Summary: TechCrunch reported on July 14, 2026, that open-weight startup Reflection AI agreed to a $1 billion compute deal with Nebius for access to Nvidia’s latest chips, with exact chip models and quantities undisclosed. The deal is Reflection’s second mega compute commitment in three weeks, after a SpaceX contract costing $150 million per month and worth up to $6.3 billion. For Reflection, the purchases turn capital into long-lived training and inference capacity, though the full supplier terms remain undisclosed.
TechCrunch reported on July 14, 2026 that Reflection AI — a startup building open-weight models — has struck a $1 billion compute agreement with Nebius, the European AI infrastructure company, securing access to Nvidia’s latest chips. That makes two mega-scale compute deals inside three weeks for Reflection: in late June it committed to a SpaceX contract paying $150 million per month, worth up to $6.3 billion in total.
For a company that only came into being in 2024, that cadence captures the rulebook of 2026: the models can be open, the compute cannot. Getting frontier-grade open-weight models trained starts with locking down a place in the supply chain.
What the Billion Dollars Actually Pays For
At its core, the deal buys access to Nvidia’s latest chips, delivered through Nebius’s AI cloud; the exact chip models and quantities went undisclosed. The scale can still be sized up against Nebius’s contract history: this past March, it agreed to a five-year deal with Meta worth up to $27 billion, committing to deliver one of the first large-scale deployments of the Nvidia Vera Rubin platform beginning in early 2027.
Put differently, Reflection is not purchasing “servers.” What it has bought is a years-long guarantee of training and inference capacity — and in a market where GPU lead times run in years, a long-dated contract of this kind is a moat in itself.
Reflection: A Second Mega-Deal in Three Weeks
The company was started in 2024 by a pair of ex-Google DeepMind researchers, styles itself an open-weight lab, and has taken in nearly $2.6 billion — Nvidia, Sequoia, and Lightspeed sit among its investors — at a last-known pre-money valuation of $25 billion.
Its compute position now rests on two pillars. The SpaceX agreement unveiled on June 22 rents Nvidia GB300 chips at the Colossus 2 facility near Memphis, Tennessee for $150 million per month through 2029, with either party free to walk away on 90 days’ notice once the first three months are over — Reflection described it as one of the largest announced open AI infrastructure commitments to date. On top of that sits the fresh $1 billion from Nebius. For perspective: on the same SpaceX compute resources, Anthropic’s contract runs $1.25 billion per month and Google’s $920 million per month. Measured against those, Reflection’s commitments look far smaller, yet for a startup they already amount to a heavy bet — and the company puts it this way: “more compute means more runway to build the world’s best open models at scale.”
Nebius: GPU Cloud’s Rising Heavyweight
Nebius was carved out of Yandex and now trades publicly as an AI infrastructure company. Within a single year its customer list has grown thick: the five-year Meta agreement worth up to $27 billion, a multi-year Microsoft contract signed last year worth up to $19.4 billion, and a $2 billion investment from Nvidia announced in March — with the shared goal of putting more than 5 GW of Nvidia systems into service by the end of 2030. Even with the Meta agreement in hand, Nebius said its 2026 guidance was unchanged, a hint at how much headroom it believes remains. The Nvidia tie-up covers the full stack, from AI factory design to fleet management, and brings early access to platforms such as Nvidia Rubin, Vera CPUs, and BlueField storage.
The setup deserves a closer look. Nebius sells no chips and builds no closed models; it takes the role of the “neutral GPU cloud” — in CEO Arkady Volozh’s framing, one of the first large-scale clouds purpose-built for AI developers. Frontier labs are one another’s competitors and hyperscale clouds each carry their own loyalties; in that landscape, an independent infrastructure provider ends up as the supplier every party can accept.
What This Says About the Compute Market
Three takeaways. First, mega compute contracts are becoming the default — billion-dollar sums, multi-year terms, named chip generations — and the buyer roster now stretches from giants to valuation-backed startups. Second, the open-weight route competes only as well as its compute allows: Reflection is swapping raised capital for GPU hours, wagering that open-model quality is, in the end, a function of scale. Third, supply keeps spreading out — SpaceX leases its own facilities, Nebius widens the neutral cloud, hyperscalers pour more into self-builds — so buyers, for the first time, truly hold options. The caution for infrastructure teams is plain: these contracts tie up capacity for 2027 through 2029. The shortage you are looking at today was settled two years ago.