DeepSeek CEO Liang Wenfeng's Leaked Investor Meeting Reveals China's Compute Deficit — and the White House Crackdown That Followed

In the last week of July 2026, two events reshaped the public understanding of U.S.-China AI competition. A leaked four-hour investor meeting transcript from DeepSeek founder Liang Wenfeng laid out China's compute deficit and argued it could be closed within a year. Days earlier, White House Office of Science and Technology Policy Director Michael Kratsios publicly accused Chinese startup Moonshot AI of evading semiconductor export controls and copying U.S. AI models at industrial scale. Together, they form a sharply paired narrative about where the U.S.-China AI rivalry stands today.

Liang Wenfeng's Key Claims: Compute Is the Only Real Bottleneck

A lightly edited transcript of a roughly four-hour closed-door investor meeting with DeepSeek founder Liang Wenfeng was published by Tencent Tech and widely circulated starting July 23, 2026 . His core argument: the gap between Chinese and American AI is entirely about compute resources, not talent.

Talent is not the bottleneck. Liang stated that "the biggest gap between us and the US lies in resources," arguing that Chinese and American AI talent come from the same global pool and are equally capable . "Talent is not the bottleneck; computing power is" .

China has used roughly 1/20th of U.S. compute to nearly close the capability gap. Liang assessed the overall China-U.S. AI gap at roughly one year, but said China has achieved that using about 1/20th of the other side's compute .

DeepSeek's own compute is very limited. DeepSeek currently has about 20,000 "H-equivalent" GPU compute units, most of which arrived only in the last month or two . By contrast, major U.S. labs operate clusters many times larger.

U.S. models are an order of magnitude larger. The largest U.S. models now activate around 800 billion parameters during inference, while Chinese models sit at tens of billions .

Huawei's chips lag Nvidia significantly. Liang said Huawei is about two years behind Nvidia and that one Nvidia GB300 GPU equals roughly four Huawei Ascend 950s in performance .

Nvidia's CUDA moat is disintegrating. Liang argued that China's move off Nvidia hardware has hit a turning point, and that chip supply should no longer be a bottleneck within a year .

AGI over profit. Liang stated DeepSeek prioritizes AGI research over commercialization, seeks only "reasonable profits," and will likely keep its top models open-source .

The White House Accuses Moonshot AI: Illicit Chips and Mass Model Distillation

On July 22–23, 2026, Michael Kratsios, director of the White House Office of Science and Technology Policy, publicly accused Chinese AI startup Moonshot AI of two coordinated violations :

1. Illicit chip access via third countries. Moonshot "acquired GB300-equipped servers and has accessed GB300s in Thailand, likely to train its AI models" — routing advanced Nvidia hardware through a third country to evade U.S. export controls that ban such chips from reaching China .

2. Mass model distillation of U.S. AI. Moonshot was accused of copying Anthropic's Claude models on a massive scale — using fraudulent accounts, proxy networks, and access-restriction evasion — to build its own Kimi K3 system, which had stunned the tech industry with its advanced capabilities .

The accusation turned what might have been a behind-the-scenes compliance matter into a high-profile policy confrontation . This followed earlier enforcement actions: in March 2026, the DOJ charged three individuals with conspiring to divert U.S. AI servers to China , and in April 2026, House chairmen announced a joint investigation into national security risks from PRC AI models using distillation via fraudulent means .

What These Concurrent Developments Mean for U.S.-China AI Policy

Taken together, these two events reveal several tensions in the current U.S. approach to AI competition with China:

Export controls face a credibility problem. Liang Wenfeng's central argument — that China's compute gap, while severe, can be closed in about a year through domestic alternatives — directly challenges the premise that restricting Nvidia chip exports keeps China permanently behind. If Chinese firms can substitute Huawei Ascend chips and break CUDA dependency within 12 months, the long-term leverage of export controls erodes .

The "compute gap" narrative cuts both ways. Liang's admission that China uses 1/20th the compute to achieve near-parity actually strengthens the case for maintaining export controls (China is highly efficient even with scarce chips), but it also suggests that even draconian controls may only delay rather than prevent Chinese AI progress .

Enforcement is intensifying but playing catch-up. The White House's public naming of Moonshot AI — a rare escalatory step — signals frustration that existing controls are being systematically circumvented via third-country routing and model distillation . The DOJ indictments and House investigations show a multi-agency push, but the Moonshot case suggests evasion is already happening at scale.

The "distillation loophole" is becoming a central policy battleground. The White House specifically called out Moonshot for using U.S. models to train a competitive Chinese system . This puts pressure on Washington to decide whether to regulate distillation techniques, tighten API access, or impose liability on U.S. AI companies whose models are used this way — measures that could reshape how U.S. AI firms deploy their technology globally.

A strategic paradox for U.S. policymakers. If Liang is right that (a) China's compute deficit is closing fast via domestic chips and (b) Chinese teams can match U.S. performance with vastly less compute, then the window in which export controls can meaningfully slow China's AI progress may be narrower than previously assumed. The Moonshot accusation, landed the same week, illustrates that Chinese firms are not waiting for domestic chip substitution — they are actively acquiring banned hardware and copying U.S. models right now. U.S. policy is thus caught between tightening enforcement today and facing a diminishing strategic return on export controls in the medium term.