The AI Race Nobody Is Watching

Two Races. Different Finish Lines.

Every AI analysis published in the past two years has been a version of the same argument. Chips, compute, scaling laws, AGI timelines, safety versus acceleration. The frontier model competition between OpenAI, Anthropic, Google, and their Chinese counterparts has generated more strategic commentary than any technology development since the space race. And like the space race, most of the commentary has been watching the wrong competition.

There are two parallel AI races happening simultaneously, with almost completely different player rosters, different strategic priorities, and — critically — different definitions of what winning means.

The American race is a frontier capability competition. The bet is that scaling toward general intelligence produces a system so comprehensively capable that it transforms every industry simultaneously. Build the most powerful model possible. The economic value of AGI, if it works, is measured in tens of trillions. OpenAI, Anthropic, Google DeepMind, xAI, Meta AI — the roster is familiar, the capital expenditure is staggering, and the objective is a system that can do everything better than any human at any cognitive task.

The US AI ecosystem is diverse, but its leading labs have concentrated investment in compute-intensive frontier models, betting that superior hardware resources will yield transformative capabilities. China — constrained to some extent by US export controls on advanced semiconductors but backed by sustained state support — has organised around open development and rapid deployment, integrating AI across its economy under a rubric of general AI.

The Chinese race is a deployment competition. The bet is that AI which works well enough today, deployed everywhere immediately, and improved incrementally, wins the manufacturing buildout, the export market, the trade balance, and the labour cost equation. The finish line is not a system that can think better than any human. It is a robot on a factory floor in Shenzhen that costs $13,560 and can be retrained for a new task in an afternoon.

Only one of these races has shown up in the factory yet. It is not the American one.

The Frontier Layer: What the US Has and What It Costs

The American frontier model position is real, documented, and currently ahead. DeepSeek’s own technical paper concedes that V4’s reasoning and agentic capabilities trail state-of-the-art frontier models by approximately three to six months. The newest US models announced in April — Anthropic’s Claude and OpenAI’s GPT-5.5 — both show significant performance gains over their predecessors. The actual gap may also be widening, as US AI firms use AI to accelerate next-generation model development.

The United States leads on raw frontier capability. That lead is real. It is also increasingly expensive to maintain, increasingly difficult to translate into industrial deployment at scale, and increasingly contested by Chinese approaches that achieve near-frontier performance at a fraction of the cost. Chinese AI labs offer their models at increasingly competitive prices while narrowing the performance gap with US counterparts. Kimi K2.5 costs four times less than OpenAI’s GPT-5.2 as of January 2026, while matching it on benchmark performance.

The cost asymmetry is not incidental. It is strategic. Chinese frontier models are designed for mass deployment across an economy where price sensitivity is high and the government is actively subsidising adoption through API access and pre-trained model licences. The open-source strategy that DeepSeek pioneered with R1 — trained in two months for under $6 million, releasing capabilities that matched proprietary American models — is not primarily a research methodology. It is a distribution strategy. Open-source models propagate faster, embed deeper into industrial ecosystems, and create switching costs that are harder to reverse than proprietary subscription relationships.

The open strategy has been one key channel through which China aims to compete with the US, by rapidly scaling up adoptions and rolling out real-life applications in various sectors from e-commerce to robotics. DeepSeek’s V4, trained in partnership with Huawei’s Ascend AI processors, allows AI systems to be built and deployed without relying solely on Nvidia, which is why V4 could ultimately have an even bigger impact than R1 — accelerating adoption domestically and contributing to faster global AI development overall.

The American response to this has been export controls on advanced semiconductors — a policy that has produced the exact innovation pressure it was designed to prevent. Denied access to Nvidia’s most advanced chips, Chinese labs developed algorithmic efficiencies that allow near-frontier performance on constrained hardware. The export controls that were supposed to maintain American AI supremacy produced DeepSeek R1. The policy instrument designed to slow China accelerated its efficiency innovation.

The Industrial Layer: Where the Race Is Actually Being Run

The frontier model competition is the AI race that the technology press covers. The industrial deployment race is the one that determines who staffs the reshored factories.

Industry estimates show more than 13,000 humanoid robots shipped globally in 2025, with roughly 87% to 90% of them produced by Chinese companies. With over 140 domestic manufacturers and more than 330 humanoid models unveiled, China’s push into embodied AI is no longer experimental — it is commercial.

87 to 90 percent of global humanoid robot shipments in 2025 came from Chinese companies. The United States — home to the most heavily capitalised AI research ecosystem in history — produced the remainder. This is not a gap in frontier capability. It is a gap in deployment velocity, manufacturing scale, and the integration of AI software with physical industrial hardware that the frontier race has not adequately prioritised.

Unitree, China’s largest humanoid robot company, sold 5,500 humanoid robots in 2025, making it the world’s top seller. Shanghai-based Agibot came second with 5,168 units. Unitree and Agibot each sold more units than Tesla’s overall production target of 5,000 humanoid robots for 2025, which Tesla did not meet.

Tesla has been developing its Optimus humanoid programme for over five years, backed by some of the most sophisticated AI research and manufacturing capability on the planet. Unitree — a Chinese startup whose G1 humanoid robot is priced at $13,560 — outsold Tesla’s entire humanoid production target without meeting particularly extensive coverage in American technology media. Elon Musk, speaking at Davos, acknowledged the competitive reality directly: “China is very good at AI, very good at manufacturing, and will definitely be the toughest competition for Tesla. To the best of our knowledge, we don’t see any significant competitors outside of China.”

The Chinese industrial AI ecosystem is not competing on a single product. It is competing on a strategy. China’s 15th Five-Year Plan, covering 2026 to 2030, has called for the forward-looking development of future industries, positioning embodied intelligence as a new engine of growth. Beijing has launched major state policies including the Robot+ initiative and the AI plus Manufacturing roadmap, aiming to build humanoid robot pilot lines and double China’s manufacturing robot density by 2030. The Ministry of Industry and Information Technology set up the Standardisation Committee for Humanoid Robots. At least 15 Chinese automakers entered humanoid robotics in 2025, including GAC, SAIC, XPeng, Chery, and Xiaomi, leveraging existing supply chains for motors, batteries, and control systems.