Reflection Pre-Bought $6.3 Billion of Colossus Compute Without a Shipped Model. The Open-Source Frontier Just Got a Procurement Story.

The CNBC scoop landed Monday and the rest of the wires confirmed it inside an hour: Reflection AI signed with SpaceX for access to Nvidia GB300 capacity at Colossus 2, paying $150 million a month starting July 1, 2026, for an initial three-month period and then rolling on 90-day termination notice through 2029. Run the term out and the deal totals roughly $6.3 billion. Reflection is the open-source frontier lab last valued at $25 billion, founded by ex-Google DeepMind researchers Misha Laskin and Ioannis Antonoglou, with Department of Energy Genesis Mission and Pentagon AI work already on the customer list. It has not yet shipped a publicly available frontier model.

That last sentence is the one that matters. The price tag is hyperscaler-tier. The customer is pre-revenue at the model layer. Reflection is doing what Anthropic did in 2023 and what OpenAI did in 2022: prepaying for capacity on a forward conviction, not on a contracted revenue line. The difference is that Reflection has committed to open-weight releases. The compute floor for a credible open-source frontier in 2026 just got priced, and the number is bigger than every Mistral funding round combined.

The Numbers

$150 million a month works out to $1.8 billion of run-rate revenue for SpaceX from one customer alone. For context, that is roughly the level CoreWeave was running across its entire book at the start of 2024. The pricing per GPU is not disclosed, but if you assume Colossus 2 stands up the publicly reported 550,000-GB300 footprint and Reflection gets a slice in the 30,000 to 50,000 GPU range, the implied per-GPU-hour rate lands in roughly the same neighborhood as bare-metal H100 rentals at the largest neoclouds did 18 months ago. Nvidia, sitting on both sides of the trade (selling the chips to SpaceX, holding equity in Reflection from the prior round), is the quiet winner whose share count never shows up in either headline.

SpaceX Is the Neocloud Now

Run the Colossus customer book in chronological order and the pattern is unmistakable. In May we wrote up Anthropic booking 220,000 GPUs at Colossus 1 under a multi-year deal that the wires now value around $45 billion through mid-2029. Google followed with roughly $30 billion of Colossus commitments on a similar window. SpaceX acquired Cursor for $60 billion all-stock on June 16, locking the largest independent coding workload to its own substrate. Six days later, Reflection signed for another $6.3 billion. The contracted demand on Colossus is north of $80 billion before the second Memphis hall finishes its first commissioning window.

What is interesting is the diversity of the customer list. Anthropic is a closed-weights frontier lab. Google runs its own silicon program. Cursor is an application-layer coding tool. Reflection is open-source. The only common factor is that they all wanted gigawatt access on a 2026 to 2029 window that the Big Three hyperscalers could not unbundle from a full-stack contract. Colossus is doing the Equinix move at the AI layer: stay neutral, take any customer, sell the floor space, let the customer worry about the workload. The difference is that the floor space here costs a billion dollars to provision and the electrons cost almost as much to deliver.

For the hyperscalers, this is the first time SpaceX shows up as a structural threat rather than a sidecar. Anthropic and Google both have their own primary clouds (AWS and Google Cloud, respectively) and are still leasing Colossus on top. That is procurement diversity on a scale the Big Three would normally be able to discount their way out of. They have not, which tells you that Colossus is offering something the hyperscalers cannot match in this window: gigawatt-class delivery on the calendar customers actually need, on a contract that is not bundled with a managed-service tax.

A $25B Lab With No Model

Reflection is the most aggressive open-frontier bet in the US. The pitch is that the weights ship publicly, the revenue comes from enterprises and governments running the models on their own infrastructure, and the moat is the model quality plus a Pentagon- and DoE-cleared deployment posture. The pitch has not been validated by a shipped artifact. The closest comparable on the open side is Z.ai shipping GLM-5.2 on Huawei Ascend silicon earlier this month, and the closest closed comparable is DeepSeek V4 continuing to push the price floor down. Neither of those labs has paid hyperscaler rates for compute. DeepSeek runs on rented capacity at a fraction of the GB300 price. Z.ai runs inside a state-supported Huawei stack. Reflection is the first open-source lab to put a contracted American compute commitment on its balance sheet at the same scale a closed frontier lab uses.

The cleanest way to read the gap is that open weights at the frontier are no longer structurally cheap. The early DeepSeek wave priced the open-source frontier near zero on marginal capex; the V3 paper put pretraining at about $5.6 million, and the assumption downstream was that anyone who copied the recipe could compete. Reflection is betting the recipe is no longer enough, that the next jump (long-horizon agentic reasoning, native tool use, very long context) requires the same gigawatt commitment a closed lab needs. $6.3 billion is the public bid for that thesis. If Reflection ships a model in 2027 that clears Claude or GPT on agentic benchmarks while keeping weights open, the bid was correct. If it does not, $1.8 billion a year of GB300 burn against zero shipped product is a Series-killer.

The Pentagon Variable

The piece that does not show up in the deal price but shapes the strategic value is Reflection's clearance posture. The lab is already shipping into DoE Genesis Mission and Pentagon AI procurement, and Colossus is a US-domiciled data center with a security envelope that an export-controlled customer can actually sign up to. Compare the picture to last week, when Washington pulled Fable 5 and Mythos 5 from non-US users in 90 minutes. An open-weights model trained inside a Pentagon-cleared facility, with the weights released to friendly jurisdictions but the training infrastructure firmly inside a US security perimeter, is exactly the deliverable the national security side of the administration has been asking for since the GLM-5.2 release.

Read that way, Reflection is the official answer to the question DeepSeek and Z.ai have been forcing for nine months: can the US ship a credible open-source frontier without ceding the training stack to China? The $6.3 billion compute lease, the Pentagon contracts already on the books, and the $25 billion valuation against zero shipped artifact all point to a procurement bet that the answer is yes, and that there is enough policy demand to underwrite a closed-stack training pipeline for open-weight outputs. The contradiction is the point. Closed compute, open weights, US-cleared.

Our Take

The most under-appreciated number in this story is the 90-day notice clause after the initial three-month period. Reflection can walk in October if a Series-extension does not land, and SpaceX can re-lease the capacity to whoever is next in the queue (the queue is long). The $6.3 billion headline is the upper bound of a contract that is structurally a three-month commit with quarterly renegotiation rights. The wires are reporting the ceiling because the ceiling makes the better headline. The floor is closer to $450 million for the first quarter, and that is the only number the Reflection board has actually underwritten.

For builders, the practical read is that the open-source tier of the model market is bifurcating. On one side, capital-efficient open weights coming out of China on state-adjacent compute (DeepSeek, Z.ai, Alibaba) at price points the wires keep underestimating. On the other side, capital-intensive open weights coming out of the US on Pentagon-cleared compute (Reflection, plausibly others soon) at price points that look structurally similar to closed labs. The token price floor we have been tracking on our pricing-floor piece is going to keep falling because the Chinese open tier keeps shipping. The procurement floor for any model labeled American open frontier is moving in the opposite direction.

Three signposts in the next ninety days. First, the October termination window: if Reflection extends past 90 days, the bid is real and the Series-extension closed. If it does not, the lease becomes the most expensive recruiting line item in the industry. Second, the first public model release: Reflection has guided to a 2026 open-weight ship, and the gap between the GB300 burn rate and the artifact ship rate is the only metric that matters. Third, whether any other neocloud lands a frontier-tier lease with a customer the Big Three would have expected to absorb. If CoreWeave, Crusoe, or a Saudi-anchored counterpart books a 2027 gigawatt-scale customer on similar terms, the SpaceX-as-neutral thesis stops being a SpaceX story and starts being a structural change in how frontier compute gets contracted. The first two answers come from Reflection. The third one decides whether the neocloud category survives the post-2027 capex reset.