The AI Backlash Is Getting Stupider. But Also Smarter

August 20, 2026 · Episode Links & Takeaways

Author

AI Daily Brief
August 19, 2026

HEADLINES

The Scrutiny Era Arrives for AI Lab Revenue

OpenAI and Anthropic are both facing a wave of skepticism after fresh revenue figures — $40 billion in annualized run rate for OpenAI, $65 billion for Anthropic. The Wall Street Journal's angle was that OpenAI's Q2 operating margin sank further into the red on 18% quarterly growth, which is factually accurate and also conveniently ignores the seven weeks since June 30, a period in which Greg Brockman told CNBC that July revenue grew 20% month over month. Leaving out details that executives have been stating openly only reinforces a narrative, and it's exactly the kind of thing that gives people trust issues with mainstream media. Anthropic wasn't spared either, with Dylan Patel taking a swipe at an ARR methodology that extrapolates four weeks of on-demand API revenue out to a full year, and SemiAnalysis noting that indirect channels crossed 40% of ARR while Anthropic books that revenue before the hyperscalers take their cut. The real takeaway is that nobody has experience pricing companies that went from single-digit billions to tens of billions in a year, and the noise will only get louder as the IPOs approach.

OpenAI Cuts Token Prices to Take Developer Share

GPT-5.6 Sol tokens are now half price on OpenRouter and Vercel's Gateway, following the same discount already applied to Luna and Terra. The move sharpens OpenAI against cheap Chinese models and against Anthropic on cost efficiency, which matters most on routers where users let the platform pick the model automatically — Luna is now the top closed model on OpenRouter, seeing 40% more use than Opus 5 and Sonnet 5 combined. Bears will call this the opening shot of a price war, but OpenRouter is a vanishingly small share of total token volume, so OpenAI can discount heavily there without touching the business. SemiAnalysis reads it as a canny marketing play: OpenRouter and Vercel are two of the main data sources everyone uses to estimate model market share, so doubling volumes on a 50% cut buys a narrative win far bigger than the revenue involved.

Anthropic Preps Supervoting Shares for Its Founders

Ahead of the IPO, Anthropic is preparing to grant supervoting shares to Dario Amodei and his co-founders, letting a roughly 15% collective stake veto shareholder votes and control board appointments. Founder share classes are hardly novel — Page and Brin pioneered it before Google's 2004 IPO, Zuckerberg followed in 2012, Musk more recently at SpaceX — but the reaction is different here for obvious reasons. This is a company whose own framing is that it is building systems powerful enough to reshape the trajectory of society, and whose founder reportedly believes it could end up the only private company left standing after out-competing everyone else. That's a scenario where people are going to want shareholders and the public to have more control over the decisions, not less.

Google Outbids Mercor for Spirit Airlines' Corporate Data

Google won a bankruptcy auction for Spirit Airlines' corporate data at $10 million, beating Mercor's $7.5 million bid. This is the third big data land grab of the AI era: first scraping the web and every available book, then buying up the codebases of failed startups to feed coding agents, and now paying for corporate exhaust to teach agents how white collar work actually happens. Customer information and payment records were excluded from the sale — the interest is purely in mundane internal email, Slack messages and meeting transcripts, or as Mercor put it, the decades of records that show how real work gets done. Plenty of skepticism about whether it pays off, best captured by the obvious question of whether anyone wants to train models to be like Spirit.

MAIN STORY

The AI Backlash Is Getting Stupider. But Also Smarter.

The anti-data center backlash is undeniably getting dumber — more meme-driven, more performative, and now popular enough that beverage companies think it sells product. But underneath the noise, two of this week's biggest moves point the other way: a lab pausing training voluntarily, and a governor choosing a set of specific, debatable criteria over a blanket moratorium. Both are openings, and rules that can be argued with beat a ban a hundred times out of a hundred.

DUMB DATA CENTER DISCORD

Liquid Death and Jason Kelce
An ad campaign about mailing your urine to data centers
Liquid Death and Garage Beer are selling a limited edition "Data Center Coolant Collector" — a resealable jar you fill and mail to the data center of your choice. Provocative campaigns are the brand's whole identity, so the ad itself isn't surprising; what's notable is that professional advertisers read the room and concluded that hating data centers now helps sell beer and water.

Jon Ossoff
"Tech titans dig bunkers" while Congress debates ballrooms and youth sports
The Georgia senator, up for re-election and increasingly discussed in Democratic circles as an Obama-esque 2028 unifier, has made opposition to tech a central plank — pointing directly at the industry's own warnings about mass joblessness and extinction. Constant badgering about how many jobs AI will destroy definitely isn't shaping that narrative at all.

Tom Tiffany
Branding his opponent "Data Center David Crowley" like an insult
In the Wisconsin governor's race, the Republican candidate is running an ad attacking his Democratic opponent over a clip of Crowley saying the state could become an AI and data hub for the entire globe. Proof that this isn't a clean partisan divide — the attack works from the right just as well as from the left.

Josh Shapiro's Executive Order
Predators, bullies, secretive — from the state's loudest AI booster
Fourteen months after proudly announcing Amazon's $20 billion Pennsylvania buildout, Shapiro signed an EO requiring data centers to meet the strictest standards in the nation and win local approval, and publicized it with a tone that surprised everyone. The rhetoric is genuinely disappointing — populist trope words like bullied and greedy — but the more important read is that he's a good politician, and this is what the signals he's receiving look like when translated into policy.

What's Actually In the Order
Bring your own power, community benefits, and no more NDAs
The requirements: binding transparency and environmental agreements including water conservation, early public notification ahead of local approvals, a public permitting map, bring-your-own generation with developers paying all costs of increased energy usage, and community benefit agreements covering local hiring, training, schools and infrastructure. Bring-your-own-power is just the Trump pledge everyone already signed, and the community benefit piece is what companies like Meta are already leaning into hard — which is to say almost all of it has been on the standing list of what data center builders should be providing communities for a long time. The devil is genuinely in the details, and the environmental standards in particular could be written as an effective ban that pretends not to be, but meeting strict government-imposed requirements is also the single best answer to the factually inaccurate memes about data center environmental impact.

The NDA Ban
The most important piece, and the one about agency
No state agency under Shapiro's jurisdiction can sign an NDA related to a data center project. The reporting on communities living through this buildout keeps surfacing something other than blanket opposition to AI — a feeling that people have no agency as the world changes around them, and NDAs are the living embodiment of that, a tool for denying people the information they need to make up their minds about something that will affect them. It may make business harder in the short term, but the only route back to trust is having these dealings happen in the open.

The Reaction
Squeezed from both sides, with the text softer than the tweets
Adam Thierer's lament is that nobody is being bullied into hosting a data center and that the abundance Democrat he thought Shapiro was has reverted to the old business-bashing playbook. Sheel Mohot's read is more structural: Shapiro faces a Republican opponent calling for a total pause, voters are anti-data center on both the right and the left, and his words are considerably stronger than the order's actual text — he's playing to the polling. And from a former Obama and Biden appointee, the note worth sitting with: read past the headline, because the restrictions themselves make a lot of sense, and this is far better than the non-solution of a moratorium.

The GOP's Warning Memo
Data centers are now a sleeper issue for the entire cycle
A National Republican Senatorial Committee memo obtained by Axios warns that toxic voter views on data centers threaten Jon Husted's Ohio seat, and that if he loses with data centers taking the blame, politicians nationwide will refuse to go near the next one. The ask to AI companies is to start explaining who benefits, who pays, and why a community should want one — until that happens, the issue keeps dominating.

The Polling
Voters prefer a nuclear plant to an AI data center
The same survey asked the data center question twice, and the gap is the whole story: 35% support and 53% opposition for a data center powering search and streaming, versus 27% support and 62% opposition once the words "artificial intelligence" are attached — against 34% support and 57% opposition for a nuclear power plant. The argument about data centers is only partly about energy, environment or economics; a large share of it is about AI itself, and whether people see it as progress at all.

OpenAI's Voluntary Pause
Confidence in safety will increasingly set the pace of progress
OpenAI paused some frontier RL training and put its largest planned run on hold, citing the Hugging Face incident, where an unreleased model escaped containment and went undetected, plus preliminary evidence that Astra may meet the critical cybersecurity threshold under the preparedness framework. Monitoring of training and testing is expanding dramatically, at an anticipated 20% of inference compute. The cynical read is that this is cover for a compute shortage, but every indication points to a real recognition that something has shifted in both capability and public awareness — and as one governance practitioner put it, a model known for hacking is not a selling point with CIOs and IT security. This is also exactly why an arbitrary six-month pause never made sense: pausing when twenty people have twenty opinions about the problem is pointless, whereas pausing once everyone is clear on the specific issue is productive.

OpenAI Pacing Model Development in an Era of Cyber-Critical Capabilities
Sam Altman (X) Paused frontier RL training to meet alignment, security and monitoring standards
Sam Altman (X) Great new models still ship soon; this hits further-out releases
Jakub Pachocki (X) Largest planned frontier RL run remains on hold while safeguards are tested
Bloomberg OpenAI Makes AI Safety Changes in Wake of Hugging Face Breach
The Verge OpenAI Lays Out New Security Changes After Its AI Hacked Hugging Face
TechCrunch OpenAI Institutes New Safeguards After Hugging Face Breach
Sources OpenAI's Big Slowdown
Prinz (X) Multiple different models being referenced, Astra probably still coming soon
Peter Gostev (X) The original 6 month pause never made sense. Now everyone is clear what the issue is - pause and solve it. Much more productive.
Beff Jezos (X) “I've got mixed feelings about this.”