Export controls are supposed to limit compute in China and thus slow down China's AI development and deployment. Do they work? One type of evidence is what Chinese AI companies directly say about compute. I've compiled several dozen relevant quotes below.
Almost all companies acknowledge compute is scarce in China. Some voice optimism about domestic alternatives for inference but are doubtful Chinese companies will be able to compete at the frontier.
(AI company and cloud provider)
Alibaba has repeatedly acknowledged compute constraints.
November 2025 — Eddie Wu (CEO), says Alibaba can’t keep up with demand because of limited compute:
“The pace at which we can add new servers is insufficient to keep up with the growth in customer orders.”
“Across all of those different links in the value chain that go to making AI servers, there is a situation of undersupply. Supply is unable to keep up with demand for all of these components globally.” Estimated constraints will last “two to three years.”
January 2026 — Qwen team lead Lin Junyang tells a panel that US compute dwarfs China’s and he thinks it's very unlikely a Chinese companies will lead in AI.
“U.S. compute may overall exceed ours by one to two orders of magnitude.”
“[We are] relatively constrained — just fulfilling delivery requirements already consumes the vast majority of our compute.” Asked the chance that the leading AI company in three to five years is Chinese, Lin put it at “20%” — “already very optimistic.”
March 2026 — Alibaba Cloud CEO Yongming Wu concedes on the Q3 FY2026 call that its chips lag foreign counterparts and that compute constraints will last for 3–5 years.
“Given that our chips still lag behind foreign counterparts and performance in various respects…”
“Over the next 3 to 5 years, global AI compute will be in a very tight supply phase.”
May 2026 — On the Q4 FY2026 call, Eddie Wu says Alibaba doesn’t have any spare capacity but will be able to address demand with in-house chips.
“There isn’t a single idle card in our servers.”
“As the only AI Cloud provider in China capable of delivering self-developed AI chips at scale, we’ve secured autonomy over our compute supply chain.”
(Alibaba disclosed T-Head had shipped only ~470,000 chips as of February 2026, a tiny fraction of the compute that American companies like NVIDIA or Google produce, especially if accounting for the much lower performance of Alibaba’s chips.)
(AI company and cloud provider)
While Baidu has also acknowledged compute constraints, they've also voiced optimism about Chinese chip supply in the medium term, at least for inference.
February 2024 — CEO Robin Li calls the sanctions impact “minimal,” citing a one-to-two-year chip reserve.
“minimal for model development, product reinventions or monetization.”
“Our AI chip reserve enables us to continue enhancing ERNIE for the next one or two years.”
February 2024 — Robin Li says Baidu may lose access to cutting-edge GPUs long-term.
“In the long run, we may not have access to the most cutting-edge GPUs, but with the most efficient homegrown software stack, net-net, the user experience will not be compromised.”
May 2026 — AI Cloud EVP Dou Shen tells the Q1 call domestic chips still trail at the frontier for training but compete for inference.
“China’s domestic homegrown AI chip market is still early, but moving fast. Domestic chips are still catching up with the most advanced global products in certain frontier training scenarios. Inference is an area where domestic chips can be highly relevant and competitive.”
(AI company and cloud provider)
ByteDance hasn't said much about compute constraints. Though that shouldn't be surprising since there's plenty of evidence that the company is using Nvidia chips on a large scale through data centers outside of China.
March 2026 — ByteDance plans ~$5.6B in Huawei Ascend purchases for 2026 (≈40bn yuan) and a separate ~$14B Nvidia H200 budget (up from ~$12B on Nvidia in 2025), after ~$7B earmarked for offshore Nvidia rental in 2025.
April 2026 — Zhang Chi, a former ByteDance Seed researcher now at Peking University, gives an insider account on the Into Asia podcast: China is falling further behind, iteration is half Google’s pace, the fastest chips are rationed to the top teams, and domestic chips aren’t used for training.
“I don’t even agree with the assumption that Chinese models are catching up — I believe we’re still far behind. I guess the gap is getting larger, very sadly.”
“Rumor has it that Google can train or perform a full round of LLM training, both pre-training and post-training, in three months… But ByteDance — probably we can only do one iteration in half a year.”
“The fastest chips are reserved for the most important teams, for example, pre-training and post-training. But other teams, I guess, we use H20.”
“Every tech company in China is facing this challenge of where to get more chips, including Alibaba and Tencent.”
“[ByteDance] has these domestic chips, but I don’t think any team trying to accelerate is actually using those chips… it’s really hard to make full use of them.” Domestic chips are “definitely not for training.”
(AI company)
DeepSeek has openly said that compute is a key challenge. They've reportedly struggled to train models on immature Chinese AI chips.
July 2024 — Liang Wenfeng (founder & CEO) says DeepSeek’s biggest problem is AI chip export controls, not money.
“We have no near-term fundraising plans. Our problem has never been money — it’s the embargo on high-end chips.”
August 2025 — Three unnamed sources claim DeepSeek wasn’t able to train its R2 model on Chinese hardware.
“DeepSeek was encouraged by authorities to adopt Huawei’s Ascend processor rather than use Nvidia’s systems after releasing its R1 model in January, according to three people familiar with the matter.
But the Chinese start-up encountered persistent technical issues during its R2 training process using Ascend chips, prompting it to use Nvidia chips for training and Huawei’s for inference, said the people.”
(AI company)
iFlytek has voiced concerns about relying on American AI chips, but claims Huawei's new chips will allow them to train a model on par with American models.
March 2025 — Chairman Liu Qingfeng says that domestic training depends on imported compute.
“Domestic large model training relies heavily on imported computing power. This is tantamount to ‘building skyscrapers on someone else’s foundation.’”
April 2026 — On the earnings briefing, Liu Qingfeng details the Ascend 910B’s hardware gap vs H200, and an Ascend-950 flagship for October.
“The Ascend 910B chip has hardware design limitations, including memory capacity significantly below the H200 and roughly 2× lower memory bandwidth.”
“This October, iFlytek will release on the Ascend 950 platform China’s first flagship large model benchmarked against the world’s most advanced mainstream models.”
(Note Huawei's 950PR, released in 2026, is only roughly half as performant as NVIDIA's H100, released in 2019)
(AI company and cloud provider)
Tencent used to express confidence about chip supply but has recently acknowledged that they, too, are constrained by U.S. export controls.
August 2024 — President Martin Lau tells the Q2 2024 call Tencent has enough training chips and options for inference.
“From our perspective, we do have enough chips for training and continuous upgrade of our existing models. We also have many options for inference chips.”
May 2025 — Martin Lau describes a “pretty strong stockpile,” enough for a few more model generations.
“We should have enough high-end chips to continue our training of models for a few more generations going forward.”
November 2025 — Tencent executives attribute a 23% YoY capex decline to chip supply-chain changes.
“a change in terms of AI chip availability… supply-chain constraints sourcing GPUs.”
January 2026 — Tencent researcher Yao Shunyu assesses:
“Whether China’s lithography machines can break through” is the real question.
(Lithography, specifically DUV and EUV are required to make the advanced logic and memory dies needed for AI chips but China’s access to lithography is restricted, thus severely constraining their ability to produce AI chips domestically.)
March 2026 — On the Q4 2025 call, Martin Lau says Tencent Cloud's 2025 revenue was constrained because Tencent prioritized its GPUs for internal use over renting them to external customers — though cloud still grew and turned a profit.
May 2026 — CSO James Mitchell tells the Q1 call capex will jump in H2 as China-designed ASICs arrive.
“Tencent expects a substantial increase in CapEx, especially in the second half of the year, as more China-designed ASICs become available.”
On the same call, Tencent management acknowledge Tencent Cloud has lacked enough GPUs, with relief pinned on domestic chips.
“Tencent Cloud has consistently lacked sufficient GPU resources, which has affected our ability to gain more revenue and market share; but in the second half of this year more domestic chips will be put into use, and we expect that as domestic GPU and ASIC supply is gradually released, the tight compute-supply situation will ease.”
(AI company)
January 2026 — Zhipu announces GLM-Image was trained end-to-end on Huawei Ascend (Atlas 800T A2, MindSpore). Some sources wrongly extended this to the GLM-5 flagship, whose report references CUDA. However, in my understanding Ascend support for GLM-5 is for inference (“Day 0 adaptation”), not training.
February 2026 — After GLM-5’s launch, Zhipu publicly asks GPU providers for support, admitting it couldn’t serve demand.
March 27, 2026 — CEO Zhang Peng tells the Zhongguancun Forum that compute is the sector’s biggest problem for the coming year.
“The biggest problem large-model companies face over the next 12 months may be compute.”
“A couple of years ago Academician Zhang Ya-Qin said: ‘Talking about [compute] chips hurts feelings; having no chips means no feelings.’ Today we’re back at that point, but it’s different — we’ve shifted to the inference stage, because demand is exploding ten- to hundred-fold and much of it is still unmet.”
March 31, 2026 — Zhipu’s HKEX annual results cite compute constraints:
“Faced with the computing power shortage that has outstripped supply since February 2026, we will continue to increase our investments, particularly in the ‘Day 0’ adaptation of domestic chips and the optimization of software-hardware integration.”
“We rely on third parties to provide computing resources to us, and any disruption of their services or fluctuation of prices could adversely affect our business, results of operations and financial condition.”
(AI company)
October 2024 — CEO Kai-Fu Lee says limited compute makes it challenging for Chinese companies to catch up.
“Shortening the gap [in AI capabilities] is extremely difficult, and I don't predict we can do so. After all, they trained their model using 100,000 GPUs, while we trained it using only 2,000 GPUs.”
My read of these quotes is that U.S. export controls have very clearly constrained Chinese compute and thus made the lives of Chinese companies a lot harder. Of course, we shouldn't just look at what Chinese companies say publicly but also directly at the capabilities they achieve, as well as a few other factors, which I'll discuss in upcoming work.
Several companies mention optimism about Chinese AI chips — will they close the hardware gap? Probably not, given they trail U.S. chips by several years, both in performance, and also in production volume.
Some companies, e.g. Moonshot, just haven't said much about compute, so I've not included them here.
I don't speak Chinese so I've relied on machine translation & Claude. Let me know if something is taken out of context!
Have another relevant quote? Let me know (see my email in the footer.)