Today’s news runs along three fault lines: AI revenue is setting records while compute supply, not demand, is the binding constraint; the open-weight stack is reshaping what frontier models can charge; and the labor-market cost of all of it is starting to show up in the data.
Records, Chip Deals, and a Forced Reversal
The money moved hard this week: a coding startup at a $3 billion run-rate, a fresh quarterly revenue figure for OpenAI, a fourth compute supplier for Anthropic, and a completed acquisition that regulators are now ordering undone.
- Cursor reaches $3B annualized revenue with a SpaceX call option looming — Cursor’s run-rate crossed $3 billion in late April with 3,000+ customers paying $100k or more. SpaceX holds a call option to buy the company for $60 billion once its own shares begin trading around June 12.
- OpenAI topped $5.7B in Q1 revenue, ahead of Anthropic for that quarter — OpenAI’s Q1 beat Anthropic’s $4.8 billion, but Anthropic projects $10.9 billion for Q2, so neither lab holds a durable revenue lead. Both are scaling faster than their compute capacity can keep up.
- Microsoft in talks to supply Maia 200 chips to Anthropic — A Maia deal would give Anthropic a fourth compute supplier alongside Amazon, Google, and SpaceX. Microsoft claims a 30% efficiency edge, though the baseline for that figure is undefined and no deal is signed.
- Manus founders weigh $1B raise to reverse Meta acquisition — Beijing has ordered Meta’s purchase of the AI agent startup unwound. The founders are raising roughly $1 billion to buy it back. A post-completion forced reversal at this scale is nearly without precedent.
The Real Question Is What AI Costs
Three analyses today converge on the economics underneath the headlines: where falling prices actually come from, whether the compute buildout can continue, and what kind of moat a hyperscaler really has.
- AI’s falling prices are a software story, not a hardware one — Software efficiency, including quantization, distillation, and inference optimization, is the main force compressing AI costs. Open-weight models on commodity hardware are now eating the low-to-mid tier of the market.
- AI compute is growing faster than frontier labs can absorb it — Epoch AI estimates the top labs hold under 30% of global AI compute today, but their growth could push that to 80% within five years. Past that point, scaling needs an economic transformation to justify the bill.
- Google’s AI moat is distribution, but the Search ad risk is real — Google is shipping Gemini across Search, Workspace, Android, and YouTube on the strength of its install base. The structural tension: every AI answer that satisfies a query without a click erodes the ad inventory that funds it.
Models, Agents, and the Machinery Underneath
A new agent-foundation model, a caution about how we read what models represent internally, and an engineering breakdown of the runtime that long-running agents keep converging on.
- Qwen3.7-Max arrives as a closed-weight agent foundation model — Alibaba’s Qwen team claims benchmark leadership across Terminal-Bench, SWE-Pro, SciCode, and four more, plus consistent performance across agent harnesses. The notable shift: the weights that made Qwen a developer default are now closed.
- Goodfire finds SAE features are map fragments, not atoms of meaning — Sparse autoencoder features each capture only a slice of curved neural geometry, so a single feature is not a clean unit of meaning. Teams relying on feature-by-feature interpretability may be reading a confident but partial story.
- Cursor’s cloud-agent architecture converges on four durable primitives — After a year building cloud agents, Cursor’s architecture (isolated environments, durable execution, state separation, self-healing infrastructure) matches what Google independently shipped in Agent Executor. That convergence is the story.
AI and the Shape of Work
Two data points landed today that describe the same shift from opposite ends: AI is now writing most of the code, and the entry-level jobs that used to teach people to write it are thinning out.
- State of Web Dev AI survey: 56% of developer code is now AI-written — A 7,258-developer survey found the share of AI-generated code doubled year over year, from 28% to 56%. Claude is the model developers pay for most, and monthly AI spend per developer is climbing as subsidized pricing unwinds.
- The entry-level job gap is the AI benchmark that actually counts — The unemployment gap between entry-level and experienced workers has widened sharply in AI-exposed occupations. Cutting junior roles now is a deferred talent-debt trade, and the bill arrives when AI workflows hit the limits of judgment only tenure produces.
Today’s Quick Hits
- Microsoft pulls Claude Code licenses, steers developers to Copilot CLI — Microsoft’s Experiences and Devices division must switch off Claude Code by the end of June, with fiscal year-end cost-cutting cited as a driver. Microsoft is also a reported $5 billion investor in Anthropic, which makes the move a clean separation of investment posture from internal tooling.
- Anthropic’s consulting arm makes its first acquisition — Anthropic’s still-unnamed consulting venture has taken Fractional AI as its operational centerpiece, a move toward capturing the enterprise deployment-services margin that Accenture and Deloitte have long owned.