The Global AI Race: China's Rise, U.S. Challenges, and the EU's AI Crisis

Executive Summary

As of mid-2026, the AI sector stands at a critical inflection point. The five largest U.S. hyperscalers—Amazon, Alphabet, Microsoft, Meta, and Oracle—are on pace to spend roughly $660–700 billion in capital expenditures in 2026, while global data-center CAPEX (including non-hyperscaler and sovereign spending) is projected to surpass $1 trillion, according to Dell’Oro Group.

China’s comparable spending lags substantially. Morgan Stanley estimated that Chinese cloud providers’ 2026 CAPEX would reach roughly $105 billion, while UBS estimated that Chinese internet giants’ AI CAPEX totaled US$59 billion last year, according to reporting by the South China Morning Post. Broader Chinese AI investment, including state funding, is estimated at around $125 billion. The spending gap is real and massive. What has surprised analysts is that the gap in model capabilities has not widened proportionally.

Stanford HAI’s 2026 AI Index provides the most rigorous data point available. It found that the benchmark performance gap between the leading American and Chinese models had narrowed to roughly 2.7 percentage points, down from 31.6 to 17.5 percentage points in 2023—despite the U.S. investing approximately 23 times more in private AI in 2025, $285.9 billion versus $12.4 billion. The 2026 Index should serve as the analytical anchor rather than isolated, one-off model comparisons.

U.S. AI Competitiveness

Chinese labs—including DeepSeek, Alibaba’s Qwen, Moonshot’s Kimi, and Zhipu’s GLM—have moved from laggards to credible alternatives, particularly in open-weight models. DeepSeek’s R1, which the company claimed was trained for under $6 million, sparked debate after several analysts, research firms, and technology publications challenged the figure as misleading. The DeepSeek inflection point triggered a roughly $593 billion single-day decline in Nvidia’s (NVDA) market capitalization in January 2025 after the model matched several U.S. reasoning benchmarks, reshaping assumptions about the relationship between compute spending, total cost of ownership, and model performance.

North American frontier labs—including OpenAI, Anthropic, and Google DeepMind—still maintain a narrow lead, and Stanford HAI research suggests that recent trends increasingly favor China’s pace of advancement. Chinese models appear significantly cheaper to operate; some estimates suggest costs 7 to 50 times lower in certain comparisons. These figures vary widely by source, task, model architecture, and provider, so they should be viewed as directional rather than precise.

A further complication arose in June 2026, when the U.S. Commerce Department, under Secretary Howard Lutnick’s signature, ordered Anthropic to suspend foreign-national access to its newest and most powerful models, Claude Fable 5 and Claude Mythos 5, following a national-security review prompted by a reported jailbreak incident. Anthropic’s inability to reliably distinguish between foreign and domestic users in real time resulted in a three-week global suspension of both models; access was restored on July 1, 2026. The episode highlighted a new risk: U.S. frontier models can be removed from the market through regulatory action with little or no advance notice, intensifying concerns about reliance on U.S. artificial intelligence infrastructure. It also triggered contract cancellations and reviews involving U.S. data analytics firm Palantir across the European Union. By contrast, Chinese open-weight alternatives—which can be hosted independently, deployed locally, and retrained on premises—are less exposed to this particular regulatory vulnerability.

Financial And Investment Implications

The U.S. AI investment case still rests on the expectation that today’s massive CAPEX will translate into durable, defensible revenue. That case looks weaker than it did a year ago. Leaked OpenAI financials, verified by the Financial Times, show a 2025 operating loss of roughly $20.9 billion on $13.1 billion in revenue; the net loss, including one-time accounting charges related to its for-profit conversion, was $38.5 billion. Losses are widening even as revenue triples year over year—the opposite of the typical software model, where scale generally improves margins.

Affordable, capable Chinese open-weight models compound this risk in two ways. First, they offer enterprises a lower-cost, high-performing alternative that could erode pricing power for U.S. proprietary, closed-weight models. Second, they allow enterprises to experiment with AI and validate use cases before committing to paid models. Neither effect guarantees a valuation reset for U.S. AI leaders, hyperscalers, or Nvidia. Demand for frontier compute remains supply-constrained, and hyperscalers report that their AI businesses are already profitable at the margin—but these trends do argue for greater scrutiny of unit economics, not less. The true cost of ownership of AI solutions is multifaceted and highly context-dependent. Factors such as open-source versus commercial solutions, support requirements, internal expertise, and hardware infrastructure needs can significantly alter the economics depending on the business case, scale, and complexity of deployment. Therefore, claimed cost-saving multiples should be viewed as directional estimates rather than fixed truths.

Strategic And Geopolitical Consequences

Export controls remain the U.S.’s primary lever, and 2026 has highlighted both their limits and their potential long-term unintended consequences. Restrictions on advanced chips have incentivized some Chinese developers to pursue domestic silicon alternatives, such as Huawei’s Ascend line, while accelerating the adoption of more compute-efficient architectures, including mixture-of-experts models that activate only a fraction of parameters per query. In effect, U.S. pressure may have contributed to faster progress in computational efficiency and applied AI research—areas where the controls were intended to preserve a strategic and commercial advantage.

Anthropic's Fable 5 and Mythos 5 Odyssey prove that U.S. export-control decisions can harm American firms’ own products and economic interests, unsettling trusted allied governments and customers who relied on those models. European commentary in particular raised questions about AI sovereignty and over-reliance on any single U.S. vendor, despite limited access to European AI alternatives.

Expect continued U.S. efforts on multiple fronts: tighter and more specific rules attempting to classify model weights as controlled technology, incentives for domestic chip and data-center capacity, and corporate hedging—with enterprises increasingly building multi-model, multi-vendor architectures precisely so that no single regulatory action, access restriction, or model retirement can disrupt their operations. This U.S. strategy requires significant investments in grid infrastructure, power generation capacity, and supply-chain resilience for essential metal commodities, among other areas. Europe is lagging behind the United States and China in frontier AI development, given its more limited capacity to develop, invest in, commercially deploy, and operate large-scale AI systems.

The Future AI Race

China is on a clear trajectory toward major AI influence and is almost certainly becoming a durable AI superpower. Its unique combination of industrial scale, state-backed investment, engineering talent, and open-weight model strategy gives it a credible path to global leadership—even without matching U.S. private investment dollar-for-dollar. Recent high-performing Chinese models, such as Qwen3.7, GLM-5.2, and DeepSeek V4, demonstrate this momentum.

The battle for AI leadership remains indecisive. Compute and energy capacity continue to favor the U.S., though that advantage may narrow as China rapidly expands its power infrastructure, semiconductor capabilities, supply chain resilience, open-model distribution, and competitive pricing for AI compute. Meanwhile, enterprise monetization and durable business models remain unsettled across the industry, while regulatory uncertainty has emerged as a relative weakness for the U.S. and EU.

In the mid-term, a two-tiered AI ecosystem is likely to emerge: the U.S. leading in the highest-performing proprietary frontier systems, while Chinese models become a dominant open-source foundation for cost-efficient global deployment. This view should be treated as a strategic assessment rather than a precise forecast, as the balance remains fluid and could shift rapidly—potentially decisively—following a major breakthrough or transformative model release from either side.

The Bottom Line

China has not yet surpassed U.S. frontier AI systems. It has narrowed the measurable benchmark gap faster than the investment gap alone would predict. The U.S. remains the leading AI power, supported by advantages in research, compute infrastructure, semiconductor access, capital, and commercialization—though that lead isn't guaranteed: high development costs, uncertain returns, shifting regulation and policies, and weaker confidence among allies could all erode it over time.

China's strengths lie in efficient model development, open-weight distribution, lower per-query costs, and rapid industrial deployment—a credible path to broad global influence even without leading every frontier benchmark.

The EU is likely to remain influential in AI governance and standards-setting. It currently lacks the investment scale, infrastructure, talent pipeline, and commercialization ecosystem to compete directly for frontier leadership. Its constraint is resources, but also coordination and political commitment.

The race is still open. Long-term success will depend not only on model capability or spending, but on the ability to deliver affordable, scalable, and widely adopted AI systems.

GlobAS Group empowers you throughout your AI journey.