BOD Weekly | June 9, 2026

BOD Weekly | June 9, 2026

Google Just Became a Compute Renter. The AI Infrastructure Thesis Just Got Bigger.

Google, the world's largest single owner of AI compute, will pay SpaceX $920 million a month for GPU capacity. The biggest owner just became a renter. If they cannot build enough, nobody can.

Last Thursday, Google disclosed in a regulatory filing that it will pay SpaceX $920 million per month from October 2026 through June 2029. The deal covers approximately 110,000 NVIDIA GPUs plus CPUs, memory, and other components. Total contract value: roughly $11 billion.

This came less than three weeks after Anthropic signed a similar deal with SpaceX for $1.25 billion a month at the same facility. Colossus 1 near Memphis, Tennessee, was originally built for xAI's own AI work. It is now the world's largest AI compute rental property, with two anchor tenants paying a combined $2.17 billion per month.

The Anthropic deal made structural sense. That company was compute constrained. It raised usage limits the same day it announced the deal. It had no data center infrastructure. Renting was the only path.

Google is a different animal. Epoch AI estimates Google is the largest single owner of AI compute globally. This is the company that builds custom TPUs, runs its own cloud, and deploys AI inference at a scale no other enterprise can match. Google Cloud is the platform where other companies rent compute. The idea that Google would become a net renter of GPU capacity from a third party is an inversion of the entire AI compute ownership model.

The company frames this as a bridge: "This is a short-term, timely agreement to ensure we have bridge capacity to meet surging customer demand for our agent platform, Gemini Enterprise, which has been even higher than we expected." Alphabet has already committed more than $180 billion in capex this year and expects that to significantly increase in 2027.

Bridge capacity. $11 billion. Over 33 months.

If that is a bridge, the river is wider than Google's own capex projections can span.

The Thesis That Just Got Validated

For PE firms evaluating AI infrastructure, the Google-SpaceX deal confirms a hypothesis we have been watching since Blackstone announced its $5 billion Google TPU cloud JV on May 18. The premise was that enterprise AI demand would create a structural shortage of specialized compute and that owning capacity during that shortage would command a premium. The deal target was 500 megawatts of AI compute capacity sold as a managed service.

Seven weeks later, Google itself is renting third-party GPU capacity at a price that implies an annualized run rate of $11 billion from a single data center. That same facility, Colossus 1, is already generating $15 billion annually from Anthropic. Combined, two tenants are paying $26 billion a year for capacity at one location.

These numbers do not look like rent anymore. They look like infrastructure tolls. If a single colocation facility in Memphis can generate $26 billion in annual lease revenue from two tenants, the asset class that Blackstone, DigitalBridge, and a growing roster of infrastructure investors are building toward is measured in hundreds of billions, not tens.

The compute ownership thesis now has three tiers.

Tier one is the hyperscale builders: Microsoft, Alphabet, Amazon. These are the companies that own and operate their own data centers at cloud scale. Even these players are now supplementing with third-party leases. Google's SpaceX deal is the proof point. The builder tier cannot meet its own demand internally.

Tier two is the infrastructure owners: Blackstone, DigitalBridge, and the PE firms building compute capacity as an asset class. These firms sit between the builders and the renters. They own the physical infrastructure that resells capacity. Their economics improve when tier one demand exceeds tier one supply. The Google deal confirms that dynamic.

Tier three is the capacity renters: Anthropic, OpenAI, and the thousands of enterprises that consume AI compute but do not own it. Their costs are set by the spread between what tier two charges and what tier one would charge if tier one had spare capacity. That spread is widening.

The Google deal tightens the spread in tier two's favor. It also puts a floor under what compute infrastructure can charge per GPU-hour, because the alternative, self-build, is now provably insufficient even for the company with the most compute on earth.

What SpaceX Becoming a Compute Landlord Means

The deal also changes how you read SpaceX itself. SpaceX files to go public next week at a $1.75 trillion valuation. The IPO narrative has been dominated by Starlink revenue, Starship milestones, and the xAI merger. The compute leasing business is now a material revenue stream. Colossus 1 generating $26 billion a year from two tenants, with Colossus 2 in development. That is not a rocket business. That is an infrastructure business inside a rocket company.

SpaceX's S-1 will force public market investors to decide what multiple a compute leasing business inside a space company deserves. A pure-play data center REIT trades at a different multiple than an aerospace manufacturer. The Colossus revenue stream sits somewhere between the two, but SpaceX's ownership structure means the entire package gets priced as one equity.

For PE firms, this creates an interesting arbitrage question. If SpaceX can generate $26 billion a year from compute leasing as a side business to rockets, what is a pure-play AI infrastructure company worth? Blackstone's TPU cloud JV is a pure-play compute business. DigitalBridge's ArcLight acquisition at $1.05 billion, announced May 27, is a pure-play infrastructure business. If the market prices SpaceX's compute revenue at a blended aerospace multiple, the pure-play infrastructure firms should trade at a premium. Or if they go public, they should price higher.

This also raises the question of whether Colossus itself eventually gets spun out. A standalone compute infrastructure company with $26 billion in annual contracted revenue and a growing capacity pipeline would rank among the largest infrastructure companies in the world. SpaceX shareholders may or may not prefer to own that exposure alongside rockets. PE firms that understand how to structure infrastructure spinouts are already running the math.

The Google Counterparty Question

There is one wrinkle worth flagging. Alphabet is both a SpaceX compute customer and a SpaceX investor. Its stake is expected to be worth more than $100 billion after the IPO. The two companies are also reportedly in talks to build orbital data centers, which would represent an entirely new category of compute infrastructure.

When your compute landlord is also one of your largest shareholders, the lease negotiation is not arm's length in the traditional sense. Google's $920 million monthly payment flows from Alphabet's operating budget to SpaceX's revenue line and then, partially, back to Alphabet's investment returns. The circular economics are complex enough that the headline number may not reflect a true market-clearing price for GPU capacity.

But that is a footnote, not a rebuttal. Even if Google received below-market terms because of the ownership relationship, the fact that it needed to negotiate for third-party capacity at all confirms that internal build capacity is insufficient. And Anthropic's $1.25 billion monthly rate at the same facility provides an independent data point. Colossus capacity commands north of a billion dollars a month from a tenant with no ownership stake in SpaceX. The market rate is real.

What This Means for Portcos

For PE portco operators, this story has a practical implication. The cost of running AI at scale is going up, and the ability to source capacity is becoming a strategic procurement function.

A mid-market portco deploying AI today typically consumes compute through a cloud provider. AWS, Azure, or Google Cloud at metered rates. The assumption has been that cloud compute is effectively infinite and that prices trend down over time as hardware improves.

The Google-SpaceX deal challenges both assumptions. If Google itself cannot meet demand from its own infrastructure, the cloud providers are capacity constrained at the aggregate level. That constraint flows downstream to enterprise customers in the form of higher prices and tighter availability. The portco that locks in a compute contract today may pay less than the portco that waits six months.

This shifts AI deployment from a software procurement decision into an infrastructure procurement decision. Portco operators who treat AI compute as a strategic input, something you contract for, hedge, and plan capacity around, will have an advantage over those who treat it as a variable cloud expense. The operating partners who understand this distinction will outperform the ones who do not.

The broader thesis remains what it has been: the data layer is the only defensible portco position. The model ships. The compute is rented. But the data infrastructure, the warehouse, the feature store, the integration layer, is what the portco owns and what compounds over time. In a world where even Google cannot buy enough compute, the portco's proprietary data becomes the one input that compute capacity amplifies but cannot replace.

More From This Week

OpenAI files confidentially for IPO. One week after Anthropic's filing, OpenAI submitted its own draft S-1. The race between the two labs is now an explicit IPO queue, with SpaceX expected to go first at $1.75 trillion. Whichever lab prices first sets the comp for the other. OpenAI disclosed $122 billion raised in its March round at $852 billion valuation. The WSJ reports OpenAI expects to burn $85 billion in 2028 even after doubling sales. The public market is about to get a line-item view into AI economics that will set valuation benchmarks for every AI-adjacent PE investment.

Anthropic confirms: rent compute, don't build it. At Bloomberg Tech, Daniela Amodei explained Anthropic's compute philosophy. "We would much prefer to be on the side of having a little bit more demand for the product than we are able to serve than the inverse." Annualized revenue crossed $47 billion in May, up from $9 billion at end of 2025. The compute-renter thesis from a soon-to-be-public $1T company directly validates the infrastructure-owner thesis that Blackstone and others are pursuing.

Thoma Bravo takes Kneat private for $466M. Life sciences validation software. PE continues targeting compliance-heavy verticals where AI can compress audit and regulatory workflows. A smaller deal, but it signals the AI-adjacent thesis is active in regulated verticals.

Bain midyear: dealmaking sharply down. A&M reports dual-track IPO prep surging. PE deal volumes are contracting while IPO readiness surges among GPs. The AI lab IPO wave is pulling exit strategy toward public markets even as private deal activity softens. Zombie funds surpass $1 trillion. Twenty percent of global PE assets held seven-plus years. The IPO window opening for AI companies may be one of the few large-scale exit channels available.

BOD Weekly publishes every Monday. PE/AI/value-creation lens. One thesis per issue, anchored in the last seven days.