Photo: Data photo of China's AI company DeepSeek. (Song Biling/Dajiyuan)
[People News] On July 25, 2026, a meeting summary purportedly from Liang Wenfeng, the founder of DeepSeek, and investors began circulating online. This document, written in clear internal language, provides detailed insights into the scale of computing power procurement, delivery timelines, willingness to allocate funds, methods for self-building clusters, perspectives on open source and 'assisting competitors,' as well as the progress of Huawei's chip ecosystem and the de-CUDA process.
Following the leak, reports suggested that Liang Wenfeng expressed dissatisfaction, resulting in a temporary halt to the second round of financing. Regardless of the eventual commercial repercussions, the strategic and enforcement implications of this summary have already become evident: it has shifted previous U.S. officials' and Congress's allegations against Chinese AI companies—accusing them of 'bypassing export controls, acquiring advanced computing power, and amplifying capabilities'—from vague suspicions into a verifiable chain of evidence.
The U.S. Department of Commerce's Bureau of Industry and Security (BIS) has long been focused on numerous sensitive issues, including abnormal procurement volumes, market premiums, third-country transshipments or deployments, final beneficiaries based in China, and the potential use of advanced chips to enhance large models and related military and intelligence capabilities. These concerns are addressed in this summary with specific figures and timelines.
The leaked material combines rumors about the investigation into companies like Moonlight (Kimi) using overseas GB300 servers in the United States, along with Congress's examination of DeepSeek, Moonlight, Ali, and MiniMax under the 'Chinese Open Source Model Risks' framework. This indicates that the implications of the leak extend far beyond the financing issues of a single startup.
It confirms the central allegations against the Chinese Communist Party's (CCP) strategy to catch up in AI: acquiring advanced computing power chips from NVIDIA through irregular or grey channels, narrowing the gap using distillation and efficient training techniques, and utilizing open source and ecological collaboration to magnify the benefits of limited computing power across the entire Chinese AI network. This ultimately poses a challenge to the United States' global leadership in cutting-edge large models.
The irregularities in computing power procurement, from 'demand' to 'traceable chains'.
The most notable aspect of the minutes is Liang Wenfeng's account of the current state of computing power and procurement plans. Currently, there are about '20,000 H-equivalent computing power', with most of it arriving in the last month or two. A significant number of machines are expected in the coming months, and basically all are NVIDIA. They are willing to pay a premium; if they spend around 20 billion yuan on procurement within a year, it would be considered an outstanding performance for the procurement department. All clusters are self-built. These statements are not mere vague plans but detailed operational strategies with specific timelines, scales, and price flexibility.
In the context of U.S. export controls, advanced NVIDIA chips, particularly the Hopper series and the subsequent Blackwell series products, which excel in high bandwidth storage and interconnect capabilities, are subject to strict licensing restrictions for supply to entities in China. The Bureau of Industry and Security (BIS) has clearly indicated that even if the chips are deployed outside of China, they typically cannot easily bypass licensing requirements through overseas subsidiaries if the entity or ultimate parent company is based in China.
The minutes refer to 'deliveries in the last month or two', 'large quantities expected in the coming months', 'mostly NVIDIA', and 'willing to pay a premium', which directly align with typical red flags in enforcement due diligence: unusual procurement patterns, high acceptance of premiums, and concentrated deliveries. Enforcers no longer need to infer computing power scale from model capabilities; they can directly trace back along the 'time—quantity—supplier—logistics—financing—electricity usage and serial number' pathway. NVIDIA, server manufacturers, cloud and data center operators, logistics providers, and related financing entities may all be included in the scope of cross-verification.
Simultaneously, rumors regarding the investigation of Kimi have further escalated the intensity of U.S.-China AI competition. U.S. officials allege that the Dark Side of the Moon has acquired and utilized GB300 servers in locations such as Thailand, potentially for model training. The GB300, as a high-performance training server configuration, possesses a computing power density and interconnect capability that far surpass ordinary inference cards. If confirmed to be used for cutting-edge large model training, it would directly address the core regulatory concerns of preventing advanced computing power from being converted into enhanced strategic capabilities.
The procurement guidelines outlined in Liang Wenfeng's summary support this type of case: Chinese AI companies are not simply caught in the narrative of 'chip shortages'; rather, they are actively and aggressively expanding their computing power related to NVIDIA, adopting a clear strategy of 'buying as much as possible.' Funding is not the bottleneck; the real issue is the inability to purchase; converting money into chips quickly is more cost-effective than keeping it in the bank. While this perspective may be commercially sound, it serves as clear evidence of evasion within a regulatory framework.
Distillation, open source, and ecological mutual assistance: Strategic ambitions with limited computing power
Merely acquiring chips is just the beginning. Another key theme emphasized repeatedly in the summary is the commitment to open sourcing, aiding in reproduction, and a willingness to assist competitors such as Alibaba (Ali), Zhipu (Zhipu), and the Dark Side of the Moon (Yue Zhi An Mian). The strongest models may also be open-sourced; not only will the weights be made public, but support will also be provided for third parties to deploy, reproduce, and lower costs; core technologies and experiences will disseminate throughout the entire ecosystem. From the perspective of industry practitioners, this can be interpreted as a vision of 'restraint,' 'goodwill,' and 'market expansion'; from the standpoint of law enforcement and strategy, this represents a pathway for capability diffusion.
Knowledge distillation is a prevalent technical approach in the current competition among large models: it involves using a powerful teacher model to generate data or soft labels to train smaller, more efficient student models, thereby enabling them to approach or even locally exceed the performance of the teacher model under limited computing power. Although the summary does not explicitly use the term 'distillation,' its focus on training efficiency, low-cost deployment, and open-source support for reproduction aligns closely with the principles of distillation and efficient post-training.
When a company acquires high-end computing power through grey channels and successfully trains robust models, it can then share open-source weights, deployment guidance, and experiences, enabling other companies to replicate or distill usable models at a lower cost. This process exponentially amplifies the strategic value of a single restricted GPU. The U.S. Congress has examined several Chinese open-source model companies under the same risk framework, driven by concerns regarding this 'shared training and deployment experience—common data or computing power sources—personnel and technology flow.'
The minutes also indicate that DeepSeek has utilized NVIDIA hardware in its V3 training, yet it 'almost does not rely on the NVIDIA ecosystem.' Instead, the company rewrites operators using its self-developed advanced language, TileLang, and is deeply involved in the Huawei chip ecosystem. The objective is to port the existing software stack to Huawei cards, ultimately allowing Huawei's super nodes to serve as a cost-effective alternative to configurations like GB200/GB300 at the task level.
This series of statements clearly outlines a strategic implication: advanced GPUs obtained through overseas or gray channels are not only used for completing a single training session but can also function as 'bridge computing power' to train more powerful models, accelerate the development of AI compilers and domestic software stacks, transfer optimization experiences to the Huawei platform, and subsequently spread the migration costs to more Chinese companies through open-source initiatives. Each GPU acquired by circumventing restrictions could play a role in helping China gradually reduce its long-term reliance on advanced U.S. GPUs.
The United States is currently in a dilemma: the stricter the restrictions imposed, the greater the motivation for domestic alternatives and de-CUDAization; conversely, if the restrictions are too lenient, it will directly accelerate the scaling of Chinese models, ongoing learning exploration, and the progress towards AGI. Analysts suggest that U.S. enforcement targets are likely to extend beyond single GPU purchases to include overseas data centers, cloud computing resources, HBM, networking equipment, server maintenance, financing channels, compiler technology, and even the circulation of model weights. The Bureau of Industry and Security (BIS) has cited the use of large models for military and intelligence capabilities as one of the national security justifications for advanced semiconductor controls. Liang Wenfeng's memorandum provides an ideal narrative foundation for elevating isolated procurement cases to an 'ecological-level investigation.'
The Chinese Communist Party's strategy to close the gap: narratives of efficiency and strategic trade-offs under resource constraints.
Liang Wenfeng acknowledges that the primary gap with the United States lies in computing power resources, rather than talent. He asserts that domestic talent has 'almost no gap' and may even have a larger base; the true bottleneck is the insufficient number of GPUs, which results in fewer experimental opportunities, an inability to train the largest models, and inadequate research. Current and upcoming resources are only sufficient for conducting more experiments at the scale of 'tens of B activations,' leaving a significant gap compared to the 800B scale in the United States. Training models of equivalent scale requires approximately 50,000 GB300 cards or 200,000 of Huawei's latest cards (for training only, excluding research). Thus, the strategy is to purchase as many resources as possible within an affordable range; first, to maximize efficiency at the scale that can be trained; and simultaneously to advance the domestic chip ecosystem while waiting for production capacity to catch up.
At the heart of this narrative is the idea of 'approaching or even partially surpassing with less computing power and in a shorter time.' The minutes suggest that China could potentially use only one-twentieth of the computing power of the United States to limit its technological lag to one or two years, with ambitions to further reduce this gap to six months or even less. This could create a structural advantage in terms of cost and product experience. Following breakthroughs achieved through continuous learning, AI could then be leveraged to accelerate its own research and development, resulting in nonlinear growth. The strategy of open sourcing and 'only earning reasonable profits,' such as recovering equipment costs within ten months, is framed as a strategic restraint aimed at increasing the likelihood of achieving Artificial General Intelligence (AGI), rather than merely focusing on commercial profit-sharing.
This strategic vision underscores the necessity of U.S. regulations: if advanced computing power continues to enter through grey channels and is utilized efficiently and amplified ecologically, the time and capability gap between China and the U.S. in cutting-edge models will be systematically narrowed.
Notably, the minutes also stress that team stability is the sole core interest, driven by vision rather than key performance indicators (KPIs). The focus is on the main trajectory of CoT → Agent → continuous learning → self-iteration → embodied AGI, while exercising restraint regarding consumer traffic and short-term business-to-business (B2B) monetization.
While these management and cultural statements do not inherently violate any regulations or laws, the underlying strategy of 'once resources are in place, organizational and talent advantages can be quickly realized' crosses the red line of U.S. regulation. The crux of U.S. accusations has never been that China lacks talent or vision, but rather whether advanced computing power and related technologies have been acquired and transformed into strategic capabilities through illicit means.
This represents a significant challenge to the U.S.'s global leadership position.
In summary, this leaked document confirms the United States' allegations regarding China's AI competition under the leadership of the Chinese Communist Party from three perspectives.
Firstly, regarding the acquisition of computing power. The unusual procurement patterns, willingness to pay premiums, self-built clusters, and the assertion that 'essentially all are NVIDIA' align closely with the BIS red flag list and support the rumors surrounding the overseas use of GB300. Law enforcement now has a clearer path for reverse investigation.
Secondly, concerning capability amplification. The strongest open-source models, assistance to competitors in replicating technologies, reducing reliance on CUDA through tools like TileLang, and deep engagement in the Huawei ecosystem indicate that limited high-end computing power can benefit the entire Chinese AI network through distillation, deployment optimization, and experience sharing, rather than being restricted to a single company.
Thirdly, in terms of strategic closed-loop systems. Advanced GPUs are utilized as bridging computing power to train models, accelerate domestic software stacks, and ultimately promote domestic hardware alternatives, directly countering the U.S. strategy of maintaining an edge through computing power bottlenecks. The stricter the restrictions, the stronger the drive for alternatives; the more relaxed the restrictions, the faster the scaling. This encapsulates the core of the U.S. dilemma.
For the Chinese Communist Party, this memorandum reveals a significant shortcoming: resources remain a critical bottleneck, and both domestic production capacity and ecological maturity require time. It is unrealistic to expect a complete surpassing given the substantial gap in computational power. However, for the United States, the real risk does not stem from whether China has already achieved full superiority, but rather from the systematic reduction of the gap, the shrinking of the lead window, and the ongoing testing and circumvention of the regulatory framework. If law enforcement and policies fail to respond promptly to the complete cycle of 'procurement-training-open source diffusion-domestic transfer', the costs of maintaining the United States' global leadership in AI will continue to escalate.
The leak of internal communications from Liang Wenfeng is not merely an ordinary incident of commercial information loss. It connects the previously scattered suspicions in the United States—such as illegal chip acquisition, the use of advanced GB300-type servers, distillation and efficient training, the amplification of the open-source ecosystem, and the promotion of domestic alternatives—into a verifiable chain of evidence. Consequently, the focus of U.S. accusations against the Chinese Communist Party shifts from policy discussions to more concrete law enforcement and strategic responses.
The essence of the AI strategic competition between the U.S. and China has never been about the fluctuations in the ranking of a single model, but rather a long-term contest involving computational power, technology diffusion, and institutional frameworks. This memorandum simply brings the real tension of this contest to light in an unexpectedly clear manner.
Some commentators suggest that this memorandum is akin to handing a knife to the United States; the issue is not whether the knife is handed over, but that the parameters of the Chinese Communist Party's AI development path closely align with the profile of a target under U.S. regulatory scrutiny. If you are not a thief, why would you secretly share experiences in theft and openly discuss visions of stealing?
(First published by the People News) △
News magazine bootstrap themes!
I like this themes, fast loading and look profesional
Thank you Carlos!
You're welcome!
Please support me with give positive rating!
Yes Sure!