Published: 2026-08-16 | Reading Time: ~11 minutes | Channel: techminute
Here's the number that should make every AI investor's coffee go cold: TrendForce now projects that Chinese domestic suppliers will capture nearly 90% of China's high-end AI chip market in 2026, leaving NVIDIA, AMD, and every other overseas vendor fighting over roughly 10%. For context, NVIDIA's data-center GPUs were, until recently, the default accelerators for essentially the entire Chinese AI industry. The same research firm says domestic high-end AI chip shipments are growing about 83% year-over-year — and Huawei, the company the US has spent four years trying to strangle, is on course to be the single biggest winner of all.
Then there's the other headline, the one with a human voice attached. On the Dwarkesh Podcast this spring — and re-circulated hard this week — NVIDIA CEO Jensen Huang said the quiet part out loud: "The day that DeepSeek comes out on Huawei first, that is a horrible outcome for our nation." On a recent Prof G Markets segment, analyst Ed Elson called the exchange "the first time I've seen him as defensive."
When the CEO of the world's most valuable chip company — a company sitting on roughly a $5.18 trillion market cap — publicly fears a Chinese open-weights lab running its models on Chinese silicon, the AI chip war has officially changed shape. It's no longer about who builds the fastest GPU. It's about who owns the software stack that makes GPUs useful, and whether a parallel Chinese ecosystem can grow up without CUDA at all.
This is the story of how that happened, what the numbers actually say, and why the scariest part for NVIDIA isn't the chip Huawei builds — it's the software Chinese developers are rewriting around it.
Let's be precise about what TrendForce actually said. The research firm — Taiwan-based, one of the most widely cited names in semiconductors — published the projection in early August 2026: Chinese domestic solutions will take almost 90% of high-end AI chip sales in China this year, with NVIDIA, AMD, and other foreign suppliers holding about 10%. DIGITIMES headlines the same trend as "Huawei on course to eclipse Nvidia."
That's a stunning reversal for a market NVIDIA effectively pioneered. And it's driven by a few compounding forces:
Export controls have a cafeteria effect. Since April 2025's H20 ban, US policy on China AI chips has zigzagged — "off, on, off, on," as analyst China Alice Han put it. NVIDIA's own CFO confirmed earlier in 2026 that even the handful of H200 product shipments approved for China-based customers produced zero revenue. Chinese buyers learned the lesson: betting your AI infrastructure on a chip that can legally be unapproved tomorrow is a losing proposition.
Domestic capacity is scaling fast. TrendForce's earlier estimates put Chinese domestic AI chip shipments at around five million units in 2026, with SMIC and Shanghai Hua Hong foundry production growing at a ~50% CAGR for the next couple of years. Not chips that are as strong as NVIDIA's best; chips that are actually available.
China's hyperscalers build their own. The 90% isn't all Huawei. TrendForce specifically names custom ASICs from Baidu, Alibaba, and Tencent, plus merchant silicon from Huawei and Cambricon. When your biggest cloud customers are designing their own accelerators, the foreign share erodes from two directions at once.
Throw in a 2026 market that is already majority domestic — one analysis cited by KuCoin suggests Huawei alone could exceed 50% of China's AI chip market — and the 90/10 split stops looking like a ceiling and starts looking like the floor.
Here's what most coverage gets wrong: the story isn't mostly about silicon performance. It's about the CANN vs. CUDA software battle.
NVIDIA's genuinely durable moat was never just the H100 — it's CUDA, the 20-year-old software framework that every AI researcher, every startup, every hyperscaler by default writes code for. Even when chip import controls forced Chinese labs to use alternative hardware, their software was still written for CUDA, so the dependency on the NVIDIA ecosystem survived.
DeepSeek's V4 broke that dependency. It's the first frontier-class Chinese large model that runs natively on Huawei's Ascend 950PR processor, released Q1 2026. To get there, DeepSeek spent months rewriting its core code for Huawei's CANN framework, sidestepping CUDA for the first time at frontier scale. According to EE Times/IoT via TrendForce, Huawei plans to produce ~600,000 Ascend 910C packages in 2026 — double 2025 — and lift total Ascend capacity to ~1.6 million units. A follow-on, the Ascend 950DT, is being pulled forward to August precisely because demand is so hot.
The hardware gap is still real — let's not romanticize it. The Ascend 910C, the predecessor to the 950 series, delivers roughly 60% of an H100's inference performance, and the US-China performance gap was about 5x as of April 2026, with some analysts projecting 17x by 2027. Huawei's entire targeted 2026 shipment volume of ~750,000 AI accelerators equals only 3–5% of NVIDIA's aggregate compute. DeepSeek's own earlier R2 model famously struggled with training stability on Huawei parts and had to revert to NVIDIA GPUs for training runs.
Raw compute is not, however, what's driving the 90/10 forecast. The market is moving on availability, regulatory certainty, price, and scale. A slightly slower chip you are allowed to buy a million of beats a faster chip you aren't allowed to import. And at inference time — where the commercial revenue actually is — Huawei's chips are good enough that the value proposition now works.
The deeper point is software. If DeepSeek can run a frontier model natively on CANN, then every Chinese lab, startup, and cloud has a new reference architecture to copy. That's how ecosystems tip.
| Metric | NVIDIA / West | China Domestic | Source |
|---|---|---|---|
| China high-end AI chip share, 2026 (TrendForce) | ~10% | ~90% | Wccftech, DIGITIMES |
| Huawei share of China AI chip market | — | >50% (one analysis) | KuCoin-survey cite |
| High-end domestic shipments YoY | — | +83% | Wccftech |
| Domestic AI chip shipments 2026 | — | ~5 million | TrendForce/Wccftech |
| NVIDIA H20 ban accounting hit | $4.5B write-down | — | 24/7 Wall St/Yahoo |
| NVIDIA Q1 FY27 guidance | ~$78.0B, zero China DC revenue | — | 24/7 Wall St/Yahoo |
| China AI accelerator TAM (Huang's own estimate) | ~$50B | ~50B | Huang/Yahoo |
| NVIDIA FY2026 revenue | $215.94B (+73% Q4 YoY) | — | 8-K via Yahoo |
The standout row is the middle: for the first time, NVIDIA's official forward guidance assumes zero China data-center compute revenue — this from the company that generated $17 billion from China in a previous year. And the write-down on the H20 ban ($4.5B) is the visible bump on top of a much longer ledger of lost sales.
Let's unpack that "horrible outcome" quote properly, because it deserves more context than the clickbait version.
Huang said this on a podcaster's show in April 2026 regarding DeepSeek's plan to launch V4 on Huawei. But it resurfaced the week of Aug 13 when 24/7 Wall St. covered the Prof G Markets commentary — and that's why it's fresh news. Two moments sharpen it:
On NVIDIA's Q1 FY2026 earnings call, Huang made the strategic stakes explicit: "The question is not whether China will have AI, it already does. The question is whether one of the world's largest AI markets will run on American platforms." He pegged the China AI accelerator TAM at roughly $50 billion — a market NVIDIA is now effectively locked out of.
The threat vector isn't that Huawei beats NVIDIA head-to-head. The threat vector is that DeepSeek proves a completely alternative stack works. V4 already handles a 1-million-token context on par with Gemini and leading US labs. If a frontier model runs well enough on CANN + Ascend, three things happen at once:
That's the nightmare on the whiteboard. Not a faster Chinese GPU. A parallel Chinese AI ecosystem that simply doesn't need NVIDIA anywhere in its supply chain.
For NVIDIA: The $5.18 trillion market cap and $215.94B FY2026 revenue are not in immediate danger — demand outside China is still nuclear. NVIDIA's answer to the specific China problem is a compliance-friendly RTX Pro 5000 with GDDR7 memory variant, reportedly coming for the Chinese market. This is the same strategic playbook as the old H20: ship a good-enough sanctioned chip, fight for the scraps.
For the rest of the industry: Microsoft's Maia, Meta's Iris, Google's TPUs, Amazon's Trainium — the whole custom-silicon wave — just got a validation boost. If China can break away from CUDA with software plus adequate silicon, the "franchise" argument for home-built accelerators in the US gets even stronger. The post-NVIDIA world was already started; Thursday's forecast is one more big tailwind.
For Chinese cloud providers: The 90% scenario makes the domestic stack strategically immune. Alibaba, Tencent, ByteDance have reportedly raced to secure Ascend 950 orders after V4 launched (Reuters, April 2026). They don't have to wait for US permission slips anymore.
For the CUDA narrative everywhere: This is the first real crack in the universal-write-for-CUDA story. DeepSeek R2's painful training failures on Huawei, and V4's... workable launch, tell you exactly where the shift happens: training remains NVIDIA's territory, but inference — where the money is — is defecting.
Everything is forecast. The 90/10 split is TrendForce's 2026 projection, not a delivered result. Three months ago it was an "almost entirely domestic" projection; it's the direction that's the news, not the precise 90.0.
Huawei >50% share is not TrendForce's headline claim. That figure comes from a separate survey-cited analysis (via KuCoin). Treat it as directional, not ground truth.
The performance gap is still enormous. 5x today, possibly 17x by 2027. China is winning market share long before winning performance. That's not necessarily a comfort — markets, not benchmarks, are what drive ecosystems — but it's an honest hedge against the "Huawei has arrived" narrative.
The R2 cautionary tale is recent. DeepSeek shared frontier training runs as recently as 2026 repeatedly struggled to complete on Huawei hardware. The production track record for this transition is still thin; Ascend 950DT deployment is only just beginning.
Geopolitical tilt both ways. TrendForce is Taiwan-based; DIGITIMES likewise. Both have a domestic-comprehension reflex in their coverage. And Huang's "horrible outcome" framing is, at minimum, pricing-motivated — NVIDIA stands to gain if policy interprets his warning as "keep export controls strict." Report, don't internalize.
The 90% forecast is not a funeral for the GPU era. It's the calendar for a new one — one where AI has two software ecosystems instead of one. Like every great competition, the real battlefield is the stack you never see. And the guy who built the first stack just told us, on the record, that he's worried.
All claims verified against Silver-tier sources (Wccftech/TrendForce, DIGITIMES, Digital Citizen, 24/7 Wall St via Yahoo, The Next Web, Reuters, CNBC), each scraped and confirmed accessible on 2026-08-16 (DIGITIMES partial-paywalled after lede; KuCoin snippet-level). Market share figures are TrendForce forecasts for 2026; performance-gap figures are as of April 2026. Last verified: 2026-08-16.
