Washington wanted to wall China off from Nvidia. Instead, it walled Nvidia out of China — and handed the world's second-biggest AI chip market to the one company America spent a decade trying to contain. "It's pretty hard to collect data about the market share of Nvidia in China," Huawei's rotating chairman Eric Xu said this week, with a straight face. "But based on the data we have collected, Ascend has surpassed Nvidia."

Huawei's Best Salesman Works in Washington

"It's pretty hard to collect data about the market share of Nvidia in China. But based on the data we have collected, Ascend has surpassed Nvidia." — Eric Xu, Huawei rotating chairman, September 2026

The claim — and what it isn't

Speaking at Huawei Connect, Xu said the company's Ascend AI accelerators now outsell Nvidia in China, by both revenue and market share. Let's be honest about the fine print: he offered no audited numbers, the data is Huawei's own, and counting Nvidia's share in China is admittedly a strange exercise — because officially, there isn't any.

But here's the thing. When the incumbent CEO says his China market share went from 95% to zero, and the challenger says it's now number one, you don't need Huawei's spreadsheets to know something big happened in between. You just need a calendar.

How you lose a market without losing a game

The timeline reads like a slow-motion own goal:

October 2022 — Washington bans chips at Nvidia's A100 class or better. Nvidia designs China-only workarounds (A800, H800). Washington closes those too.
2023–2024 — Nvidia ships the H20, a chip stripped down so hard it barely deserves the logo. Chinese buyers buy it anyway, because the alternative is worse.
April 2025 — Washington bans the H20 too. Jensen Huang calls it "deeply painful": roughly $15 billion in revenue evaporates.
Early 2026 — A brief H200 thaw, then new complications. Nvidia books zero China data center revenue in its February and May 2026 results. Huang says China share fell from 95% to 0%.
September 2026 — Trump and Xi talk. Chip policy emerges "unchanged": the US guards its AI lead, China doubles down on self-reliance.

Nvidia didn't lose China to a better chip. It lost China to a law.

The boomerang

Here's what the policy assumed: no Nvidia chips, no Chinese AI. Here's what actually happened: no Nvidia chips, no Nvidia revenue — and a Chinese AI buildout that SemiAnalysis now ranks second-largest on Earth, with domestic capex racing toward $100 billion a year.

Demand didn't disappear. It re-routed. Bernstein estimates relayed by The Economist put Nvidia at roughly 40% of China's AI chip market in 2025, collapsing to around 8% in 2026 — with Huawei rising to about half. DeepSeek, China's breakout AI lab, now co-designs and open-sources software that runs its models on Ascend silicon. Every model trained on Ascend is a future that never buys an H200.

Brute force is a strategy

Huawei can't buy EUV machines, so its chips are made on older nodes with lower yields. A single Ascend still loses a footrace with a single Nvidia flagship. So Huawei changed the race. Its answer is arithmetic: strap more chips together.

The CloudMatrix 384 system chains 384 Ascend chips into one monster compute node.
The upcoming Atlas 950 SuperPod links 8,192 chips; the Atlas 960 scales to 15,488.
Xu's own framing at the conference: make "one million processors work as one computer."
The roadmap runs on schedule: Ascend 950PR entered mass production in April 2026 (target: around 750,000 units this year), the 960 — with double the compute and memory — lands in 2027, the 970 in 2028.

Is this inefficient? Enormously. More chips, more watts, more racks. But China has cheap electricity, deep pockets, and — thanks to export policy — a completely captive customer base. Brute force is only stupid when you have a smarter option available.

The gift that keeps giving

Here's the part Washington may not have priced in. Xu says demand for Ascend chips outstrips supply — Huawei can't even produce enough for China, which is why it's limiting overseas sales to a trickle. And once Chinese AI labs build their training pipelines on Huawei's chips, its CANN software stack, and its CloudMatrix interconnect, they don't switch back the day an H200 clears customs. Distribution moats die fast. Stack moats compound.

Xu expects many of China's leading model developers to begin training on Ascend systems next year. If that lands, the substitution isn't a workaround anymore — it's the default.

The honest counterargument

Skeptics — and there are serious ones — say none of this proves Huawei has "caught" Nvidia. The Council on Foreign Relations argues Huawei still lags badly on per-chip performance, HBM memory supply, and manufacturing yield, and that easing controls would hand China a two-to-three-year compute windfall. Analysts like Noah Smith note Huawei doesn't expect a single chip matching the H200 until late 2027 at the earliest.

Also true: Nvidia has never been richer — about $194 billion in data center revenue last fiscal year. China was 17% of its revenue in fiscal 2024; now it's in the low single digits. Nvidia didn't shrink. It just sold its future customers to the only company on Earth with both the motive and the mandate to replace it.

What to watch

Ascend 960 (2027) — does double-per-chip compute close the single-chip gap with H200-class silicon?
HBM supply — China's homegrown memory (CXMT and friends) is the real bottleneck for 750K-chip years. Watch yield news, not press releases.
Washington's next swing — if H200 or B30A-class exports reopen, the test is whether Chinese buyers still want them. Early signs: the door may be reopening onto an empty room.
Xu's 2027 prediction — Chinese AI developers training on Ascend at scale. If it happens, the substitution becomes structural.
Overseas deployments — Huawei's rare foreign sales are the first honest test of Ascend outside a captive market.

The bottom line

The point of export controls was to buy time — keep China's AI a few years behind. Instead, Washington ended up as the silent partner in Huawei's domination of the world's second-largest AI chip market: zero revenue for Nvidia, near-total substitution for Huawei, and a full-stack competitor that now has funding, demand, and a captive test bed no startup could ever buy.

The wall was supposed to keep the chips out. Turns out it mostly kept the competitor in — and made it strong.

Sources: DIGITIMES (Oct 1, 2026); Reuters (Sep 17–18, 2026); The Register (Sep 30, 2026); South China Morning Post (Sep 18, 2026); CNBC; SemiAnalysis China Datacenter Model (2026); Bernstein estimates via The Economist; ITIF, "Backfire: Export Controls Helped Huawei and Hurt U.S. Firms" (Oct 2025); Council on Foreign Relations (Dec 2025); Nvidia FY2026 results.