Inside China's AI Manhattan Project: Export Control Gaps and the Race to Build Sovereign AI

When Western analysts describe China's national AI programme, they increasingly use the phrase "AI Manhattan Project." The analogy captures something real: this is a state-directed, resource-unlimited programme to develop a technology that its leadership believes will determine which nation dominates the 21st century — and potentially determine the outcome of any military conflict with the United States.

China's 15th Five-Year Plan (2026-2030), approved by the National People's Congress in March 2026, elevates AI to the same strategic priority level as nuclear weapons, space, and quantum technology. The plan commits to making China an "AI powerhouse" and establishes specific targets: domestic AI chip self-sufficiency, frontier AI model capability parity with Western systems, and deployment of AI across military, government, and industrial sectors.

Understanding the gaps in Western export controls — and how China is exploiting them — is essential for any developer, investor, or policymaker thinking about AI competition over the next decade.

The Scale of China's AI Investment

The numbers are difficult to verify precisely because much of China's AI investment flows through state-owned enterprises, provincial government funds, and military channels that do not require public disclosure. Best estimates:

National AI Fund: China established a 1 trillion RMB (approximately $140B) national AI investment fund in 2024, managed by the Ministry of Science and Technology and the National Development and Reform Commission. Deployment is over 5 years.

State-backed AI companies: Baidu, Alibaba, Tencent, Huawei, ByteDance, and iFlytek all receive direct state subsidies, preferential access to government data, and regulatory protection. The "national team" framing means these companies operate partly as state instruments for AI capability development.

Military AI: The PLA's Strategic Support Force has a dedicated AI division. AI is integrated into the C4ISR (command, control, communications, computers, intelligence, surveillance, and reconnaissance) programmes. Specific budgets are classified.

Data advantages: China's population of 1.4 billion, combined with relatively limited privacy restrictions and mandated data sharing by private companies under national security laws, gives Chinese AI developers access to training data at a scale that no Western company can legally replicate.

The Export Control Strategy — and Its Gaps

The US export control regime for AI chips is built around controlling FLOP (floating-point operations per second) thresholds. Chips above a certain performance level require export licences to China; chips below the threshold can be sold freely.

This creates several exploitable gaps:

The H20 / A800 gap: Each time the US raises the performance threshold for restricted chips, Nvidia designs a compliant product just below the line. The H20 (restricted in 2024) was specifically engineered to have maximum China-permissible performance while staying under the export control threshold. Before restriction, Nvidia shipped approximately $2B in H20 chips to China. China accumulated a meaningful stockpile.

The chip smuggling network: Chainalysis, Silverado Policy Accelerator, and US Commerce Department enforcement teams have documented extensive chip smuggling operations routing restricted chips through Singapore, Malaysia, Thailand, and UAE intermediaries. In 2024, Commerce's Bureau of Industry and Security (BIS) added 37 entities to the Entity List for chip transshipment. The volume of smuggled chips is not publicly quantified but is believed to be substantial.

The cloud access gap: US export controls restrict physical chip exports but do not prevent Chinese companies from renting compute on US cloud platforms. Until new rules in late 2024, Chinese companies could access H100-equivalent GPU compute via AWS, Azure, and GCP for training frontier models. Post-2024 rules require cloud providers to perform enhanced due diligence on customers training AI models at scale, but enforcement gaps remain.

The model weight gap: Once a frontier AI model is trained, the weights can be downloaded, copied, and run anywhere. Export controls on chips affect training but not deployment of already-trained models. DeepSeek released V3 weights under an open licence. Any Chinese organisation now has access to frontier-capability model weights regardless of chip restrictions.

The allied country gap: Export controls are most effective when allied nations enforce them consistently. Japan, the Netherlands, South Korea, and Taiwan have agreed to restrict advanced chip and equipment exports to China. But enforcement consistency varies. Reports of Japanese, Korean, and Taiwanese companies finding workarounds through third-country subsidiaries persist.

China's Domestic AI Stack

Despite export controls, China has assembled a meaningful domestic AI hardware and software ecosystem:

Huawei Ascend series: The 910B (manufactured at SMIC 7nm-equivalent) is China's primary AI training chip. Less energy-efficient than Nvidia H100 but capable of frontier model training (as demonstrated by GLM-5 and DeepSeek V4 development). The 910C is in production for 2026.

Cambricon: China's other significant AI chip company, focused on inference accelerators. Lower performance ceiling than Ascend but important for domestic edge AI deployment.

Biren Technology: GPU startup developing high-performance training chips. Production in small volumes.

Domestic software stack: Huawei's CANN (Compute Architecture for Neural Networks) is the Ascend equivalent of NVIDIA CUDA — the programming framework that makes the hardware useful for model development. DeepSeek's willingness to optimise V4 for Ascend hardware validates that CANN is mature enough for frontier development.

DeepSeek's cost efficiency innovation: The DeepSeek R1 and V3 models demonstrated that frontier capability can be achieved with significantly less compute than US labs use, through architectural innovations (MoE, efficient attention mechanisms, synthetic data). This directly undermines the compute restriction strategy — if China can achieve parity with less compute, restricting compute access is less effective.

The Military Dimension

The clearest articulation of why this is treated as an "AI Manhattan Project" is the military application timeline. PLA doctrine explicitly incorporates AI for:

• Autonomous drone swarms (demonstrated at scale in multiple exercises)
• ISR fusion (combining satellite, SIGINT, HUMINT, and OSINT data in real-time)
• Cyber operations automation (AI-assisted malware development, vulnerability discovery)
• Decision support for military commanders (AI-assisted course of action analysis)
• Logistics optimisation (applying supply chain AI to military logistics)

US military AI research (DARPA, Joint AI Center, now CDAO) is substantial but operates in a more fragmented, slower-moving institutional environment than the PLA's centralised AI programme.

The 2027 PLA modernisation deadline — the year Xi Jinping has set for the PLA to be capable of winning a "regional war" — is explicitly linked to AI capability milestones. AI-enabled ISR and decision-making are considered force multipliers that could compensate for quantity disadvantages relative to the combined US-allied military.

What Developers and Enterprises Need to Know

The AI stack is bifurcating permanently. If you are building AI applications at enterprise scale, the question of whether your compute stack is US-controlled or China-accessible will increasingly matter for compliance reasons, not just technical ones.

Open-weight models are ungovernable. DeepSeek V3 and V4 weights are publicly downloadable. Any Chinese (or other) actor wanting to deploy frontier AI capability can do so without needing any chip or cloud access. The export control strategy has significant limitations against open-weight model releases.

China will have competitive AI capability regardless of export controls. The combination of domestic chip development, open-weight model access, architectural efficiency innovations, and state investment means China will maintain frontier AI capability. The question is relative efficiency and cost, not binary access.

Compliance exposure is growing. US entities that knowingly assist Chinese AI military applications — through chip sales, cloud services, research collaboration, or model licensing — face growing legal exposure under export control law and the NDAA. Legal counsel should review any significant China AI business before 2026 year-end.