The Information reports Liang Wenfeng has called Huawei chip training one of DeepSeek’s biggest bets β but the fuller record shows this is a retry after a documented failure, which is a structurally different kind of risk.
The Information reported on September 21, 2026 that DeepSeek chief executive Liang Wenfeng told investors that training on Huawei chips is one of the company’s biggest bets, with Huawei set to deliver a new batch of training chips as early as the fourth quarter of 2026. That is The Information’s reporting, not a release from DeepSeek or Huawei, and nothing here vouches for it beyond that attribution.
A correction to earlier framing is owed before the analysis begins. Coverage here earlier the same day treated the Huawei commitment as a forward bet defined by a delivery window. The fuller record makes it a second attempt after a documented failed one β which is a materially different kind of bet. An earlier training effort on Huawei’s Ascend 910C failed, with Huawei engineers reportedly working on site; DeepSeek reverted to Nvidia for that model, with reported failure modes including stability problems, slower inter-chip connections and weaker software tooling.
The number that organises everything else is Liang’s own: he estimated that a frontier-scale run would need roughly 50,000 Nvidia GB300 processors or roughly 200,000 Huawei Ascend processors β his estimate, not a vendor specification, and one that explicitly excludes the many experiments that precede a final training run. The reported funding context includes a round of roughly 50 billion yuan (approximately $7.4β7.5 billion) at a valuation of roughly 500 billion yuan (approximately $74β75 billion), with roughly $4.5 billion reportedly earmarked for compute. The round is reported, not confirmed closed.
The reported record runs as follows. An earlier training run on Huawei’s Ascend 910C failed and DeepSeek reverted to Nvidia, with reported failure modes of stability, inter-chip connections and software tooling, and Huawei engineers reportedly working on site. Then, at a closed-door meeting on 20 May 2026 — per a transcript that circulated online and was posted to GitHub on 22 July — Liang said: “For us, we can buy some noncompliant cards”, and in that same reported transcript he named no chips, no suppliers, no purchasing route and none of the rules he had in mind. No unlawful conduct is alleged here by anyone.
After that transcript leaked, DeepSeek paused its fundraising, and investors at the most recent meeting were reportedly asked to hand over their phones before proceedings began. On 21 September 2026, The Information reported that Liang told investors Huawei chip training is one of DeepSeek’s biggest bets, with a new batch of Huawei training chips reported for delivery as early as the fourth quarter of 2026. Nothing here claims those chips will arrive, work or fail.
The key insight: Framing this as a forward bet on new chips misses the most important structural fact: the bet has already been placed once and failed. A second attempt against the same opponent, with a harder unit constraint, inherits all the failure modes of the first plus the compounding penalties of greater scale.
Four times the units is not four times the problem β every extra node is another thing that can stop a run.
The Structural Read
Most coverage of the 4-to-1 unit ratio treats it as a linear cost: four times the hardware, four times the expense. That framing understates the problem in a way that matters structurally. When a substitute needs more units to perform the same synchronised job, each additional node adds interconnect paths, power draw, physical footprint and one more component capable of failing mid-run. In any long synchronised job β and a frontier training run is exactly that β the probability that something somewhere fails rises with the number of parts. Critically, the cost of a failure is not one node’s worth of work; it is whatever the entire job has to redo. That is almost certainly why the reported failure modes in the earlier attempt were stability and inter-chip connections, not raw arithmetic speed.
The second structural point is the one that almost every catch-up story misses. Observer Theo Jaffee put it precisely: “Huawei is not catching up to a constant. Nvidia’s improving as well, so Huawei has to catch up to where Nvidia will be.” That is an opinion, not a finding, but the structure it identifies is general and underappreciated. Closing a gap against a moving target requires sustaining a faster rate of improvement than the leader, for as long as it takes. Almost every public comparison in this industry is stated in levels β this chip against that chip, this score against that score. A level comparison can look encouraging while the underlying rates of improvement make the gap widen.
The third structural layer is the one that complicates the phrase “domestic chips” most directly. As reported, Huawei’s binding constraint is high-bandwidth memory, and the HBM used in advanced processors is made by a small number of manufacturers. The general property: a substitution strategy inherits the supply chain of its scarcest input. Changing who assembles the final part does not remove a dependency that sits one layer inside the part. “Domestic” accurately describes who builds the processor; it does not necessarily describe who makes every component within it. No market-share, capacity or price figure is offered here, and no position is taken on anyone’s trade policy.
In the Map of AI framework, compute sits at the foundation layer β everything above it (models, products, distribution) depends on its reliability, not just its throughput. A 4-to-1 unit ratio stress-tests reliability at the foundation in a non-linear way. DeepSeek’s reported characterisation from Liang β one to two years behind US competitors while operating on roughly a twentieth of the computing resources β is his own characterisation, not a measurement, but it frames the stakes accurately: the constraint is foundational, not feature-level.
Theo Jaffee β Observer Opinion
“Huawei is not catching up to a constant. Nvidia’s improving as well, so Huawei has to catch up to where Nvidia will be.”
Three Implications
IMPLICATION 1 β The Retry Premium
A second attempt at a failed training migration is not the same strategic signal as a first attempt. It means the organisation already has direct evidence of what breaks, which is valuable β but it also means the prior failure is now baked into investor expectations, disclosed or not. When a company’s compute position is itself market-moving information, managing what investors know about failure history becomes part of the fundraising process rather than a detail around it. The phone-confiscation detail reported from the most recent investor meeting sits inside that dynamic.
IMPLICATION 2 β The Derivatives Problem for Every Catch-Up Story
The Theo Jaffee observation generalises beyond DeepSeek and Huawei. Any analyst, investor or operator benchmarking a challenger against a leader using level comparisons β chip-vs-chip, score-vs-score, release-vs-release β is answering the wrong question. The structurally correct question is whether the challenger’s rate of improvement exceeds the leader’s. Level data cannot answer that. This applies to every catch-up narrative currently in circulation in AI infrastructure.
The HBM dependency reported inside Huawei’s advanced processors is a supply-chain lesson that applies broadly: strategic autonomy at the system level does not automatically confer autonomy at the component level. Any organisation β government, enterprise or lab β that is evaluating supply-chain risk using the country-of-final-assembly as the primary variable is looking at the wrong layer. The binding constraint is almost always one level deeper than the one being managed.
Business Engineer Framework
The Map of AI β Where DeepSeek’s Bet Actually Lives
The Map of AI maps 200+ companies across nine layers of the AI stack. DeepSeek’s Huawei bet is a foundation-layer event: a stress test on the compute substrate that every model, product and distribution advantage above it depends on. Understanding which layer a strategic move operates on β and which layers it leaves exposed β is the core analytical move the Map of AI is built for.
Editorial disclosure: This article is not investment advice. No unlawful conduct by DeepSeek, Huawei or any other party is alleged anywhere in this piece. All figures are reported rather than independently verified. The round referenced is reported, not confirmed closed. Liang Wenfeng’s statements are attributed as his own characterisations or estimates where stated. The transcript of the 20 May meeting circulated online and is reported as such β it is not vouched for. No position is taken on export controls, sanctions, trade policy or the conduct of any government.
This is not investment advice, and nothing above alleges unlawful conduct by any person or company. The remark about “noncompliant cards” comes from a transcript of a closed-door meeting that circulated online and is reported here rather than vouched for; in it no chips, no suppliers, no purchasing route and no rules were identified, and nothing above states or implies that DeepSeek purchased any such processor, evaded any control, or broke any law. The chip counts are Liang Wenfeng’s own estimate, not vendor specifications or benchmarks, and they exclude the many experiments that precede a final training run. The characterisation of Chinese developers as one to two years behind on roughly a twentieth of the compute is his, not a measurement. Funding figures are as reported, with the round not confirmed closed, and dollar values are approximate conversions from yuan. Theo Jaffee is an observer offering an opinion. Nothing above claims the new chips will arrive, work or fail, and nothing above takes any position on export controls, sanctions, trade policy or the conduct of any government.
Gennaro Cuofano is a business analyst with an International MBA in corporate finance and business strategy (2012). He leads research at FourWeekMBA, which he created, and at The Business Engineer (businessengineer.ai), his newsletter on AI and business strategy, with over 90,000 subscribers and more than 660 published analyses. His work breaks down how companies make money, read from their own filings: revenue and profit structure, capital allocation and ownership. FourWeekMBA reached about four million business people, including C-level executives, investors, analysts and product managers, in 2022 alone. See how we source.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
