The GPU Toll Road: Jensen Huang, G20, and the Centralization of AI Infrastructure

At the G20 summit, Jensen Huang did not unveil a chip. He did not present a benchmark. Instead, he stood in front of the leaders of the world's largest economies and made a very simple, very comfortable, and very convenient request: build more AI infrastructure. More GPUs. More data centers. More compute spread thinly across the globe, like a fresh layer of asphalt for the digital century. It is a story that almost tells itself: compute equals intelligence, intelligence equals growth, growth equals national security. But after years of reading the fine print on technological utopias, I have learned to listen for what is left unsaid. There are no numbers. No efficiency targets. No mention of energy, water, land, or long-term operating costs. And crucially, no mention of who will own and govern the machine that every economy is being asked to plug into. This is not a technology announcement; it is a political intervention. Nvidia controls the overwhelming majority of the AI accelerator market, and its CUDA software ecosystem has become the assembly language of modern machine learning. Jensen Huang's business model depends on a specific belief: that AI progress is, to a first approximation, a function of compute. The more GPUs you buy, the smarter your models become. That belief is not wrong, but it is incomplete. By carrying this message to the G20, Huang is trying to turn a corporate roadmap into national budget priorities. Governments are being asked to fund infrastructure that, in almost every case, will be purchased from one company. The framing is economic growth; the mechanism is rent extraction. This is not a conspiracy; it is the geometry of incentives. The phrase "AI infrastructure" is not politically neutral. If infrastructure means roads, ports, and power lines, then yes, it is a public good. But in the current context, it means a much more specific thing: thousands of high-end accelerators, connected by high-speed networking, managed by proprietary software, and replaced every two years. That is a private good, wearing the costume of a public utility. When a head of state says "we need AI infrastructure," they are implicitly accepting a definition of the future that advantages one company and one narrow technical path. The true infrastructure of AI should include open datasets, data cooperatives, distributed evaluation frameworks, model audit standards, and energy governance. Those elements have no GPU vendor to lobby for them, so they vanish from the conversation. They vanish because they are not on anyone's invoice. Now let's look at the technical claim. The scaling law suggests that model performance improves as a smooth function of training compute. This observation has justified a decade of exponential capital expenditure. It is a real effect, not a trick. But like all empirical regularities, it has boundary conditions. Compute expansion runs into physical limits that no amount of money can remove. Electricity is the most obvious: a single hyperscale data center can draw as much power as a small city, and the grid in most countries is simply not ready. Water for cooling is now a geopolitical issue. Memory bandwidth and interconnect technology are already lagging behind what architects want. At the same time, algorithmic progress has not frozen. Sparse architectures, synthetic data, quantization, and inference-time search can reduce the compute required for a given capability by an order of magnitude. In early 2025, DeepSeek trained a strong model with far fewer H800 GPUs than many Western labs thought possible, using aggressive engineering optimizations instead of bigger clusters. The lesson is not that scaling is dead; the lesson is that scaling is a parameter, not a law. When the world's largest compute vendor meets a government agency looking for a policy program, everyone has an incentive to ignore that nuance. "Scale is the answer" becomes the easiest thing to agree on, precisely because it requires no imagination. The deeper problem is supply chain concentration. Every major AI build-out depends on a small set of chokepoints: advanced packaging from a single manufacturer, HBM memory from a handful of suppliers, and Nvidia's GPU design at the center. If any one link breaks, the global roadmap slips by a year or more. A typhoon in Taiwan, a new export restriction in another region, or a power shortage near a newly built data center can cause synchronized failure across thousands of projects. That is not a design flaw; it is a systematic risk. In decentralized systems, we spend enormous energy designing around single points of failure. We design for partition tolerance, for Byzantine faults, for the honest node that refuses to blink. Yet in the most important new infrastructure on earth, we are doing the opposite: concentrating capabilities into narrower and narrower funnels. This is not infrastructure; it is a toll bridge with a single operator. Last year, I started a small project with a group of AI researchers to design what we called Ethical Oracles — smart contracts that enforce human-centric values in autonomous transactions. The first thing we learned is that you cannot enforce values you cannot see. The second is that you cannot see through a single vendor's dashboard. When we tried to audit a pipeline of compute credits, we ended up auditing a series of aggregated invoices with no visibility into utilization, pricing, or even the source of the electricity powering the workload. The black box is not the model. The black box is the procurement contract. That experience convinced me that the governance gap in AI will not be solved by more compute; it will be solved by more transparency. And transparency is not a product Nvidia can sell. I keep coming back to something I learned in 2017, when I spent three months auditing the whitepapers of forty-two failed ICOs. The post-mortems all looked the same: a grand narrative about decentralized infrastructure, a token sale, and then a slow collapse when the unit economics stopped matching the story. The projects that survived, and the very few that thrived, were the ones that had a direct connection between price and scarcity, between the infrastructure they built and the value they returned. The same test applies to national AI policy. When governments fund infrastructure through public commitment, they are buying a promise that future applications will generate enough productivity to amortize the capital expenditure. If those applications do not appear, the private sector keeps the revenue from construction, and the public balance sheet absorbs the overcapacity. We saw this with the fiber optic bubble in 2001. We saw it with sovereign wealth funds that overpaid for real estate. Don't confuse liquidity with loyalty. Public money is a counterparty, not a customer. It can fill an order book for two fiscal cycles, but its priorities change with elections. Loyalty is what happens when the technology itself creates durable value that people choose to pay for. That is the only honest measure of infrastructure. Here is the contrarian angle. The problem is not too much infrastructure; it is infrastructure built in too uniform a pattern. The railway boom of the nineteenth century was wasteful, but it left behind a network with many uses and many operators. The risk of the current moment is not overbuilding. It is building every station with the same rails, the same signals, and the same ticket office. If we want resilience, we need heterogeneous compute: edge devices, specialized accelerators, open-source software stacks that run on multiple vendors, and training pipelines that can tolerate a sudden gap in one supply chain. That path is less glamorous, less profitable, and harder for one CEO to sell to the G20. But it is the only path that does not turn the world's most important technology into a permanent geostrategic hostage. Compute is not neutrality. If we let one architecture define the commons, we will have elected a new kind of central bank without a mandate, answerable only to its shareholders. The next decade will not be decided by how much compute we build. It will be decided by whether we can govern that compute; whether we can audit it, switch it off, and trust it when nobody is watching. Around the G20 table, leaders heard a beautiful story about scale. Some will write checks. Some will build. But the ones who ask who owns the toll road, who sets the tolls, and who reviews the toll records will actually create public infrastructure. The rest will be tenants in a GPU empire, paying a tax every time a model wakes up. I would rather build a public square: a shared place with open doors, visible rules, and no single landlord. That is the infrastructure worth fighting for.