By Dimitri Gellé

The global race for Artificial Intelligence (AI) is often portrayed as a technological contest between Silicon Valley’s ingenuity and Beijing’s state-backed ambition. However, a closer examination reveals that the true competition is structural and geopolitical, defined by contrasting economic models, regulatory philosophies, and access to fundamental resources. This structural imbalance has recently been highlighted by two highly publicised, yet contrasting, narratives: the technological insider’s warning and the financial contrarian’s bet.

Jensen Huang, the Chief Executive Officer of the AI chip manufacturer Nvidia, has issued a significant warning, stating that “China is going to win the AI race” [4]. Simultaneously, financial contrarian Michael Burry, famed for his “Big Short” bet against the 2008 housing crisis, has placed a substantial short bet against AI companies like Nvidia and Palantir, signalling a considerable risk of overvaluation [4]. The convergence of China’s state-backed structural advantages and the West’s market enthusiasm and regulatory caution suggests a fundamental imbalance in this “AI Innovation War.”

The AI Bubble Warning

Bubble fears are rising again around AI-linked stocks, and it’s not hard to see why. Investment in the sector has reached a scale that is genuinely difficult to grasp: spending on data centres alone approached $500 billion in 2024, growing at over 50% year-on-year in early 2025, with projections suggesting the bill could exceed $1.2 trillion globally by 2029 [3]. On top of that, speculative behaviour is picking up. Retail investors are piling in, and US margin debt jumped more than 32% in just six months, a pace last seen at the peak of the dot-com bubble in 2000 [3]. A handful of companies are driving almost the entire market: the ten largest US firms now account for more than 20% of global stock market value [3]. When money concentrates this fast around these few names, history tends to issue a warning.

Yet Deutsche Bank’s analysts argue that drawing a straight line to 2000 misses something important. Back then, sky-high valuations were built on little more than optimism: companies with no profits, no revenue model, and no clear path to either. Today is different. The biggest AI companies are generating real earnings, strong cash flows, and expanding profit margins of 55% to 65% in infrastructure segments [3]. Their shares trade at 25x to 40x expected earnings, elevated, but nowhere near the 80x-plus multiples that were common at the dot-com peak [3]. And crucially, unlike the late 1990s when firms borrowed heavily to fund expansion, today’s tech giants are largely paying for their investments out of their own pockets [3].

Think of it this way: imagine two property developers. One borrows everything to build a neighbourhood before a single tenant has signed a lease. That was 2000. The other builds with cash reserves, already has anchor tenants locked in on long-term contracts, and is generating rent from day one. That is closer to what companies like Microsoft or Google are doing today with AI infrastructure. The risk isn’t insolvency; it’s whether the neighbourhood fills up fast enough to justify the price of the land.

And that is precisely where Deutsche Bank sees the real vulnerability. Monetisation remains an open question: despite high usage, only around 3% to 5% of AI users are actually paying for the service [3]. Corporate adoption is still patchy, and an MIT study cited in the report found that up to 95% of enterprise AI projects have yet to deliver measurable financial returns [3]. If that gap between spend and payback widens rather than closes, even fundamentally sound companies could face a painful market correction, made worse by the sheer concentration of capital in the sector.

Deutsche Bank’s conclusion is clear-eyed: this is not a bubble in the classic sense, but “the risk could rise if investment outpaces demand, monetisation disappoints, or speculative excesses build” [3]. Burry may yet be proven right. But unlike 2000, this cycle has real earnings and solid balance sheets behind it, which means any correction, if it comes, would likely look less like a crash and more like a long, uncomfortable reckoning.

The Regulatory Brake

The question of how to regulate AI without strangling it is one that divides governments as sharply as the technology itself. And the stakes, according to the US National Security Commission on Artificial Intelligence (NSCAI), could not be higher: the commission warns that the race to develop and deploy AI “is already intensifying strategic competition,” and that China “stands a reasonable chance of overtaking the United States as the leading center of AI innovation in the coming decade” [1].

China’s approach has never been about giving AI a free rein. Since 2023, the Cyberspace Administration of China has required AI providers to align outputs with “socialist core values,” block politically sensitive content and submit algorithms for state registration [6]. In practice, this means models like Baidu’s ERNIE are trained to refuse questions about Tiananmen Square, Taiwan or criticism of Chinese leadership, a form of built-in censorship embedded at every stage of development [6]. Think of it like a car with a mandatory speed limiter: it can still move fast, but certain roads are permanently closed.

Yet the CIGI paper’s central finding is striking: these restrictions have not fundamentally crippled China’s AI performance. DeepSeek’s V3 model remains competitive with leading American systems on technical benchmarks, even while complying with China’s censorship framework [6]. The limitations are real but largely confined to politically sensitive domains; outside those boundaries, the model’s computational capabilities remain intact [6].

The United States, meanwhile, has moved in the opposite direction. President Trump’s January 2025 Executive Order rescinded Biden’s 2023 safety-first framework, explicitly prioritising innovation and speed over precaution [6]. Yet the NSCAI had long warned that this kind of reactive policymaking was insufficient. Its diagnosis was blunt: the US government “champions AI leadership in speeches and memorandums, but deploys few resources relative to commercial investment and historic funding benchmarks” [1]. Rather than simply deregulating, the commission called for a coherent, White House-led national technology strategy capable of integrating promotion and protection policies across the entire AI ecosystem [1].

What the combined picture reveals is that neither model is without cost. China’s security-first approach preserves state control but introduces content bias and risks stifling the kind of open-ended innovation that produced DeepSeek, a success the CIGI paper describes as “the exception rather than the norm” within China’s ecosystem [6]. The US, now loosening its own guardrails, faces a different gamble: whether speed without strategy leads to leadership or to vulnerability. As the NSCAI put it, China’s competitive model “should not define the US approach to innovation, but it does present an alternative model of AI development” [1], one that forces Western governments to reckon seriously with what kind of race they are actually running.

China’s Structural Edge: The Energy Factor

Beyond regulation, the AI race has an energy problem. As Brookings senior fellow R. David Edelman notes, “the fates of breakthrough technologies and the energy to power them are deeply interwoven, with progress at scale in one difficult to advance without the other” [2]. AI systems are, in his words, “notorious energy consumers” [2]: training a single cutting-edge model could consume roughly as much electricity as 5 million US households in a year, while a standard query to a large language model requires around ten times the electricity of a conventional search engine result [2].

Both Washington and Beijing understand what this means strategically. Xi Jinping has explicitly linked AI and “new energy” as twin national priorities, calling for “a major strategic deployment for building a new power system” empowered by new technologies [2]. China’s response is coordinated from the top: central planners set targets, direct infrastructure, and absorb risk that private markets would not.

The US model is different. Investment is real but market-driven. TerraPower’s recent $650 million fundraise for its advanced Natrium nuclear reactor, backed by NVIDIA’s venture arm NVentures and Bill Gates, illustrates the logic: “as AI continues to transform industries, nuclear energy is going to become a more vital energy source to help power these capabilities” [7]. TerraPower’s first plant is set to be the United States’ first commercial advanced nuclear facility, with regulatory approval expected next year [7].

The contrast is not simply public versus private capital. It is a question of speed and coordination. As Edelman concludes, “putting into place a clear strategy that recognises this interplay and harnesses the multipliers of private sector investment will confer not just strategic advantage but could also offer platforms for multilateral and even bilateral cooperation” [2]. Whether private innovation can match the urgency of a state-directed competitor remains one of the defining questions of the AI decade.

Geopolitical Flashpoint: The Chip War

Semiconductors have become the oil of the AI era, and control over their production is now as strategically consequential as control over energy reserves. No single move illustrates this shift more than TSMC’s announcement in March 2025 that it would expand its US footprint to $165 billion, adding three fabrication plants, two advanced packaging facilities and a major R&D centre on American soil, what the company itself described as “the largest single foreign direct investment in US history” [5]. For TSMC Chairman and CEO Dr. C.C. Wei, the logic was unambiguous: “semiconductor technology is the foundation for new capabilities and applications” in an AI-driven world [5]. The objective is equally clear: to complete what TSMC calls “the domestic AI supply chain,” anchoring the production of the world’s most advanced chips within US borders and away from the geopolitical fault lines of the Taiwan Strait [5].

The ambition is considerable. The execution exposes a deeper structural vulnerability. For decades, the United States has led in chip design while outsourcing the manufacturing of its most critical hardware to a single foreign partner in one of the world’s most contested regions. TSMC’s $165 billion commitment is, in effect, an admission that this model is no longer sustainable, and that reshoring it will take years, not quarters.

Beijing, meanwhile, has not been standing still. Chinese technology firms have placed orders for more than 2 million Nvidia H200 chips priced at approximately $27,000 each, a figure that already far exceeds Nvidia’s available inventory of 700,000 units [8]. The demand pressure is such that Nvidia has imposed unusually stringent commercial terms: full upfront payment, with no options to cancel, seek refunds or modify configurations after order placement [8]. The payment structure, Reuters reports, “effectively transfers financial risk from Nvidia to its customers, who must commit capital without certainty that Beijing will approve the chip imports” [8]. It is a posture born of hard experience: Nvidia previously absorbed a $5.5 billion inventory write-down after the Trump administration abruptly banned sales of the H20 chip, its most powerful product then available to Chinese buyers [8].

The episode distils the central paradox of the hardware dimension of AI competition. Washington holds the technological keys but cannot fully control the geopolitical conditions under which they are used. Beijing faces genuine hardware constraints, yet has demonstrated a capacity to adapt and innovate under pressure. On the Western side, private capital is beginning to address at least part of the infrastructure gap. Yet the contrast with China’s top-down coordination remains sharp. Where Beijing can align chip policy, energy infrastructure and industrial strategy through a single chain of command, Washington must rely on market signals, regulatory timelines and investor appetite that move at their own pace. The question of whether that model can match the urgency and the durability of a state-directed competitor is not merely academic. It may be the defining industrial policy question of the decade. 

Conclusion: A Structural Imbalance at the Heart of the Race

The global AI race is not a contest of algorithms. It is a contest of economic structures and political will. The West’s lead in innovation is real but rests on fragile foundations: monetisation remains unresolved, supply chains are being reshored at enormous cost, and private capital is filling gaps that public strategy has left open. China’s position is more constrained today but more deliberately constructed for the long term, adapting around chip embargoes, coordinating energy infrastructure from the top down, and absorbing industrial risks that markets would not.

The race will not be won by the fastest sprinter. It will be won by the actor that can sustain the marathon, and on that measure, the structural contest is far closer than the current innovation gap suggests.

Edited by Maxime Pierre.

References

[1] National Security Commission on Artificial Intelligence (NSCAI). (2021). Final Report, Chapter 9: The Global AI Race. Retrieved from https://reports.nscai.gov/final-report/chapter-9

[2] Brookings Institution. (2025, August 11). Interwoven frontiers: Energy, AI, and US-China competition. Retrieved from https://www.brookings.edu/articles/interwoven-frontiers-energy-ai-and-us-china-competition/

[3] Deutsche Bank. (2025, October 28). Artificial Intelligence – Bubble or Boom? (Perspectives Viewpoint). Retrieved from https://wealth.db.com/content/dam/deutschewealth/insights/investing-insights/asset-class-insight/2025/perspectives-viewpoint-equity-artificial-intelligence-bubble-or-boom.pdf

[4] Financial Times. (2025, November 5).  Nvidia’s Jensen Huang says China ‘will win’ AI race with US. Retrieved from
https://www.ft.com/content/53295276-ba8d-4ec2-b0de-081e73b3ba43

 [5] TSMC. (2025, March 4). TSMC Intends to Expand Its Investment in the United States to US$165 Billion to Power the Future of AI. Retrieved from https://pr.tsmc.com/english/news/3210

[6] Centre for International Governance Innovation (CIGI). (2025, November 5). AI Development and Governance in China amid US-China Competition. Retrieved from https://www.cigionline.org/documents/3593/no.338_He.pdf

[7] TerraPower. (2025, November 12). TerraPower Announces $650 Million Fundraise. Retrieved from https://www.terrapower.com/terrapower-announces-650-million-fundraise

[8] Reuters. (2026, January 8). Nvidia requires full upfront payment for H200 chips in China, sources say. Retrieved from https://www.reuters.com/world/china/nvidia-requires-full-upfront-payment-h200-chips-china-sources-say-2026-01-08/

[9] Pexels. (n.d.). A robotic arm and a chessboard [Image]. Retrieved from https://www.pexels.com/photo/a-robotic-arm-and-a-chessboard-8438868/