For the past two years, the dominant narrative about artificial intelligence ( AI ) has been one of boundless possibility. Larger models, trillion-token training runs, and record-breaking capex ( capital expenditure ) cycles have reinforced a sense of uninterrupted acceleration. But technological change is rarely so straightforward, and this time is no exception. As AI moves from experimentation to real-world applications, the limits imposed by the physical world, capital markets, and political systems clearly matter more than its theoretical potential.

The most immediate constraint is electricity. Nowhere is this more evident than in the United States, where power demand from data centres is expected to rise from roughly 35 gigawatts to 78GW by 2035. Northern Virginia, the world’s largest cloud-infrastructure cluster, has already effectively exhausted its available grid capacity. Utilities in Arizona, Georgia, and Ohio warn that new substations may take almost a decade to build. A single campus can require 300-500MW, enough to power an entire city. Silicon can be manufactured quickly; high-voltage transmission cannot.

Markets are responding with the speed and ambition one would expect. Hyperscalers ( the major tech firms building advanced AI models on the back of ever-greater computing capacity ) have become among the world’s largest buyers of long-dated renewable energy. Private solar and wind farms are being built expressly to serve cloud facilities, and some firms are exploring next-generation small modular reactors as a way to bypass slower municipal infrastructure.

These efforts will eventually expand the frontier of what is possible, but they do not eliminate the constraint so much as redirect it. The next wave of AI capacity will likely be concentrated not in Northern Virginia or Dublin, but in regions where land, power, and water remain abundant: the American Midwest, Scandinavia, parts of the Middle East, and western China. The geography of AI is being written by physics, not preference.

Mishmash of models

Silicon is the next constraint, and here the story is becoming more complicated. While Nvidia once appeared to be the universal substrate beneath all AI development globally, that era is ending. In a significant milestone, Google trained its latest large language model, Gemini 3, entirely on its own Tensor Processing Units – and Amazon’s Trainium2, Microsoft’s Maia, and Meta’s MTIA chips are all being developed for similar purposes. Similarly, in China, Huawei’s Ascend platform has become the strategic backbone for domestic model training in the face of US export controls.

Some of this shift reflects natural technological maturation. As workloads increase, specialized accelerators become more efficient than the general-purpose GPUs originally adapted for AI. But the timing is not accidental. Scarcity, geopolitical friction, and cost pressures have pushed hyperscalers to assume a role once reserved for semiconductor firms. Given that departing from Nvidia’s CUDA ecosystem carries enormous organizational costs, the growing willingness to incur it signals how severe the constraint has become. What will follow is a more fragmented hardware landscape, and with it, a more fragmented AI ecosystem. Once architectures diverge at the silicon level, they rarely reconverge.

The third constraint, capital, operates in a more subtle way. Hyperscaler investment plans for 2026 now exceed US$518 billion, a figure that has risen by nearly two-thirds just in the past year. We are already witnessing the largest private-sector infrastructure buildout in modern history. Meta, Microsoft, and Google revise their capex guidance so frequently that analysts struggle to keep pace.

Yet it is still early days for economic returns. Baidu recently reported 2.6 billion yuan ( US$369 million ) in AI-application-related revenue, driven largely by enterprise contracts and infrastructure subscriptions, and Tencent says it has lifted profitability through AI-enhanced efficiencies across its mature businesses. But in the US, most companies still bury their AI earnings within broader cloud categories.

The gap between AI adoption and monetization is wide but familiar. In past technological waves, infrastructure spending routinely preceded productivity gains by years. The constraint comes not from weak investor sentiment, but from the strategic pressure enthusiasm creates: different firms pursue different conceptions of value because their business models and cost structures demand it.

Many sectors simply cannot adopt AI at the pace that new models are being released. Large banks, for example, remain bound by security and compliance frameworks that require air-gapped, on-site, fully auditable software deployments. Such rules instantly cut them off from the most advanced frontier models, which rely on cloud-side orchestration and rapid iterations through new versions. Healthcare systems face similar limits, and governments even more so. The problem is not AI’s theoretical capabilities, but the difficulty of incorporating such tools into legacy systems built for a different era.

Diverse architectures 

Taken together, these forces suggest a future very different from the one implied by the standard media narrative. AI is not converging towards a single universal frontier. Diverse regional and institutional architectures are being shaped by different limits – from power shortages in the US to land and cooling constraints in Singapore and Japan, “geopolitical” scarcity in China ( where Western export controls limit access to advanced chips and cloud hardware ), regulatory friction in Europe, and organizational rigidities across the corporate world. Technology may be global, but implementation is local.

Fortunately, real-world constraints are not the enemy of progress. Often, they form the scaffolding around which new systems take shape. The fiber-optic glut of the late 1990s, initially derided as wasteful overshoot, later underpinned the rise of streaming, social media, and cloud computing.

Today’s constraints will play a similar role. Power scarcity is already shifting the geography of AI. Silicon fragmentation is creating new national and corporate ecosystems. Capital asymmetries are pushing firms into different strategic equilibria. Institutional limits are shaping the first real use cases.

The next decade of AI will belong not to the systems with the greatest theoretical capability, but to the ecosystems most adept at turning real-world limits into design advantages. Possibility defines the horizon, but constraint will determine the route the world ultimately takes.

Jeffrey Wu is director at MindWorks Capital. 

Copyright: Project Syndicate