For all the hand-wringing about an AI bubble, however, it’s quite possible that what we’re seeing is less a question of whether a bubble exists — and more about what kind of bubble it might be. So what does a bubble typically look like in technology? That is less apocalypse and more accounting: misaligned timelines, misunderstood risk pricing, assets that take longer to earn their keep than investors thought.
AI magnifies that mismatch. Power lines and data centers take years; models iterate in months. The result is a marketplace in which optimism about software is meeting the physics and permitting cycles surrounding concrete, copper, and cooling.
Why Bubble Math Looks So Different in AI
AI supply chains are linear and capital-heavy: land and substations, transformers and switchgear, shells and chillers, racks, accelerators, and skilled labor to stitch it all together. One miss anywhere along that chain can leave billions stranded with half-built capacity.
The risks on financing are rising quickly. Reuters has reported on a showcased New Mexico campus connected with Oracle, which secured around $18 billion in credit alongside multi-hundred-billion-dollar cloud and infrastructure commitments tied to hyperscale partnerships. Meta has telegraphed that it will make multiyear infrastructure spend in the hundreds of billions. These are essentially bets on future demand and the grid’s capacity to deliver electrons where and when they are wanted.
Unlike previous cycles, a lot of the spend is on long-lived, location-locked assets. GPUs depreciate; substations, high-voltage interconnects, and water rights do not. That adjusts both the downside and the salvage value if expectations cool.
Demand Is Existential but Uneven Across Sectors
Enterprise evangelism is broadly happening, but isn’t standing in all the right places. McKinsey’s newest survey of executives revealed that while most big companies now have pilot projects testing AI in at least one part of their business, very few have deployed it at scale or seen a clear impact on revenues and profits. Cost takeout in customer support, marketing ops, and coding assistance is what’s leading the way, and any rollouts elsewhere are slowing as risks and compliance issues come into play.
Consumer usage is similarly spiky. A handful of generative apps log eye-popping engagement, but sustainable willingness to pay is still concentrated in professional workflows. That has implications because the training-to-inference mix drives capacity demands: a world dominated by inference desires proximity to users and unrelenting reductions in cost per token; a world dominated by frontier training desires dense, power-rich campuses and the largest possible interconnects.
The Bottleneck Is Not Just Chips, It Is Power and Space
Industry leaders increasingly caution that what’s really scarce is space, power, and time. As one Big Tech chief executive recently put it, chips can be found but not enough “warm shells” ready to plug them into. The International Energy Agency forecasts that electricity demand from data centers could double globally in the next few years, with AI contributing a meaningful share. That wave propels utilities, regulators, and developers into an unusual coordination — if not yet broadly ambitious.
Interconnection queues are another drag. Studies from the Lawrence Berkeley National Laboratory indicate that multi-year delays are now standard operating procedure for new grid hookups, particularly in fast-expanding areas. Even when shells do arrive, some sites go underused because the promised megawatts come late or have constraints that prevent the latest accelerators from being fully utilized. The Uptime Institute has identified shortages of medium-voltage equipment and long lead times for critical power components as ongoing threats.
Certainly none of this is to say the buildout is a mistake. It implies that the rhythm of physical infrastructure can cap the upward slope of AI adoption, and that overbuild risks look more like stranded sockets, idle GPUs, and expensive power contracts than sudden demand collapses.
How to Measure Bubble Risk in Practice Today
- Watch utilization. Persistently sub-50% GPU utilization at scale was an early indicator that capacity had been provisioned before the demand. Monitor the difference between reserved and realized AI instance hours across the leading clouds and also cancellations or delays in long-term commitments.
- Follow the power math. Check for binding contracts, not letters of intent; megawatts with secured delivery dates on them; and the cost per delivered watt after subtracting transmission and curtailment. Revenue per watt and revenue per square foot will be as true to form as was revenue per chip.
- Monitor cost curves. If model efficiency, sparsity, and lower-precision training outrace hardware growth, then required capex per unit of capability declines — which is good for operators but harder on returns that are predicated on permanent scarcity. On the flip side, larger models or more agentic workloads could soak up capacity quicker than anticipated.
What Could Pop and What Could Stick in AI Infrastructure
Training clusters overbuilt without definite demand or power/cooling redundancy, as well as speculative campuses holding out for grid upgrades, are the most exposed. Vendor-financed hardware stacks can become hot potatoes if resale values soften and interconnect topologies don’t match.
On the other hand, assets like power plants and high-voltage transmission lines in major metros can generally hold value through full cycles. Modular-ready facilities — mixing air and liquid cooling, housing denser racks, allowing phased fit-outs — can be repurposed from training to inference or another high-performance computing function if the market for AI capacity returns to normal.
A Better Lens for Builders and Investors
Think in options, not absolutes. Build to real milestones for each stage, pair capacity with contracted workloads, don’t mind incubating alongside AI tenants that value density. Co-develop power with long-term PPAs, on-site generation, and grid-enhancing technology to reduce the time required to get interconnected.
And most importantly, decouple the software curve from the steel curve. Software may keep leaping; steel will keep walking. If the AI bubble bursts, it’s unlikely to be because demand for the technology goes south; rather, we’ll see a deflation in how quickly and by when companies plan to use an age-old bag of tricks. That’s a cooler, more useful way of thinking about the risks — and making bets that may survive not just a boom but whatever follows.