The Transcript Circulated on July 23, but No Original Audio Is Public
A 42-page PDF titled "Liang Wenfeng Investor Meeting: Audio Transcript" circulated publicly on July 23. Its cover lists May 20 as the recording date and July 16, 2026 as the editing date. The second page names an audio file, "deepseek 0520.m4a," with a duration of roughly three hours and 44 minutes. It also says the text was automatically transcribed and AI-edited without speaker labels, and that proper nouns and numbers may contain recognition errors. The document establishes that a transcript is circulating; on its own, it does not establish that every line was spoken by Liang Wenfeng.
The National Business Daily reported that an institution participating in DeepSeek's investment process verified that the closed-door meeting took place in May 2026 and that the content was credible. Bloomberg had also reported on May 22, citing anonymous sources, that DeepSeek told prospective investors it would prioritize research, open source, and AGI. Those reports support the occurrence and broad themes of the meeting, but no original audio has appeared publicly, and neither DeepSeek nor Liang has confirmed the transcript line by line. Company figures and strategic judgments below are therefore attributed to the transcript rather than treated as official disclosures.
Open Source and Low Pricing Are Framed as Strategic Restraint
The main speaker in the transcript describes DeepSeek's vision as restraint toward profit maximization, placing open source, low pricing, and a refusal to capture every user entry point within the same logic. According to the document, the early team did not join primarily to pursue a listing or personal wealth, and open source was not a concession forced by competition. In the transcript's framing, DeepSeek uses open source and low pricing to increase its chances of reaching AGI instead of first maximizing revenue from its existing user base.
The document connects that choice to a specific pricing method: API prices are set around recovering the cost of a batch of equipment in roughly ten months, and one model—whose name appears to have been mistranscribed—was reportedly cut to one-quarter of its original price. The transcript also states the other side of the equation: DeepSeek remains a company that needs revenue, financing, and sustainable operations. The ten-month payback period, price cut, and demand elasticity come from unaudited spoken material and cannot be used to infer actual gross margins or future pricing.
Post-Financing Compute Expansion Remains Supply-Constrained
The transcript claims that DeepSeek had about 20,000 "H-series-equivalent" accelerators at the time of the meeting, many of them newly delivered, and intended to keep buying after its financing. It presents available compute—not a belief that scaling has reached its limit—as the constraint on model size. The model that can be trained is first derived from the resources on hand. This fits the public timeline of DeepSeek's V4 release on April 24, but no official document substantiates the specific accelerator count or procurement schedule.
The document also describes the China-US model gap as a time lag under compute constraints: its current narrative is that DeepSeek operates with roughly one-twentieth of the compute used by US peers while lagging by one to two years, with a goal of narrowing that gap to three to six months using a larger compute share and seeking localized advantages. The "one-twentieth" ratio and "three-to-six-month" target are internal comparisons in the transcript, not performance findings verified under a common benchmark. A question near the end mentions 150B to 250B active parameters, but the transcript contains no clear corresponding answer, so it does not establish a next-generation model plan at that scale.
Domestic-Chip Cooperation and the TileLang Route Require Separate Verification
According to the transcript, Huawei planned to provide DeepSeek with 16,000 "950" accelerators, which the document equates to about 4,000 Nvidia B-series units. It says the capacity would support the current model generation but not the next one. The document also suggests depreciation periods of about five years for Nvidia accelerators and no more than three years for Huawei units, citing power consumption and product generations. No public contract or technical report supports these procurement, equivalent-compute, or depreciation figures, so they remain claims contained in the meeting material.
There is more public evidence for the TileLang portion. DeepSeek's public GitHub organization has released TileKernels, a kernel library written with TileLang, confirming that the company uses this class of high-level GPU kernel language. The transcript goes further, saying TileLang can shorten operator-development time while accepting a roughly 1% to 2% execution-efficiency loss and can work with AI to rebuild an operator ecosystem; the same passage says the route is not yet complete. The public repository confirms adoption, but it does not verify the loss figure or establish that CUDA's ecosystem has already been displaced.
Continual Learning and Coding Agents Rank Ahead of Product Expansion
The transcript identifies continual learning as the most important capability for a next-generation model while acknowledging that no mature solution has been found globally. It argues that today's Agents are constrained partly because they cannot learn effectively over time. A model that accumulates capabilities during use could first help DeepSeek develop its successor and then accelerate research toward general intelligence. In the document's technical priority order, continual learning ranks ahead of product-line expansion, while Coding Agents rank ahead of vertical Agents for finance or healthcare.
That priority does not amount to a delivery timeline for continual learning. The transcript repeatedly describes the methods as exploratory, with promising internal ideas that have not yet worked end to end. Its application strategy is also conditional: general Agents and Coding Agents come first for now, while the business model for vertical sectors remains unsettled. Embodied intelligence appears as a longer-term direction aimed at physical-world labor rather than an immediate product commitment.
A Flat Organization Is Adding Necessary Hierarchy
The document divides DeepSeek's research management into two tracks. Company-wide tasks such as model releases use top-down "formal" organization, while researchers can explore ideas bottom-up during the rest of their time. The internal guideline cited in the transcript is that formal assignments should ideally occupy no more than half of an employee's time. This structure is not intended to remain purely flat forever: the document explicitly says that some departments need stricter hierarchy as headcount grows, while others can remain loose.
The commercialization discussion is equally conditional. The transcript says team stability is a core company interest and that financing and equity options reduce retention risk. If enterprise, consumer, and API demand continue growing, the company could approach profitability; the fallback would be to operate the API business well. These are management judgments under stated assumptions, not revenue guidance or a listing plan. Removing the conditions would turn scenario discussion at an investor meeting into a financial commitment.
The Document Offers Clues, Not a Substitute for Formal Disclosure
Public reporting, DeepSeek-V4's April 24 release, and DeepSeek's public TileKernels repository provide cross-checks for the meeting's timing and some technical themes. They do not validate every number, model name, or causal claim in the transcript. The source document itself warns about automated-transcription errors, and terms such as "DDCP" and "MILES" show obvious recognition uncertainty.
Until the original audio or formal minutes become available, the document is better treated as a clue to DeepSeek's strategic trade-offs than as a replacement for company announcements, financial data, or technical reports. Its emphasis on open source, research priority, and AGI aligns with previous public statements, while chip procurement, equivalent compute, payback periods, revenue projections, and next-generation model scale still require traceable first-party confirmation.