The AI Compute Squeeze Is Getting Real

If your product or roadmap assumes AI tokens will keep getting cheaper and capacity will always be sitting there when you need it, this week gave you three reasons to revisit that assumption. A large model provider hit a compute wall and paused new sign-ups to its top tier. A hyperscaler committed to tripling its data center footprint. And the Federal Reserve raised interest rates for the first time in three years, which changes the math on the debt paying for all of it. For anyone budgeting AI spend or designing systems that depend on a specific model at a specific price, the planning window just got shorter.

The subsidy era is running out

The clearest signal was operational. After the release of its newest flagship model, one major lab reported demand it described as unprecedented and, within days, paused new subscriptions to its $200 professional plan. The stated reason was that those users put the most strain on the systems, and the company wanted to protect service for existing customers while it added capacity. Existing accounts and API access were left untouched.

The reaction from practitioners was blunter than the announcement. One widely shared take was that the era of subsidized tokens is ending, and teams should prepare accordingly. Another predicted the $200 tier would not come back in its current form, on the theory that a small number of power users maxing out limits were being carried by everyone else. Whether or not that specific prediction holds, the direction is worth taking seriously: the price you pay per token today may reflect a land-grab phase, not the steady-state cost of running these systems. If a workflow you are building only pencils out at current pricing, treat that pricing as temporary.

The buildout is accelerating anyway

Supply is being thrown at the problem at a scale that is hard to intuit. Microsoft told Bloomberg it plans to reach 38 gigawatts of global capacity by 2032, up from around 12 gigawatts today. Only about 2 gigawatts of its current fleet is dedicated to AI compute, and the plan is to grow that share to roughly a third. This is notable because Microsoft had been the most conservative of the hyperscalers, scaling back leases in early 2025 in a move that rattled the market. The reversal reads as a bet that AI infrastructure is nowhere near overbuilt.

Nvidia's Jensen Huang reinforced the point at a Goldman Sachs conference, reaffirming roughly 70% growth and saying demand runs well above that, with supply the only limiter. His framing of the hardware is the part worth holding onto: a single rack-scale GPU system, connected over NVLink with around 2 million parts, now runs about $8.5 million and draws on the order of 250,000 kilowatts. He said orders for those rack-scale systems are growing about 27% month over month. These figures are reported as his, and they explain why capacity is rationed rather than instant. You cannot conjure a $8.5 million, multi-ton system on demand.

Now the money gets more expensive

The financing layer is where the two threads meet. Data center construction has increasingly been funded by debt, with Moody's projecting $240 billion in hyperscaler bond issuance this year. The Fed's rate hike, its first in three years, was unanimous, with two more increases expected by the end of 2027 and a real chance of another before this year closes. Higher rates raise the cost of every bond that funds a new building.

One former Bloomberg opinion writer put the tension plainly: talk of tariffs and oil as the rationale for hikes distracts from the reality that taming inflation may require hurting the stock market or AI capex, and that is uncomfortable for a lot of people. The Fed is caught between segments of the economy, with the 30-year mortgage back above 7% signaling rates that are already high in some places while still not high enough to cool the buildout. If borrowing gets expensive enough to slow construction, the capacity shortage that paused subscriptions this week gets worse before it gets better.

My take: Plan for AI cost and availability to move against you over the next 12 to 18 months, not in your favor. If you are building on a single provider's flagship at today's promotional pricing, that is a concentration risk worth naming now: design your architecture so you can route work to cheaper or smaller models where quality allows, and keep a fallback provider wired up rather than theoretical. Watch the financing signals as closely as the model benchmarks, because the cost of debt, not the pace of research, is the variable most likely to decide what you can actually run.