INSIDERS

What the people running the race say

The people building AI see the bottlenecks first. We use what they say as a guide to what to measure, and check each claim against the data on this site.

Key points

  1. Americans talk about power; Chinese talk about chips. Each side is worried about its own weak spot.
  2. Musk's 'China will far exceed the world in AI compute' is not in the data yet: China's share of new compute fell to 4.7%. His power point holds (2.3x US generation).
  3. Huang's 'nanoseconds behind' is really months (4.4), and Chinese power is plentiful rather than cheap.

On power

Feb 2026 · Power

Elon MuskSpaceX (after buying xAI)

Power on the ground cannot meet AI demand, so go to space and use sunlight

Global electricity demand for AI simply cannot be met with terrestrial solutions, even in the near term, without imposing hardship on communities and the environment.

What the data sayThe IEA expects US data-centre use to rise by about 240 TWh by 2030, about 40 TWh a year. US generation grew 128 TWh in the latest 12 months, so data centres alone would take about 31% of the growth.

Watch: US data-centre power needs versus how fast US generation grows — Power

Source: Al Jazeera, Musk merges SpaceX and xAI firms, plans for space-based AI data centres (3 Feb 2026)

Nov 2025 · Power

Satya NadellaMicrosoft

The problem is power, not chips: GPUs sit in inventory with nowhere to plug them in

The biggest issue we are now having is not a compute glut, but it's power. You may actually have a bunch of chips sitting in inventory that I can't plug in. It's not a supply issue of chips; it's actually the fact that I don't have warm shells to plug into.

What the data sayUS generation rose only +2.9% in the latest 12 months against +5.7% in China. The binding constraint in America is getting power to the site, which is what 'warm shells' means.

Watch: How fast US generation grows — Power

Source: Tom's Hardware, Microsoft CEO says the company doesn't have enough electricity to install all the AI GPUs in its inventory (November 2025, on the BG2 podcast)

May 2025 · Power

Sam AltmanOpenAI

In the end the cost of AI converges to the cost of energy

Chips will get cheaper and cheaper. But an electron is an electron. Eventually the cost of intelligence, the cost of AI, will converge to the cost of energy.

What the data sayIf he is right, the country with more electricity wins in the long run. China added 532 GW of generating capacity in 2025, the US 44 GW. But Chinese chips need about 2.4x as much power per unit of compute, which eats into that lead.

Watch: Power prices and volumes in both countries — Power

Source: Tech Policy Press, Transcript: Sam Altman testifies at US Senate hearing on AI competitiveness (8 May 2025)

Apr 2025 · Power

Eric SchmidtFormer Google chairman

AI's natural limit is electricity; the US needs 92 GW of new generation by 2030

The PRC understands the foundational power of AI and energy, pouring resources into AI R&D while building the world's leading renewable energy capacity and modern grid infrastructure.

What the data sayThe US added 44 GW of capacity in 2025, mostly solar that runs only part of the day. At that pace, 92 GW of firm power by 2030 is far off; China added 532 GW in the same year.

Watch: How many gigawatts the US adds each year — Power

Source: Written testimony of Eric Schmidt, House Energy & Commerce Committee (9 Apr 2025)

Feb 2024 · Power

Elon MuskTesla, xAI (then)

The shortage moves on: chips, then transformers, then electricity

A year ago the shortage was chips. The next shortage will be voltage step-down transformers, and the next will be electricity. Next year you will see they just can't find enough electricity to run all the chips.

What the data sayUS generation grew +2.9% over the latest 12 months (+128 TWh). The shortage he predicted shows up as waiting lines and prices, not as blackouts: see the transformer and connection queues on Power Watch.

Watch: How fast US generation is growing — Power Watch: the grid

Source: New Atlas, Elon Musk: AI will run out of electricity and transformers in 2025 (March 2024, at Bosch Connected World)

On both

Jan 2026 · Power and chips

Elon MuskTesla, SpaceX, xAI

China will win on power, sort out the chips, and far exceed the world in AI compute

China's going to have more power than anyone else and probably will have more chips. Based on current trends, China will far exceed the rest of the world in AI compute. People are underestimating the difficulty of bringing electricity online.

What the data sayOn power he is right in direction, though not yet at 3x: in 2025 China generated 2.34x as much as the US. On compute the data point the other way so far: China's share of new AI compute fell from 11% in 2023 to 4.7% in 2025.

Watch: China's generation vs the US, and whether China's compute is catching up — Chips

Source: Business Insider, Elon Musk says China will 'far exceed the rest of the world in AI compute' (7 Jan 2026, on Moonshots with Peter Diamandis)

Nov 2025 · Power and chips

Jensen HuangNvidia

China will win (later: China is 'nanoseconds behind'); energy in China is almost free

China is going to win the AI race. … (statement hours later) As I have long said, China is nanoseconds behind America in AI. It's vital that America wins by racing ahead and winning developers worldwide.

What the data say'Nanoseconds' is closer to months: China's best model (Kimi K3) sits at a level US models first reached 4.4 months earlier. On price, OIES finds Chinese data-centre power costs US$ 0.08–0.14/kWh on the coast, near Virginia's US$ 0.10; the cheap power comes from local subsidies of up to 50%.

Watch: How many months apart the best models are, and whether Chinese power is really cheap — Models

Source: CNBC, Nvidia's Jensen Huang softens his 'China will win the AI race' remark to FT (6 Nov 2025)

Jun 2025 · Power and chips

Ren ZhengfeiHuawei founder

Chips are a generation behind, but clustering many of them closes the gap in practice

Our single chip is still a generation behind the US. We use mathematics to make up for physics, non-Moore's-law methods to supplement Moore's law, and cluster computing to make up for single chips.

What the data sayThe data bear out the method and its cost: China's big AI data centres use about 2.3 MW per 1,000 H100s' worth of compute, against 0.9 MW in the US, roughly 2.4x the power. Clustering works because China has the power to spare.

Watch: How much more power the same compute takes — Chips

Source: ChinaTalk, Huawei founder on US v China and basic research (translation of the People's Daily interview, 10 Jun 2025)

On chips

Jan 2026 · Chips

Dario AmodeiAnthropic

The US is many years ahead in making chips; selling them to China is a big mistake

We are many years ahead of China in terms of our ability to make chips, so I think it would be a big mistake to ship these chips. It's a bit like selling nuclear weapons to North Korea.

What the data sayPer chip, the best US design (B300) does 3.3x the work of China's best (Ascend 910C). In 2025 all China-bound chips together added 636,000 H100s' worth of compute, against 12.9 million elsewhere.

Watch: The per-chip gap and the gap in compute shipped — Chips

Source: Axios, Anthropic CEO on Nvidia chips to China (20 Jan 2026, Davos)

Jan 2026 · Chips

Lin JunyangTech lead, Alibaba Qwen

Under 20% chance a Chinese firm overtakes the US in 3–5 years; US compute is 10–100x larger

Computational resources in the US are one to two orders of magnitude larger. OpenAI and others pour compute into next-generation research, while in China we are stretched to the limit just meeting daily demand.

What the data say'One to two orders of magnitude' means 10–100x. The largest training runs in 2026 differ by 50x, inside his range; and China's share of newly shipped AI compute is 4.7%.

Watch: The ratio between the largest training runs — Models

Source: SCMP, China AI has 'less than 20%' chance to exceed US over next 3 to 5 years: Alibaba scientist (January 2026, at Tsinghua's AGI-Next)

Jul 2024 · Chips

Liang WenfengDeepSeek

Money is not the problem; the ban on advanced chips is

Money has never been the problem for us; bans on shipments of advanced chips are the problem.

What the data saySince then DeepSeek and others have stayed about 4–8 months behind the US frontier, with a fraction of the compute: the largest US training run in 2026 used 50x the compute of China's largest.

Watch: The gap in compute used for the largest training runs — Models

Source: ChinaTalk, DeepSeek CEO interview (translation of the 36Kr/Waves interview, July 2024)

This page organises information; it does not recommend buying or selling anything. Data belongs to each publisher and may contain errors or delays. Quotes are taken from reports and transcripts; see the sources for the originals. Make your own investment decisions.

Updated 29 Sep 2026 12:24 JST · US power to Jun 2026, China power to May 2026 · models to 9 Sep 2026