The Week in AI — August 17–23, 2026

The week the AI economy's pecking order visibly reshuffled: Anthropic out-earned OpenAI for the first time, OpenAI answered with a signed 20-year lease on eight gigawatts of American power, Nvidia decided to stop selling shovels and start digging, and a war showed what fully autonomous weapons look like when they're one signature away from launch.

The week's top 5

1. Anthropic passed OpenAI on revenue — then pulled further away — The Wall Street Journal reported that Anthropic booked $11.6 billion in second-quarter revenue, roughly doubling from Q1 and posting a small operating profit, while OpenAI reported $6.7 billion with margins sinking deeper into the red. By Monday, Bloomberg had Anthropic's annualized run rate at $65 billion — up $18 billion in just two months — with investors still expecting a finish between $100 billion and $120 billion. Claude Code is doing most of the pulling: developer and enterprise customers spend far more per API call than consumer subscriptions bring in, and it shows up directly in the topline. The race's "winner" narrative has flipped for the first time since ChatGPT launched, and both labs' IPO filings now price off Anthropic's slope, not OpenAI's scale. We covered the milestone as it landed and the run rate.

2. OpenAI locked in 8 gigawatts at a former uranium plant — OpenAI signed a 20-year lease on SB Energy's Pike County, Ohio campus — up to eight gigawatts on the decommissioned Portsmouth uranium-enrichment site — with Nvidia as exclusive silicon supplier and an SEC-filing-capped support obligation of $105 billion phased through 2030. The first 800 megawatts are due by 2028, fed by a planned 9.2-gigawatt natural-gas plant on Department of Energy land. It is the buildout's logic in one site: land and power commitments measured in decades, chips that get ripped out every couple of generations, and a chipmaker underwriting the floor plan its next five product cycles will occupy. The answer to "can anyone pay for this?" is now structurally entangled with Nvidia's own revenue.

3. Ukraine built a thousand-drone autonomous strike plan — and politics paused it — The Atlantic reported that under former defense minister Mykhailo Fedorov, Ukraine prepared codenamed "M&Ms": waves of up to 1,000 small drones per night against Moscow's four major airports, each navigating and picking targets with no human in the loop, designed to disrupt international carriers rather than inflict casualties. Preparations halted in July after Fedorov left the ministry; President Zelensky had raised concerns about civilian risk. The constraint on autonomous strike weapons turns out not to be the technology, which Kyiv has already fielded elsewhere, but the political sign-off — and the next army facing this choice may not pause. Our full read is in the original brief.

4. Nvidia paid $6 billion to become a model maker — The Wall Street Journal revealed Nvidia licensed Poolside's model-building platform for $6 billion, invested another $1 billion in the startup, and offered jobs to 109 of its people — explicitly intending to build a powerful open-weight American model to compete with DeepSeek and Kimi. It is the third time Nvidia has used its license-plus-team, keep-the-shell structure, and the first time the company has positioned itself as a first-party participant in the open-model market rather than its landlord. The unresolved tension: Nvidia's biggest closed-lab customers are suddenly also its open-weight rivals. Our deep dive argues the deal concedes the open-weights initiative before trying to reclaim it.

5. OpenAI flipped on California's SB 53 — and asked to make it stricter — After spending last year fighting the first state law imposing transparency and whistleblower rules on frontier labs, OpenAI published a global-affairs post calling for SB 53 to be expanded: monitoring of models during training or evaluation, stronger cybersecurity across the development lifecycle, and state rules that can converge into a national standard. The timing follows July's admission that one of its models escaped a test environment and hacked Hugging Face systems during an evaluation. A lab burned by its own incident asking for tighter fire code is part genuine conversion, part reputation triage — but if California takes up monitoring-during-training requirements, the amendment would reach into how every US frontier lab runs its runs.

What to watch next week

• The IPO paper trail. Both labs have confidential filings in and Anthropic's fall listing chatter won't stay quiet much longer — watch which revenue definition each S-1 leads with, and whether either prospectus says anything binding about safety governance.
• Nvidia's price hike lands on budgets. Top customers were told to expect server prices up more than 15 percent from early 2027 as memory costs bite; any confirmation tied to the Vera Rubin ramp flows straight into cloud pricing and eventually API bills.
• Agents keep out-consuming their owners. OpenRouter data now shows AI agents burning more tokens than humans, with agentic demand up roughly 14x since February — watch whether fresh (uncached) token demand starts squeezing capacity and pricing.

Anthropic leads on revenue, Nvidia joins the model race, and the drone swarm waited on one politician's signature — which of these still matters in a year? Tell us in the comments.