Anthropic confidentially filed for a US listing on 1 June, and reporting since has put the eventual valuation anywhere from $965 billion to $2 trillion, depending on who’s counting and when.
The more consequential question sitting inside Anthropic’s pre-IPO paperwork is a strategic one: is the company still a tenant of hyperscaler compute, or has it quietly become a builder and owner of physical infrastructure in its own right?
The answer, based on the deals stacking up over the past six months, is increasingly the latter. And that shift has implications for colocation operators, GPU cloud lenders and power markets well beyond San Francisco.
From tenant to landlord
Until recently, Anthropic’s compute story was a rental story: capacity from Amazon, Google, Microsoft and, more unusually, from Elon Musk’s SpaceX. That’s changed. In June, Apollo and Blackstone structured a $35 billion chip-backed special purpose vehicle for Anthropic, with the first gigawatt of capacity deployed through UK neocloud provider Fluidstack rather than any of Anthropic’s existing hyperscaler partners. Capacity’s own reporting on the deal flagged this as the detail that mattered most: a company with an $11 billion AWS campus already running in Indiana and multi-cloud access to Google TPUs and Microsoft Azure chose a smaller, dedicated infrastructure partner instead, prioritising delivery certainty and custom-built facilities over an established balance sheet.
That wasn’t a one-off. Anthropic has since pledged $50 billion to US computing infrastructure, building purpose-designed sites with Fluidstack in Texas and New York, work expected to generate 800 permanent jobs and 2,400 in construction. Dario Amodei, Anthropic’s chief executive, framed the spend as necessary for infrastructure that can “support continued development at the frontier”. Fluidstack’s own leadership were similarly direct about what the arrangement unlocks, telling Capacity the firm is proud “to partner with frontier AI leaders like Anthropic”.
Layer that onto a $150 billion chip agreement with Google, and Anthropic starts to look less like a lab that buys compute and more like an infrastructure company that happens to train models. For colocation and wholesale operators, that’s the whitespace worth watching ahead of any IPO roadshow: if the biggest, best-funded AI labs are increasingly building rather than renting, does the addressable market for the world’s data centre landlords shrink just as the capital to expand it becomes cheaper and more plentiful than ever?
The circularity question
The second thread running through Anthropic’s pre-IPO infrastructure is harder to untangle, and it only became visible because of somebody else’s paperwork. When SpaceX filed its own IPO documents in May, they confirmed that Anthropic will pay $1.25 billion a month to rent roughly 300MW of capacity at xAI’s Colossus data centres in Memphis through May 2029, a deal previously known only in outline. It’s a striking arrangement given the public rivalry between Musk and Anthropic’s founders, and it sits alongside a separate $30 billion compute agreement between SpaceX and Google running through mid-2029.
The pattern that emerges is a web of compute-for-cash and compute-for-equity relationships between a small number of counterparties, Anthropic, Google, Amazon, Microsoft, SpaceX and Nvidia, that overlap and reinforce one another. Nvidia’s own move to formalise this as a financeable model landed this month, when the chipmaker confirmed a $500 billion financing platform with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR, treating GPUs as collateral in a way Jensen Huang described as the first time that technology chips have become “an investable asset class”. Blackstone’s president and chief operating officer, Jon Gray, welcomed the scale of the tie-up, noting the firm continues “to be enormous investors globally across the Nvidia ecosystem”.
For data centre financiers, the open question isn’t whether these deals are individually sound. It’s whether public market disclosure, arriving through Anthropic’s own S-1 rather than a rival’s, will be the first moment regulators and analysts can properly map how much of the AI infrastructure boom is being financed by capital that ultimately flows in a circle between the same half-dozen companies. Anthropic’s IPO won’t answer that on its own, but it’s likely to be the most detailed single data point yet.
Power becomes the new bottleneck
If ownership and circularity are the finance-side risks, power supply is the physical one, and it’s arguably the least resolved. Capacity’s own coverage this month of the financing gap behind nuclear power deals for AI data centres set out why signing a power purchase agreement and actually getting a reactor built, insured and delivered on schedule are two very different problems. Nvidia’s exposure now extends well past chips and into power directly, through its $3 billion stake in Texas developer Lancium, the same site behind the Stargate campus in Abilene.
That matters for Anthropic specifically because its Fluidstack-built US sites and its Amazon and Google capacity commitments all depend on power and grid interconnection timelines that sit outside Anthropic’s direct control. Analysts expect Anthropic’s eventual prospectus to formally list compute and power availability as a risk factor, which would mark one of the first times grid interconnection risk has been priced into a frontier AI company’s public filing rather than treated as background noise in a press release.
The neocloud sector offers a preview of what happens when that risk isn’t managed well. CoreWeave’s total debt reached $35 billion at the end of Q2, more than double its long-term debt from the previous quarter, as the company financed a $9.4 billion capital expenditure programme. Blackstone senior managing director Jasvinder Khaira has previously pointed to the appeal of this model for lenders, describing “significant investment in specialised GPU cloud infrastructure” as a core driver of the debt market’s growth. Whether Anthropic follows a similar path into asset-backed debt, rather than pure equity, once it owns rather than rents its facilities is a natural follow-up question for its IPO roadshow.
What to watch next
None of this will be fully resolved before Anthropic’s shares, if they arrive, start trading. But the direction of travel is clear: a frontier AI lab building its own data centres, financing them through increasingly complex vehicles, and doing so against a power market that hasn’t caught up. It’s a combination that data centre financiers, colocation operators and grid planners will be picking through for months.
Much of that conversation is expected to continue in person next month. The debt, ownership and power questions raised here sit squarely on the agenda at Metro Connect Fall x Datacloud USA, running in Austin from 1 to 3 September, where neocloud, hyperscaler, investor and fibre operator executives will address capital strategy and long-term AI infrastructure expansion directly. Given how much of this picture has only become visible through other companies’ filings, that room may end up knowing more than Anthropic’s prospectus ultimately discloses.
Related stories
