Compute Is Now a Roadmap Risk, Not Just a Bill

For a while, the compute your AI features run on felt like a utility. You paid the bill and moved on. That is changing. A state just froze new data center construction, chip access shifted overseas, and the power to run all this got its own political fight. Let me catch you up on what moved and what it means for your roadmap.

The state that said no to servers

New York became the first state to hit pause on new data centers. Gov. Kathy Hochul signed an executive order that stops the state from approving permits for projects 50 megawatts or larger. That could affect more than a dozen projects. The freeze lasts until an environmental review wraps, which she expects to take about a year.

This is not a one-off. Maine's legislature passed a similar pause before the governor vetoed it. More than 230 organizations called for a national pause back in December. The public mood turned: two-thirds of people in one poll worried about data centers raising their power bills, and another survey found people would rather have an Amazon warehouse next door than a data center.

Here is why it lands on your desk. New capacity is getting bigger fast. Nearly a quarter of new data centers built through 2030 will top 500 megawatts, per BloombergNEF, driven by AI. Bigger builds mean bigger fights, and bigger fights mean slower supply in the places that push back.

Where the chips are allowed to go

Compute supply is not just about buildings. It is about who gets the chips. China plans to let its top AI firms, including Alibaba, ByteDance, and DeepSeek, buy Nvidia's H200 chips, a reversal after it had withheld approval even with US authorization.

At the same time, Trump signed an order pushing hard on quantum computing, in part to keep adversaries from getting there first. As one writer put it plainly, China already holds 60 percent of global quantum patents. The theme is the same across both stories: chip and compute access is now a policy lever governments pull.

You do not control any of this. But the vendor you depend on does answer to it. When access rules shift, pricing and availability shift too. That belongs in your planning, not in your surprise pile.

The power problem hiding under the AI story

The real bottleneck behind data centers is electricity. That is why four microreactors hitting a milestone matters. Four US companies got their reactors to criticality, a test that shows a reactor can sustain a chain reaction, Casey Crownhart reported. Aalo Atomics hit the mark in the early hours of July 4, barely beating the deadline.

Do not read this as clean power arriving soon. All four reached zero-power criticality, which one former nuclear official said "can be achieved without making real engineering progress on fuel or design." These reactors still need cooling systems and years of work before they feed a grid. Aalo says 10 megawatts by 2027 to power an on-site data center, but nuclear timelines slip.

The signal is that power supply is the constraint AI has to solve, and the fix is still years out. Hochul is even weighing a fund to make data centers pay into the grid. Cheap, easy power is not the default anymore.

Skip the specialists, not the network

If you do move compute in-house to dodge cloud bills and supply risk, there is a catch waiting. Building AI infrastructure is harder than buying servers. Networking has become one of the biggest hurdles. A GPU cluster is the most expensive asset a company buys, and every day it sits idle waiting on the network is money burning.

That is the pitch from Hedgehog, a Seattle startup building open-source software to run GPU networks. Its founder was surprised that the buyer is rarely a network engineer. It is platform and DevOps teams handed thousands of GPUs and told "you own the network now." They do not want to learn BGP. They want it to behave like the rest of their stack.

The practical read: owning compute trades a cloud bill for a staffing and setup problem. Know which trade you are making before you sign off on a private buildout.

The deep cut

The thing to catch is that these four stories are one story. Where servers get built, which chips you can buy, whether there is power to run them, and who can even operate the network are all moving at once, and all outside your team's control. That means your compute assumptions have a shelf life now. If your roadmap for the next 18 months assumes cheap, available, predictable compute in every region, it is quietly resting on a bet that just got shakier. Put a real line item in your next review for compute supply risk. Ask what happens to your launch if capacity in a key region freezes for a year, because in New York it just did.

Three questions for your team

1. If a state or region froze new data center capacity for a year, which of our shipping plans would slip, and do we have a fallback region?
2. Are we assuming one chip vendor and one cloud in our cost model, and what breaks if their pricing or access changes on policy grounds?
3. If we are considering moving compute in-house, who on the team actually owns the network, and are they set up to run it without a room full of specialists?