STORY RECORD
DeepSeek founder's contradictions exposed
Analyst Teortaxes▶️ (70K followers, rising attention) flags contradictions in Liang Wenfeng's leaked investor call: Wenfeng says he won't do chips 'if possible' but has been working on chips for years; claims government gave 'not a cent' while China's National AI Industry Investment Fund invested $1B with voting rights; and cites a ~$3B compute cap that doesn't match US export limits or his own job listings.
Teortaxes points to DeepSeek's job posting for an 'AI Computing Cluster Performance and Reliability Engineer' that says 'on your first day, you'll be managing a 100,000 card cluster' — far exceeding the ~20,000 H-equivalent cards Wenfeng acknowledged.
The analyst argues DeepSeek is building a tightly coupled heterogeneous system (Nvidia for weight updates, Huawei SuperPoDs for inference rollouts, eventually in-house Whale Compute), and could reach ~200K Ascend 950-equivalent capacity by end of 2026 Q4 — enough for serious research but not immediate OOM-scale pretraining.
Zephyr (166K followers) adds nuance: the 4:1 exchange ratio Wenfeng cited is training-FLOPS-only; on memory bandwidth it's 2:1, and decode throughput per GPU difference is estimated at 1.3x-1.7x.
The call transcript is single-source (behind WeChat paywall), unconfirmed by DeepSeek, and the original audio hasn't surfaced — but Reuters via Yicai independently confirmed the call's existence on July 23.
Why It Matters
This is the first time DeepSeek's founder has been caught in a pattern of public contradictions about the company's most sensitive strategic question: how much compute does China's most celebrated AI lab actually have, how is it paying for it, and what hardware is it really running?
The gap between Wenfeng's 'we have 20K H-equivalent cards, can't spend more than $3B' and the job listing for a 100K-card cluster suggests either intentional misdirection to investors or a far more ambitious heterogeneous buildout than publicly acknowledged.
The China AI Fund investment (with voting rights) contradicts his 'no state money' claim.
If Teortaxes' analysis is correct, DeepSeek is quietly building a hybrid Nvidia-Huawei-Whale infrastructure that could reach meaningful scale by late 2026 — reframing the US-China compute gap narrative from 'China can't scale' to 'China is scaling on non-Nvidia hardware faster than export controls anticipated.'
The Facts
10A leaked investor call transcript attributed to DeepSeek CEO Liang Wenfeng is circulating on X, with the original full transcript behind a WeChat paywall. The original audio has not surfaced. Reuters, citing Chinese state-owned media Yicai, independently confirmed the call's existence on July 23, 2026.
moderate · confidence 0.6
In the call, Wenfeng stated DeepSeek had approximately 20,000 H-equivalent GPUs as of early June 2026, with most arriving in the preceding 1-2 months.
moderate · confidence 0.6
Wenfeng claimed DeepSeek prefers NVIDIA over Chinese chips, is actively buying NVIDIA cards including 'non-compliant' cards at a premium under export controls, and plans to spend aggressively on NVIDIA hardware.
moderate · confidence 0.6
Wenfeng said Huawei's 950 SuperNode can fully substitute for NVIDIA's GB200/GB300 in performance and price, with a 4:1 card exchange ratio (four Huawei cards = one NVIDIA card) and a two-year lag.
moderate · confidence 0.55
Analyst Teortaxes▶️ (70K followers) noted contradictions: Wenfeng says he won't do chips 'if possible' but has been working on chips for years; says government 'not a cent' while China National AI Industry Investment Fund invested $1B with voting rights; cites $3B compute cap inconsistent with US export limits of 75K units per entity.
weak · confidence 0.45
DeepSeek posted a job listing for 'AI Computing Cluster Performance and Reliability Engineer' that states: 'on your first day, you'll be managing a 100,000 card cluster.' The listing mentions NCCL, NVLink, RDMA/InfiniBand, and training-oriented whole-system stability concerns about different hardware batches. Other listings mention Ascend and 'CPU/GPU/NPU' work.
weak · confidence 0.45
Zephyr (166K followers, AI chips expert) corrected the 4:1 ratio: it applies to training FLOPS only; on memory bandwidth the ratio falls to 2:1 (8TB/s vs 4TB/s), and decode throughput per GPU difference between Ascend SuperPOD and GB300 is estimated at 1.3x-1.7x.
moderate · confidence 0.55
Teortaxes▶️ argued DeepSeek may achieve ~200K Ascend 950-equivalent capacity by end of 2026 Q4, based on a tightly coupled heterogeneous system using Nvidia for weight updates and SuperPoDs for inference, eventually supplemented by in-house Whale Compute.
weak · confidence 0.35
Huawei's Atlas 950 SuperPoD debuted at WAIC 2026, built from up to 8,192 Ascend NPUs with Huawei's UnifiedBus interconnect. One full SuperPoD (8,192 NPUs) = 2,048 GB300s ≈ 28 NVL72 GB300 racks ≈ 3.8 MW of power.
moderate · confidence 0.5
Wenfeng said the US-China AI gap is approximately 12-18 months, with China using roughly 1/20th the compute of US labs. He said to train a model as large as the top US one would need 50,000 GB300s or 200,000 Huawei 950s.
moderate · confidence 0.55
Still Open
1- ContradictionContradiction note is limited to snapshotted evidence.contradicted