The AI Chip Wars: Compute as the New Oil

October 29 2025 NVIDIA becomes the

October 29, 2025. NVIDIA becomes the first company in history to close above $5 trillion in market capitalisation. Five years earlier, the company had been worth about $320 billion — a sixteen-fold increase driven almost entirely by one thing: the demand for the specialised chips that power artificial intelligence. NVIDIA does not manufacture these chips. It designs them. The manufacturing is done by TSMC (Taiwan Semiconductor Manufacturing Company), in Taiwan, using machines made by ASML, in the Netherlands, with memory chips made by SK Hynix and Samsung, in South Korea. The supply chain stretches across four countries on three continents, and every link in it is a geopolitical chokepoint — a single point whose disruption stops everything. The AI revolution runs on silicon, and the silicon runs through a supply chain that is the most strategically important — and the most fragile — in the world.

AI Chip Wars

The geopolitical competition to control the design and manufacture of advanced AI accelerators — principally NVIDIA GPUs, AMD MI-series, Google TPUs, and custom silicon from Cerebras, SambaNova, and Tenstorrent. The AI chip wars became a defining feature of US-China relations after the October 2022 US export controls on advanced GPUs to China.

The Supply Chain: Four Links, Four Countries

To understand the AI chip wars, you have to understand the supply chain that produces AI chips. The supply chain has four critical links, and each link is dominated by a small number of companies in a small number of countries. If any link breaks, the entire chain stops.

The first link is design. AI chips are designed by companies that specialise in chip architecture. The dominant designer is NVIDIA — an American company founded on April 5, 1993 by Jensen Huang (a Taiwanese-American engineer), Chris Malachowsky, and Curtis Priem, and headquartered in Santa Clara, California. NVIDIA originally focused on making GPUs (graphics processing units) — chips originally designed for rendering video-game graphics, which turned out to be remarkably good at the matrix multiplication that neural networks require. By the 2010s, NVIDIA had shifted from gaming to AI. Today, NVIDIA controls an estimated 80–95% of the AI accelerator chip market (figures vary by methodology; data-center GPU share is most often cited around 88–92%). Its main competitors are AMD, another American company, and Intel, a third American company that has struggled to compete in AI. The design layer is dominated by the United States.

The second link is manufacturing. Chip designs are just blueprints. To turn a blueprint into a physical chip, you need a fab (fabrication plant) — a semiconductor manufacturing factory that can print the design onto silicon wafers with extraordinary precision. Building a new fab costs $10–20 billion and takes three to five years. The dominant manufacturer is TSMC, based in Hsinchu, Taiwan. TSMC was founded in 1987 by Morris Chang, a Taiwanese-American engineer who had worked at Texas Instruments for a quarter-century before returning to Taiwan to start TSMC. Chang pioneered the pure-play foundry model — a chip-manufacturing business that only manufactures chips designed by others, with no in-house design business of its own.

Pure-play foundry

The “pure-play foundry” model was TSMC’s foundational innovation. Before TSMC, chip companies both designed and manufactured their own chips (the “integrated device manufacturer” or IDM model, exemplified by Intel). TSMC’s insight was that this dual role created a conflict of interest: a foundry’s design customers worried that the foundry would steal their designs. By manufacturing only — never designing its own competing chips — TSMC removed that conflict and won the trust of fabless chip designers like NVIDIA, Apple, AMD, and Qualcomm. The pure-play foundry model is now industry standard for advanced chipmaking.

Today, TSMC manufactures about 92% of the world’s most advanced logic chips — the chips (at the smallest process nodes, under ~10 nanometres) that power AI systems, smartphones, and computers. NVIDIA, Apple, AMD, Qualcomm, and many others all rely on TSMC to manufacture their chips. The manufacturing layer is dominated by Taiwan.

The third link is equipment. To manufacture advanced chips, TSMC needs machines that can print patterns on silicon wafers at the nanometre scale. The most advanced of these machines use a technology called EUV (extreme ultraviolet) lithography — a chip-manufacturing technology that uses light with a wavelength of 13.5 nanometres to etch patterns on silicon. EUV is the most complex manufacturing technology ever developed by humans. It took decades of research and billions of dollars of investment to create. And it is made by exactly one company: ASML, headquartered in Veldhoven, the Netherlands. ASML is the sole supplier of EUV lithography machines in the world. There is no alternative. If ASML stops selling EUV machines, TSMC cannot manufacture the most advanced chips, and the AI industry stops. The equipment layer is dominated by the Netherlands.

The fourth link is memory. AI chips do not work alone. They need high-speed memory to store the data they are processing. The memory used in AI accelerators is called HBM (high-bandwidth memory) — a specialised memory type that stacks multiple memory chips vertically, connected by microscopic wires called through-silicon vias (TSVs), to provide much higher bandwidth than conventional memory. The dominant producer of HBM is SK Hynix, a South Korean company founded in 1983. SK Hynix supplies HBM directly to NVIDIA, and its dominance is so great that it overtook Samsung Electronics in quarterly operating profit in Q4 2024, in annual operating profit for FY2025, and in market capitalisation on June 22, 2026 — ending Samsung’s roughly 25-year run as South Korea’s most valuable company. The memory layer is dominated by South Korea.

These four links — design (US), manufacturing (Taiwan), equipment (Netherlands), and memory (South Korea) — form the supply chain that produces the chips that power AI. Each link is a chokepoint. If Taiwan is blockaded, TSMC stops manufacturing. If the Netherlands restricts ASML’s exports, TSMC cannot get the machines it needs. If South Korea has a problem, the memory supply dries up. The entire AI industry depends on a supply chain that is concentrated in four countries, any one of which could be disrupted by war, natural disaster, or political conflict.

NVIDIA: From Gaming to Trillion-Dollar AI Empire

NVIDIA’s rise is one of the most remarkable business stories of the twenty-first century. The company was founded in 1993 by Jensen Huang, Chris Malachowsky, and Curtis Priem. Its original business was making graphics chips for video games. For most of its first two decades, NVIDIA was a successful but not extraordinary company — a leader in the gaming GPU market, but not a player in the broader technology industry.

The turning point came in the 2010s, when researchers discovered that GPUs — the chips NVIDIA had designed for rendering video-game graphics — were remarkably good at training neural networks. Neural networks require massive amounts of matrix multiplication (multiplying large grids of numbers together), which is exactly the kind of mathematics that GPUs are designed to do. As deep learning took off, researchers began buying NVIDIA GPUs by the thousands. NVIDIA noticed, and it began to design chips specifically for AI workloads.

The A100 → H100 → H200 → Blackwell B200 progression

NVIDIA’s successive generations of AI-accelerator chips have roughly doubled performance per generation. The A100, launched May 2020 with 54 billion transistors, became the foundation of the large-language-model era (though GPT-3 itself was trained on the earlier V100 architecture). The H100, announced March 2022 and shipped late 2022 with 80 billion transistors, was designed specifically for the transformer architecture and powered the ChatGPT boom — when ChatGPT launched in November 2022, demand for H100s exploded. The H200 followed in 2023, and the Blackwell B200, announced at GTC in March 2024 with 208 billion transistors, is the current frontier. Each generation was in such demand that NVIDIA sold every chip it could make.

The financial results were staggering. NVIDIA’s revenue went from about $27 billion in fiscal year 2023 (ended January 2023) to about $61 billion in fiscal year 2024 (ended January 2024) — a 126 percent year-over-year increase. Its market capitalisation went from about $320 billion in 2020 to $5 trillion on October 29, 2025 — the first company in history to close above that milestone. The increase was driven almost entirely by the AI boom. NVIDIA had become, by some measures, the most valuable company in the world, and it had done so by selling the picks and shovels of the AI gold rush.

NVIDIA’s dominance is not just about chip design. It is also about software. NVIDIA’s CUDA software platform — in development for nearly two decades — allows developers to program its GPUs and has become the standard programming environment for AI research. CUDA is the source of NVIDIA’s moat — a durable competitive advantage that protects a company from competitors: even if a rival builds a chip that matches NVIDIA’s hardware performance, developers need to rewrite their software to use it, and most are reluctant to do so. This software moat is one of the reasons NVIDIA’s market share has remained so high despite the entry of competitors like AMD and Intel.

Taiwan: The Geopolitical Fault Line

The concentration of chip manufacturing in Taiwan is the single most concerning geopolitical vulnerability in the AI supply chain. TSMC’s most advanced fabs are in Hsinchu and Tainan, both in Taiwan. If Taiwan were to be disrupted — by a Chinese invasion, a blockade, a natural disaster, or a political crisis — the global supply of advanced chips would be severely constrained, and the AI industry would face an existential crisis.

The vulnerability is well understood by policymakers. The RAND Corporation estimated in a 2024 report that TSMC produces about 92% of the world’s most advanced logic chips (those under ~10nm). A Chinese invasion or blockade of Taiwan would cut off this supply, with consequences that would extend far beyond AI — affecting smartphones, computers, cars, military systems, and virtually every industry that uses advanced electronics.

The Taiwan vulnerability has led to some extraordinary policy discussions. Some US policymakers have reportedly discussed the “broken nest” scenario — the idea that, in the event of a Chinese invasion, the US should ensure that TSMC’s facilities are destroyed, so that China cannot capture them and use them to gain control of the global chip supply. This idea is controversial and has not been officially adopted as policy, but the fact that it is discussed at all is a measure of how strategically important TSMC has become.

The vulnerability has also led to efforts to diversify chip manufacturing away from Taiwan. TSMC is building fabs in Arizona (supported by the CHIPS Act), in Japan, and in Germany. But these efforts are slow — a new fab takes three to five years to build — and they will not eliminate the dependence on Taiwan in the near term. The most advanced chips will continue to be made in Taiwan for years to come.

US Export Controls: The Chip Embargo

The United States has used its position in the chip supply chain — as the dominant designer and as the home of many of the companies that make chip design software — to restrict China’s access to advanced AI chips. The export controls are the most consequential technology embargo of the twenty-first century, and they are the central instrument of the US-China technology competition.

The first major export controls were imposed on October 7, 2022 by the BIS (Bureau of Industry and Security) — the US Commerce Department agency that administers export controls. The controls restricted the sale of advanced AI chips to China, citing national security concerns. The specific threshold was based on chip performance: chips that exceeded certain interconnect bandwidth and compute performance metrics were banned.

NVIDIA responded by creating modified versions of its chips that fell below the thresholds. The A800 and H800 were downgraded versions of the A100 and H100, with reduced interconnect bandwidth, that were legal to sell to China under the October 2022 rules. NVIDIA sold them in large quantities. Chinese AI companies, including DeepSeek (the Chinese AI lab founded by Liang Wenfeng — see B57 for the full profile), stockpiled H800 chips while they were still available.

On October 17, 2023, the Biden administration updated the controls (effective November 17, 2023), closing the A800/H800 loophole. The new rules banned the A800 and H800 and introduced a broader definition of the chips that were restricted. NVIDIA responded again, creating the H20, L20, and L2 — further downgraded chips that complied with the new rules. But the H20 was significantly less capable than the H100, and it was not clear that it would be competitive enough for Chinese AI companies to use.

The effectiveness of the export controls is debated. Epoch AI, an independent research organisation that analyses AI trends, estimates that the controls have given the United States a hardware lead of around four years over China for training frontier AI models (but essentially no lead in serving those models to users). This is a significant lead, but it may be shrinking. The DeepSeek-R1 release in January 2025 demonstrated that a Chinese AI company could build a frontier model using H800 chips that it had acquired before the October 2023 ban. The release triggered a debate about whether the controls were working, and it led to calls for tighter restrictions.

Some analysts argue that the controls are counterproductive. They argue that the controls hurt NVIDIA’s revenue (NVIDIA has lost billions of dollars in Chinese sales), that they incentivise China to develop its own chip industry (which is exactly what has happened with Huawei’s Ascend chips), and that they accelerate the very self-sufficiency that the controls are designed to prevent. Others argue that the controls are working as intended — they have slowed China’s AI development, and they have preserved a US lead that would otherwise have been eroded.

The truth is probably that the controls have had both effects. They have slowed China’s access to the most advanced chips, but they have also incentivised China to develop its own alternatives. Whether the net effect is positive or negative for US interests depends on how quickly China can develop competitive domestic chips, and on whether the US lead is large enough to matter.

The CHIPS Act and Stargate: America’s Industrial Response

The United States has not relied on export controls alone. It has also invested in its own chip manufacturing capacity, through two major initiatives: the CHIPS and Science Act and the Stargate Project.

The CHIPS and Science Act was signed into law by President Biden on August 9, 2022. The Act provides about $52 billion in subsidies for semiconductor manufacturing in the United States, along with a 25 percent investment tax credit for semiconductor manufacturing equipment. The Act also authorises about $200 billion for scientific research and development. The total authorised spending is about $280 billion over ten years.

The CHIPS Act was a response to the recognition that the United States had become dangerously dependent on foreign chip manufacturing. In 1990, the United States accounted for about 37% of global semiconductor manufacturing capacity. By 2022, that figure had fallen to about 12%. The CHIPS Act was designed to reverse this decline by subsidising the construction of chip fabs in the United States.

The results have been mixed. TSMC is building a fab in Arizona, supported by CHIPS Act funding, but the project has been delayed and has faced labour disputes. Intel — the largest single recipient of CHIPS Act funding — ousted its CEO Pat Gelsinger in December 2024 amid struggles with its foundry business. Samsung is building a fab in Texas. The CHIPS Act has catalysed significant investment — over $240 billion in private sector semiconductor investment has been announced since the law’s passage — but it will take years for the new fabs to come online, and it is not clear that they will be competitive with TSMC’s most advanced facilities.

The Stargate Project, announced by President Trump on January 21, 2025, is a different kind of initiative. Stargate is not about manufacturing chips — it is about building the data centres that use them. The project is a joint venture between OpenAI, Oracle, SoftBank, and MGX (an investment company from the United Arab Emirates). The total planned investment is $500 billion over four years, with an initial $100 billion tranche in 2025. The project plans to build AI data centres with a total capacity of about 10 gigawatts — comparable to the output of several nuclear power plants.

Stargate is roughly ten times the size of the CHIPS Act, and it represents an unprecedented private investment in AI infrastructure. The project was announced on Trump’s first full day in office, and it signalled that the new administration intended to prioritise AI development — through infrastructure investment rather than through the regulatory approach of the Biden administration.

Both the CHIPS Act and Stargate reflect a recognition that the AI race is, in significant part, an infrastructure race. The country that has the most chips, the most data centres, and the most compute — the processing power available to run AI workloads, measured in FLOPs (floating-point operations per second) or in chip-hours — will have a significant advantage in developing AI. The United States is investing heavily in that infrastructure, both through government subsidies (CHIPS Act) and through private investment (Stargate).

China’s Response: Huawei and SMIC

China has not been passive in the face of US export controls. It has been developing its own chip industry, and the most visible results of this effort are Huawei’s Ascend AI chips and SMIC’s (Semiconductor Manufacturing International Corporation, founded April 3, 2000) manufacturing capabilities.

Huawei, the Chinese telecommunications giant, designs AI chips through its subsidiary HiSilicon. The Ascend series — including the 910B and 910C — is designed to be a domestic alternative to NVIDIA’s chips. The Ascend 910C entered mass production in the first quarter of 2025, and Huawei plans to produce about 600,000 Ascend 910C chips in 2026, roughly double the 2025 level. The Wall Street Journal has reported that Huawei is poised to ship 800,000 Ascend 910B/C chips in total.

The Ascend chips are manufactured by SMIC, China’s largest foundry. SMIC has been subject to US export controls on EUV lithography equipment, which means it cannot buy the ASML machines that TSMC uses to manufacture the most advanced chips. Despite this, SMIC achieved a significant milestone in September 2023, when a teardown of the Huawei Mate 60 Pro smartphone revealed a 7-nanometre chip — the Kirin 9000s — manufactured by SMIC. This was a surprise to Western analysts, who had believed that SMIC could not produce 7nm chips without EUV. SMIC achieved it using older DUV (deep ultraviolet) lithography — the previous-generation technology, using 193nm light, that is not subject to the same export controls — combined with a technique called multi-patterning.

What is multi-patterning — and why is it a workaround?

Lithography works by shining light through a patterned mask onto a silicon wafer, etching tiny features. EUV’s 13.5nm wavelength lets you print fine features in a single exposure. DUV’s 193nm wavelength is too coarse for 7nm features in one pass. Multi-patterning is the workaround: you print the pattern in multiple exposures, each shifted slightly, building up finer features than a single pass could achieve. The cost is complexity, time, and yield (the percentage of chips on a wafer that work correctly) — multi-patterning dramatically increases the chance that something goes wrong on each wafer. SMIC’s 7nm DUV production is believed to have low yield, which is part of why it is not yet cost-competitive with TSMC’s EUV-based 7nm.

The 7nm achievement was significant, but it came with caveats. The yield is believed to be low, and the cost is believed to be high. SMIC is not yet competitive with TSMC on either quality or cost. But the trajectory is clear: China is investing heavily in its domestic chip industry, and it is making progress, even under the constraints of US export controls.

The question is whether China can close the gap fast enough to matter. Epoch AI estimates that the US hardware lead is about four years. If China can develop competitive chips within that timeframe, the export controls will have bought time but not a permanent advantage. If China cannot, the export controls will have preserved a significant US lead. The answer will depend on how quickly SMIC can improve its manufacturing capabilities, how quickly Huawei can improve its chip designs, and whether the US tightens its controls further.

The Memory Layer: SK Hynix and the HBM Squeeze

The least-discussed but increasingly important link in the AI chip supply chain is memory. AI accelerators need HBM to function, and HBM is made by a small number of companies, led by SK Hynix of South Korea.

HBM’s three-dimensional structure — multiple memory chips stacked vertically and connected by through-silicon vias — allows it to provide much higher bandwidth than conventional memory, which is essential for AI workloads that need to move large amounts of data quickly. SK Hynix is the dominant producer of HBM, particularly the HBM3E generation that is used in NVIDIA’s H100 and H200 chips. SK Hynix supplies HBM directly to NVIDIA, and its dominance is so great that it overtook Samsung Electronics in quarterly operating profit in Q4 2024, in annual operating profit for FY2025 (47.2 trillion won vs Samsung’s 43.6 trillion won, reported January 2026), and in market capitalisation on June 22, 2026. This was a remarkable shift: Samsung had been South Korea’s most valuable company for roughly 25 years, and SK Hynix’s rise was driven almost entirely by the AI boom.

The HBM supply chain is under strain. Samsung and SK Hynix have both warned that AI-driven memory shortages could persist until 2027 and beyond, with customers pre-booking capacity years in advance. The shortage is a reminder that the AI chip supply chain is not just about the chips themselves — it is about all the components that go into them, and any bottleneck in any component can slow the entire industry.

What the Chip Wars Mean

The AI chip wars are, in essence, a competition for control of the most important strategic resource of the twenty-first century. Compute — the ability to process information at scale — is to the AI age what oil was to the industrial age: the essential input that powers the economy, and the resource that nations compete to control.

The comparison to oil is not perfect. Oil is a natural resource that can be found in specific geological formations; compute is a manufactured product that can, in principle, be made anywhere. But the geopolitical dynamics are similar. Just as the oil supply chain had chokepoints — the Strait of Hormuz, the Suez Canal, the refineries — the compute supply chain has chokepoints: TSMC in Taiwan, ASML in the Netherlands, SK Hynix in South Korea. Just as oil-producing nations had geopolitical leverage, chip-producing nations have geopolitical leverage. And just as the United States used its military and economic power to secure oil supplies, it is now using its export controls and industrial policy to secure compute supplies.

The chip wars will not be resolved quickly. The supply chain is too complex, the investments are too large, and the geopolitical tensions are too deep. The United States will continue to try to restrict China’s access to advanced chips, and China will continue to try to develop its own alternatives. Taiwan will remain the most dangerous flashpoint, and the Netherlands and South Korea will remain critical, if less visible, players.

The outcome of the chip wars will shape the development of AI for decades. If the United States maintains its lead, it will continue to dominate the AI industry, and the AI systems of the future will be built primarily on American-designed chips, manufactured in American-allied countries. If China closes the gap, the AI industry will become more bifurcated, with separate ecosystems built on separate chip technologies. And if the supply chain is disrupted — by war, by political conflict, or by natural disaster — the AI industry could face a crisis that would make the DeepSeek-R1 market panic look mild by comparison.

The chip wars are not just about technology. They are about power, sovereignty, and the future of the global economy. The chips that power AI are the most strategically important objects in the world, and the competition to control them is the defining geopolitical struggle of our time.