← Back to feed

A transcript of a private investor meeting held by DeepSeek CEO Liang Wenfeng on July 22, 2026 appeared on GitHub and circulated widely across Chinese and Western AI communities within hours. DeepSeek has since paused its ongoing fundraising process. The document, uploaded by a GitHub user who translated and transcribed the session, contains Liang’s candid assessment of the compute gap between DeepSeek and frontier US labs — a gap he described as structural and tied directly to US export controls.

The transcript has not been officially confirmed by DeepSeek, but multiple corroborating accounts and the specificity of the technical discussion have led researchers to treat it as credible.

What the Transcript Says

Liang reportedly told investors that DeepSeek’s algorithmic efficiency work — the R-series architecture, MLA attention, and MoE routing improvements that made DeepSeek V3 and R1 globally significant — was partly a response to hardware scarcity rather than a philosophical preference. The core constraint: DeepSeek cannot legally obtain NVIDIA H100s or Blackwell GPUs in the volumes that US frontier labs run, and Huawei’s Ascend 910C chips, while improving, remain behind on per-chip compute density.

On the timeline question — when China might close the gap — Liang is described as deliberately avoiding a specific forecast, saying only that the gap is “measurable and meaningful” but not insurmountable.

The compute capacity comparison implied by the discussion aligns with published estimates. AISI’s July 2026 assessment put open-weight frontier models 4-7 months behind closed-weight US systems on cyber capability benchmarks. Moonshot AI’s Kimi K3, which debuted at 93.4% SWE-bench Verified and reached Arena code leaderboard top-10, has prompted debate about whether that gap is closing faster than official estimates suggest.

Why the Fundraise Paused

DeepSeek was in the process of assembling an additional round on top of the $7 billion it closed at a $50 billion valuation in June. Tencent and Alibaba had both moved into the earlier raise. New institutional investors — reportedly including international LPs — were in diligence.

Candid hardware limitation disclosures in a leaked founder meeting create two problems for an active raise. First, they hand regulators and geopolitically cautious LPs a documented admission that US export controls are working as intended — language that can be used to justify further restrictions. Second, they expose the gap between DeepSeek’s public narrative (efficiency-first, not compute-constrained) and Liang’s private framing.

The pause is described as temporary. DeepSeek has not commented publicly.

The Broader Context

The leak lands at an awkward moment for the China-vs-US AI narrative. Kimi K3’s 2.8 trillion-parameter open-weight model with a $3/$15 pricing has analysts at SCMP calling the gap “a matter of weeks” for some capability classes. Meanwhile Liang, one of the most technically credible figures in Chinese AI, appears to be telling investors something more cautious in private.

The gap that exists is real and hardware-driven. DeepSeek’s published efficiency work shows the team has found ways to compete at less compute, but Liang’s comments suggest the company internally distinguishes between “closing the algorithmic gap” and “closing the compute gap.” The first is achievable. The second depends on factors outside DeepSeek’s control.

Key Numbers

• DeepSeek fundraise status: paused following transcript leak
• Prior round: $7B closed at $50B valuation (June 2026)
• Compute gap estimate: AISI pegs open-weight models 4-7 months behind closed frontier
• DeepSeek V4 Pro: 80.6% SWE-bench Verified, MIT license
• Kimi K3 (Moonshot): 93.4% SWE-bench Verified, 2.8T open-weight — sets the benchmark for Chinese frontier access