Amazon's $2B Gilroy data center bypassed the community vote
Two infrastructure stories and one research story today: a California town discovers an Amazon data center it never got to vote on, an independent researcher ships latent reasoning as a real model, and London's King's Cross cements its status as a global AI hub.
Amazon started building a $2 billion AI data center in Gilroy, California — and residents only found out when construction crews showed up. Gilroy, about 30 miles southwest of San Jose, watched the project slip through under zoning rules set 45 years ago that required no public meetings and no council vote. The Wall Street Journal reports AWS first applied to build in 2020, and the public comment window quietly lapsed in 2024 — before data centers became the local flashpoint they are today. When resident Rosa Rodriguez tried to raise concerns about water use in a region hit by drought, she was told the window had already closed. Amazon says the site went through a long approval process with public notices and comment periods, and points to an open house it hosted plus a wastewater-recovery system it's building to supply the campus.
The pattern here is the story: data center developers don't need to fight community opposition if the rules never required community consent in the first place. Gilroy's property was zoned for general industrial use in 1981, so a project that would have drawn protests almost anywhere else sailed through on paperwork alone. It's the same Amazon infrastructure push we covered this week — the company is backing a 7.65 GW gas plant for an off-grid Texas AI campus — Amazon backs a 7.65 GW gas plant for an off-grid Texas AI campus. Expect more towns to discover, mid-construction, that their own zoning codes predate the AI buildout.
An independent researcher packaged DeepSeek V4 Flash with a latent-reasoning head and shipped it as a complete, runnable model. Nicholai Mitchko's DeepSeek-V4-Flash-0731-Latent-Reasoning on Hugging Face does its "thinking" in compressed latent space instead of emitting token-by-token chains of thought: a small 35.7-million-parameter head reads the backbone's hidden state, compresses it into a 1024-dimensional latent, and decodes it back into the model, with a learned stop head deciding when reasoning is done — one latent step stands in for roughly six reasoning tokens. The model reports a BIG-Bench Hard aggregate of 0.94 (0.88 by the more conservative reading), with NVFP4 quantization that needs Blackwell-class hardware, and a custom vLLM fork to serve it since upstream vLLM can't drive the latent loop.
It's a research demonstration, not an official DeepSeek release — but it's a meaningful one. Latent reasoning is one of the more active frontiers right now, and most examples live in papers rather than on a model card you can actually pull down and serve. The caveats are honest: evaluation is BBH-only, syntax-heavy tasks are a known weak spot, and the latent space remains opaque. We covered the independent run confirming DeepSeek V4 Flash's 82.7% benchmark score this morning — Independent run confirms DeepSeek V4 Flash's 82.7% score.
London's King's Cross — once one of the seediest districts in the city — is now one of the world's top AI hubs, rivaled only by San Francisco and Beijing. TechCrunch reports the neighborhood that hosted a notorious red-light district in the 1980s now hosts DeepMind, OpenAI, Anthropic (in a 158,000-square-foot office), Meta, Isomorphic Labs, Wayve, Synthesia and Sierra, with Jeff Bezos' Prometheus reportedly in talks to move in. London has roughly 3,600 AI startups that have raised about $12.1 billion, and AI companies have leased more than a million square feet of office space in the city since June, pushing King's Cross vacancy down to 0.9%.
The transformation started in 2016 when DeepMind moved in, and it's now feeding a genuine sovereignty debate: after Anthropic cut off European access to Mythos and Fable this summer, the local consensus is that Europe can't just rent AI capability — it has to build it. Founders say the mood has shifted from "London catching up" to "London as a default place to start a serious AI company."
What to watch: whether Gilroy-style zoning loopholes become the next battleground in the data center wars — and whether cities start rewriting 40-year-old industrial rules.
Should communities get a real vote on AI data centers in their backyard? Tell us in the comments.
Sources: Tom's Hardware · Techstory · Silicon Valley Business Journal · Baxtel · Nicholai Mitchko's blog · Hugging Face model page · Hacker News discussion · TechCrunch
