SpaceX’s Colossus data centre is now the world’s largest commercial AI compute platform, and Wall Street punished the company the day it proved it. On June 22, 2026, SpaceX signed its fourth external compute lease, this time with open-source AI startup Reflection AI, pushing committed external revenue at the Memphis facility past eighty billion dollars through 2029. Reflection AI, which has shipped no public product, agreed to pay one hundred and fifty million dollars a month for access to Nvidia’s newest GB300 chips starting July 1. SpaceX stock, public for barely two weeks under the ticker SPCX, fell about ten percent on the news, its worst day since its Nasdaq debut.
Colossus was built to train Grok for Elon Musk’s xAI, which merged into SpaceX in February 2026. Less than two years after that internal purpose was set, the facility has become an AI compute platform serving four outside tenants, Anthropic, Google, Cursor, and now Reflection, none of whom compete with each other in the same product category but all of whom need the same scarce thing: guaranteed access to frontier-grade chips.
How an Internal Training Cluster Became an AI Compute Platform
Anthropic signed the first major external lease on this AI compute platform, committing to roughly one point twenty-five billion dollars a month at the original Colossus 1 facility, a deal reported at approximately forty-five billion dollars through mid-2029. Google followed with a commitment of around nine hundred and twenty million dollars a month at Colossus 2, described as bridge capacity while its own data centres catch up with demand. Cursor, the AI coding tool, signed next, and SpaceX has since moved to acquire Cursor outright for sixty billion dollars in stock.
Reflection AI, founded in 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou, completes the current roster on the AI compute platform with a deal that could reach 6.3 billion dollars if it runs its full three-year term, according to TechCrunch’s reporting on the agreement.
The contract includes an unusual exit clause: either party can terminate with ninety days’ notice after the initial three months, a flexibility that lets both companies hedge against an AI compute platform market moving this fast. Understanding the scale of these commitments requires context around broader AI infrastructure spending strategies across the industry: Microsoft has raised its 2026 AI capital expenditure guidance to roughly one hundred and ninety billion dollars, and Google’s own figure sits between one hundred and seventy-five and one hundred and eighty-five billion. Colossus is now competing in that league, not as a hyperscaler building for itself but as a dedicated landlord to the handful of organisations that need the most powerful compute on the planet.
Why a Pre-Revenue Lab Signed a $1.8 Billion Annual Bill
Reflection AI has never shipped a public frontier model, yet it raised two billion dollars at an eight billion dollar valuation in October 2025 and was reportedly seeking a new round near twenty billion dollars by this spring. The company is building toward an explicitly national security posture, with reported ties to the Department of Energy’s Genesis Mission and Pentagon AI programmes, positioning itself as an open-weight American alternative in a field where Chinese open-source models have gained wide adoption among developers.
That framing matters for why a company with no revenue would commit nearly one point eight billion dollars a year to an AI compute platform before it has a product to sell. As Forbes contributor Jon Markman put it in his analysis of the deal, a lab will only sign a bill like that if it believes capacity, not capital, is the scarce constraint, a dynamic detailed in Forbes’s coverage of the transaction and SpaceX’s stock reaction. The compute comes first for Reflection; the model comes second.
Why the Stock Fell on What Looked Like Good News
SpaceX went public on June 11, 2026, at a valuation near one point seventy-seven trillion dollars, a number professional investors spent the roadshow calling detached from a business still losing money on rockets and satellites. Ten days later, the same company booked a multibillion-dollar recurring compute contract, and its stock dropped anyway. That reaction says less about the Reflection deal than about how difficult conventional valuation models find it to price a business that is simultaneously an aerospace company, a satellite operator, and now the landlord for the AI industry’s most sought-after hardware.
The scepticism has a specific author. Investor Michael Burry, who closed his fund earlier this year, has been warning publicly that AI capital expenditure sits dangerously concentrated on four balance sheets, Microsoft, Google, Amazon, and Meta, and has placed bets against Nvidia and the broader chip index on that thesis. Reflection’s willingness to queue up behind established hyperscalers for the same GB300 chips complicates that argument: the AI compute platform buyer base is widening to include venture-funded labs and government-adjacent programmes that a four-company concentration story does not capture.
The Physical Constraint Behind the Financial Story
None of this works without power. The energy demands of AI infrastructure are a hard limit on how many Colossus-scale facilities can exist anywhere, and Memphis was chosen specifically because of its available grid capacity. Satellite imagery has reportedly shown cooling equipment at Colossus 2 capable of managing considerably less than the one gigawatt of capacity Musk has claimed, and the company was said to be deploying natural gas turbines on site as recently as May 2026 to close that gap. The combination of physical location, power access, and chip supply is what turns Colossus into a genuine competitive moat rather than just a large building full of GPUs.
What This Means for You
If you use an AI product built on Claude, Gemini, or an open-weight model from a lab like Reflection, the compute arrangements being signed on this AI compute platform this year are part of what determines which capabilities exist and at what cost over the next several years. Anthropic’s commitment funds part of what trains Claude. Google’s funds part of what trains Gemini. These are not abstract financial transactions sitting on a balance sheet somewhere; they are the physical infrastructure decisions shaping which AI systems the rest of the world gets to use.
The concentration of that infrastructure in a handful of facilities, owned by a small number of entities and now including SpaceX’s own orbital data centre ambitions, is efficient in the way any specialised infrastructure is efficient. It is also a structural vulnerability, to power supply disruption, to the financial decisions of a small number of individuals, and to whatever happens the next time a tenant’s ninety-day exit window actually gets exercised rather than quietly renewed. Every unexercised exit window between now and 2029 is itself a data point on whether this AI compute platform represents durable demand or a very expensive bet still waiting to be tested.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, emerging technology, and their impact on business, society, and geopolitics. LiveAIWire publishes daily AI news and analysis at liveaiwire.com.