Two numbers landed on Huawei's ledger this year, and they point in opposite directions.
The first: roughly $12 billion in AI chip revenue expected in 2026, up from $7.5 billion in 2025 — a jump of at least 60%, and not a forecast built on optimism but on orders the company says it has already booked, according to a Financial Times report that Reuters said it could not independently verify.
The second: a 36% collapse in first-half net profit, to 23.81 billion yuan (about $3.54 billion), disclosed on Monday, Aug. 31. That is a steeper drop than the 32% fall recorded in the same period a year earlier. Revenue climbed 9.6% to 467.82 billion yuan, roughly $69.6 billion. Growth is real. It is simply not reaching the bottom line.
The gap between those two numbers is the price of building a domestic AI stack from the silicon up, and Huawei is paying it in cash.
The Cost of Replacing America
Research and development spending rose 25.2% in the first half to 121.38 billion yuan — 25.9% of revenue, well above the 21.8% Huawei devoted to R&D across all of 2025. The cost of making its products rose 12.4%, outpacing revenue growth, with rising memory chip prices squeezing margins in the consumer division that sells smartphones. Administrative costs also climbed sharply.
The cash flow statement is starker still. Huawei's day-to-day operations consumed 39.9 billion yuan in the first half, against 31.2 billion generated a year earlier — a swing of roughly 71 billion yuan. Inventories rose 42% from the end of 2025, the signature of a company stockpiling components and building ahead of demand it expects but has not yet billed.
Huawei said the results were in line with its own forecasts. It also said its full-year outlook remains under review, citing external uncertainty and higher input costs.
The privately held Shenzhen company, which discloses results voluntarily, has been here before. U.S. sanctions and export controls cut its access to advanced chips and Google's Android, contributing to a 29% revenue collapse in 2021. Revenue has since recovered to 880.9 billion yuan in 2025, its second-best year ever. What is new is that the recovery is now being financed by an R&D budget growing faster than sales.
DeepSeek Flipped the Switch
The demand surge has a specific trigger and a specific date. On April 24, DeepSeek released its V4 model, optimized to run on Huawei's Ascend architecture and its CANN software framework rather than Nvidia's CUDA. Huawei engineers reportedly worked directly with DeepSeek ahead of launch. The company confirmed its full Ascend SuperNode line supported V4 on day one, and that its chips were used to train the V4-Flash variant. Alibaba Cloud and Tencent Cloud deployed V4 services within hours.
Then the buying started. Alibaba, ByteDance and Tencent are all named among the firms racing to secure supply. Prices for the Ascend 950PR — which entered mass production in March and has captured the majority of this year's orders — have reportedly risen about 20% on demand. An upgraded, training-focused 950DT is due in the fourth quarter.
The technical fit matters. V4 uses a mixture-of-experts design with up to a trillion total parameters but roughly 37 billion active per inference pass, which rewards efficient inference hardware over raw training throughput. The 950PR is currently the only Chinese-made AI processor supporting FP8. Huawei is separately reported to be roughly doubling output of its older Ascend 910C, toward about 600,000 units, while targeting roughly 750,000 950PR units this year — a figure expected to fall short of demand.
Huawei's rotating chairman, Eric Xu, framed the strategy plainly when he laid out the Ascend roadmap: "Computing power is — and will continue to be — key to AI. This is especially true in China." His stated goal is to meet long-term compute demand using the process nodes actually available on the Chinese mainland.
Why This Matters
Eighteen months ago Nvidia supplied the overwhelming majority of AI silicon used by Chinese cloud providers. Chief executive Jensen Huang now says the number is nothing at all: "In China, we have now dropped to zero," he said on a Special Competitive Studies Project podcast, calling U.S. export policy a strategy that "has already largely backfired." Nvidia's FY2026 10-K states the company is "effectively foreclosed from competing in China's data center computing market."
Into that vacuum, Bernstein projects Nvidia's China AI GPU share falling to roughly 8%, from 66% in 2024. TrendForce expects China's high-end AI chip market to grow more than 60% in 2026, with domestic suppliers taking about half. Morgan Stanley sizes the market at $67 billion by 2030.
Huawei's $12 billion is therefore less a company milestone than a measurement of how fast a parallel AI stack has assembled itself — model, chip, compiler, cloud — with a Chinese model as the forcing function rather than a Chinese chip. The margin damage is the tell: this is a market being built, not harvested.
What to Watch
Three things. First, whether SMIC can deliver. Its N+3 7nm-class node runs without EUV lithography, yields trail TSMC's, and JP Morgan estimates a wafer-to-packaged-chip cycle time near eight months versus roughly three at TSMC. Second, the 950DT in Q4 — inference wins are the easy half, and training remains where Chinese firms are most constrained. Third, Huawei's full-year outlook, currently under review. If R&D keeps compounding faster than revenue, 2026 will be remembered as the year the domestic stack arrived and nobody made money on it.
“Computing power is - and will continue to be - key to AI. This is especially true in China.”— Eric Xu, Deputy Chairman and Rotating Chairman, Huawei