Silicon, Not Software, Will Decide the AI Race

By Antonin Bergeaud and Robin Rivaton, Project Syndicate | Jul 31, 2026

PARIS—Did the AI boom begin when it did simply as a result of scientific breakthroughs? At first glance, it would seem so. The transformer architecture—the foundation of today’s large language models—was introduced in 2017. By 2020, researchers had established the basic logic of scaling: more computation and data produce predictably better models. Then, in late 2022, the public launch of ChatGPT revealed the mass-market appeal of these technologies, setting off a global race for AI leadership.

Scientific advances alone, however, do not trigger investment booms. Rather, AI delivered a market-wide shock to the semiconductor and computing ecosystem. The transformer opened a promising technological path; ChatGPT created the mass market that justified large-scale investment; and systems such as Codex extended demand into applications like software development.

What followed was an unprecedented wave of investment, fueled by a combination of algorithmic innovation, hardware advances, and the emergence of a mass market for machine intelligence. For all the hype surrounding frontier models, the future of AI depends on processors, networking equipment, cooling systems, electricity, cloud platforms, and the organizational capacity to coordinate them all.

Economists have long recognized that market size shapes the trajectory of innovation. In a seminal 2004 paper, Nobel laureate economist Daron Acemoglu and Joshua Linn showed that larger potential markets—driven by demographic change—stimulated pharmaceutical innovation.

The same logic applies to AI, albeit with an important qualification. In pharmaceuticals, the demand shock came from outside the industry as populations aged. With AI, by contrast, successive breakthroughs have created demand for even greater computing power. ChatGPT created the market that now finances the hardware on which its successors will be trained, and the prospect of a vast market for machine intelligence has made innovation far more profitable. This is true not only for semiconductors, software, networking, and data-center infrastructure, but also in sectors like energy and construction, which have long struggled to raise productivity.

Yet technologies do not improve at the same pace indefinitely. As they mature, they face diminishing returns, making further advances progressively harder. This helps explain why technological progress often slows after an initial burst of rapid acceleration. Occasionally, however, a new technological paradigm emerges that resets the “innovation clock”—a mechanism described in recent work by Nobel laureate economist Philippe Aghion and his co-authors.

AI is such a reset, but with a distinctive twist. Rather than simply inaugurating a new technological paradigm, it has restarted the semiconductor innovation cycle by creating a new and seemingly insatiable demand for computing power.

The Limits of Moore’s Law

The AI boom is therefore best understood not as an isolated technological event but as the latest chapter in a decades-long cycle of innovation and investment. Since the invention of the transistor at Bell Labs in 1947, the tech industry has been transformed by successive waves of change: mainframes, personal computers, the internet, mobile devices, cloud computing, and now AI.

These waves are often portrayed as the inevitable consequence of Moore’s law—the observation that the number of transistors on an integrated circuit doubles every two years. But this interpretation overlooks a powerful feedback loop: Moore’s law operated as a virtuous cycle in which technological progress and market expansion reinforced one another. Cheaper compute enabled new applications, expanding the market for computing and making continued investment in chips, manufacturing tools, memory, software, and digital infrastructure profitable.

Unlike earlier computing waves, AI places enormous strain on physical infrastructure, from parallel processing and memory bandwidth to high-speed interconnects and data-center capacity. Compounding the problem, its emergence came after the breakdown of the energy bargain that sustained Moore’s law.

For decades, Moore’s law was complemented by the Dennard scaling principle: as transistors became smaller, they could operate at lower voltage. This kept thermal loads under control, allowing each new generation of chips to deliver more computing power without requiring commensurate investments in electricity, cooling, and physical infrastructure.

That bargain began to unravel in the mid-2000s. Although transistors continued to shrink, voltage could no longer drop as quickly, preventing processors from running faster without overheating. To sustain progress, innovation expanded beyond the transistor itself to encompass the entire computing stack: multicore processors, graphics processing units, specialized accelerators, high-bandwidth memory (HBM), advanced chip packaging, hyperscale cloud infrastructure, and high-speed networking.

Generative AI is the first mass-market application to make full use of this post-Dennard architecture. A chatbot may appear weightless to users, but behind its interface lies an enormous physical apparatus. Training frontier models requires vast amounts of parallel computing, high-speed data transfer, dense clusters of specialized chips, industrial-scale cooling, and abundant, reliable electricity.

The AI rupture, then, is not merely about software; it is also about physical capital, energy, and industrial capacity. According to Stanford’s 2025 AI Index, the amount of computing power used to train advanced AI models has been doubling roughly every five months. Epoch AI estimates that computing power devoted to training frontier models has increased nearly fivefold each year since 2020.

But compute is only one of AI’s essential inputs. The other is the immense body of digital information created by earlier computing waves. The internet transformed decades of human text, code, images, audio, and video into a vast digital archive, while the collapse in storage costs—another consequence of relentless advances in semiconductor manufacturing—made it economical to preserve virtually all of it. Pre-training is, in effect, the process of compressing all that data into a statistical model.

The result is a self-reinforcing dynamic reminiscent of the old Moore’s law feedback loop: better models increase the value of additional computing power; higher expected returns justify greater investment in accelerators, memory, cloud capacity, and electricity; and those investments expand frontier AI models’ capabilities by easing the bottlenecks that constrain them, enabling new applications that further boost demand for compute.

Against this backdrop, it is clear why today’s dominant AI firms are those that coordinate the complementary investments required to turn scientific advances into economic value. Nvidia’s competitive advantage, for example, rests on the decades-long development of CUDA—the platform through which developers use its chips—together with its software libraries, technical expertise, advanced packaging, and supply-chain coordination.

TSMC’s dominant position likewise reflects the value of its manufacturing expertise. Amazon, Microsoft, and Google, by contrast, convert dispersed demand for compute into vast cloud platforms and infrastructure. Their competitive advantage rests not on any single technology, but on their ability to coordinate capital, hardware, software, and electricity at an extraordinary scale.

Industrial Policy for the AI Age

The clearest sign that silicon remains at the heart of today’s technological revolution is that every major power now treats control of the semiconductor supply chain as a strategic asset. AI may look like a software revolution, but its constraints remain physical: processors, memory, manufacturing equipment, packaging capacity, and electricity. That has transformed chipmaking from a private industry into a geopolitical battleground.

The United States is a case in point. For several decades, dispensing with fabs allowed US companies to specialize in chip design while outsourcing manufacturing to overseas producers. This division of labor helped the US retain its technological lead even as advanced production became concentrated in Taiwan.

But rising geopolitical tensions have led policymakers to rediscover the strategic importance of manufacturing capacity. The 2022 US CHIPS and Science Act marked a sharp break with decades of fabless complacency, committing roughly $52 billion in public funding and catalyzing more than $450 billion in announced private investment in domestic semiconductor production.

The effects are already visible. TSMC has recently increased its planned investment in its Arizona megaproject to $265 billion, with Apple reserving more than half of the company’s initial two-nanometer (2nm) production capacity. Intel’s Fab 52 plant in Arizona is already producing 18A chips, which Microsoft reportedly plans to use for its next-generation Maia accelerator.

Meanwhile, Taiwan’s decades-long gamble on TSMC continues to yield extraordinary dividends. The company is effectively running its 3nm production lines at full capacity, with demand for its advanced-node chips reportedly exceeding supply by roughly three to one. Taiwan’s unrivaled manufacturing capabilities have made its so-called silicon shield a strategic reality: the island’s security is now a matter of vital strategic interest to the US and its allies.

South Korea’s experience highlights another AI chokepoint: memory. South Korean policymakers understood long ago what their European counterparts still struggle to recognize: memory remains a strategic asset even when it appears readily available on global markets.

With that in mind, South Korea launched an industrial-policy push that, in 1983, made it the third country in the world, after the US and Japan, to produce 64K DRAM chips. Samsung and SK Hynix built on that early lead to form, together with the US-based Micron, an oligopoly that controls more than 90% of the global DRAM market. SK Hynix has also emerged as the leading supplier of HBM, a critical input for training frontier AI models.

China offers the most striking demonstration of what ambitious industrial policy can achieve. Over the past few decades, it has climbed the semiconductor value chain from the bottom up, beginning with outsourced semiconductor assembly and test (OSAT), which comprises the final stages of chip production after fabrication. Western firms largely dismissed this segment as a low-margin business, but China treated it as a strategic entry point.

That bet has paid off. Companies like JCET, Tongfu, and Huatian have expanded from contract packaging into chiplet integration and 2.5D packaging, which allows processors and memory to communicate at much higher speeds. JCET is now the world’s third-largest OSAT firm, with annual revenues of $5 billion.

From packaging, China moved into memory. CXMT’s DDR5 chips—the latest generation of DRAM—reportedly reach 8,000 mega-transfers per second, enabling Chinese producers to challenge the Samsung-SK Hynix-Micron oligopoly.

The next step was fabrication. SMIC, China’s leading foundry, is now producing Huawei’s Ascend AI accelerators using deep-ultraviolet multi-patterning, a technique that extends the capabilities of older lithography systems through a series of additional manufacturing steps.

By late 2025, China had reportedly reached the industry’s final frontier: lithography. After much trial and error, a Chinese consortium developed a prototype of an extreme-ultraviolet (EUV) machine—the most advanced chipmaking technology and the hardest to replicate. Officials hope to bring the technology into production by 2028, although some people close to the project consider 2030 more realistic.

The numbers speak for themselves: China now produces roughly 35% of its semiconductor equipment domestically, with local firms supplying over 40% of the tools used in key manufacturing processes such as etching and thin-film deposition. Although the country still lags in lithography, design software, and photoresists, it has built domestic capabilities across nearly every stage of the semiconductor value chain.

These investments would make little sense if semiconductors were a mature industry on the verge of being displaced. Instead, the rise of AI has made semiconductors a central battleground in the contest for technological and military supremacy.

The Next AI Bottleneck

As the wall of separation between chip designers and chipmakers that defined the 1990s and 2000s crumbles, the fabless model is giving way to a new form of vertical integration. From Google’s TPU and AWS’s Trainium to Microsoft’s Maia, the hyperscalers that dominate cloud computing are now designing their own AI accelerators.

But these companies continue to rely on TSMC to fabricate their leading custom accelerators. Vertical reintegration, it turns out, stops at the foundry door. With the strategic bottleneck shifting from chip design to manufacturing and advanced packaging, governments around the world are spending tens of billions of dollars to expand domestic chip manufacturing.

The implications extend far beyond semiconductors. The International Energy Agency projects that data-center electricity consumption will more than double by 2030, to roughly 945 terawatt-hours—just under 3% of global electricity use. Contrary to popular belief, the main challenge is not building more data centers but securing the land, equipment, grid capacity, and permits required to operate them.

The lesson for policymakers is straightforward. Frontier models matter, but they are only one layer of the AI stack. Technological leadership also requires computing power, electricity, cloud infrastructure, advanced semiconductors, skilled engineers, and markets large enough to justify sustained investment.

These capabilities are unevenly distributed across the world’s major economies. The US combines a vast continental market, deep capital markets, world-leading cloud providers, and firms willing to deploy new technologies at scale. China’s advantages, by contrast, lie in its ability to achieve self-sufficiency across the semiconductor value chain, from packaging to lithography. Europe, for its part, has struggled to turn scientific excellence into industrial leadership, treating innovation as a research problem and deployment as an afterthought.

In industries where innovation depends on scale, physical capacity matters as much as scientific ingenuity. AI is no exception, as competition shifts from developing better models to inference—that is, running them efficiently once they have been trained. While training remains essential, meeting growing demand at low cost may be the greater challenge.

The cost of running an AI model ultimately comes down to its physical inputs: semiconductors, electricity, cooling, and the capital invested in data centers. Today, industry estimates put servers at around 60% of the annualized cost of operating a large data center, while energy accounts for 7%. But as hardware becomes cheaper and more standardized, electricity will account for a growing share of the cost, making access to abundant, reliable power the factor that separates winners from losers.

The chip race is therefore also an energy race. After all, a less efficient processor is not a decisive disadvantage if electricity is abundant and cheap. That, in turn, reshapes the geography of AI: because data centers can be built almost anywhere, reliable, low-cost electricity matters more than proximity to cities.

In this regard, China’s decade-long push to expand and modernize its electricity system could prove as consequential as the South Korean and Taiwanese bets on semiconductors a generation ago. But sustaining an AI economy requires both institutional reform and a profound physical transformation.

On the institutional front, China is building a more integrated national electricity market. Every province is required to operate a wholesale electricity market, while the share of electricity traded across provincial borders has risen from 18% in 2017 to roughly 24% in 2024. The physical challenge is even greater: per capita electricity demand reached 7.5 megawatt-hours in 2025—almost double the global average and up from barely one MWh in 2000. A vast build-out of generation and transmission infrastructure will be required.

Whether China succeeds or not, the nature of technological competition has fundamentally changed. For decades, it was viewed as a race to push the frontiers of science and semiconductor design. Today, it is a race to build the physical infrastructure that makes scientific breakthroughs commercially viable. Now that the silicon clock has been reset, the next innovation cycle will be measured as much in watts as in nanometers.

Antonin Bergeaud is Professor of Economics at HEC Paris.

Robin Rivaton is CEO of Stonal, a European technology company, an AI sherpa to the French business confederation MEDEF, and an affiliate of the Paris-based think tank Fondapol. He is the author of eight books.

Copyright Project Syndicate