Nvidia concedes China AI chip market to Huawei

The headline most outlets ran with — 'Nvidia concedes China to Huawei' — buries the more striking half of the story: China is also conceding Nvidia. After the Trump administration cleared H200 exports under a 25% surcharge in December 2025 and formalized the policy by proclamation on January 14, 2026 ^[4], Beijing instructed domestic firms to stop buying the previous-generation H20 over national-security concerns ^[5]. The result is a zero on Nvidia's quarterly China data center line that no longer reflects what Washington forbids — it reflects what Beijing refuses. Nvidia reported $0 in China-related revenue versus $4.6 billion in the same quarter a year earlier, and shipped no Hopper-family products into the country at all ^[2].

This matters because it changes the policy lever. If China were a willing buyer blocked only by US rules, partial relicensing (H200 with a surcharge, H20 with a license) would unlock real revenue. Instead, the demand side has been politically captured. Industry commentary on the May 21 print converged on the same read: Stacy Rasgon at Bernstein told CNBC the ban 'effectively hands' the AI market to Huawei ^[3], while Huang's own line to investors — 'expect nothing' on near-term approvals — implicitly acknowledges that even if the door reopens, the room is empty ^[1]. Morgan Stanley pegs China's AI chip market at $67 billion by 2030 ^[6], and Nvidia is now structurally outside it.

CFR's Chris McGuire is correct in narrow terms: Huawei's per-chip performance trails Nvidia roughly 5x today and could widen to 17x by 2027 on raw FLOPS ^[7]. But that framing misses where the lock-in war is actually being fought. Huawei's CANN software stack is increasingly positioned as a drop-in replacement for CUDA, and the 950PR is projected to take roughly 60% of China's domestic AI accelerator demand ^[8]. SMIC remains the practical ceiling — yields reported at 50-60% versus TSMC's 80-90% ^[8]— but inside that ceiling, the software story is moving faster than the silicon story.

The per-chip gap also gets narrowed by cluster-scale engineering. Huawei's Atlas 950 SuperCluster paired with the HCCS interconnect closes part of the gap at training-cluster scale, and the DeepSeek lineage compounds this: DeepSeek-V4 was reportedly tuned for Ascend 950PR ^[8]. The strategic implication is that CUDA's moat — the real reason Nvidia commanded 95% China share before controls — is being deliberately retired inside China's borders. Bernstein's January 2026 estimate already had Nvidia and Huawei tied at ~40% each in China ^[3]. By Q1 FY2027 that balance has snapped, and the snap is software-driven as much as silicon-driven.

Nvidia's stock didn't break on the China zero because the company simultaneously reframed itself as a CPU company. Huang announced that the Vera CPU line opens a 'brand new' $200 billion total addressable market 'we have never addressed before, and every major hyperscaler and system maker is partnering with us to deploy it' ^[1]. CFO Colette Kress guided to $20 billion in 2026 Nvidia CPU revenue ^[2], and the company paired the print with an $80 billion buyback authorization ^[1]. Total revenue still printed $81.6 billion, up 85% year-over-year ^[2]. The market read: lose a $4.6B-per-quarter China datacenter business, replace it with a multiple of that in datacenter CPUs sold into the same hyperscalers Nvidia already owns on GPU.

This is the cleanest explanation for why Bernstein's Rasgon can simultaneously say China was 'handed to Huawei' and that Nvidia stock keeps working ^[3]: non-China demand is supply-constrained, not demand-constrained, so the dollars Nvidia can ship anywhere it wants are still gated by capacity. The Vera pivot is also a strategic answer to the lock-in story above — if CUDA is being decoupled inside China, Nvidia's response is to deepen the rack-level integration story (CPU + GPU + interconnect + NVLink) outside it, where lock-in still compounds. The China line item becomes a rounding error against a $20B-and-growing CPU business that didn't exist a year ago.

The Taiwan detentions are not an isolated enforcement event. They are a new node in a smuggling map that has surfaced in the last fourteen months. The US DOJ arrested Super Micro co-founder Wally Liaw in March 2026 over an alleged $2.5 billion ring routing Super Micro/Nvidia servers through Taiwan, Thailand and Hong Kong to mainland China between 2024 and 2025 ^[9].

Against that backdrop, Taiwan's Keelung District Prosecutors' Office raiding 12 sites, seizing ~50 AI servers and NT$9 million in cash, and moving to detain three men named You, Wang and Chen on May 22 ^[10]reads as the first formal Taiwan-side enforcement of US controls on its own soil — a sovereignty signal as much as a criminal case. Huang's response, telling reporters in Taipei that Super Micro 'has to run their own company' and urging it to 'enhance and improve their regulation compliance' ^[11], is the public part of what is almost certainly a private message to every Nvidia OEM partner: the perimeter is being audited, the prosecutors are now on it, and your access to allocation depends on your paperwork. The grey market that propped up Nvidia's de facto China shipments through 2024-2025 is being closed from the supply side at the same time the demand side disappears.

Two competing readings of the same data dominate expert commentary. CFR's McGuire frames the outcome as proof the controls are working: Huawei is bottlenecked by SMIC yields reported at 50-60% versus TSMC's 80-90% ^[8], the per-chip performance gap is widening not narrowing, and pushing China onto a slower domestic stack is precisely the desired outcome ^[7]. Reuters/IDC data via Medium shows Nvidia's China accelerator share fell from 95% to 55% across 2025 ^[12]— under the McGuire reading, that is exactly what successful denial looks like. The American Action Forum framing of the 25% H200 surcharge as a 'political tax' ^[4]adds a fiscal dimension: even where sales are allowed, they're now revenue events for the US Treasury.

The contrarian read is that controls accelerated the very capability they were meant to suppress. Huawei's $7.5B-to-$12B revenue jump year-over-year ^[6]and the DeepSeek-V4-on-950PR pivot ^[8]describe an ecosystem that has industrialized rather than stagnated. Even Bernstein's January 2026 read of a 40/40 split ^[3]implies parity reached faster than the McGuire timeline assumed. The honest synthesis is that both can be true: per-chip leadership remains American, but ecosystem control inside China has shifted irrevocably, and the question for the next four years is whether that ecosystem stays bounded by Chinese borders or starts winning third-country deals — which is the scenario in which McGuire's read fails and the contrarians' wins.