The takeaway: DeepSeek has told prospective investors in its second funding round that it is suspending the deal for now, Bloomberg reported on 25 July, after remarks widely attributed to founder Liang Wenfeng went viral. The document at the centre of it is a leaked transcript of a meeting held on 20 May, published by Tencent Tech in late July, and its own header states that it was auto-transcribed by speech recognition, organised by AI, and that "individual proper nouns and numbers may contain recognition errors". The striking figures now circulating from it — a 4:1 chip ratio, a one-twentieth compute claim, a two-year gap — are precisely the class of detail that warning covers.

What is actually confirmed

Less than the headlines imply.

According to people familiar with the matter, cited by Bloomberg, DeepSeek verbally informed some would-be backers that they would not be signing investment agreements in the coming days, "as they had expected". The suspension "stemmed in part from Liang's frustration over online reports about his comments to investors during his first financing deal", which closed in June.

Note the hedge: in part. Bloomberg does not present the leak as the sole cause, and it keeps two objects carefully separate — the viral posts, and the transcript those posts concerned.

The rest is heavily qualified in the reporting itself. Negotiations "remain fluid" and the company may still proceed. It is unclear whether DeepSeek has communicated its intentions to all prospective investors. The sources asked to remain anonymous to discuss private discussions. DeepSeek did not immediately respond to an emailed request for comment about the transcript or the fundraising.

The document, and how it was made

This is the part most coverage has skipped, and it changes how everything else should be read.

Bloomberg's own formulation is worth quoting in full, because the qualifier is load-bearing: "Bloomberg hasn't verified the authenticity of those posts, which concerned a transcript of a meeting Liang held with unidentified parties."

The transcript itself was a leak from a closed-door investor meeting. Multiple Chinese outlets ran differing versions; Tencent Tech's, at 118 numbered remarks, is the most complete, and the WeChat links have since been taken down. The newsletter Hello China Tech, which analysed it on 23 July, states plainly: "DeepSeek has not confirmed the record."

And the document carries a preamble that deserves to travel with every quotation lifted from it:

Audio "deepseek 0520.m4a", total length approximately 3 hours 44 minutes. This transcript was auto-transcribed from speech recognition and organized by AI, without distinguishing speakers; the content in brackets indicates the audio timestamp position; individual proper nouns and numbers may contain recognition errors, please refer to the original recording as authoritative.

Three things follow. The meeting was held on 20 May — the filename says so, and two independent accounts of the document give the same date. It surfaced roughly two months later. No human is vouching for the wording. And the transcript explicitly warns that numbers may be wrong.

Nothing below should be read as something Liang said. It is what a machine-assisted transcript, unconfirmed by the company, records him as saying.

Why the disclaimer bites

The numbers doing the most work in the coverage are the ones the warning specifically covers.

The timeline is the clearest case. In a single exchange (#57) DeepSeek is described as 12 months behind, then 12 to 18 months, then 6 to 12 months, before settling on "to put it simply, two years behind, using one-twentieth of the compute." Hello China Tech reads this as conversational approximation rather than a calibrated estimate, and notes the transcribed text appears muddled at that point. With a speech-recognition disclaimer attached, that may account for some of the individual figures — though not for the shape of the exchange itself. Misrecognition garbles a numeral; it does not manufacture a coherent sequence of four different ones.

The hardware claims are the same shape. The transcript has Liang calling Huawei's 950 super-node a full replacement for Nvidia's GB200 and GB300 "in performance and price", and asserting that "four Huawei chips match one Nvidia chip" (#67, #68). Hello China Tech's objection stands on its own terms — no workload, no specific parts, no baseline, so a 4:1 ratio is not a measurement. Add machine transcription of spoken numerals and it is weaker still.

What the transcript argues

With that carried all the way through, the thesis is coherent and blunt.

Its central claim is that one variable explains the field: "All the differences we see, including talent, model capability, and applications, can be attributed to differences in compute resources." (#56)

The talent point cuts against the prevailing Western reading: "Talent is not the bottleneck. Resources are the biggest bottleneck. Resources first affect talent development: less compute means fewer experimental opportunities, so our talent base overall is weaker than America's. The talent gap is fundamentally a compute gap." (#43)

On scale, it rejects the idea that DeepSeek's efficiency is a philosophy rather than a constraint: "We believe in scaling. Bigger is always better. What stops us from scaling is compute, not desire. We train a model at this size not because it is enough, but because that is all our resources allow." (#59, #60)

Since January 2025 the standard account of DeepSeek has been an efficiency story — a comparatively small lab showing that frontier performance could be approached without frontier compute budgets. The transcript reframes that record as adaptation to scarcity rather than a discovery about what is necessary.

Read straight, it is also a pitch. A company two years behind on a twentieth of the compute is a company unusually good at converting resources into capability, and that is the thesis underneath the valuation.

What it does not say

The compute thesis has been read in some quarters as an argument that money is beside the point. The document does not support that, and it is worth being precise about why.

The transcript names capital directly, in the same breath as chips: "The biggest gap between us and the US is in resources. On one hand, we can't buy enough chips domestically. On the other, our capital investment is far less than America's. The salary share is small. The bulk is compute." (#55)

It also supplies a second, non-compute rationale for raising. On team stability — described as the company's "single biggest core interest" (#37) — it has Liang saying: "This risk has been substantially relieved by our recent fundraise. Everyone received fairly large option grants." (#39) That is past tense in a meeting held before the June round closed, which fits Bloomberg's description of remarks made "during his first financing deal" — a process, not a single event. Hello China Tech's conclusion is that the transcript confirms the fundraise was intended to reduce retention risk.

And Bloomberg states the purpose of the current round plainly: "The startup is now chasing more funds to support an ambitious expansion plan, including an increase in computing capacity."

Compute being the binding constraint is an argument for capital, not against it. The transcript does note that money alone will not conjure chips: "Spending Rmb 20bn this year would mean our procurement team did an exceptional job. It is extremely difficult to spend that much. You can't buy that many chips, and the prices are high." (#54) — a spoken figure, and subject to the same disclaimer as the rest. Both things hold at once, and the document does not resolve them.

The figures have never agreed

Anyone tracking the round should know how wide the reported spread is.

For the current round, two outlets have carried different numbers for a fortnight. Bloomberg reported in mid-July that DeepSeek was seeking at least ¥10bn at a pre-money valuation of at least ¥480bn — about $71bn — alongside IPO preparations. Reuters reported on 20 July, which we covered at the time, a raise of as much as ¥50bn at a valuation of about ¥500bn. Bloomberg's 25 July story restates its earlier figures rather than revising them.

If Reuters' ~¥500bn is a post-money figure, it and Bloomberg's ¥480bn pre-money describe much the same valuation — but neither outlet states which basis it is using, so the apparent agreement is an inference rather than a reconciliation. The unambiguous gap is the size of the round: a ¥10bn floor against a ¥50bn ceiling.

The June round has been reported across a wider spread still. Reuters put it at about ¥50bn ($7.4bn) at a post-money valuation of ¥350–400bn ($52–59bn); filings by two Chinese investors later implied ¥350.88bn, around $52bn; Reuters itself on 20 July used about ¥450bn post-money; and Bloomberg says $7bn against a "roughly $50 billion price tag". Bloomberg sits at the low end of that spread and our June coverage used the top of it. No single figure is settled, and Reuters has carried two of them itself.

One number is firmer, because it comes from a maintained index rather than a deal leak: Liang's net worth more than doubled after that round, to $36bn from about $16.7bn, on the Bloomberg Billionaires Index — placing him above Anthropic's Dario Amodei and OpenAI's Greg Brockman among creators of AI models.

What to watch

Two things, both checkable.

Whether the round resumes, and on what terms. Bloomberg is explicit that it may. A pause conveyed verbally to some investors is not a withdrawal.

And the transcript's one dated claim. It has Liang saying: "There is a historic window for domestic AI chip substitution. We believe that within the next year, something will be verified: the domestic chip ecosystem has no problems at all." (#64) Because the meeting is dated 20 May 2026, that points to a deadline of roughly May 2027.

The same disclaimer applies here as everywhere else: "within the next year" is a spoken quantity in a machine-transcribed document, and it is offered as the only claim in the record with a testable shape — not as a verified deadline. But if the compute thesis is right, the domestic substitution timeline is the story, and it is the one thing a funding round cannot accelerate.

Background: we have covered DeepSeek's June raise, the follow-on round and its STAR Market plans, and how DeepSeek cut its own running costs and published the method.