A daily brief from the AI frontier. This week the AI story stopped being about models and started being about money and megawatts: a rocket company turned itself into a cloud landlord, a chip-stock rout reopened the "is this a bubble?" debate, and a photo giant that once sued AI firms decided to get paid by one instead. The through-line — compute, capital and content are now the contested ground, and everyone is repricing them at once.
SpaceX becomes an AI landlord
The single most telling deal of the week didn't come from a model lab. SpaceX signed an agreement worth up to $6.3 billion to rent compute from its Memphis "Colossus 2" data center to Reflection AI, an open-source startup founded by former Google DeepMind engineers. Reflection will pay roughly $150 million a month from July 2026 through 2029 for immediate access to Nvidia's top-end GB300 chips, CNBC reported. With Anthropic, Google and Cursor already on its client list and committed outside revenue now topping $80 billion through 2029, SpaceX has quietly become one of the largest third-party compute platforms on earth. The lesson: in 2026, owning the megawatts can be a better business than owning the model.
The chip rout that reopened the bubble debate
Then the bill came due — at least in investors' heads. A debt-and-Fed-driven repricing of AI capex sent the Nasdaq down about 2.2%, Micron roughly 13%, and the Philadelphia Semiconductor Index nearly 8%, while Korea's KOSPI tripped a circuit breaker in one of the sharpest single-session chip selloffs on record, NPR reported. Unlike earlier wobbles, this leg was specifically about funding: hyperscalers building data centers with borrowed money, against a Fed in no hurry to cut. The market's question has shifted from "how big can AI get?" to "who pays for the capex, and what happens if the returns lag the spending?"
Qualcomm takes the fight to Nvidia's software moat
If the rout was about who funds the hardware, Qualcomm made a move about who controls the software on top of it. The chipmaker agreed to buy Modular for around $4 billion, acquiring the hardware-agnostic Mojo language and MAX inference stack — and Swift and LLVM creator Chris Lattner — in a bid to loosen Nvidia's CUDA grip on millions of developers, TechStartups reported. CEO Cristiano Amon called it a "pivotal moment." It is a different bet than buying more silicon: a credible, portable alternative to CUDA is arguably the single biggest variable in whether any non-Nvidia accelerator ever wins real workloads.
Getty stops suing and starts selling
The week's most surprising turn came from outside the AI labs entirely. Getty Images, which had taken generative-AI companies to court, announced a multi-year display partnership with OpenAI to surface more than 400 million licensed assets — premium sport, entertainment and news editorial, plus iStock content — across ChatGPT's search and discovery experiences, Getty said. Investors loved the pivot from litigation to licensing: Getty's stock roughly doubled on the news. It is a template other rights-holders will study — a revenue-share that pays content owners for showing up inside AI products rather than fighting them.
The agents move into Slack
Finally, the enterprise-AI fight relocated to where work actually happens. Anthropic launched Claude Tag, replacing its old Slack app with a persistent, shared AI teammate that learns a company, watches channels and acts on its own — governed by admin token caps and per-action audit logs, TechCrunch reported. In the same week, OpenAI gave ChatGPT the ability to take real actions inside Slack. Two frontier labs racing to embed always-on agents in the same collaboration surface marks a shift from personal chatbots toward shared "virtual employees" — and a quiet challenge to every piece of seat-priced workflow software an agent can absorb.
The money view, and what to watch
The connecting thread is repricing: compute is being valued as the scarce asset (SpaceX's lease), capital is getting pickier about who funds it (the chip rout), the contest is climbing the stack toward software (Qualcomm/Modular), and content is becoming a paid input rather than a lawsuit (Getty/OpenAI). Watch three things next: whether the selloff deepens or steadies as megacaps report capex; whether Modular/Mojo and other open stacks make real dents in CUDA; and whether more rights-holders follow Getty into license-don't-sue deals. Signal, not advice; no live prices.