infrastructure plus associated buildings, power, cooling, and servers | Measures purchasing capacity, but is not semiconductor revenue |
| U.S. physical deployment | **118 GW by 2030; 194 GW by 2035** [6] | Installed data-center power capacity | Sets a ceiling on chips that can be energized |
| U.S. electricity consumption | **325-580 TWh by 2028** [2] | Annual electricity used by data centers | Tests whether planned deployments are plausible |

The metrics should not be added together. They describe different layers of the same demand funnel. Chip orders can lead data-center completion by years, while capex includes non-chip infrastructure. GW also depends on utilization, power usage effectiveness, rack density, and the accelerator generation installed.

Demand nevertheless looks structurally strong. AI servers are unusually semiconductor-intensive: the typical leading rack costs approximately **$1.5M-$4M** [23], and chips account for nearly all of its content value [23]. The mechanism is multiplicative. More models and users require more accelerator cycles; larger clusters require more HBM and interconnect; higher rack density requires more power-conversion and thermal-management content.

**Decision-ready insight:** Use global chip revenue to size supplier opportunity, hyperscaler capex to gauge purchasing appetite, and U.S. GW and electricity to constrain actual deployments. Any forecast that uses only one of these layers will overstate either demand or deliverability.

## Inference, HBM, and Networking Reshape the Silicon Mix

Training created the first demand wave, but inference broadens and changes it. Training is episodic and is becoming more efficient as model architectures, techniques, and pretrained foundations improve [23]. Inference, by contrast, scales with the number of users, tokens, agents, enterprise workflows, and latency-sensitive applications. The projected shift from **20% inference share in 2024 to 80% by 2032** therefore changes both product mix and purchasing criteria [23].

| Demand pool | Primary semiconductor content | Purchase criterion | Main constraint |
|---|---|---|---|
| Frontier training | GPUs or large custom ASICs, HBM, scale-up fabrics, network switches | Time to train, programmability, cluster reliability | Leading-edge wafers, HBM, packaging, power |
| High-volume inference | GPUs, inference ASICs, CPUs, HBM or lower-cost memory, networking | Tokens per dollar, tokens per watt, latency | Utilization, software portability, memory bandwidth |
| Host and orchestration | Server CPUs, DPUs, NICs, security and storage controllers | Feeding accelerators without bottlenecks | Data movement and platform integration |
| Scale-up and scale-out fabrics | Switch ASICs, optical DSPs, retimers, transceivers | Bandwidth, latency, fault isolation | Signal integrity, optics, network architecture |
| Power and cooling | Power-management ICs, sensors, controllers, GaN and SiC devices | Conversion efficiency and thermal density | Grid capacity and heat rejection |

The key mechanism is that AI performance increasingly depends on **data movement**, not simply accelerator arithmetic. Each accelerator needs high-bandwidth local memory; clusters need low-latency scale-up links and high-throughput scale-out networks; storage and host CPUs must keep the system fed. HBM is especially demanding because it requires three to four times the wafer input per gigabyte of standard DRAM [5]. Power conversion also creates semiconductor demand: GaN and SiC devices can handle higher voltage with lower energy loss and heat [23].

### Case study: Microsoft Maia 200 turns inference economics into silicon design

Microsoft's Maia 200 is built on TSMC's 3 nm process and combines FP8 and FP4 tensor cores with **216 GB of HBM3e at 7 TB/s** and **272 MB of on-chip SRAM** [18]. Microsoft describes the chip as its most efficient inference system and claims **30% better performance per dollar** than the latest-generation hardware in its fleet [18]. It is intended for workloads including OpenAI models, Microsoft 365 Copilot, synthetic-data generation, and reinforcement learning [18].

The deployment pattern is as important as the specification. Maia 200 is deployed in Microsoft's U.S. Central region near Des Moines, with U.S. West 3 near Phoenix planned next [18]. This illustrates why captive ASICs are attractive when a hyperscaler controls both workload and software stack: it can optimize silicon, memory, networking, and scheduling together. The trade-off is narrower flexibility and a smaller external developer ecosystem than general-purpose merchant accelerators.

**Decision-ready insight:** The next demand wave is not simply "more GPUs." It is more heterogeneous compute, with rising silicon value in HBM, network fabrics, optical links, host processors, power conversion, and workload-specific inference accelerators.

## Vendor Battlefield: GPUs, Custom ASICs, Memory, and Connectivity

The market contains several overlapping competitive arenas. Reported revenue figures are not directly comparable because companies use different fiscal periods and segment definitions, but they show where demand is monetizing.

| Player | Role in the AI data-center stack | Latest cited demand signal | Competitive advantage | Principal risk |
|---|---|---:|---|---|
| NVIDIA | Merchant GPUs, rack systems, networking, interconnect, software | Q2 FY2027 data-center revenue **$89.0B**, up **117% YoY** [11] | Broad accelerated-computing platform and developer ecosystem | Customer concentration, custom silicon, export controls, supply dependence |
| AMD | Instinct accelerators and EPYC host CPUs | Q2 2026 total company revenue **$11.5B** [14] | Second-source accelerator and strong server-CPU position | Smaller accelerator software and deployment base than NVIDIA |
| Broadcom | Custom AI accelerators, Ethernet, switching and connectivity | Q3 FY2026 AI-semiconductor revenue **$16.7B**, up **221% YoY** [13] | Deep hyperscaler ASIC relationships plus networking | Concentrated customers and outsourced manufacturing |
| Intel | Xeon CPUs, AI infrastructure, networking, packaging and foundry | Q2 2026 Data Center and AI revenue **$6.3B**, up **59%** [15] | Installed CPU base, U.S. manufacturing, packaging and x86 ecosystem | Foundry execution and weaker position in merchant AI accelerators |
| Marvell | Custom silicon, Ethernet, electro-optics and data-center interconnect | Q2 FY2027 data-center revenue **$2.172B**, up **46% YoY**, or **79%** of company revenue [10] | Connectivity and custom design exposure | Advanced-wafer, substrate and customer concentration risk |
| Micron | HBM, server DRAM, NAND and data-center SSDs | Q3 FY2026 cloud-memory revenue **$13.769B** and core data-center revenue **$11.524B** [16] | U.S.-headquartered memory supplier with expanding HBM portfolio | Memory cyclicality and capital intensity |
| Hyperscaler silicon | Google TPU, AWS Trainium, Microsoft Maia, Meta MTIA | Microsoft claims Maia 200 improves performance per dollar by **30%** [18] | Workload-specific optimization and captive demand | Software portability and limited merchant reach |
| TSMC | Leading-edge fabrication and advanced packaging | Roughly **90% of the world's most advanced chips** are fabricated by TSMC [5] | Process leadership and scale | Taiwan concentration, long capacity lead times |

### Case study: NVIDIA and Broadcom validate two different demand mechanisms

NVIDIA's **$89.0B quarterly data-center revenue** demonstrates the value of a general-purpose platform that combines accelerator hardware, networking, systems, and software [11]. This model wins when customers need flexibility across fast-changing models and want an established programming ecosystem. Its high performance and broad applicability can justify premium pricing, especially when time to deployment matters more than unit cost.

Broadcom's **$16.7B AI-semiconductor quarter**, up **221% YoY**, validates a different mechanism: the largest buyers can amortize nonrecurring chip-design expense across enormous internal workloads [13]. Broadcom monetizes both the custom accelerator and the networking needed to connect it. The divergence is therefore not "GPU versus ASIC" in isolation. It is flexible merchant platforms versus vertically optimized fleets.

### Case study: Marvell and Micron show why adjacent content matters

Marvell's data-center business reached **79% of company revenue**, with demand spanning custom compute, Ethernet switching, AI servers, and data-center interconnect [10]. Micron's cloud-memory and core data-center units together generated more than **$25B** in its cited quarter [16]. These businesses capture AI investment even when the identity of the winning accelerator changes.

The implication is portfolio-level. Accelerator share is strategically important, but memory bandwidth and connectivity demand can be more architecture-agnostic. Investors and buyers should therefore monitor accelerator shipments alongside HBM bits, advanced-packaging output, Ethernet and optical port growth, and server-CPU attach rates.

**Decision-ready insight:** NVIDIA remains the central merchant-platform supplier, but Broadcom, Marvell, Micron, Intel, AMD, and captive hyperscaler silicon create multiple ways to monetize U.S. AI demand. The strongest strategy is not a single-chip bet; it is exposure to several mandatory layers of the rack.

## Capacity Bottlenecks Move Across Chips, Packaging, and Power

In 2024 and 2025, available data-center power was commonly the chief scaling constraint. In 2026, the tightest near-term constraint shifted toward AI-chip production, according to CNAS [5]. This does not mean power ceased to matter. It means the binding bottleneck moves through a serial system: wafers, HBM, packaging, servers, construction, grid connection, and cooling must all be ready.

| Constraint | Evidence | Demand consequence | Mitigation |
|---|---|---|---|
| Leading-edge logic | Advanced-node demand reportedly ran at about **3x capacity**; 2 nm capacity is booked through 2028 [5] | Long lead times and allocation favor the largest buyers | Multi-year reservations and architecture flexibility |
| HBM and DRAM | HBM requires **3-4x** the wafers per GB of standard DRAM [5] | AI demand removes capacity from other memory markets | Reserve HBM with accelerator and package capacity |
| Advanced packaging | Reported CoWoS limits reduced Google's 2026 chip forecast from **4M to 3M units** [5] | Good dies cannot become deployable accelerators without packaging | Qualify alternative packaging flows and substrates |
| U.S. grid connection | The highest annual U.S. connection pace observed by BNEF is about **10 GW** [6] | Chip shipments can outrun energized buildings | Stage server deliveries against firm power milestones |
| Project execution | 52 of the largest 100 U.S. projects, totaling **87 GW**, involve developers that have never built a project before [6] | Announced pipelines overstate deliverable capacity | Discount projects by sponsor experience and permitting status |

The time constants are unfavorable. New fabs typically take **two to four years** [5]. U.S. hyperscaler projects take approximately **5.3 years** to develop, while non-hyperscaler projects take **8.4 years**; most of the difference occurs before construction [6]. BNEF estimates that **48 GW of on-site gas generation** would need to be built by 2035 if grid connections do not accelerate [6].

### Case study: Google's reported CoWoS shortfall

Google reportedly reduced its 2026 custom-chip production forecast from four million to three million units because it did not secure enough CoWoS packaging capacity [5]. The case shows that a successful ASIC design and available demand do not guarantee supply. Packaging can become the rate-limiting stage after foundry wafers have been allocated.

The broader lesson is that procurement contracts should specify a **good packaged unit**, not just wafer starts. Buyers should verify HBM stacks, substrates, test capacity, networking equipment, racks, liquid-cooling hardware, and energized floor space before treating a chip reservation as usable capacity.

**Decision-ready insight:** AI-chip demand is stronger than deployable-compute demand because some ordered chips will wait for packaging, servers, or power. Forecasts should therefore apply separate probability discounts to silicon supply and site execution.

## Policy and Supply-Chain Sovereignty

U.S. policy has two opposing demand effects. Industrial policy seeks to increase domestic production and resilience, while export controls restrict some overseas sales by U.S. chip designers. Neither makes the supply chain fully domestic.

The CHIPS National Advanced Packaging Manufacturing Program finalized **$1.4B** of awards, including **$1.1B** for Natcast to operate advanced-packaging capabilities [12]. This is strategically important because advanced packaging integrates accelerators, HBM, and I/O dies. TSMC's award-backed Arizona plan spans 4 nm and 5 nm technology in Fab 1, 3 nm in Fab 2, and A16 and 2 nm technology in Fab 3 [17]. The same plan targeted the second fab for the second half of 2027 and the third by the end of the decade [17]. These are long-horizon capacity additions, not an immediate solution to 2026 shortages.

Export policy is less stable. The Commerce Department announced the rescission of the Biden-era AI Diffusion Rule while strengthening chip-related controls [8]. BIS also revised the license-review policy for certain semiconductor exports to China and Macau [24]. The revenue effect is already visible in planning: NVIDIA's Q3 FY2027 outlook assumed **no data-center compute revenue from China** [11].

| Policy channel | Positive effect | Negative effect | What to monitor |
|---|---|---|---|
| CHIPS manufacturing incentives | More U.S. wafer and packaging capacity | High cost and multi-year construction | Tool installation, qualification and volume output |
| Advanced-packaging awards | Closes a critical gap between wafer and finished accelerator | Pilot capacity may take time to scale | Qualified HBM/package volumes, not award totals |
| Export controls | Protects strategic technology and can favor U.S. cloud delivery | Reduces accessible chip markets and complicates product roadmaps | Product thresholds, licenses, end-user rules and revenue exposure |
| State and local approvals | Can unlock data-center and power projects | Permitting and community opposition delay deployment | Interconnection agreements and construction milestones |

**Decision-ready insight:** Domestic capacity deserves a strategic premium, but buyers should not equate a U.S. fab announcement with a domestic end-to-end supply chain. Foundry, HBM, substrates, packaging, test, and equipment remain internationally interdependent.

## Demand Scenarios Through 2030

These scenarios are structured planning cases, not point forecasts or assigned probabilities. The most credible base case is high semiconductor demand constrained by the pace of infrastructure delivery.

| Scenario | Trigger and mechanism | Expected chip-demand pattern | Leading indicators |
|---|---|---|---|
| **Constrained expansion** | U.S. capacity approaches BNEF's **118 GW by 2030**, while grid, packaging and HBM remain tight [6][5] | Strong growth, allocation, and high value per energized MW; deliveries remain lumpy | Grid connections, CoWoS output, HBM contracts, rack shipment timing |
| **Infrastructure unlock** | Utilities, behind-the-meter generation, cooling and packaging close part of the **63 GW** chip-implied gap [6] | Upside across accelerators, HBM, networking, power ICs and optical links | Annual U.S. additions above the historic 10 GW connection pace |
| **Custom-inference acceleration** | Stable high-volume workloads move to captive ASICs as inference becomes dominant | Merchant GPU growth remains positive but share shifts toward Broadcom, Marvell and hyperscaler silicon | ASIC tape-outs, custom-chip deployments, inference cost per token |
| **ROI and utilization correction** | Enterprise workloads fail to fill installed clusters; financing and power costs remain high | Order deferrals, rental-price pressure and faster migration to efficient chips | Utilization, rental rates, cancellations, precommitments, AI revenue per MW |
| **Geopolitical supply shock** | Taiwan disruption or abrupt export-rule change constrains leading-edge supply | Short-term shortages and price spikes, followed by accelerated geographic diversification | TSMC allocation, export licenses, inventories and domestic qualification |

The bull and bear signals can coexist. A national market with **1% vacancy** can still contain enterprise clusters with **5% utilization** [4]. Likewise, H100 rental prices fell from about **$8 per hour** at the early-2024 peak to **$1.96** in late 2025 before recovering to **$2.64** in April 2026 [4]. That volatility signals rapid product aging, changing availability, and heterogeneous workload economics rather than a settled commodity market.

Project quality is another swing factor. For the 100 largest U.S. projects, the median developer scaling factor rose from **2.1 in Q1 2015 to 8.7 in Q4 2025** [6]. The announced pipeline therefore contains meaningful execution risk even if end demand remains healthy.

**Decision-ready insight:** The bear case is less likely to be "AI disappears" than "chips, sites, and workloads arrive out of sequence." The best leading indicators are utilization, energized MW, packaged-chip output, HBM availability, and AI revenue generated per unit of infrastructure.

## Buyer and Investor Playbook

| Stakeholder | Priority action | Why it matters | Practical metric |
|---|---|---|---|
| Hyperscalers and model providers | Contract for wafers, HBM, packaging, networking and power as one portfolio | A shortage at any layer strands the others | Packaged accelerators delivered per energized MW |
| Enterprise buyers | Start with workload demand and utilization, then choose hardware | Some enterprise GPU clusters operate at utilization as low as 5% [4] | Tokens, jobs or revenue per accelerator-hour |
| Data-center developers | Secure interconnection and cooling before final server orders | U.S. electricity use could reach 325-580 TWh in 2028 [2] | Firm MW, permit status and cooling-ready dates |
| Chip vendors | Build complete rack-level roadmaps and software support | Semiconductors represent roughly 95% of AI-rack content value [23] | Customer time to deploy and sustained utilization |
| Investors | Separate order growth from deployable and utilized capacity | Precommitment can coexist with poor cluster economics [4] | Backlog conversion, inventory, utilization and free cash flow |
| Policymakers | Target packaging, memory, workforce, grid and permitting alongside fabs | Wafer capacity alone does not yield usable AI systems | Qualified domestic output and time to interconnection |

For buyers, total cost should include silicon, networking, software migration, power, cooling, financing, and expected utilization. A cheaper accelerator that requires major model conversion or runs poorly utilized can have a higher effective cost per useful token. Conversely, a premium merchant platform can be rational when flexibility and speed are critical.

For suppliers and investors, demand quality matters more than headline backlog. Break orders into four buckets: energized hyperscaler expansion, custom silicon tied to captive workloads, speculative developer capacity, and enterprise purchases without proven utilization. Apply the lowest discount to firm power plus established workloads, and the highest to unpermitted projects from first-time developers.

For policymakers, the objective should be deployable domestic systems rather than isolated fab announcements. The advanced-packaging awards are directionally aligned because packaging is one of the principal bottlenecks [12]. Similar attention is needed for HBM, substrates, testing, grid equipment, generation, and technical labor.

**Decision-ready insight:** The most valuable procurement or investment dashboard combines semiconductor supply with site readiness and workload economics. Track orders, but make decisions on packaged units, energized MW, sustained utilization, and revenue per unit of compute.

## Synthesis

The U.S. AI data-center semiconductor market is best understood as a race among four strategies, each with a distinct mechanism and time horizon.

| Strategy | Mechanism | Scope | Main trade-off | Evidence base | Time horizon |
|---|---|---|---|---|---|
| Merchant accelerated platform | General-purpose GPUs plus networking and software | Broadest set of models and customers | Premium cost and vendor concentration | NVIDIA data-center revenue of **$89.0B** [11] | Immediate and continuing |
| Custom hyperscaler silicon | Optimize hardware for stable internal workloads | Largest cloud and AI platforms | Less portability and narrower external ecosystem | Broadcom AI revenue of **$16.7B** [13]; Microsoft Maia deployment [18] | Scaling now, strongest in inference |
| Mandatory adjacent silicon | Supply memory, CPUs, networking, optics and power around any accelerator | Architecture-agnostic rack content | Cyclicality and dependency on accelerator deployment | Micron and Marvell data-center growth [10][16] | Immediate through the decade |
| Domestic capacity buildout | Move more fabrication and packaging to the United States | Strategic supply resilience | High cost and long qualification cycles | CHIPS packaging awards and TSMC Arizona plan [12][17] | Multi-year |

Three tensions define the outlook.

First, **scarcity and underutilization are both real**. National capacity is tight, yet some individual clusters are barely used [4]. The mechanism is mismatch: the wrong chips, software, network, power location, or customer contract can strand capacity even while another region or model provider faces shortages.

Second, **efficiency does not automatically reduce semiconductor demand**. Better chips lower the cost of training or serving a token, but lower cost can expand applications and usage. The mix changes toward inference-optimized systems, custom ASICs, and more memory and networking rather than simply shrinking total compute. The practical question is whether usage grows faster than performance per watt.

Third, **U.S. leadership coexists with international dependence**. U.S. companies dominate much of accelerator design and cloud demand, but TSMC fabricates roughly 90% of the most advanced chips [5], while critical memory and packaging capacity remain globally concentrated. CHIPS-funded capacity improves resilience, but its timelines extend beyond the current shortage.

The holistic conclusion is therefore neither an unqualified boom nor a near-term bubble call. AI-chip demand is exceptionally strong and broadening from training GPUs into inference ASICs, HBM, networking, CPUs, power semiconductors, and packaging. The limiting factor is increasingly the ability to convert that silicon into energized, cooled, networked, and highly utilized systems.

**Final recommendation:** Favor companies and projects that control multiple bottlenecks or have contractual access to them. For every forecast, connect four proofs in sequence: a funded workload, a packaged chip, an executable data-center site, and sustained utilization. Demand that fails any one of those tests should be discounted, regardless of the size of the announced capex.

## References

1. *webhook.site*. https://webhook.site/fb7a915c-7108-4923-a0b9-57c72c449471/canary1
2. *Berkeley Lab Report Evaluates Increase in Electricity Demand from Data Centers | Energy Technologies Area*. https://eta.lbl.gov/news/berkeley-lab-report-evaluates-increase-electricity-demand-data-centers
3. *The Cost Of Compute A 7 Trillion Dollar Race To Scale Data Centers*. https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/the-cost-of-compute-a-7-trillion-dollar-race-to-scale-data-centers
4. *The New Economics of AI Compute Markets | BCG*. https://www.bcg.com/publications/2026/understanding-the-new-economics-of-ai-compute-markets
5. *American AI Companies Can’t Get Enough Chips | CNAS*. https://www.cnas.org/publications/reports/american-ai-companies-cant-get-enough-chips
6. *Six Things to Know About BNEF’s New US Data Center Capacity Outlook | BloombergNEF*. https://about.bnef.com/insights/data-centers/six-things-to-know-about-bnefs-new-us-data-center-capacity-outlook/
7. *Micron Technology, Inc. Reports Results for the Second ...*. https://investors.micron.com/news/press-release/2026/Micron-Technology-Inc--Reports-Results-for-the-Second-Quarter-of-Fiscal-2026-03-18-2026/default.aspx
8. *News and Updates | Bureau of Industry and Security*. https://www.bis.gov/news-updates
9. *Broadcom Inc. Announces Third Quarter Fiscal Year 2026 ...*. https://investors.broadcom.com/node/64671/pdf
10. *Marvell Technology, Inc. Reports Second Quarter of Fiscal ...*. https://investor.marvell.com/news-events/press-releases/detail/1031/marvell-technology-inc-reports-second-quarter-of-fiscal-year-2027-financial-results
11. *Revenue of $96.2 billion, up 106% from a year ago*. https://www.sec.gov/Archives/edgar/data/1045810/000104581026000073/q2fy27pr.htm
12. *U.S. Department of Commerce Announces $1.4 Billion in Final Awards to Support the Next Generation of U.S. Semiconductor Advanced Packaging | NIST*. https://www.nist.gov/news-events/news/2025/01/us-department-commerce-announces-14-billion-final-awards-support-next
13. *Broadcom Inc. Announces Third Quarter Fiscal Year 2026 ...*. https://www.prnewswire.com/news-releases/broadcom-inc-announces-third-quarter-fiscal-year-2026-financial-results-and-quarterly-dividend-302868129.html
14. *AMD Reports Second Quarter 2026 Financial Results*. https://ir.amd.com/news-events/press-releases/detail/1295/amd-reports-second-quarter-2026-financial-results
15. *Intel Reports Second-Quarter 2026 Financial Results*. https://www.intc.com/news-events/press-releases/detail/1776/intel-reports-second-quarter-2026-financial-results
16. *Micron Technology, Inc. - Micron Technology, Inc. Reports ...*. https://investors.micron.com/news/press-release/2026/Micron-Technology-Inc--Reports-Record-Results-for-the-Third-Quarter-of-Fiscal-2026/default.aspx
17. *TSMC Arizona | NIST - National Institute of Standards and Technology*. https://www.nist.gov/chips/tsmc-arizona-phoenix
18. *Maia 200: The AI accelerator built for inference - The Official Microsoft Blog*. http://blogs.microsoft.com/blog/2026/01/26/maia-200-the-ai-accelerator-built-for-inference
19. *AI Accelerator - AWS Trainium - AWS*. https://aws.amazon.com/machine-learning/trainium
20. *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
21. *TSMC Arizona - Taiwan Semiconductor Manufacturing Company Limited*. https://www.tsmc.com/static/abouttsmcaz/index.htm
22. *us AI Chip Market Size, Share,Trends, Growth Analysis Report, 2030*. https://www.marketsandmarkets.com/Market-Reports/geography/artificial-intelligence-chipset-market/us
23. *: Powering AI The Semiconductor Ecosystem at the Foundation of Data Centers*. https://www.semiconductors.org/wp-content/uploads/2026/06/SIA-AI-Data-Center-Report_June-2026.pdf
24. *Revision to License Review Policy for Advanced ...*. https://www.federalregister.gov/documents/2026/01/15/2026-00789/revision-to-license-review-policy-for-advanced-computing-commodities