In partnership with
In Todayโs Issue:
๐๏ธ Community opposition has stalled $130 billion of data centers
โก Cerebras says its new CS-4 runs up to 30x faster than GPUs
๐ต๏ธ A stealth model called Ox Alpha is outscoring the frontier
โ๏ธ The Justice Department steps into xAI's fight with Minnesota
โจ And more AI goodnessโฆ
โก The Signal
The AI industry has spent the past year winning the argument in Washington and losing it street by street.
Community opposition blocked or delayed $130 billion of data center projects in the first quarter of this year alone, roughly as much as in all of 2025, and organized groups now operate in 49 states. Set that beside the rest of today's issue and a shape appears. Cerebras just doubled its inference speed again, which means more racks, more megawatts and more sites in more counties. OpenAI has stood up a team to think about what happens when a state no longer needs its citizens' labor or their consent. Anthropic is trying to teach millions of people to use AI before they decide how they feel about it. The capability curve keeps bending upward. The consent curve is bending the other way, and this week Senate Republicans put a price on the gap: one Ohio seat.
All the best,
Kim Isenberg
(The CS-4 rack. Image: Cerebras via ServeTheHome)
โก Cerebras Unveils the CS-4, Its Fastest Machine Yet
Cerebras introduced the CS-4 on August 18, a rack-scale system built from three of its new Wafer Scale Engine 3 Turbo chips. Each wafer carries four trillion transistors and 900,000 cores across 46,225 square millimeters of silicon, and the full rack delivers 750 PFLOPs. On the open model GPT-OSS-120B, Cerebras clocked more than 4,400 tokens per second per user, which it says is up to 30 times faster than GPU-based systems on identical prompts, at 10x the throughput per watt of the CS-3. First shipments are due this quarter.
๐ tl;dr: These are the vendor's own numbers on one model. If they hold, the payoff lands in agents, which can pack far more reasoning, verification and tool calls into the same wall clock.
(Dean Ball, left, who leads OpenAI's Strategic Futures team, alongside Noam Shazeer. Image: TechCrunch)
๐๏ธ OpenAI Starts Publishing on Who Holds Power After AI
OpenAI launched AI Futures on Thursday, the blog of its new Strategic Futures team, led by Dean Ball, the former White House adviser who helped publish America's AI Action Plan. The opening essay argues that political power has always rested on human beings: soldiers and police to project force, bureaucrats to run the machinery, wage earners to tax. If force can be projected by autonomous systems and revenue collected from data center output instead, that bargain weakens, and the public is left with a small seat at the table or none at all. The team calls this category of risk "the largest, most serious, and most conceptually challenging to address."
๐ tl;dr: A frontier lab is now publishing on how its own technology could make citizens optional to a state. Worth reading, and worth watching what it changes inside OpenAI.
(Anthropic)
๐ Anthropic Publishes How It Teaches People to Use AI
Anthropic set out its approach to AI instruction on Thursday, built around Claude Academy and the 4D AI Fluency Framework it already uses on its own staff. New hires are taught from day one to decide deliberately which tasks to hand to a model and which to keep, and to learn the mistakes models tend to make so they can review AI output properly. After onboarding, the company runs what it calls "ever-boarding", always-on training, on the grounds that continuous learning is now expected of every role.
๐ tl;dr: The curriculum implies something Anthropic never says outright: getting real value out of these models is a trained skill, and it does not arrive with the subscription.
๐ฌ Watch This
Open models are sold as roughly ten times cheaper than frontier APIs. This episode asks what that number actually costs once it reaches production.
Kim Isenberg and Peter Thum sit down with Jonathan Cornelissen, co-founder and CEO of DataCamp, and Yusuf Saber, DataCamp's Chief AI Officer and the founder of Optima, to walk through the model decisions behind an AI tutor used by more than 19 million learners. They stay concrete throughout: how tasks get routed between open-weight and frontier models, what self-hosting really costs once caching, latency and reliability are priced in, how you tell whether a learner actually understood a concept, and how a tutor avoids confidently teaching something wrong. The useful question is where open models genuinely win today, and where the frontier still has to do the work.
"...whatever the hell this is: 80% ... I am very confused."
โ Ben Davis, developer and co-host of the Nerd Snipe podcast
In a 10-task spot check posted early on Friday, an unnamed model handed out free on OpenCode passed 8 of 10, against means of 65% for Fable 5, 62% for GLM-5.3 and Grok 4.6, and 52% for GPT-5.6 Sol. Davis flagged his own caveat: it is a small subset.
Nobody will say who built the model that just outscored the frontier on a developer's coding test. OpenCode is handing out free access to a stealth model called Ox Alpha, with a 1M-token context window, multi-modal input and zero data retention, and says it has "capacity for 100T tokens per day". No lab has yet put its name to it.
(Ox Alpha against four named frontier models on 10 DeepSWE tasks. Image: Ben Davis / X)
The Data Center Backlash Stopped Being Local
The Takeaway
๐ Local opposition blocked or delayed 75 data center projects worth about $130 billion in the first quarter of 2026 alone, roughly matching the whole of 2025 in three months.
๐ Active opposition groups more than doubled in that quarter and now span 49 states. Maine came within one House vote of the first statewide ban in US history.
๐ The power bill is the engine: PJM's capacity auction went from $28.92 to $329.17 per megawatt-day in two cycles, and ratepayers absorb it.
๐ The industry's own fix, a voluntary Ratepayer Protection Pledge signed in March, has not moved opinion. Now Senate Republicans say it is costing them Ohio.
The fight over data centers stopped being a zoning story this year and turned into a national one with a price tag. Data Center Watch counted 75 projects worth roughly $130 billion blocked or delayed by community opposition in the first quarter of 2026, about as much as all of 2025 in three months. Active opposition groups more than doubled over the same period and now operate in 49 states. More than 300 state data center bills were filed in the first six weeks of the year, statewide moratorium proposals turned up in 14 legislatures from both parties, and Maine came within one House vote of the first statewide ban in US history. The researchers call it a structural shift rather than a spike.
(A resident protesting a proposed data center. Image: Getty via Newsweek)
The grievance is mostly the power bill. In PJM, the grid serving 13 states including Ohio and Virginia, the capacity auction cleared at $28.92 per megawatt-day for 2024/25 in most zones, then $269.92, then $329.17 for 2026/27, a rise IEEFA attributes in large part to data center demand and one that reaches households. Add the water use and the very small permanent workforce, and the industry's pitch stops landing. The companies saw it coming: on March 4 Amazon, Google, Meta, Microsoft, OpenAI, Oracle and xAI signed a White House Ratepayer Protection Pledge to build, bring or buy their own power and pay the full cost of the infrastructure. It is voluntary and unenforceable, and opinion has kept moving against them since.
(Sen. Jon Husted, whose seat the memo is worried about. Image: Getty via ABC News)
This week the backlash reached the part of the system that reacts fastest. The National Republican Senatorial Committee sent a private memo, headlined "Ohio Data Center Risk" and obtained by Axios, telling top AI companies that data centers are sinking Sen. Jon Husted against Sherrod Brown, and warning that if voters' perceptions are not fixed "the campaign against them will expand far beyond Ohio". It hands the repair job back to the industry: explain "who benefits, who pays, and why a community should want one". Trump went the other way a day later, telling a White House crypto event that any governor should want an AI plant, before allowing that the industry "can use a little public relations help".
Why it matters: The buildout's license has rested on the assumption that opposition stays local, disorganized and cheap to outlast. $130 billion of blocked projects in a single quarter, organized groups in 49 states and a Senate campaign arm warning its own donors are three different ways of saying that assumption has expired.
Sources:
๐ https://www.motherjones.com/politics/2026/08/gop-memo-republican-senate-data-center-ai-elections-trump/
๐ Axios: Scoop on the NRSC's "Ohio Data Center Risk" memo
๐ Data Center Watch: Q1 2026 report on blocked and delayed projects
๐ IEEFA: Data center growth spurs PJM capacity prices
Winning Startups Aren't Bigger. They're Leaner.
Founders deploying AI agents across sales, marketing, and customer service are closing more with smaller teams. 65% more sales leads. Less headcount.
Download the free Practical Guide to Agentic GTM for Startups and start building the stack your competitors dream of.
The chart: Heatmap Pro has put the same question to Americans four times in a year, with the wording unchanged: would you support or oppose a data center being built near where you live. In August 2025 the country split 43% support to 42% oppose. By February a bare majority was opposed, by May it was 70%, and in the latest round, run by Embold Research from August 8 to 13 among 2,045 registered voters, 75% say they would oppose one and 61% say they would strongly oppose it.
The lesson: Twelve months is very fast for settled opinion to move this far, and there is no group left to build a defense on. Heatmap reports the same swing across age, gender, income, party and the rural-urban divide. That is why the NRSC memo reads the way it does.
The caveat: This is one pollster's tracking series, with a margin of error of about 2.3 points, and "near where you live" invites a harder answer than "somewhere in your state". It measures how people feel about a hypothetical neighbor, which is a different question from how they vote in November.
โ๏ธ Washington Just Stepped Into Minnesota's Fight Over Nudify Apps
โก Bottom line: Minnesota banned apps that let users strip the clothes off real people's photos. Elon Musk's xAI sued.
๐ก Why it matters: It is the first court test of whether a state can limit what an AI tool is able to produce.
๐ What it means: The administration's push to preempt state AI laws has arrived in court, in an awkward case.
The law in one sentence: if your app lets someone take a real person's photo and generate an image of their intimate parts, Minnesota can fine you up to $500,000, whether or not you meant it to happen. That is HF 1606, signed by Governor Tim Walz on May 7 and in force since August 1. It passed the state House 132 to 1 and the Senate 65 to 0. The penalty lands on the platform, never on the user.
(Grok, the tool at the center of the case. Image: Getty via TechCrunch)
xAI sued on July 27, five days before the law took effect, arguing that the statute violates the First Amendment. At Wednesday's hearing its lawyer, Robert Dunn, said the statute carves out nothing for consent, art or parody. Minnesota's Assistant Attorney General, Janine Kimble, answered that terms of service are no defense while the company is still reporting tens of thousands of users generating these images, and that "the repercussions only happen after the fact": somebody has to see the picture, report it, and find it first.
(Elon Musk, whose xAI brought the suit. Image: Getty via Austin American-Statesman)
Then the federal government showed up. On August 18 the Justice Department filed a statement of interest saying Washington shares Minnesota's goal but objects to how the law is written. Federal statutes such as the TAKE IT DOWN Act, which Trump signed in 2025, require knowing conduct and apply only where an adult did not consent. HF 1606 has neither limit. Federal law would not reach an AI image of a shirtless man in a pool, the filing notes, but Minnesota's might, because "intimate part" covers the inner thigh.
(Minnesota Attorney General Keith Ellison, the named defendant. Image: Law Commentary)
It also warns that states "must be careful not to impose excessive measures that hinder American national and economic security", citing an executive order that calls for a single national AI standard rather than "50 discordant state ones". That last sentence is what the filing is actually for. After months of failing to preempt state AI laws through Congress, the administration is making the same argument in court, in the least comfortable case available: a child-protection bill that 197 of 198 Minnesota legislators voted for. Judge Donovan Frank heard both sides on Wednesday and said he would rule soon.
50 Excel Hacks That Make Spreadsheets Work For You
Most Excel users waste hours on tasks that take experts minutes. The difference is not talent. It is knowing the right shortcuts.
These 50 Excel hacks from Kenji Explains cover the functions, formulas, and workflows that make spreadsheets work for you instead of against you.
Each one comes with step-by-step instructions and ready-to-use templates. Subscribe to Marketing Against the Grain and get all 50 free.
Here's what you'll get:
Time-saving shortcuts that eliminate the most common formula frustrations
Interface essentials most users never discover on their own
Game-changing functions for data analysis, automation, and visualization
Guided practice template so you can apply every hack immediately
The people who are fast in Excel are not smarter. They just stopped fighting the tool.
Get 50 Excel Hacks free when you subscribe to Marketing Against the Grain today.
