The race in Artificial Intelligence is no longer measured solely by who has the most capable model. Increasingly, it hinges on a less showy but more decisive question: who can sustain the financial resources necessary to train, serve, and scale these systems. In this contest, the U.S. and China seem to be adopting increasingly divergent strategies.
The United States maintains an advantage through its industrial muscle, access to the most advanced accelerators, and control over much of the software supporting modern AI. China, constrained by export restrictions and a strained supply chain, is pushing towards more efficient models, MoE architectures, domestic chips, and a product culture accustomed to operating with less hardware margin. The difference is not just technological; it’s geopolitical, energy-related, and corporate as well.
The U.S. bets on scale, chips, and control of the stack
The American strategy revolves around a well-known idea: control over infrastructure grants significant market power. NVIDIA has built its dominant position not only with GPUs but through a complete stack integrating accelerators, networking, software, libraries, validated systems, and architectures for what the company calls “AI factories.” Blackwell exemplifies this vision: it’s not just marketed as a chip but as an architecture for training, inference, reasoning models, and large-scale enterprise deployments.
This position also makes NVIDIA a central piece in geopolitical pressure. U.S. export restrictions on advanced chips to China aim to preserve Washington and its allies’ computing advantage. But control is never perfect. Recently, Reuters reported information from Bloomberg indicating U.S. authorities suspect NVIDIA chips found in servers reaching China via Thailand, with Alibaba named as a potential final destination. Alibaba denied any link to the companies involved or the use of prohibited GPUs.
The other major front is manufacturing. The Trump administration has pushed to bring some advanced production back to U.S. soil, with Intel as a key player aiming to regain relevance against TSMC’s dominance. A preliminary agreement between Apple and Intel, reported by Reuters from the Wall Street Journal, aligns with efforts to bolster local manufacturing and diversify a supply chain overly dependent on Taiwan.
Within this context, the project known variously as “Terafab,” linked to Elon Musk and Intel, has emerged. Forbes reported a $25 billion initiative to manufacture AI chips, with local Texas sites near College Station being considered among potential locations. These details should be viewed cautiously: no comprehensive public confirmation exists from all involved companies, and the published figures vary by source.
Their logic is clear. If Tesla, xAI, SpaceX, and other Musk-linked projects want to compete in autonomous driving, humanoid robotics, AI agents, and data centers, they need assured computing capacity. The GPU shortage and manufacturing concentration in Asia have made access to silicon a matter of business independence.
China responds with efficiency, open models, and domestic hardware
China cannot play the same game exactly. U.S. restrictions have limited access to the most advanced chips, forcing Chinese companies to maximize every available cycle. A different strategy is emerging: do more with less, reduce inference costs, leverage expert architectures, and accelerate the adoption of local chips like Huawei Ascend.
DeepSeek V4 exemplifies this approach. The company announced DeepSeek-V4 Preview in April 2026, with Pro and Flash versions, supporting up to a million tokens and designed with an architecture offering an aggressive cost-performance ratio. While it doesn’t eliminate the competitive edge of large closed U.S. systems, it pressures pricing and makes clear why clients should pay more for certain capabilities.
Moonshot AI follows a similar path with Kimi K2.6, pitched as a multimodal model focused on programming and agents. Meanwhile, Anthropic has released Claude Opus 4.6, and OpenAI has introduced GPT-5.5, highlighting that the closed frontier continues to advance. The gap isn’t simply that “China has caught up with the U.S.,” but that Asian open or semi-open models are reducing the practical distance in many enterprise use cases.
Efficiency has become a pressing necessity. MoE architectures allow activating only parts of a model per query, lowering computational costs compared to dense designs. Quantization, compression, caching, and kernel improvements are now as vital as model size. In a market where inference may outweigh training costs in final expenses, this focus on cost becomes a competitive advantage rather than a limiting factor.
Huawei’s Ascend 910C reflects this response. Its chips lack the global software support CUDA provides or the access to the most advanced manufacturing nodes, but China is actively promoting their adoption as a national alternative. U.S. officials acknowledge Huawei’s limited volume but also note that China is closing the gap in AI capabilities.
Open source is no longer just philosophy
The third front is open models. Meta has long argued that Llama democratizes access to advanced AI. Its official page boasts over 1.2 billion downloads, illustrating widespread community adoption.
Yet, the debate is more nuanced. The Open Source Initiative has questioned whether Llama’s licenses truly qualify as open source under their definition, citing restrictions on use and limited transparency regarding training data. It’s more precise to say these are often “open weights” models with proprietary licenses, rather than fully open-source models in the strictest sense.
Why would a company give away part of its technology? Strategically, making models more accessible shifts value towards infrastructure, proprietary data, distribution, product integration, and scalable operation. Meta can afford this because it monetizes attention, advertising, and platforms. Chinese firms use open models to gain international adoption, build community, and offset hardware disadvantages.
The result is a less comfortable market for closed labs. If an open or low-cost model can solve 80% of tasks at a fraction of the cost, many companies might opt not to always pay for the top-tier models. AI decisions are increasingly about architecture choices—cost per token, latency, privacy, data sovereignty, supplier dependency, and ease of tuning—more than benchmark scores.
The key question isn’t whether Chinese mathematical ingenuity will completely surpass the U.S. industrial muscle. It’s which layers of the stack this competition will unfold in. The U.S. still leads in accelerators, cloud, low-level software, and capital. China gains ground in efficiency, open models, and cost pressures. Together, these forces are shaping a new silicon order, where the most valuable AI isn’t always the largest, but the most sustainably deployable.
Frequently Asked Questions
How do U.S. and Chinese AI strategies differ?
The U.S. emphasizes advanced chips, large data centers, proprietary software, and supply chain control. China responds with more efficient models, MoE architectures, domestic chips, and aggressive pricing.
Why is NVIDIA so central to this race?
NVIDIA’s dominance isn’t just GPUs; it’s a combination of hardware, CUDA, networking, enterprise software, and comprehensive data center solutions for AI.
Are Chinese open models already better than U.S. closed models?
Not necessarily. They can perform similarly on specific tasks and often beat on price, but closed models from OpenAI, Anthropic, and Google still set the frontier in advanced capabilities.
Is Meta’s Llama truly open source?
It depends on the definition. Meta promotes it as open, but the Open Source Initiative considers its licenses not fully compliant with strict open source standards. It’s more accurate to call them open-weight models with proprietary licenses.