Texas can't agree on who gets to say no to a data center
The compute story keeps drifting away from chips and toward utilities. Today: a Texas county can't legally stop a data center it doesn't want, and a desktop-class rig just matched a server's prompt-processing speed.
Texas' data center boom is stalling on water, not power — and counties say they have no way to stop it. The state is sitting on 248-plus planned AI data center projects and roughly 474 GW of grid interconnection requests, more than five times its own peak demand, which is why Governor Greg Abbott froze new connections and ordered a statewide audit in August. The squeeze is already local: San Marcos' city council rejected a $1.5 billion facility in February after a hearing that ran to 2 AM and then banned new data centers inside city limits, while Hays County paused approvals for water-intensive development through the end of 2026, citing the Edwards Aquifer during an active drought. In the Rio Grande Valley, a Scottish developer holds an option on 1,775 acres near Harlingen for a possible 2 GW campus, and Brownsville has yet to schedule its first hearing on a 90-day moratorium.
The part that matters is jurisdiction, not volume. A state senator has warned the attorney general that counties lack authority to impose development moratoriums, and one county has already been sued for more than $100 million for trying — which is why Hays County stalled on making its pause permanent. That legal chill, plus a legislature that has promised Senate Finance Committee hearings grounded in research rather than "rumors," leaves the actual decision with whoever can afford the litigation. Power shortages get solved with capital and a two-year build order; an aquifer over a drought does not. Expect the water-right fights to be the template other states copy, not the exception.
Qwen3.8 Flash Next just crossed 1,300 tokens-per-second prefill on a desktop chip. An open-source contributor rewrote llama.cpp's handling of the model's sparse attention on AMD's Strix Halo and posted 1,358 tok/s at a full 131,072-token prompt — about 96 seconds to read a maximum-length context, against the 400 tok/s community forks managed before and the 100 tok/s one commenter reports on their own rig. The work targets the memory-bandwidth wall that limits consumer hardware, and the branch is headed upstream, which should carry over to GLM 5.3 Flash's similar sparse-attention architecture. Notably, the parity target was a closed-source inference server that advertised the same 1.2k tok/s figure. There is an irony in the timing: OpenAI just retired its own 1,000-tokens-per-second coding experiment as a hosted service, while that speed class arrives on a machine that fits in a backpack — see OpenAI retires its 1,000-tokens-per-second coding experiment. Decode speed is still the weak spot, with owners reporting 15-25 tok/s, so long-context agents remain the winning local use case rather than fast chat.
What to watch: whether the Texas legislature preempts county moratoriums outright, which would end the local-veto question for the whole state.
Should a county be able to block a data center over its aquifer, even when the state wants the capacity? Tell us in the comments.
