Inside the AI machine: How the tech race binds nations into dependence

ChatGPT and Claude may be the face of AI, but behind the screens lies a complex ecosystem of chips, data, energy and humans.Across the Asia Pacific, control over those layers is reshaping power balances, binding allies and rivals into intricate forms of dependence.This is an artificial intelligence (AI) chip.*Chips like these power the devices and platforms we use every day: from ChatGPT and Google recommendations, to parking assistance and smartphone cameras.The most advanced ones contain hundreds of billions of transistors: microscopic switches performing trillions of calculations a second, making them some of the most complex objects ever made.With parts only a few atoms thick, meaning "a speck of dust" can ruin hundreds of thousands of dollars' worth of product, only a handful of companies can produce chips like this.Their sudden importance has transformed niche technology manufacturing businesses into matters of foreign policy and national security. AI applications like ChatGPT depend on the physical system beneath them.But chips are just one piece of a larger machine, without which, they are effectively useless.AI is often portrayed through Silicon Valley and soaring valuations, but these are just the surface layers of what insiders and industry leaders like NVIDIA CEO Jensen Huang have popularly described as "AI's layered cake": an international ecosystem of chips, data, energy, models and applications (or some variation of this).Each layer is dependent on the others: models need data and compute power; computing requires chips; chips operate in data centres; and all of it requires energy. But here’s the twist: no country controls every layer.In fact, critical capabilities are so globally dispersed yet concentrated in only a few places, that geopolitical allies and rivals alike are often forced to depend on one another: a dynamic some say has prevented any one country from dominating AI, while also raising the cost of conflict. The AI ecosystem is made up of: chips, data, energy, models and applications."The user interface — ChatGPT, Claude — is really just the tip of the iceberg," Robyn Klingler-Vidra, an AI political economy expert who authored a new report on key Asian players, says."There are so many enormously valuable layers of the tech stack, spread across 'niche superpowers' in the Asia Pacific, and companies playing with them, that AI underscores something people may not intuitively see: why governments are so concerned about deep trade and critical minerals."As part of a wider year-long ABC project, we spoke to people across the AI landscape: from everyday users to past and present representatives of major global companies and emerging players — SK Hynix, TSMC, OpenAI, Google, 01.AI and Microsoft — with many others agreeing to speak on background.Pieced together, their insights tell a story much more nuanced than chatbots and stock prices: a tangled system of power and dependence shaped by resources, diplomacy, and remarkable accidents — and with consequences for Australia.Underneath the AI hood: Laying it on thinTo understand the AI race, we first need to understand how the physical system works and what countries are competing over.There is no universally agreed version of the so-called AI “stack” — some call it a chain, cake, or ecosystem — but for simplicity, think of it as chips, data, energy and software, with the first three being largely physical.At the heart of the physical stack are chips, whose importance is difficult to overstate: when you prompt Google or ask AI a question, it’s chips doing the computing behind the scenes. NVIDIA's Blackwell AI processor. The large central section contains the GPU — the chip designed to perform the huge numbers of calculations required for AI.(Supplied: NVIDIA) NVIDIA's GB200 Grace Blackwell Superchip combines two Blackwell AI GPUs with a Grace CPU on a single computing board.(Supplied: NVIDIA) NVIDIA DGX SuperPOD infrastructure built from multiple GB200 systems. Large AI models can require many racks of processors connected through extremely fast networks.(Supplied: NVIDIA)"Put simply, without exponential advances in AI chips, we would not have today's AI capabilities," Chris Miller, author of the 2022 book Chip War, says.While US company NVIDIA designs many of the world’s leading AI chips, it does not manufacture them.That work falls largely to companies such as the Taiwan Semiconductor Manufacturing Company (TSMC), which makes the majority of the world’s most advanced AI processors. But even TSMC cannot operate alone, and relies heavily on specialised technology and materials from other countries to manufacture them.“Advanced chip manufacturing is not one thing: Taiwan may make the world’s most advanced AI chips, but it depends on lithography machines from the Netherlands, ultra-pure materials and chemicals from Japan, memory components from South Korea," Gregory C Allen, AI policy expert and founder of Decision Tree, says. Exterior of TSMC's Fab 2 Semiconductor manufacturing facility in Taiwan.(Supplied: TSMC) A worker examines semiconductor material through a microscope inside TSMC's Fab 3.(Supplied: TSMC) The square object is a photomask, also called a reticle: a highly precise template used during chipmaking to project tiny circuit patterns onto a silicon wafer using light.(Supplied: TSMC)These companies have spent decades, out of the limelight, becoming extraordinarily specialised in different, but often complementary, parts of the chain: relationships that require enormous trust."We do not compete with our customers," a TSMC spokesperson says. "That trust allows us to work closely with the world's leading innovators, aligning our manufacturing roadmaps with their product designs years before production begins."That division of labour has allowed a swathe of companies to become highly specialised, while also concentrating capabilities in a handful of places. ASML's lithography machines use light to print microscopic circuit patterns onto silicon wafers, enabling the production of advanced chips.(Supplied: ASML) SK Hynix HBM4 high-bandwidth memory, used alongside advanced AI processors to move large volumes of data rapidly.(Supplied: SK HYNIX)As AI's capabilities grew more powerful, these supply chains became increasingly political, with governments restricting access to chips and the technology needed to make them: as the US famously did in 2022 when it began restricting advanced chip exports to China, just weeks ahead of ChatGPT's public release."In a parallel universe where the US never imposed these export controls, China would almost certainly already be the leader," Allen says. Those restrictions have since evolved, while new ones have spread up and down the supply chain — from Japan, to the Netherlands, to China — profoundly shaping AI's development.In other words, the geography of today's AI race isn't simply the product of technological competition. It has, in part, been deliberately engineered: a key point we'll return to shortly.But chips are only one part of the AI stack, meaning even if you secure access to them, they actually do very little on their own. xAI's Colossus 1 data centre in Memphis, Tennessee — now in partnership with Anthropic — houses more than 220,000 NVIDIA GPUs similar to the sort pictured above.(Supplied: Steve Jone/Flight by Southwings for SELC)The power of advanced processors is only fully realised when thousands are combined with high-speed memory and networking, assembled into servers and racks, and connected together inside vast data centres, where the calculations behind our search queries, Netflix recommendations, and map directions actually take place.Data centres have existed for decades, but the generative-AI boom has rapidly increased their scale, visibility and strategic importance."I imagine data centre developers were ambitious, but demand has probably outstripped even what they anticipated," says Klingler-Vidra."It's plausible now that data centres were being planned without anyone fully imagining just how data-hungry AI would become."These facilities are effectively giant computers: warehouses filled with processors, networking equipment and cooling systems, working together to train and run increasingly powerful AI models. NVIDIA's GB200 NVL72 is a rack-scale AI computing system designed to sit inside a data centre. Each rack contains 72 Blackwell GPUs and 36 Grace CPUs.(Supplied: NVIDIA) A view of the supercomputer housed inside the European Centre for Medium-Range Weather Forecasts data centre in Italy.(Reuters: Ali Withers)Similar to chips, however, running data systems at scale creates another bottleneck and dependency: energy. Every task takes a sip of power, so data centres require enormous amounts of electricity, along with grid connections, land, cooling and, often, water.Wherever data centres are built, their appetite for power is reshaping energy systems around them, a pressure countries including Singapore and the US were already confronting before the AI boom.“China is far ahead in the scale of its electrical system and available power capacity, and computation requires enormous amounts of electricity,” says Ya-Qin Zhang, chair of AI studies at Tsinghua University and former chairman of Microsoft China, noting that China generates more than twice as much electricity as the US and over a third of the world’s supply. Some proposed AI data centres in Australia are already approaching the electricity demand of entire cities.(ABC News: Carl Saville)In Australia, similar concerns are only starting to get louder. As data-centre development surges, experts warn the sector could require as much electricity as currently required to power millions of households within a decade.Nonetheless, only once these systems are in place — chips, data centres and energy — do we reach the layers most people recognise as AI: models, the trained systems that generate predictions and responses, like OpenAI’s GPT, and applications, the software that people use, like ChatGPT.Companies such as OpenAI and Anthropic operate largely at this visible layer, while depending on the broader international system. Anthropic's AI model Claude is one of the most recognisable frontier models.(Getty: Matteo Della Torre) OpenAI runs ChatGPT.(AP: Richard Drew)"Partners and suppliers are central to AI progress because compute supply is not only scarce, but must be secured and managed years in advance," a spokesperson for OpenAI told the ABC, noting that the company previously relied more heavily on individual compute providers."We now work across a diversified ecosystem, which improves resilience and creates compute certainty, [which] directly affects whether AI can scale."Zooming out, powerful AI depends as much on specialised chemicals and energy as on computer scientists. Where those dependencies become difficult to replace, technology intersects with economic leverage and geopolitical strategy — a reality reflected in documents like the 2025 US AI Action Plan.A redrawn Asia Pacific: Divided by politics, united by necessity China is dominating the AI race in terms of robotics, industrial manufacturing and energy generation.(Reuters: Tingshu Wang) The US is leading the AI race with frontier model development and private investment.(Reuters: Nathan Howard)On raw numbers alone, the AI balance of power appears fairly straightforward.The US dominates in data centres, private investment and leading models, while China leads in electricity generation, robotics, research output and manufacturing capacity.At face value, that suggests a familiar two-power contest; but the concentration of critical capabilities across the stack tells a different story. AI chip manufacturing capabilities are predominantly centred in Taiwan.We've seen that while the US might dominate in terms of software or chip design, it's Taiwan that dominates advanced-chip manufacturing. While manufacturing is based in Taiwan, it heavily depends on chemicals, components and machines sourced from the Netherlands, South Korea and Japan.Taiwan's role, however, depends heavily on specialised equipment, materials and chemicals from the Netherlands, Japan and South Korea.So while the US might lead in design, a shortage of Japanese chemicals or a critical Dutch lithography component can grind the system to a halt.This isn't theory: it's happened before, and continues to happen.In 2019, a stoush between Japan and South Korea choked supplies of chipmaking materials; in 2022, Russia’s invasion of Ukraine disrupted half the world's neon supply, sending prices soaring. Singapore and Australia are heavily investing in becoming strong in data.Closer to home, Singapore and Australia may seek to become regional data hubs, but those ambitions depend on software, chips and enormous amounts of energy.In 2019, Singapore had to pause new data centre development for a few years as demand strained limited power, land and water resources. Development resumed in 2022 under tighter efficiency rules, the same year Singapore began importing electricity from Laos via Thailand and Malaysia.AI power therefore doesn't simply belong to whoever has the most money or data centres: it can sit inside an obscure chemical or trade agreement that nobody notices until it disappears."What happens to the world if Taiwan magically disappears? Economic apocalypse. What happens to the semiconductor world if the Netherlands magically disappears? Less sudden, but still a catastrophe. What happens if you lose the US? Same thing. What happens if you lose Korea? Same thing," Allen explains."There are multiple essential nodes in the supply chain, much like there are multiple essential organs in your body: you need all of them."Those dependencies can also cut across alliances. For example, South Korea, a US ally, supplies critical semiconductor technology while remaining deeply tied to China, even as Washington restricts China's access to some of the most advanced technologies.This geography reflects a mix of deliberate policy, commercial decisions, political necessity and historical accident.For example, observers argue that the 2022 export controls intended to limit Chinese AI development have also encouraged greater efficiency and investment in domestic alternatives, an effect NVIDIA CEO Jensen Huang has described as the controls “backfiring”."It's a double-edged sword," Kai-Fu Lee, founder of Chinese AI firm 01.AI and former head of Google China, says. "When you don't get [top-end NVIDIA chips], you have to figure out ways to make things faster, you're forced to become more efficient."Dynamics like these — present across the supply chain, from rare earth refining and specialised talent to subsea cables and proprietary data — complicate the idea of AI as a simple "race"."China refines more than 90 per cent of the world’s rare earths: that represents a huge source of power in a system predicated on their availability," Klingler-Vidra says. Rare earth minerals are critical to AI hardware, robotics and power systems — and increasingly a source of geopolitical leverage.(Reuters: David Becker) A humanoid robot demonstrates its capabilities during the launch of Galbot's unmanned convenience store in Hong Kong in August 2026.(AFP: Leung Man Hei)"No country controls every layer, so being extremely difficult to replace in one layer may matter more than trying to dominate the entire thing."For smaller economies, that creates opportunities."Instead of trying to do everything and ending up mediocre at everything, find the things you can be world-class at and use them to get everything else you need," Allen says."What Australia needs is strategic relevance, and ideally, strategic indispensability, which it can achieve not by building everything itself, but by controlling something the wider system needs."But strategic importance cuts both ways: becoming indispensable to the stack, also means inheriting the risks and costs.Opportunities, transparency, and the price of getting it wrong A person holds a copy of Il Foglio AI. The AI-produced supplement ran as a month-long experiment in March 2025. (Getty Images: Ivan Romano) Not a day goes by without AI featuring in the news.Nearly four years since ChatGPT's release, public discourse around AI remains dominated by frontier companies — from breakthrough capabilities to warnings about existential danger — with far less attention paid to the systems beneath them.Observers point out that while some of the opacity is intentional, much of it is a consequence of how the industry evolved: companies may have legal reasons to avoid appearing dominant, while governments may fear economic retaliation from trading partners for speaking bluntly about AI.“The self-censorship is a real challenge,” Allen says.“Politically, economically, it's understandable; but if the public don't hear the plain truth, how are they going to know who to vote for or which policies to support?" Anthony Albanese delivers a speech titled AI in Australia's Interests at The University of Sydney.(AAP: Dan Himbrechts) Construction walls at the M3 NEXTDC data centre facility in West Footscray, Melbourne.(ABC News: Steven Viney)In July, Australian Prime Minister Anthony Albanese declared AI a national priority. But some were quick to point out why turning that ambition into policy is much harder than simply attracting investment or building more infrastructure.No country controls everything it needs. For Australia, the question is therefore not how to become technologically self-sufficient, but where to build strength, and which dependencies it is willing to accept.One current AI strategy is to capture domestic value through the jobs, skills and infrastructure built to support foreign AI models, like data centres. AirTrunk — owner of the SYD3 data centre under construction in Western Sydney — was acquired in 2024 by US private equity giant Blackstone and the Canada Pension Plan Investment Board.(Reuters: Hollie Adams)But some experts warn investments of this sort risk turning Australia into an "infrastructure enclave" that consumes domestic land and electricity, without guaranteeing the much-needed intellectual property or economic leverage.They maintain that Australia already has advantages in energy, critical minerals, and industries like mining and agriculture: areas where it could apply AI to capitalise on existing strengths."Australia has a lot of data centres, but many are foreign-owned or funded through outside investment, which might make us strong in that layer, but won't necessarily result in the value or leverage that other areas might provide," Vikas Kumar, who authored a recent AI chain competitiveness framework, says."The opportunity at hand is to combine AI with sectors where Australia already has deep, in-demand knowledge, to create exportable products and services. South Korean tech giant Samsung supplies advanced AI memory globally, including to China, despite Seoul's close US alliance.(Reuters: Kim Hong-Ji) US technology giant Google develops some of the world's leading AI models, while relying on a globally distributed physical supply chain.(AP: Thibault Camus)"For instance, if I'm going to Korea, building that relationship would be my number one motivation: securing key components, while offering applications or energy they need in return."This approach makes AI policy as much about diplomacy as technology and private investment: trade policy, alliances, immigration and energy decisions can shift technological leverage almost as much as breakthroughs in a laboratory.Several Chinese and American AI companies told the ABC on background they were already working towards new partnerships with major Australian industries, including mining; OpenAI also pointed to its December 2025 initiative with CommBank, Coles and Wesfarmers.By the same token, observers were careful to point out that interdependence across the stack does not automatically equate to security, as Europe learnt from its reliance on Russian energy.“It turns out it wasn’t cheap, we just weren’t counting correctly,” Allen says, pointing to similar dependencies like Chinese rare earths, which may be costlier than assumed once “strategic vulnerability” is factored in. In 2025, China's DeepSeek stunned competitors with a powerful new model trained using NVIDIA chips acquired before tighter US export controls. (Reuters: Dado Ruvic)Influence over critical parts of the AI system therefore matters not only to those building powerful AI models and applications, but also to those with the leverage to constrain them.For now, as major powers bet that leadership in AI will directly translate into wider economic and geopolitical power, the rate and scale of investment across the chain — from energy to chip design to frontier software — continues to be extraordinary, reflecting the sheer size of that bet."In the US, we’re spending more on AI every 10 months than the entire Apollo Moon program cost from start to finish over 13 years: it is an astonishing outlay of capital into a technological question mark that could potentially backfire," Allen says."The Industrial Revolution reshaped the global balance of power and almost every aspect of human life: you could literally see its impact from space. My prediction is that, with the level of change occurring across the entire AI stack, this is the kind of moment we’re living through now, hence why it feels so ominous."For Australia, like much of the rest of the world, there is no single technological race to win. There is only a web of choices to negotiate: what to build, what to trade, whom to trust, where to become indispensable, and where dependence becomes dangerous.Read the story in Chinese (中文版) or Indonesian.CreditsReporting & Producing: Steven VineyAudience Engagement & Additional Reporting: Doug DingwallGraphics/Design & Additional Reporting: Jarrod FankhauserAdditional Media: TSMC, SKHYNIX, NVIDIA, Getty, ReutersTranslations: Natasya Salim & Michael Li*Editor's Note: Some visuals and information have been simplified or adapted for clarity. Where information was unavailable or requests not met, we relied on expert analysis and reasonable approximations. The rotating chip visual is an amalgamation of several leading AI chip schematics, rather than a representation of any single model.