The queue is the grid
On 3 August the Governor of Texas ordered an audit of every data centre advancing through the state grid's interconnection process, and the transmission study that would have moved the next batch forward was postponed while it runs. The number behind the order is the interesting part. ERCOT is considering approximately 474 gigawatts of connection requests, more than five times the state's record peak demand, and about 90 percent of them come from data centres.
A queue that size is not a forecast. It is a pile of intentions, and nobody can serve a request from an intention. This post is about what the codex assumes about where compute lives, and why that assumption looks different this month.
Concentration has a lead time
Centralised inference needs a building, and a building needs an interconnection agreement. That agreement is the scarce input, it is granted by one authority in one jurisdiction, and this month that authority said it would not grant more until it had checked who was asking and why. Nothing in the order concerns models. The constraint is entirely upstream of them.
The codex does not forecast this, and it should not be read as if it did. What it specifies is a tier that does not stand in that queue. The validator chapter puts the Light tier on a single consumer GPU with 24GB or more, and says why: the commitment is that participation should not require institutional capital. A node of that size can serve the language surface in its distilled variant and take the audit role for every surface. Nobody applies for a gigawatt to run one.
What this does not claim
It would be easy to stretch that into something false, so three limits.
First, consumer GPUs still draw power. Moving the load from one substation to many sockets does not make it free; it makes it small, separate and non-queued, which is a different property.
Second, the codex is explicit that not everything runs on a spare card. The same chapter reserves full training and frontier inference for the Cluster tier, which means multi-node hardware with a high-bandwidth interconnect, and that tier lives in exactly the kind of facility the Texas queue is full of. A network that depended only on Light nodes would be a smaller network. The design is that the frontier work is bounded to the tier that can do it, and the rest does not wait on it.
Third, none of this says the network is large. The roadmap defines Phase 1 as exiting only when a published count of independent validators exists across at least two jurisdictions, and that criterion is the codex's own answer to the same worry: a network that sits inside one regulator's reach inherits that regulator's queue.
Why it belongs in a post about a tenet
P8 says compute is paid for, not promised: the protocol issues no obligation it cannot meet from the current period's revenue plus its reserve. The Texas queue is the opposite shape, hundreds of gigawatts of obligations that cannot all be met from anything that exists today, which is why an audit was the only available move.
The lesson is not that data centres are a mistake. It is that a design whose every step waits on one kind of permission has one kind of failure, and a design with a tier that needs no permission has at least one step that keeps moving when the others stop.