The Signal — August 24, 2026

The AI buildout keeps running into questions it treated as settled. This week the chip supplier decided to compete with the labs it sells to, and the governors who courted the data centers started apologizing for them.

Nvidia is paying $6 billion to compete with its own customers

Nvidia is spending $6 billion to license Poolside's technology and move more than 100 of the startup's employees inside the company, where they will work on Nemotron, Nvidia's family of open-weight models. The chipmaker is separately putting about $1 billion into Poolside at a $12 billion pre-money valuation. The Wall Street Journal reported the arrangement on Saturday evening, sourced to people familiar with it. Nvidia has not published terms.

The stated target is Chinese open-weight leaders like DeepSeek and Kimi K3. The unstated one is harder to miss. Open-weight models are cheap to run and easy to customize, which makes them a direct alternative to the closed frontier models sold by OpenAI and Anthropic, two of Nvidia's largest customers. This is not entirely new ground: Reuters reported on August 11 that Nvidia was already building a roughly trillion-parameter Nemotron 4, with VP of generative AI Kari Briski saying that "every company and every country needs accessible frontier open models to strengthen safety and security, accelerate innovation, and provide a foundation they can rely on from one generation to the next."

Nvidia's incentive here is structural rather than sentimental. If a small number of labs control the models, they eventually gain leverage over what hardware gets bought and at what price. A plural, cheap, open model layer keeps compute demand broad and keeps that leverage from concentrating. What changed this week is the directness. Funding research and releasing weights is one thing. Buying a competitor's stack and hiring its engineering team to build a rival to your best customers is another, and it is the sort of move that lawyers for somebody eventually read out loud in a deposition.

Sources: Wall Street Journal · Reuters · Techmeme

Abbott says data center developers dug their own grave

Texas Gov. Greg Abbott told ABC's "This Week" that data center companies "basically dug their own grave" and "got the backlash they deserve," faulting developers for moving into communities without building local support first. Nine months ago Abbott stood next to Google to announce a $40 billion investment across three campuses and called Texas the epicenter of AI development. In June he directed state regulators to make data centers pay the full cost of the electrical infrastructure they require and pushed to phase out their tax incentives. Earlier this month he ordered an audit of projects seeking to connect to the grid. He also noted that fewer than 10% of data center companies had answered a state request for the information regulators need to forecast power demand.

The reversal is not a Texas story or a Republican one. Pennsylvania Gov. Josh Shapiro imposed new restrictions last week, a year after celebrating a $20 billion Amazon investment. New York Gov. Kathy Hochul ordered a one-year moratorium on hyperscaler data centers in July. Annenberg Public Policy Center polling released August 11 found 61% of Americans oppose new data center construction in their area, up from 49% in the survey that ended in March. That 12-point move was the largest shift on any AI question the center has measured, and it crosses party lines at 69% of Democrats, 54% of Republicans and 53% of independents. President Trump is pushing the other way, telling the Texas Tribune that saying no to data centers is a mistake because the industry "could be bigger than oil."

We wrote on August 10 about the count of local bans passing 500 towns and counties. What is different now is the altitude. When the objection lived at the zoning board, developers could route around it by picking a friendlier county. Governors control permitting timelines, grid interconnection queues and the tax incentives that made many of these projects pencil out in the first place. The constraint on frontier compute is no longer chip supply or capital. It is whether the people who live near the substation will put up with it.

Sources: Axios · Annenberg Public Policy Center

A benchmark built from 393 real coding sessions

SemiAnalysis released AgentX 1.0 today under an Apache-2.0 license, describing it as the first fully open-source multi-turn agentic coding inference benchmark at one-million-token context. The firm says it spent more than $3 million building the dataset and runs the matrix on roughly 2MW of continuously operated compute across more than a thousand chips, covering MI355X, GB300 NVL72, GB200 NVL72, B300, B200, MI325, MI300X, H200 and RTX Pro servers.

The interesting part is what the benchmark is arguing about. Standard inference tests feed a model 8,000 tokens and ask for 1,000 back, then report how fast that went. Agent traffic looks nothing like that. Sessions run for dozens or hundreds of turns, context piles up as they go, most of the input on any given turn was already processed on the previous turn, and subagents spin up and die carrying their own separate context. The published methodology puts numbers on it: across 393 opt-in Claude Code sessions, the median request carried 142,000 input tokens and produced 444 output tokens, and 44% of sessions used subagents.

Sit with that ratio for a second. Roughly three hundred tokens read for every one written. Under that load, serving a model is mostly a question of whether the cached version of a conversation can be found and reused, which means the cost drivers are how key-value tensors move between machines, whether a session gets routed back to the node already holding its prefix, and what happens when the cache outgrows fast memory. Those are storage and networking problems wearing a machine learning costume, and the conventional benchmark was never measuring them.

Two limits worth stating plainly. SemiAnalysis notes that AgentX measures serving performance and cannot evaluate model quality, since the replayed traces are synthetic tokens standing in for the original prompts and code. And the firm's own comparative results, the ones that answer the headline question about whether Nvidia's software advantage survives this workload, sit behind a paywall. We are not reporting numbers we cannot check. The open artifacts are the benchmark itself, the methodology and the dataset profile, and those are enough to be useful.

Sources: SemiAnalysis · InferenceX methodology · GitHub

On the Editor's Desk

We looked again at the letter asking the FTC to investigate AI companies that buy books, scan them and destroy the originals. We told you yesterday why we were holding it: it is a request from advocacy groups rather than an actual investigation, and we had already covered the underlying practice. Nothing has changed since, so it stays held. If the FTC opens a case, that is a story and we will run it. The Illinois voice-training lawsuits are in the same position for the same reason, waiting on the court rather than on us.

DeepSeek's experimental vision model looked promising, but the only benchmark results available are DeepSeek's own, measured on DeepSeek's own evaluations, and we would rather wait for someone else to run it. We also passed on a report that a well-known researcher moved from OpenAI to Meta. It traces back to a single scoop repeated elsewhere, and a hire with no stated research direction attached does not tell you much.