AI’s Next Bottleneck Is Not Chips. It Is a Place to Plug In.
Texas just put every new data center through a grid audit, and Anthropic bought ten billion dollars of compute in Norway. Same day, same story. Here is the map change that follows, and the one number worth watching.
Texas has about 474 gigawatts of projects waiting in line to connect to its grid. Roughly 90 percent of that is data centers, many of them tied to AI and crypto.
On August 4, Governor Greg Abbott directed ERCOT and the Public Utility Commission to audit every new data center project before it can connect. Not a moratorium. A filter. Texas has spent two years as the default answer to "where do we build," and it has just started asking applicants to prove they are real.
The same day, Anthropic committed roughly 10 billion dollars over six years to Volta, an AI cloud provider running Nvidia hardware out of a 133 megawatt Bitdeer data center in Norway.
Read those two together and you have the next eighteen months. The scarce input in AI stopped being silicon somewhere in the last year. It is now interconnection: the right to draw power in a specific place at a specific time.
## Why the queue number beats the deal sizes
Anyone can read a 10 billion dollar contract. Far fewer people read interconnection queues, and that is where the constraint actually shows up.
A 474 gigawatt queue is not a forecast of 474 gigawatts of demand. It is a sign that filing costs almost nothing, so speculative projects and funded ones sit in the same line. An audit is what a grid operator does when it can no longer tell them apart.
For an AI lab, that ambiguity converts into scheduling risk. If you cannot know when you get power, you cannot know when you get capacity, and compute contracts are how labs plan model releases.
## Three escape routes, all visible in one week
The first is geography. Norway sells cheap, cold, low-carbon power with spare headroom. Anthropic's deal is not a climate gesture. It is the purchase of a queue position that no longer exists in Texas.
The second is mobility. Runware launched Sonic Inference Pod, a transportable modular data center unit already deployed at more than ten sites across the US, Europe and Asia Pacific. If you cannot move power to compute, move compute to power.
The third is connectivity. EON raised 10.75 million dollars from General Catalyst and Andreessen Horowitz to build a twenty satellite laser network linking data centers across continents. Seed stage and speculative, and worth noting anyway, because it only pencils out in a world where your compute is scattered across whichever jurisdictions will have you.
Notice what none of these are. None of them is a chip story.
## The part I would hedge
Offshore and sovereign capacity solves power. It does not solve latency, data residency, or politics.
A US lab training in Norway is unremarkable. A US lab serving regulated American customers from Norway is a different conversation, and export control policy has been written for chip sales, not for leasing time on someone else's cluster abroad. That gap will get argued about before it gets closed, and the arguing will be slower than the building.
## What to do with this
Watch interconnection queues, not announcements. When ERCOT's audit results land, they will be the first honest count of how much AI data center demand is contracted versus merely optioned. That number tells you more about real 2027 capacity than any funding round will.
And if you are evaluating a compute provider, ask one question before the spec sheet: where is your power, and what is your queue position? "We have Nvidia allocation" was the 2024 answer. It is no longer the binding one.
We track these as they land on the AI deal tracker.