Table of Contents
On December 3, 2025, Jensen Huang sat down at CSIS and described AI as a “five-layer cake: energy, chips, infrastructure, models, and applications.” Then he delivered a verdict that should concern anyone paying attention: “China has twice the amount of energy we have as a nation.” At the chip layer, he was more confident—“we are generations ahead”—but he immediately added the caveat that matters: “Don’t be complacent. Anybody who thinks China can’t manufacture is missing a big idea.” (CSIS conversation with Jensen Huang)
The math here is disorienting. David Sacks, Trump’s AI czar, has said China is “three to six months” behind in AI capability. (MIT Technology Review) Huang himself has wavered between “nanoseconds behind” and, briefly, “China is going to win.” The whiplash isn’t indecision—it’s an honest reflection of a competition where different metrics tell radically different stories.
Here’s the uncomfortable truth: the United States leads in proprietary frontier models, compute infrastructure, and chip manufacturing. China leads in open-source AI, energy capacity, build velocity, and the sheer scale of industrial AI deployment. Neither advantage is decisive. Both are eroding. And the outcome of this contest will shape not just which companies profit, but which values govern how AI reshapes economies, militaries, and societies worldwide.
This is not a race with a finish line. It’s a restructuring of global power on a timeline measured in quarters, not decades.
The five-layer stack and who owns each floor
Huang’s five-layer framing is the clearest lens for understanding where the competition actually stands. At each layer, the balance of power looks different—and the interdependencies create fragilities that neither side fully controls. The uncomfortable reality is that America dominates some layers while China dominates others, and the layers interact in ways that make simple scorecards misleading.
Energy sits at the foundation, and here China’s lead is structural. China generates roughly twice the electrical power of the United States, with grid reserve margins in some regions sitting at 80–100%, compared to around 15% in much of America. (Medium analysis) Chris Miller, author of “Chip War,” testified before the Senate that the trinity of AI leadership is “computing power, brain power, and electrical power”—and that “America has a substantial lead in computing power, but China leads in electrical power.” (CNBC)
Energy isn’t just about training models. It’s about the velocity of data center deployment. Huang put it starkly: “If you want to build a data center here in the United States, from breaking ground to standing up an AI supercomputer is probably about three years. They can build a hospital in a weekend.” (Fortune) That construction differential compounds over time. Every year of permitting delays and grid bottlenecks in America is a year China can use to scale.
Federal projections show data centers alone could triple their share of national power consumption to 12% by 2028. The implications are sobering: if America cannot build the physical infrastructure to house AI, then chip superiority becomes a stranded asset—like owning the world’s best engines but lacking runways. China’s approach integrates energy planning with AI deployment; America’s treats them as separate problems to be solved by different agencies and different markets.
Chips remain America’s trump card—for now. The United States controls roughly two-thirds of global AI compute and maintains a manufacturing advantage that estimates place at 5 to 15 years ahead of China’s domestic capability. (Atlantic Council) NVIDIA’s Blackwell GPUs achieve commercial yields estimated at 60–80%, while Huawei’s Ascend chips struggle with yields between 5% and 20%. (Tom’s Hardware)
The yield differential is crucial because it translates directly into cost and scale. A 60% yield means six working chips per ten manufactured; a 10% yield means one. Multiply that across millions of chips, and China faces a compounding cost disadvantage that no amount of state subsidy can entirely overcome. But this advantage depends on a supply chain that runs through Taiwan, South Korea, the Netherlands, and Japan—allies, but not territories. A disruption at any node ripples through the entire Western AI ecosystem.
But China is adapting faster than skeptics predicted. Huawei plans to produce roughly 600,000 of its 910C Ascend chips next year, doubling this year’s output. More significantly, SemiAnalysis found that Huawei’s CloudMatrix system—despite using chips that individually deliver only one-third the performance of NVIDIA processors—outperformed NVIDIA’s competing system on some metrics by linking five times as many chips together. (Bloomberg) Architecture can compensate for process node disadvantages, at least partially. The lesson from CloudMatrix is that China is not trying to replicate America’s approach—it’s trying to route around it.
Baidu recently switched on a computing cluster with 30,000 of its own Kunlun chips, which it claims can train DeepSeek-like models. The company’s chip business won a $139 million order from state-owned China Mobile. (Rest of World) Cambricon Technologies reported 4,000% revenue growth in early 2025 and achieved profitability for the first time. The Chinese domestic chip ecosystem is nascent, but it exists—and it’s accelerating under the pressure of export controls that were designed to prevent exactly this outcome.
Infrastructure is where China’s build velocity becomes operational advantage. China’s “East Data, West Compute” strategy places compute centers in resource-rich western provinces while serving data-dense eastern markets—a coherent national plan executed at scale. The United States, by contrast, is discovering that around 70% of AI industry executives now identify grid capacity as their biggest obstacle, a profound shift from a few years ago when chip supply dominated concerns. (Medium analysis)
CoreWeave’s CEO framed it bluntly: “America’s ability to lead in AI hinges on a simple but urgent question: Can we build the computing infrastructure fast enough to unleash AI’s full potential?” (CoreWeave) The United States has focused on winning the chip race; China is focused on winning the build-out race. Those are not the same contest.
Models present the most counterintuitive picture. The United States leads in proprietary frontier systems—GPT-5.2, Claude Opus 4.5, Gemini 3 Pro—and the performance gap at the very top remains real. But China has taken a fundamentally different approach, betting heavily on open-source. Huang, speaking in Beijing this summer, was unambiguous: “Models like DeepSeek, Alibaba, Tencent, MiniMax, and Baidu Ernie bot are world class, developed here and shared openly.” He called Qwen and DeepSeek “the best open reasoning models in the world today.” (CNBC)
The numbers bear this out. Chinese open-source models grew from 1.2% of global usage in late 2024 to nearly 30% in 2025, according to a study covering 100 trillion tokens. (South China Morning Post) The performance gap between the best Chinese and U.S. models shrank from 9.3% in 2024 to 1.7% by February 2025. (Wilson Center)
DeepSeek’s January 2025 release of R1 and V3 disrupted the industry’s core assumptions. These models achieved performance comparable to leading Western models at dramatically lower costs, proving that effective AI development doesn’t require $100 million training runs. The company reported training costs of roughly $5.6 million—a fraction of what OpenAI spent on GPT-4. That efficiency advantage matters because it democratizes access: startups, universities, and developing nations can now participate in frontier AI development using Chinese open-source models in ways they cannot using American proprietary ones.
Applications are where the rubber meets the road. China operates roughly 2 million industrial robots and installed about 295,000 more in 2024 alone—more than the rest of the world combined. The U.S. installed approximately 34,000. (The AI Insider) China’s Ministry of Industry and Information Technology estimates that by the end of 2025, over 60% of large Chinese manufacturers will have adopted some form of “AI + Manufacturing” integration. The United States leads in AI software; China is pulling ahead in AI-as-industrial-infrastructure.
The application layer is where abstract model capability becomes economic output. A marginally better language model matters less than a deployed factory automation system. China’s advantage here reflects its manufacturing base: 1.4 billion people and the world’s largest middle class generate data at a volume unmatched elsewhere, while the factory floor provides endless opportunities for AI deployment. America builds the best tools; China has more places to use them.
Follow the money—and the missing money
Private investment tells one story: the United States attracted $109 billion in AI investment in 2024, nearly 12 times China’s total. If the Stargate initiative materializes, OpenAI alone will deploy $500 billion over four years to build new AI infrastructure. (RAND) America’s venture capital ecosystem remains unmatched for funding frontier research and scaling startups. The United States boasts all ten of the world’s top AI firms by market value, plus 37 of the top 50; NVIDIA alone became the first company valued at $5 trillion. China has just four AI firms in the top 50—the same as Israel. (The Conversation)
But the investment advantage masks structural weaknesses. American AI spending concentrates at the top of the stack—models and applications—while infrastructure and energy remain bottlenecked. China’s spending is more vertically integrated, flowing from rare earth extraction through chip manufacturing to deployment. The U.S. imports 70% of its rare earth minerals from China, a dependency that export controls have not addressed. (American Security Project) Beijing has wielded rare earths as economic weapons through export controls for years; there’s no reason to expect restraint if competition intensifies.
The investment math also obscures different national priorities. American capital flows toward frontier model development and consumer applications—chatbots, image generators, coding assistants. Chinese capital flows toward industrial applications, robotics, and infrastructure. Goldman Sachs estimates every $1 in AI investment generates $4.90 in economic output, but that multiplier depends on what you’re investing in. (Bloomberg) An AI system that automates a factory creates different economic value than one that writes marketing copy.
Talent presents a similar mixed picture. The U.S. outdoes China in top AI research talent, but the gap is narrowing. In 2019, 59% of the world’s top AI researchers worked in the United States versus 11% in China; by 2022, those figures were 42% and 28%. (The Conversation) In 2020, China ranked first globally with 3.57 million STEM graduates, compared to 820,000 in the U.S. (CNBC) According to a MacroPolo report, nearly half of the world’s top AI researchers completed their undergraduate studies in China—many of them currently working in America.
China’s Thousand Talents Plan actively recruits overseas experts, and the flow of researchers back to China is accelerating. Immigration policy, research funding, and academic freedom all shape which direction talent moves. America’s advantage in attracting global talent depends on remaining attractive—something that cannot be taken for granted. The researchers training GPT-6 and Claude 5 may have completed their undergraduate degrees in Beijing or Shanghai; whether they stay in America to build those systems is a policy choice, not a law of nature.
The strategic question isn’t who has more money or more researchers today. It’s who has the pipeline to sustain advantage over the next decade. China produces more PhDs in the sciences than anywhere else, ensuring a deep pool of expertise. More data, more talent, and more coordinated investment create a self-accelerating loop. The compounding effects favor whoever maintains momentum.
The open-source gambit and why it matters
China’s open-source strategy isn’t just clever engineering—it’s a geopolitical play. By releasing models like DeepSeek and Qwen for free, Chinese firms have accomplished something American export controls were designed to prevent: they’ve made Chinese AI the default for developers who can’t afford OpenAI’s API prices or don’t want vendor lock-in. This is soft power through software—an approach that echoes America’s own playbook from the internet era, now turned against its originator.
DeepSeek-V3.2 illustrates the economics. At $0.27 per million input tokens, a complex task costing $15 with GPT-5 costs approximately $0.50 with DeepSeek. (Sokada) For startups in Brazil, Nigeria, Indonesia, and India—countries representing billions of future AI users—the cost differential isn’t marginal. It’s determinative. A founder in Lagos or Jakarta choosing between building on GPT-5 or DeepSeek isn’t making an ideological choice; they’re making a financial one. And they’re choosing China.
Huang captured the strategic logic: “China’s open-source AI is a catalyst for global progress, giving every country and industry a chance to join the AI revolution.” That’s not altruism. When developers worldwide build on Qwen and DeepSeek, they build familiarity with Chinese AI ecosystems, create demand for Chinese cloud infrastructure, and generate training data that flows back to Chinese labs. Open source was “extremely powerful” for AI innovation, Huang said in Beijing, and Chinese firms offer the “best open reasoning models.” (NVIDIA Blog)
The strategic implications extend beyond market share. Open source allows not just the contribution of individual companies but the combined resources of an ecosystem—what Huang called “very clever about open source engineering here in China.” While OpenAI guards its weights, DeepSeek shares them. While Anthropic requires API calls, Alibaba lets developers download Qwen. The philosophical difference becomes competitive advantage when the rest of the world decides which ecosystem to join.
The United States has noticed. OpenAI finally released an open-source model in August 2025, breaking from its proprietary stance for the first time since GPT-2 in 2020. (Fortune) But the American ecosystem remains predominantly closed, and the most capable American models require expensive API access or enterprise contracts. Meta’s Llama represents a partial counterweight, but Llama’s licensing restrictions limit commercial deployment in ways that DeepSeek and Qwen do not.
Huang’s framing cuts to the strategic heart: “The question is not whether China will have AI—it already does. The question is whether one of the world’s largest AI markets will run on American platforms.” He argues that allowing Chinese models to run on NVIDIA chips gives American companies insight into global AI development. (Calcalist) That may be true, but it’s a defensive posture—preserving market share rather than commanding it.
The irony is profound. For decades, American technology companies preached openness as a competitive virtue—open protocols, open standards, open source. That philosophy built the internet and established American dominance over global digital infrastructure. Now China has adopted the playbook, and America has retreated behind proprietary walls. The student has learned from the teacher.
The ways this contest could fracture
The optimistic scenario involves what researchers call “splinternet” formation: two parallel digital ecosystems—one American, one Chinese—competing for the allegiance of the rest of the world. That’s unpleasant but manageable, a return to Cold War spheres of influence adapted for the digital age. New tech alliances are already emerging: the “Chip 4” alliance comprising the U.S., Japan, Taiwan, and South Korea coordinates semiconductor strategy, while China cultivates partners through its Digital Silk Road. (World Economic Forum)
The pessimistic scenarios are worse. Yoshua Bengio, the Turing Award laureate, has warned that “if major powers such as the US and China treat AI development as a race, competitive pressures could ultimately harm everyone.” (Shanghai Gov) The logic is straightforward: racing incentivizes speed over safety, corner-cutting over caution, deployment over alignment. The Center for AI Safety cites malicious use, competition leading to rushed development, organizational risks, and rogue AI as primary drivers that could undermine human-first approaches. (World Economic Forum)
AI-enabled autonomous weapons represent the most acute risk. In November 2024, Washington and Beijing publicly pledged to exclude AI from nuclear command and control systems—a rare moment of alignment. (RAND) But conventional applications of AI to military systems proceed without such constraints, and the economic pressures of the competition make restraint expensive. Each side fears falling behind in military AI more than it fears the consequences of racing ahead.
For developing nations, the competition creates a forced choice reminiscent of the original Cold War’s Non-Aligned Movement. Both the U.S. and China are aggressively promoting their respective AI technology stacks in emerging markets. Chinese firms like Huawei and Alibaba Cloud offer comprehensive AI solutions bundled with Belt and Road infrastructure projects. American tech giants backed by government initiatives push their own platforms. (ICTworks) Countries that once navigated between superpowers now face mounting pressure to align their digital infrastructure with one ecosystem or the other.
The rise of regional tech hubs in India, Kenya, and Brazil suggests that a new technological non-alignment may be possible. These countries are developing their own AI applications adapted to local needs while selectively engaging with both American and Chinese technologies. But the costs of non-alignment are rising. India’s attempt to regulate AI while deploying both American and Chinese systems faces pressure from both sides. Indonesia’s digital infrastructure investments require choosing between Huawei equipment and American alternatives.
There’s a counterfactual worth considering: fragmenting the global AI ecosystem into competing blocs reverses historic achievements in transnational innovation. Instead of a shared global innovation frontier, the world faces slower, duplicated, and less efficient technological evolution. (National Interest) The 2025 Nobel laureates in economics emphasized that societies embracing openness prosper; those that isolate themselves risk stagnation. The Cold War’s space race produced innovations, but it also meant two separate programs developing redundant capabilities—resources that could have been pooled.
The domestic risks matter too. AI-influenced campaigns pose threats to democratic processes, with potential to destabilize governments and influence election outcomes. (APCO Worldwide) China has already pioneered AI-enabled authoritarian systems; the question is whether American democracy can develop effective countermeasures without becoming what it opposes. The tools that protect elections can also be used to surveil citizens. The algorithms that detect disinformation can also censor dissent. Competitive pressure makes these tradeoffs harder to navigate thoughtfully.
What it would take to win—and what winning might mean
The Georgetown Journal of International Affairs published a framework for the AI competition in November: “Compete, Counter, Cooperate.” (Georgetown) Competition means maintaining leads where they exist—chips, proprietary models, top talent. Countering means addressing vulnerabilities—energy infrastructure, supply chain dependencies, build velocity. Cooperation means preserving channels for mutual benefit, particularly on safety. The challenge is executing all three simultaneously while under competitive pressure that incentivizes the first at the expense of the second and third.
The factors that will decide the outcome cluster around a handful of variables:
Energy and infrastructure velocity. America cannot win an AI competition if it takes three years to build what China builds in months. Permitting reform, grid modernization, and strategic deployment of nuclear and renewable capacity are prerequisites, not nice-to-haves. The Biden administration’s approach treated infrastructure as climate policy; the Trump administration’s treats it as competition policy. Neither has yet produced the velocity required. Huang’s Taiwan partnership, bringing TSMC expertise to Arizona fabs, represents a template: allied cooperation that accelerates domestic capability. (CSIS)
Talent retention and immigration. Half the world’s top AI researchers trained in China and currently work in America. That pipeline can flow in either direction, and it responds to policy. Immigration restrictions, research funding cuts, and academic freedom concerns all shape which direction talent moves. Stephen Roach at Project Syndicate has argued that cuts to basic research could enable China to win the AI race—not because China’s researchers are better, but because America’s best researchers might leave or never arrive. (Project Syndicate)
Open-source strategy. The American ecosystem’s proprietary bias cedes the global middle market to China. A more aggressive open-source posture—perhaps coordinated across American and allied labs—could change that calculus. The UK, France, Germany, and Canada all have AI research capabilities; coordinated open-source releases could match Chinese scale while preserving Western values around safety and transparency.
Manufacturing resilience. Taiwan produces the world’s most advanced chips through TSMC, a concentration of capability that represents a single point of failure for the entire Western AI stack. Diversification through Arizona fabs and allied partnerships is progressing, but slowly. Huang acknowledged Taiwan’s crucial role at CSIS: “Taiwan really needs to have some acknowledgment for the incredible effort that they’re putting in place to help us reindustrialize the United States.” He noted that a recent TSMC event in Arizona was “two-thirds Taiwanese and one-third American.” (CSIS) The dependency is real and cannot be wished away.
Safety coordination. The irony of the competition is that both sides face similar risks from the technology they’re racing to deploy. Joint research, incident reporting, and red-team testing could reduce duplication, prevent accidents, and maintain global trust—much like U.S.-Soviet collaboration on nuclear safety during the original Cold War. (RAND) The November 2024 pledge on nuclear command and control proves such coordination is possible even amid competition.
The Congressional Budget Office, ignoring AI, projects long-run U.S. growth at 2%—a doubling every 36 years. Goldman Sachs estimates AI investment alone could reach $200 billion globally by year-end, with projections suggesting every $1 in AI investment generates $4.90 in economic output. (Bloomberg) Even pessimistic AI scenarios add meaningful growth; optimistic scenarios transform the trajectory. The winner benefits from faster economic growth, a more capable military, and powerful new tools of economic statecraft. The stakes explain why neither side can afford to disengage.
The White House’s AI Action Plan frames winning as ushering in “a new golden age of human flourishing, economic competitiveness, and national security.” (White House) That’s not wrong, but it’s incomplete. The question isn’t just whether America or China leads. It’s whether the competition produces AI systems that broadly benefit humanity, or ones that concentrate power, automate conflict, and lock in existing inequalities. The outcome will decide more than market share. It will shape global norms for how AI is used, who controls access to the most capable systems, and whether computational power becomes concentrated in a handful of states or distributed more widely.
The reason for hope isn’t that either side is clearly winning. It’s that both sides face the same fundamental constraint: AI systems that destabilize their societies, alienate their allies, or accelerate global catastrophe undermine their own interests. Competition creates pressure to cut corners, but it also creates pressure to demonstrate that your approach produces better outcomes for more people. Work with the UK’s National Health Service improved operational efficiency, boosted staff morale, and delivered better patient outcomes. (World Economic Forum) Wildlife agencies use AI to predict wildfire hotspots; conservation groups deploy it to identify ocean plastic accumulation. The technology can improve lives—if the competition doesn’t prevent us from using it wisely.
Eric Schmidt, former Google CEO, captured the logic in Shanghai: “China and the US should cooperate to jointly maintain world stability and ensure that humanity can control these AI tools.” (Shanghai Gov) That’s not naiveté—it’s recognition that a race to the bottom benefits no one. Some analysts argue you cannot win an AI race, but you can promote better outcomes by including more people and ensuring the benefits are broadly shared. The competition may be inevitable; the mutual destruction is not.
The outcome will depend on choices made over the next few years: whether to prioritize speed or safety, openness or control, competition or cooperation. The technology itself is agnostic. What matters is whether the humans racing to deploy it remember that winning means something only if the world that results is worth living in.
The most hopeful reading of the current moment is that both sides understand this, even if neither can say it publicly. The most dangerous reading is that competitive pressure will overwhelm that understanding, producing a race neither side can afford to lose and neither can afford to win. Six months is not much of a lead. What we do with that margin will define the century.
Whoever wins, the rest of the world doesn’t have to lose. The original Cold War ended not with nuclear exchange but with integration—a reminder that competitions can resolve in ways that benefit everyone. AI could raise living standards, cure diseases, accelerate scientific discovery, and solve problems that have bedeviled humanity for centuries. The question is whether the race to build it leaves room for the wisdom to deploy it well. That depends on whether six months of lead time buys six months of reflection—or just six months of faster sprinting toward a finish line nobody can afford to cross alone.