By Stuart Kerr, Technology Correspondent, LiveAIWire
Starting today, July 1, 2026, Reflection AI is paying SpaceX one hundred and fifty million dollars a month for access to Nvidia GB300 chips at the Colossus 2 data centre in Memphis, Tennessee. The deal, which runs through 2029 and totals approximately 6.3 billion dollars if it runs its full term, makes Reflection the fourth major external tenant at a facility that Elon Musk built for an entirely different purpose.
Colossus was constructed to train Grok, the AI model developed by Musk’s xAI. It is now the world’s largest commercial AI compute platform by committed external revenue, with four tenants — Anthropic, Google, Cursor, and now Reflection — generating more than eighty billion dollars in committed revenues through 2029. The transformation from internal model-training facility to commercial compute marketplace has happened in less than two years, and it has significant implications for how AI infrastructure will be built, owned, and monetised as the technology scales.
How Colossus Became a Commercial Platform
The Colossus supercomputer came to public attention in late 2024, when SpaceX revealed the Memphis facility as one of the largest GPU clusters ever assembled. At the time, its sole purpose was training Grok models for xAI. Within months, Musk had pivoted. Rather than use Colossus exclusively for internal workloads, SpaceX began leasing compute capacity to external organisations — a decision that has proved extraordinarily lucrative.
Anthropic signed the first major external lease, committing to approximately forty-five billion dollars through mid-2029 at the original Colossus 1 facility. Google followed with a thirty-billion-dollar commitment at Colossus 2, paying around nine hundred and twenty million dollars per month. Cursor, the AI coding tool, signed a smaller agreement. Reflection AI, a startup building open-source frontier models, completes the current tenant roster with a deal that begins today.
The aggregate committed external revenue now exceeds eighty billion dollars through 2029. That figure does not include whatever SpaceX continues to spend on Colossus for its own Grok training workloads. As a commercial infrastructure play, it represents one of the fastest pivots from internal tool to revenue-generating platform in technology history.
Understanding the scale requires context around AI infrastructure spending strategies. Microsoft raised its 2026 AI capital expenditure guidance to one hundred and ninety billion dollars. Google’s equivalent figure is between one hundred and seventy-five billion and one hundred and eighty-five billion dollars. Those are the organisations building at the frontier of AI compute. Colossus is now competing in that space, not as a hyperscaler but as a dedicated landlord to the organisations that need the most powerful compute available.
What Reflection AI Is and Why It Matters
Reflection AI is an open-source AI startup building frontier models that it intends to release publicly rather than gate behind a commercial API. The company’s decision to pay one hundred and fifty million dollars per month for compute access — more per month than many established technology companies spend on cloud infrastructure in a year — signals how serious well-funded open-source AI development has become.
The Reflection deal structure includes a notable exit clause: either party can terminate with ninety days’ notice after an initial three-month period. That provision is unusual for a commitment of this scale and suggests that both parties wanted flexibility as the AI compute market continues to move rapidly. It also means that the eighty-billion-dollar headline figure carries some optionality: tenants can leave, and new ones can arrive.
The access Reflection receives is to Nvidia GB300 chips, the current generation of GPU hardware designed specifically for large-scale AI training and inference. These are not general-purpose data centre chips. They are purpose-built for the kind of computationally intensive work required to train frontier models, and their availability has become the critical bottleneck for every organisation attempting to compete at the AI frontier. SpaceX, through Colossus, now controls access to a significant concentration of that hardware on behalf of external customers.
The Implications for Nvidia and the AI Hardware Market
One of the less obvious consequences of Colossus’s commercial pivot is what it implies for the relationship between compute access and chip manufacturers. Nvidia remains the dominant supplier of AI training hardware, and the GB300 chips at Colossus were purchased by SpaceX. SpaceX is now effectively acting as a compute intermediary: buying chips from Nvidia, building the infrastructure to run them at scale, and leasing access to organisations that either cannot or prefer not to build equivalent infrastructure themselves.
This intermediary model is not new — it is, in essence, what Amazon Web Services, Microsoft Azure, and Google Cloud do. But those hyperscalers are platform businesses serving thousands of customers at multiple price points. Colossus is doing something narrower: serving a small number of very large customers with very specific requirements for frontier AI compute.
The energy demands of AI infrastructure are a significant constraint on how many Colossus-scale facilities can exist. Memphis was chosen partly because of its power availability. As AI model training continues to require more energy per training run, the combination of physical location, power infrastructure, and chip access becomes a genuine competitive moat. SpaceX has that moat in Memphis, and the tenant lineup suggests that some of the most important AI organisations in the world agree.
What This Means for xAI and Grok
The commercial success of Colossus as a leased infrastructure platform raises an interesting strategic question for Musk’s own AI ambitions. xAI built Colossus to give Grok a compute advantage over competitors. That advantage is now being shared with Anthropic, Google, and Reflection — three organisations that are directly competing with Grok in different market segments.
The commercial logic is clear: leasing idle or surplus capacity generates revenue that helps fund continued infrastructure expansion. The strategic logic is more complex. Google, which is paying SpaceX nine hundred and twenty million dollars per month at Colossus 2, is simultaneously one of the companies Musk has publicly characterised as a competitor to xAI. The transaction is a reminder that infrastructure economics and competitive dynamics do not always align neatly.
For the AI industry more broadly, the Colossus model offers a data point on what the compute layer of AI development might look like at scale. Not every AI organisation can or should build its own data centre. The economics of frontier model training increasingly favour concentration of compute in a small number of very large facilities, with access provided on a lease or usage basis to the organisations doing the actual model development.
What This Means for You
The Colossus story matters to anyone who uses AI tools or builds products on AI models, because the compute arrangements being made today determine what models will be available and at what cost over the next several years. Anthropic’s forty-five-billion-dollar commitment to Colossus is part of what funds the training runs that produce Claude. Google’s thirty-billion-dollar commitment is part of what funds Gemini. These are not abstract financial transactions: they are the physical infrastructure decisions that shape which AI capabilities will exist.
The concentration of that infrastructure in a small number of facilities also creates risk. The SpaceX data centre ambitions extend beyond Colossus to orbital infrastructure, but the terrestrial facilities represent the near-term reality: a handful of very large compute clusters, owned by a small number of entities, serving the organisations building the models that the rest of the world uses.
That concentration is efficient. It is also a structural vulnerability — to power supply disruptions, to geopolitical intervention, to the financial decisions of a small number of individuals. The Reflection AI deal starting today adds another layer of commitment to that structure. It also adds another reason to watch what SpaceX does next with the world’s most commercially successful AI supercomputer.
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.