For three years the artificial intelligence race has been narrated almost entirely through the lens of silicon. Who designs the fastest accelerators, who can buy them, who is blocked from buying them. It is a tidy story, and the United States plainly wins it. But the story is incomplete, and increasingly the people building the largest AI systems on the planet are telling a different one. The binding constraint, they say, is no longer just the chip. It is the electricity to run the chip.
That reframing matters because it changes who is winning and who is merely ahead. On the most useful version of the map, the US holds the chips but is short on power; China holds the power but is short on chips. Both are real constraints. The interesting question — for founders, infrastructure investors, and policymakers from Washington to New Delhi — is which one bites harder, and how soon.
The framing
Compute has two physical prerequisites, and we have spent most of our attention on only one of them. The first is the processor itself — the leading-edge GPU or custom accelerator that does the matrix maths. The second is the megawatt-hours required to keep racks of those processors fed, cooled, and running around the clock. For most of the modern computing era, power was a background cost. With frontier AI, it has moved to the foreground.
US technology leaders have begun to concede this openly. Elon Musk has argued that “the limiting factor for AI deployment is fundamentally electrical power,” warning that the United States would soon be able to produce more chips than it can actually power — adding the pointed exception, “except for China.” When the people most invested in American chip supremacy start talking about the grid, it is worth listening. Electricity has quietly become an AI variable, sitting alongside transistor density and memory bandwidth on the list of things that determine how big a model you can train and how cheaply you can serve it.
This is the core asymmetry. The US bottleneck is energy and the physical infrastructure to deliver it. China’s bottleneck is access to the most advanced silicon, courtesy of export controls. Each side is strong precisely where the other is weak.
China’s power edge
China’s advantage is unglamorous but profound: cheap, abundant electricity and the institutional capacity to build generation and data-centre capacity at remarkable speed. Where a new substation or transmission line in the West can take years of permitting and litigation, China’s state-directed model can stand up power infrastructure on a timeline that Western operators can only envy. The result is that, for a data centre operator inside China, power is closer to a given than a constraint.
The scale of the build-out reflects this. According to Rystad Energy, China’s data-centre capacity is projected to reach roughly 60 gigawatts by 2030 — a figure that, if realised, would represent one of the largest concentrations of compute-serving power capacity in the world. That is the optimistic reading of the power-edge case: when energy is plentiful and cheap, you can afford to throw it at AI workloads and worry about efficiency later.
But the honest version of this story carries a caveat, and it is a significant one. Cheap power and fast construction do not automatically translate into useful compute. Much of China’s data-centre capacity reportedly sits underused, with utilisation rates estimated in the region of 20 to 30 percent. Even the chief of SMIC, China’s most important chipmaker, has warned that newly built capacity could simply sit idle. Some of that is a build-quality and location problem — facilities erected for political or speculative reasons in the wrong places, without the chips or demand to fill them. So China’s edge is real, but it is an edge in potential energy more than in delivered AI. Power without enough advanced silicon is a half-finished engine.
The US chip edge
The American position is the mirror image. The United States, through its companies and its allies, still controls the design and manufacture of the world’s best AI silicon, and it has used export controls to keep the frontier of that silicon out of Chinese hands. On chips, the lead is genuine and structural.
The problem is everything around the chip. Power and permitting have become the choke points. Connecting a hyperscale data centre to the grid in parts of the US now involves multi-year interconnection queues, local opposition, and a transmission system that was not designed for gigawatt-scale point loads appearing in a single county. Projects that look fully financed on paper stall in practice — not for lack of capital or chips, but for lack of a guaranteed electron supply on the timeline the business case demands.
And the capital is staggering. US Big Tech — Amazon, Microsoft, Meta, and Alphabet — is projected to spend around $630 billion on data centres in 2026, by Morgan Stanley’s estimate. That spending is increasingly debt-funded, which raises the stakes considerably. When you are borrowing at that scale to build infrastructure whose output depends on power you cannot yet guarantee, the energy bottleneck stops being an operational nuisance and becomes a financial risk. The chips may be the best in the world; if they cannot be powered on schedule, the returns on that capex get pushed out, and the debt does not wait.
So which constraint bites harder? The fair answer is that they bite on different timescales. China’s idle-capacity problem is downstream of an export-control regime that could, in principle, be eased or routed around, and its domestic chip industry is improving. America’s power problem is downstream of physics, planning law, and grid economics that move slowly regardless of who is in office. A chip shortage can be solved with a policy change or a new fab. A power shortage is solved with steel, copper, turbines, and a decade of planning reform. On the evidence, the energy constraint looks structurally stickier.
Where it leaves India
If the binding constraint on AI is shifting from silicon to electricity, then the strategic map opens up for countries that are neither the US nor China — and India is unusually well placed to read it.
India’s opportunity is not to win the chip race; that is not a realistic near-term ambition. The opportunity is to position itself as a power-and-place magnet for compute. India is adding renewable capacity at scale, has large tracts of land, a deep engineering workforce, and a fast-growing domestic AI demand base. For global operators hunting for sites where clean power can be secured at reasonable cost and reasonable speed, India can credibly compete — provided it fixes the same grid-interconnection and land-acquisition frictions that are now hobbling the US.
The strategic choice for New Delhi is one of emphasis. Chasing chip-fabrication self-sufficiency is expensive, slow, and contested; chasing an energy-and-data-centre strategy plays directly to India’s structural advantages. The smarter posture is to treat chip access as something to negotiate — through partnerships, assured supply arrangements, and a welcoming stance toward hyperscalers — while making the country’s clean-power build-out the genuine differentiator. India does not need to own the silicon to host the compute.
The broader lesson of the chips-versus-power frame is that the AI race is not a single contest but two overlapping ones, and the second — energy — is where the laggards of the chip race can still compete. The US has the silicon and a grid problem. China has the power and a silicon problem. For everyone else, the winning move is to be honest about which constraint you can actually relieve, and to build there. For India, that constraint is power — and that is a far better hand to be dealt than it first appears.
