Preface Executive Briefing

Week of May 30, 2026

Moore made chips smaller. Tau makes them wait less.

China's bid to win on cost, not chip-for-chip.

Huawei's new Tau Scaling Law reframes the chip race from raw power to removing waiting time, and a domestic price war is collapsing the cost of AI from both sides. This briefing explains what that shift means for the market, in business terms.

The race is changing shape

From raw model performance to the lowest cost per useful answer. China is racing on a different axis to the one everyone was watching.

Whose model is smartest?

Whose model is cheapest per useful answer?

Smaller parts, or less waiting?

For sixty years, chips got faster mainly by getting smaller. Last week, Huawei put a different question on the table.

“How small can we make the chip?”

Fit more, smaller switches into the same space, so each calculation gets faster and cheaper. This drove sixty years of progress.

“How much waiting can we remove?”

Huawei’s new idea: speed also comes from cutting the time chips spend waiting, for signals, for memory, and for each other.

Tau’s question points here: every layer of the system spends time waiting.

  1. You

    wait for the answer to appear

  2. Model serving

    waits for tokens to be generated, one by one

  3. The cluster

    servers wait to coordinate the work between them

  4. Chip to chip

    waits for data to travel between chips

  5. Processor and memory

    the processor waits for memory to deliver data

  6. The chip

    waits for signals to travel across it

Tau does not replace Moore. It asks a bigger question: not just how small the parts are, but how little the whole system waits.

Not one better chip, a whole system

China compensates for weaker chips with packaging, interconnect, more chips, local software and lower cost. The bet is the whole system, not any single part.

Where it is working

  • A full domestic stack is forming. Chinese chips, Chinese AI models like DeepSeek, Qwen and MiMo, and Chinese cloud platforms to run them on.

  • It competes on scale, not single-chip power. Stringing many domestic chips together reaches top-tier total computing power, even when each chip is weaker.

  • The engineering is real. Huawei says its newest chip design squeezes out meaningfully more performance, though it has not shipped yet.

Where it still lags

  • Chip for chip, it is still behind. China's best AI chip is reported at roughly a third of the top US chip's power.

  • That scale costs energy. Matching the output reportedly takes around four times the electricity.

  • The hard part is repeating it. Reliable manufacturing at volume, and software that is easy to build on, are still unsolved.

≈⅓

the power of China's best AI chip versus the top US chip, one for one

so it needs more chips to do the same work

≈4×

the electricity to match that computing power at scale

it competes by scale, not by efficiency

Comparisons are Huawei's latest AI hardware against NVIDIA's, based on independent analyst estimates and Huawei's own pre-launch claims, not lab benchmarks.

Two price wars at once

Cheaper domestic chips underneath and cheaper domestic models on top, each one reinforcing the other.

Cheaper compute underneath

China cannot always buy the best NVIDIA chips. But domestic compute that is available, good enough, and cheaper at the system level lets Chinese model makers avoid the scarcity premium on imported frontier chips.

Cheaper models on top

Chinese labs are slashing API prices. DeepSeek, Alibaba's Qwen and Xiaomi's MiMo now serve frontier-class models at a fraction of US prices.

≈2.9%

of GPT-5.5's price, for the same output

On 27 May, Xiaomi cut its MiMo model to match DeepSeek exactly: about $0.87 per million output tokens, versus $30 for GPT-5.5.

Cheaper chips make cheaper models possible. Cheaper models pull more demand onto domestic chips. Each war feeds the other.

What ‘cheaper’ actually looks like

Western frontier pricing against the Chinese cost-collapse, side by side.

Each dot is a model. Further left is cheaper; higher is more capable. The Chinese models sit mid-table on capability but far cheaper than the US models near them.

Intelligence
60
55
50
45
Opus 4.8
GPT-5.5
GPT-5.4
Qwen3.7-Max*
Kimi 2.6
MiMo V2.5 Pro
Gemini 3.1 Pro
Sonnet 4.6
DeepSeek V4 Pro
DeepSeek V4 Flash
Gemini 3.5 Flash
$0$5$10$15$20$25$30

Output price per million tokens

US / WesternChinese

The Chinese frontier sits near the price floor while staying mid-table on capability: far cheaper, and good enough for most work.

Output price observed 30 May 2026. Capability is the Artificial Analysis Intelligence Index (third-party). *Qwen3.7-Max is shown at Alibaba’s limited-time $3.50 output price; it returns to $7.50 on 22 June 2026.

The question is no longer which model is smartest

Moore made chips smaller. Tau tries to make chips wait less. China's AI price war tries to make intelligence cheaper.

01

Tau reframes the race

For sixty years, faster chips meant smaller chips. Huawei's Tau adds a second question: how much waiting can we remove across the whole computing system? It does not end Moore's Law, it wraps a bigger optimisation question around it, and that reframing is what lets weaker chips still compete.

02

The edge is the system, not one chip

Chip for chip, China still trails the US by roughly three to one and burns far more power. But by combining many domestic chips with local models, software and cloud, it reaches good-enough scale at lower cost. The bet is coordination across the whole stack, not winning the single-chip race.

03

The price war is the proof

Cheaper domestic chips underneath and aggressively cheap domestic models on top now reinforce each other, with the top Chinese models serving output at roughly 3% of the leading US price. For buyers, that quietly changes the question from 'which model is smartest?' to 'which stack gives the lowest cost per useful answer?'

FAQ

No. Chip for chip, China's best AI processor is still about a third as powerful, and matching that scale costs far more energy. What changed is that China is competing on the whole system and on price, not on beating the single best chip.