Understanding the AI Race: Labs, Chips, Power and the U.S.-China Contest

The AI race in numbers

A plain-English scoreboard of the artificial intelligence race: money raised, revenue, chips, data-center spending, power demand and model rankings. Every figure links to its source and carries an as-of date. Figures marked “reported” come from people familiar with private talks and have not been disclosed by the companies.

$89.0BNvidia data center revenue in one quarter, up 117% from a year earlierQuarter ended July 26, 2026 · Nvidia, Aug. 26, 2026

$720–745BCombined 2026 capex guidance from Amazon, Alphabet, Microsoft and MetaAs of July 2026 calls · DataCentres.com summary of July 2026 earnings calls

$965BAnthropic’s post-money valuation after its $65B Series HMay 28, 2026 · Anthropic, May 28, 2026

$1.25TValuation OpenAI is reportedly targeting in a new ~$30B round (not final)Reported Oct. 22, 2026 · Bloomberg, Oct. 22, 2026

945 TWhIEA base-case projection for global data-center electricity use in 2030, up from 415 TWh in 2024IEA, Energy and AI (2025)

2.7%Lead of the top U.S. model over the top Chinese model on Stanford’s measureAs of March 2026 · Stanford HAI, 2026 AI Index

1. Frontier labs compared: funding, valuation, revenue

The “frontier labs” are the handful of companies building the most capable general-purpose AI models. Private labs raise money from investors and report revenue selectively, so the table separates company disclosures from figures reported by news organizations citing unnamed sources.

Latest valuation, U.S. dollars

Anthropic (Series H)$965B

OpenAI (March round)$852B

xAI (at SpaceX merger)$250B

DeepSeek (reported)~$52–59B

Mistral AI~$24B

Last priced or reported valuations, not live market values. Anthropic: Anthropic, May 28, 2026. OpenAI: $852B post-money in March 2026, TechCrunch, Oct. 22, 2026; it is reportedly negotiating at $1.25T pre-money (Bloomberg, Oct. 22, 2026). xAI: valued at $250B when SpaceX acquired it, Reuters, Feb. 2, 2026. DeepSeek: reported range from its first outside round, Reuters, June 3, 2026. Mistral: CNBC, Sept. 8, 2026.

Lab	Latest funding	Valuation	Revenue signal
OpenAI	$122B round, March 2026; seeking ~$30B more (reported) Bloomberg via Yahoo Finance, Oct. 8, 2026	$852B post-money (March); talks at $1.25T pre-money TechCrunch, Oct. 22, 2026	About $50B annualized at end of September; expects $70B+ by year-end (reported) Bloomberg via Yahoo Finance, Oct. 8, 2026
Anthropic	$65B Series H, May 2026 Anthropic, May 28, 2026	$965B post-money; IPO investors reportedly discussing ~$1.5T Reuters, Oct. 15, 2026	Run rate topped $65B by end of July (reported) Reuters, Aug. 17, 2026; Q3 revenue $16.8B, told investors it was adjusted-profitable (reported) Reuters, Oct. 14, 2026
xAI (SpaceX)	$20B Series E, Jan. 2026 xAI, Jan. 6, 2026	$250B in SpaceX all-stock deal, Feb. 2026 Reuters, Feb. 2, 2026	Not credibly reported; omitted
Google DeepMind	Funded by Alphabet (see capex)	Part of Alphabet; no separate value	Not broken out
Meta (Muse models)	Funded by Meta (see capex)	Part of Meta; no separate value	Meta sells no cloud capacity
Mistral AI	€3B Series D led by Samsung, Sept. 2026 CNBC, Sept. 8, 2026	~$24B (€21B+) post-money	Not credibly reported; omitted
DeepSeek	First outside round, ~$7B, Tencent and CATL among backers (reported) Reuters, June 3, 2026	Reported ~$52–59B	Not credibly reported; omitted

Plain English: “Annualized revenue” or “run rate” multiplies a recent month or quarter to estimate a full year. It is not audited annual sales, and OpenAI and Anthropic count cloud-partner sales differently, Bloomberg reported, so the two figures are not a like-for-like comparison.

2. Compute and chips: who sells the picks and shovels

Training and running AI models takes enormous numbers of specialized chips, mostly Nvidia graphics processing units (GPUs). Nvidia’s data center sales are the clearest public measure of how fast the industry is buying compute.

Nvidia data center revenue per quarter

Q2 FY2026 (to July 2025)$41.1B

Q1 FY2027 (to April 2026)$75.2B

Q2 FY2027 (to July 2026)$89.0B

Nvidia data center segment revenue by quarter. Source: Nvidia CFO commentary (SEC), Aug. 26, 2026. Nvidia guided total company revenue of $108.0B (±2%) for the next quarter and assumed no data center compute revenue from China.

Share of sales: Data center was $89.0B of Nvidia’s $96.2B total revenue in the quarter ended July 26, 2026. Nvidia, Aug. 26, 2026
Market value: Nvidia’s market capitalization was about $5.08 trillion at the Oct. 23, 2026 close. Stock values change daily. CompaniesMarketCap, close Oct. 23, 2026
Biggest reported cluster: xAI’s Colossus 2 near Memphis had 110,000 Nvidia GB200 and 440,000 GB300 chips, Elon Musk said, with more batches planned through December. Bloomberg via Yahoo Finance, Sept. 24, 2026
China sales: Shipments of Hopper data center chips to China were under 1% of Nvidia’s data center revenue in the July quarter. Nvidia CFO commentary (SEC), Aug. 26, 2026

Plain English: More chips in one cluster lets a lab train bigger models faster. But chip counts are company claims, and each new chip generation does more work per chip, so counts alone do not rank labs.

3. Big Tech capex: the $700 billion build-out

Capital expenditure, or capex, is money spent on long-lived assets such as servers, data centers and networking. The four largest U.S. cloud and platform companies gave these calendar-2026 forecasts on their July 2026 earnings calls.

2026 capex guidance (calendar year)

Amazon~$220B

Alphabet$195–205B

Microsoft~$175B

Meta$130–145B

Midpoints shown for ranges. Companies define capex differently: Amazon quotes cash capex; Microsoft and Meta include finance leases; Microsoft’s figure fell from about $190B only because of a lease-accounting change, the company said. Sources: DataCentres.com summary of July 2026 earnings calls; Platformonomics, July 2026. Third-quarter updates are due in late October 2026.

The four spent almost $172 billion combined in the second quarter of 2026 alone, up 77% from a year earlier, according to a tally of company filings by Platformonomics, July 2026. Alphabet said it expects capex to increase significantly again in 2027. DataCentres.com summary of July 2026 earnings calls

Plain English: Capex is the clearest sign of conviction: these companies are committing cash now on the bet that AI demand will pay it back later. Some of the growth reflects higher memory and component prices, not only more capacity.

4. Data centers and power demand

AI chips draw large amounts of electricity, and power supply has become one of the main limits on how fast data centers can be built.

Data-center electricity use, terawatt-hours per year

World, 2024415 TWh

World, 2030 (IEA base case)945 TWh

U.S., 2023176 TWh

U.S., 2028 (LBNL high case)325–580 TWh

Global: IEA, Energy and AI (2025) (415 TWh is about 1.5% of world electricity use). U.S.: U.S. DOE / LBNL, Dec. 20, 2024, which put data centers at about 4.4% of U.S. electricity in 2023 and 6.7%–12% by 2028.

Where the data centers are: The U.S. hosts 5,427 data centers, more than 10 times any other country, per Stanford’s 2026 AI Index. Stanford HAI, 2026 AI Index
Jobs in the plumbing: U.S. employment in computing infrastructure, data processing and web hosting was 453,800 in September 2026, down from 465,700 in May, seasonally adjusted. BLS via FRED (CES5051800001), Sept. 2026, updated Oct. 2, 2026

Plain English: A terawatt-hour is a billion kilowatt-hours. The ranges are wide because nobody knows how quickly chips will get more efficient or how many announced projects will actually be built.

5. Model performance: who is on top

LMArena ranks chatbots by having people vote between two anonymous answers. Scores are Elo-style ratings with error margins; models within a few points are statistically tied.

LMArena text score, top model per lab

Gemini 4 Argon (High) · Google1517

Claude Opus 5.5 (High) · Anthropic1509

Muse Wonder (Max) · Meta1509

Kimi K3 (Max) · Moonshot (China)1487

GPT-5.6 Sol (xHigh) · OpenAI1485

Grok 4.20 Beta 1 · xAI1475

DeepSeek-V4.1-Flash (Max) · DeepSeek1476

Mistral Large 4 · Mistral1434

Bars start at 1400 points to make gaps visible. Best-ranked model per lab, LMArena Text overall, snapshot dated Oct. 22, 2026 (about 8.9 million votes, 418 models). Margins are roughly ±4–8 points. Rankings change often. Source: LMArena Text leaderboard, Oct. 22, 2026.

Coding: On SWE-bench Verified, a widely used coding test, top performance rose from 60% to near 100% in one year, according to Stanford’s 2026 AI Index. Stanford HAI, 2026 AI Index
Agents: AI agents went from about 12% to about 66% task success on OSWorld, which tests real computer tasks. Stanford HAI, 2026 AI Index
Uneven skills: The same report notes a top model reads analog clocks correctly only 50.1% of the time. Stanford HAI, 2026 AI Index

Plain English: No single benchmark captures “best.” Voting leaderboards measure what people prefer in a chat; test suites measure specific skills. Benchmarks also go stale once models master them, which is why Stanford calls the frontier “jagged.”

6. U.S. vs. China

Private AI investment, 2025

United States$285.9B

China$12.4B

Private AI investment in 2025. Stanford notes private figures likely understate China’s total because of government guidance funds. Source: Stanford HAI, 2026 AI Index.

Performance gap: U.S. and Chinese models have traded the lead since early 2025; as of March 2026, the top U.S. model led by 2.7%, Stanford found. Stanford HAI, 2026 AI Index
New companies: The U.S. had 1,953 newly funded AI companies in 2025, more than 10 times the next country. Stanford HAI, 2026 AI Index
Talent: The number of AI researchers and developers moving to the U.S. has dropped 89% since 2017. Stanford HAI, 2026 AI Index
Chips and export controls: In January 2026 the Commerce Department moved to case-by-case license reviews for Nvidia H200, AMD MI325X and similar chips to China, with conditions including U.S. third-party testing. U.S. Commerce Dept. (BIS), Jan. 13, 2026 Nvidia’s latest outlook assumes no data center compute revenue from China. Nvidia, Aug. 26, 2026
Home-grown hardware: DeepSeek’s V4 models, released in preview in April 2026, were adapted to run on Huawei Ascend chips, Reuters reported. Reuters, Apr. 24, 2026

Plain English: The U.S. leads on money, chips and data centers; China has closed most of the gap on model quality, often with cheaper open-weight models. Export rules decide how much advanced American hardware Chinese labs can buy.

7. Boston and Massachusetts in the AI race

Massachusetts competes less on giant data centers and more on research, talent and applied AI in health care, life sciences and robotics.

State money: The state capital plan budgets $68.6 million for the Applied AI Hub from fiscal 2026 through 2030, including $10 million in FY2026 and $33.3 million in FY2027. Mass. FY26 Capital Budget, line D050
Accelerators: The Massachusetts AI Hub, IBM and Red Hat launched an accelerator for early-stage AI startups during Boston Tech Week. Massachusetts AI Hub In Kendall Square, LabCentral’s AI BioHub supports AI-first biotech startups with a $1.9 million MassTech grant. LabCentral AI BioHub
Research: MIT’s CSAIL is one of the largest university computing labs in the world. See our report on MIT’s AI model for crystal materials.
Startups and funding: Recent local raises include Ledgebrook’s $200 million and Graylark’s $10.7 million seed. Browse our guide to AI companies in Massachusetts and the Boston robotics guide.

State-level AI job counts are not published as a single official series by the Census Bureau or BLS, so we do not estimate one here.

What the numbers mean

Spending is running ahead of revenue. Big Tech’s 2026 capex guidance of $720–745 billion is many times the reported annualized revenue of the leading labs, so the build-out is a bet on future demand.
Nvidia is the toll booth. Most of that spending flows to chips; Nvidia’s data center revenue more than doubled in a year.
The lead is narrow and rotating. The top of LMArena is a near-tie among Google, Anthropic and Meta, and Stanford measures the U.S.–China gap at under 3%.
Power is the new bottleneck. Official projections have data-center electricity use roughly doubling or more by 2028–2030.
Private valuations are not market prices. They are set in negotiated rounds and can move sharply, as OpenAI’s reported shift from $1.4 trillion to $1.25 trillion shows.

What to watch

Late-October 2026 earnings from Alphabet, Microsoft, Meta and Amazon, with updated capex guidance.
Anthropic’s expected IPO filing and the valuation investors settle on (Reuters, Oct. 15, 2026).
Whether OpenAI closes its reported ~$30 billion round (Bloomberg, Oct. 22, 2026).
Nvidia’s next quarterly report against its $108 billion revenue outlook (Nvidia, Aug. 26, 2026).
How many H200 export licenses to China are actually granted under the January rule (U.S. Commerce Dept. (BIS), Jan. 13, 2026).
Colossus 2’s planned additional chip batches through December (Bloomberg via Yahoo Finance, Sept. 24, 2026).

AI race glossary

Frontier model: One of the most capable general-purpose AI models available at a given time.
GPU: Graphics processing unit, the chip type most used to train and run AI models.
Capex: Capital expenditure: spending on long-lived assets like data centers and servers.
Run rate / annualized revenue: Recent revenue scaled up to a full year; an estimate, not audited annual sales.
Post-money valuation: A company’s value including the new money just raised; pre-money excludes it.
Open-weight model: A model whose trained parameters are published so others can run or modify it.
Terawatt-hour (TWh): One billion kilowatt-hours of electricity.
Export controls: Government rules that require licenses to sell certain technologies abroad.
Benchmark: A standardized test used to compare AI systems on a task.

AI race FAQ

Who is winning the AI race right now?

There is no single winner. As of Oct. 22, 2026, Google’s Gemini 4 Argon led LMArena’s text leaderboard, closely followed by models from Anthropic and Meta, while Anthropic and OpenAI lead on private valuation and reported revenue.

How much is OpenAI worth?

OpenAI was valued at $852 billion after its March 2026 round. In October 2026 it was reportedly negotiating a new round of about $30 billion at a $1.25 trillion pre-money valuation; terms were not final.

How much is Anthropic worth?

Anthropic was valued at $965 billion post-money after raising $65 billion in May 2026. Investors in its planned IPO were reportedly discussing a valuation around $1.5 trillion in October 2026.

How much are tech giants spending on AI?

Amazon, Alphabet, Microsoft and Meta guided to roughly $720 billion to $745 billion of combined capital expenditure for 2026 on their July 2026 earnings calls, mostly for AI infrastructure.

How much does Nvidia make from AI chips?

Nvidia’s data center segment brought in $89.0 billion in the quarter ended July 26, 2026, up 117% from a year earlier, out of $96.2 billion in total revenue.

How much electricity do AI data centers use?

The IEA estimates data centers used 415 TWh worldwide in 2024, about 1.5% of global electricity, and projects 945 TWh by 2030 in its base case. In the U.S., data centers used about 4.4% of electricity in 2023.

Is China catching up to the U.S. in AI?

On model quality, largely yes: Stanford’s 2026 AI Index found the top U.S. model led the top Chinese model by 2.7% as of March 2026. The U.S. still leads heavily in private investment, at $285.9 billion versus $12.4 billion in 2025.

What is Massachusetts doing in AI?

The state budgets $68.6 million for its Applied AI Hub from fiscal 2026 to 2030 and backs accelerators, while MIT, Harvard and Kendall Square companies focus on AI research, health care, biotech and robotics.

Sources

Figures checked Oct. 23, 2026. Where a figure could not be verified from a company disclosure, official statistic or established news organization, it was left out.