Elon Musk’s SpaceX has signed a compute deal worth up to $6.3 billion with Reflection AI, an open-weight AI startup founded by former Google DeepMind researchers. Starting July 1, Reflection will pay $150 million a month for access to Nvidia’s latest GB300 chips at the Colossus 2 data center near Memphis — the same facility that already houses lucrative leases with Anthropic and Google. The difference: Reflection is betting the entire deal on open-weight AI as the answer to a frontier lab landscape that just got fenced in by export controls.
It is the largest open AI infrastructure commitment announced to date, and it arrives at a moment when the closed-versus-open model debate has stopped being academic. The US government’s ban on Anthropic’s Fable 5 and Mythos models has made open-weight alternatives look less like an ideological choice and more like a hedge. Reflection is the first company to put billions behind that hedge.
The big signal: SpaceX becomes the compute landlord for open-weight AI
The Colossus data center was originally built by xAI for Musk’s own AI ambitions. When those internal efforts faltered, SpaceX converted the facility into what is effectively a GPU rental operation — and a remarkably profitable one. Anthropic pays $1.25 billion a month. Google pays $920 million. Reflection’s $150 million is smaller, but the strategic positioning is different. Reflection is not renting compute to build a closed frontier model. It is renting compute to train open-weight models that anyone can download, inspect, and run locally.
The deal runs through 2029, though either side can cancel with 90 days’ notice after the first three months. That cancellation clause matters. It tells you that even at $6.3 billion, nobody is making a binding three-year bet on the current state of the AI hardware market. The chips, the economics, and the regulatory environment are all moving too fast for that. What Reflection is really buying is immediate access to GB300 silicon — the kind of access that open-weight labs have historically struggled to secure because the biggest compute clusters have been locked up by closed frontier labs willing to pay more.
The timing is not subtle. Reflection explicitly tied the deal to the Fable 5 export control fallout, saying in a statement that “recent events highlight how important open source is to the AI ecosystem, with more nations and enterprises recognizing the risks and costs associated with exclusively depending on closed models.” When a startup founded in 2024 can land a multi-billion-dollar infrastructure deal partly because the government banned its competitor’s product, the competitive dynamics of frontier AI have shifted in a way that deserves attention.
Groq pivots from chips to cloud after Nvidia’s $20B licensing deal
While Reflection bets on open-weight models, Groq — the inference chipmaker whose LPU technology Nvidia licensed in a December deal worth a reported $20 billion — confirmed a $650 million funding round on Monday. The round was led by Disruptive (whose founder Alex Davis serves as Groq’s chairman) and Infinitum. It comes roughly six months after Nvidia poached Groq’s founder and CEO Jonathan Ross, president Sunny Madra, and other key employees in a transaction that paid investors handsomely but left the company without its core leadership.
Groq’s response to losing its founder and its core IP is a pivot to neocloud services. Doug Wightman, the remaining co-founder, is now CEO. The company says it operates 13 data centers across North America, Europe, the Middle East, and APAC, serving over five million developers and processing trillions of tokens weekly. It has added Alan Rice (formerly xAI and Meta) as COO, and brought in Sinclair Schuller as CTO and Rakesh Malhotra as CPO — both veterans of the enterprise cloud software world.
The subtext is that Nvidia now owns the LPU IP and has launched its own Groq 3 LPX inference hardware system. Groq is competing against the company that licensed its technology, using a business model (inference cloud) that Nvidia also dominates. Whether a neocloud can survive in a market where the largest GPU provider is also the incumbent inference platform is an open question — but $650 million buys enough runway to find out.
DeepMind goes to Hollywood, OpenAI goes to open-source security
Google DeepMind announced a $75 million investment in A24, the indie studio behind “Everything Everywhere All At Once” and the recent blockbuster “Backrooms.” Demis Hassabis framed it as a “first-of-its-kind” partnership to build AI tools for filmmaking, with DeepMind receiving “feedback and guidance from leading artists.” This follows Netflix’s acquisition of Ben Affleck’s AI filmmaking company InterPositive earlier this year and Amazon MGM’s internal AI unit. The frontier labs are no longer just competing for developers — they are competing for creative industry partnerships as a distribution channel for their models.
OpenAI, meanwhile, launched “Patch the Planet,” an initiative that pairs security firm Trail of Bits with open-source maintainers to find and patch vulnerabilities before they become exploitable. OpenAI’s Codex Security tools will assist in the process. The program is a direct counter-narrative to the concern raised by Anthropic’s Mythos security tool — that AI can now automatically identify bugs in codebases and generate exploits. OpenAI is turning that capability defensive, and it is hard not to read the move as a competitive positioning play against Anthropic while addressing a genuine need: open-source projects are the foundation of commercial software, and they are chronically under-resourced on security.
Open-source watch
Agentic loops go mainstream. Boris Cherny, the creator of Claude Code, told Meta’s @Scale conference that “loops are just as important and as big a step” as the shift from hand-written code to AI agents. The idea: instead of managing agents through discrete check-ins, you authorize persistent agent loops that run continuously — one refactoring architecture, another unifying duplicated abstractions, both submitting pull requests like any other contributor. The stopping condition is non-deterministic, which is a polite way of saying the model decides when it is done. This is the architecture that makes always-on agentic systems possible, and it is moving from experiment to production faster than expected.
Ollama v0.30.10 shipped on June 17, continuing a near-weekly release cadence (v0.30.8 on June 12, v0.30.9 on June 15). vLLM v0.23.0 landed on June 15, the first stable release since v0.22.1 on June 5. Neither is a landmark release, but the sustained cadence matters: local inference and open model serving are getting consistent engineering attention, which is what makes practical local AI deployment viable for small teams.
What builders should take from this
Three things are happening simultaneously, and they matter for anyone building with AI right now.
Compute access is becoming the real moat — and open-weight labs just got a seat at the table. The Reflection deal proves that the export control fallout is not just a regulatory story. It is reshaping who can buy compute and why. If you are building on open-weight models, the infrastructure story behind those models just got more serious.
Inference economics are diverging from training economics. Groq’s pivot to neocloud services and Nvidia’s launch of its own Groq 3 LPX hardware signal that the inference market is where the next competitive battle will play out. For teams running agents in production, the cost and availability of inference compute — not just raw model quality — is becoming a strategic decision.
Agentic loops change the operating model. If agents prompting agents in continuous loops is becoming production-grade, the way you design guardrails, observability, and human oversight has to change. You cannot manually review every PR from an agent that never stops running. The approval layer, the rollback path, and the cost monitoring all need to account for systems that operate autonomously and indefinitely.
The practical takeaway
The AI infrastructure story of the day is not a model launch or a benchmark. It is a $6.3 billion compute deal that signals open-weight AI has arrived as a serious infrastructure buyer — not because open-weight models caught up on capability, but because the regulatory environment made closed-model dependency a risk. When the government bans a frontier model, the compute does not disappear. It finds a new tenant with a different thesis.
For investors, the signal is that open-weight infrastructure is now a funded thesis at frontier scale. For builders, it means the open models you are integrating today may have significantly more compute behind them tomorrow. For anyone running agentic systems, the loops are coming — and your oversight architecture needs to be ready before the agents stop waiting for you to check in.
Leave a Reply