Why Anthropic and Google are renting from SpaceX

GM, this is Milk Road AI, your shortcut to understanding the most important developments in AI.

A rocket company just became the most important infrastructure player in AI and here's how it happened in 122 days.

Speaking of infrastructure, today’s partner is Arc by Circle.

Arc is an open Layer-1 blockchain built for real-world financial applications. Public testnet is live, App Kits are here, and mainnet is getting closer. Start building on Arc.

THE SILENT LANDLORD OF THE AI ARMS RACE

In 1956, the U.S. government needed to find out what the Soviet Union was building behind the Iron Curtain.

The military had satellites, analysts, and a multi-billion-dollar intelligence apparatus.

But they had one problem: none of it mattered without someone who could actually get the film back.

That someone was Lockheed, specifically, a division called Skunk Works, operating out of a secret facility in Burbank, California.

They built the U-2, the SR-71, and quietly became the company that delivered whatever the U.S. government needed most.

Source: MakeaGIF

SpaceX just did the same thing to AI, and the two deals that confirmed it were disclosed within 15 days of each other, both buried in an IPO filing that most people skimmed for the valuation number and closed.

Let's back up because every major AI company on earth is running into the same problem: power.

Data centers need electricity, and AI data centers need obscene amounts of it.

By some estimates, data center energy consumption could hit 1,050 terawatt-hours in 2026.

That's the equivalent of an entire nation's power grid, consumed by servers.

This is exactly why one of our core positions inside Milk Road PRO is Bloom Energy (BE).

If AI is the biggest infrastructure buildout in history, the companies supplying reliable power stand to be some of the biggest winners.

So far, that thesis has played out well. Bloom Energy is up roughly 83% from our average entry, but we still believe we're in the early innings of the AI power buildout.

Source: Milk Road/Analysts

If you want to catch opportunities like this before they become obvious to everyone else, come join us inside Milk Road PRO.

Anyways, but here's the catch: even if you have the money to build more capacity, you can't just plug in.

Power grid connections take 5–10 years, and GPUs take up to a year to arrive, there's no shortcut around the bottleneck.

Which brings us to SpaceX.

Anthropic signed a deal to lease all of SpaceX's Colossus 1 data center for $1.25B per month through 2029, a contract worth more than $40B.

Two weeks later, Google followed suit, signing a deal worth $920M per month for access to 110,000 NVIDIA GPUs through June 2029.

Source: @SawyerMerritt

Together, those two deals represent over $26B in annualized contracted revenue for a company that did $18.7B in total revenue in all of 2025.

Google is the world's largest single owner of AI compute, and they have spent hundreds of billions building data centers.

They have custom chips, the most sophisticated infrastructure team in the industry, and more engineers per square foot than most countries have per capita.

And they still had to rent capacity from a rocket company.

Google's own explanation was almost comically diplomatic.

They called it a 'short-term, timely agreement to ensure we have bridge capacity to meet surging customer demand for Gemini Enterprise.'

Translation: we didn't see it coming, we can't build fast enough, and SpaceX had the only warehouse big enough to hold what we needed.

What SpaceX actually built

Colossus 1 started life as an xAI training cluster in Memphis, Tennessee, and was built in roughly 122 days, a world record for AI data center deployment speed.

At peak, it housed over 200,000 NVIDIA H100 GPUs, 194 petabytes per second of memory bandwidth, and more than 1 exabyte of storage.

Then xAI moved its Grok 5 training to Colossus 2, a repurposed Electrolux factory they built in 122 days, packing in 550,000 NVIDIA GB200 and GB300 GPUs with gigawatt-scale power capacity.

Colossus 1, suddenly vacant, was exactly what Anthropic needed: 300 megawatts of ready-to-use inference capacity. They could start paying in July.

This is the Lockheed play, SpaceX didn't just build a rocket company but rather built the infrastructure that the entire AI industry needs to function, at a moment when nobody else could build it fast enough.

And now they're renting it out at $2.17B per month.

The companies growing the fastest are also the ones running out of compute the fastest.

Anthropic's growth numbers are genuinely hard to believe when you see them laid out.

Anthropic exploded from $1B ARR in January 2025 to $35B by May 2026, a 30x revenue increase in just 12 months, one of the fastest growth ramps in software history.

Source: @JoshKale

Claude's daily active users more than doubled from February to March 2026 alone.

Claude Code hit $2.5B ARR in under nine months, and over 1,000 business customers now spend more than $1M per year.

The problem with growing that fast is that your infrastructure projections are wrong before the ink dries.

Anthropic had Amazon as its primary cloud provider and Google as a $13B investor (through Google's partnership and investment deal with Anthropic), yet neither could provide capacity fast enough to keep up with Anthropic's growth.

Because Anthropic blew past every internal projection so violently that there was simply no capacity reserved for them at that scale.

So they called SpaceX, and there's also a strategic angle here that's easy to miss.

By signing with SpaceX, Anthropic now has negotiating leverage over both Amazon and Google simultaneously.

Neither investor can hold the infrastructure relationship over their head anymore.

Anthropic is paying $15B a year, roughly half its current ARR, for compute costs.

That's a bet that growth will continue to accelerate faster than the compound costs compound.

Given the last 15 months, it's hard to argue they're wrong.

None of these contracts is permanent.

Both the Google and Anthropic deals include 90-day termination clauses after December 31, 2026.

If AI demand softens, if a competitor delivers a better model for cheaper, or if the neocloud market gets crowded, SpaceX could see those revenue lines evaporate almost as fast as they appeared.

The AI division is losing $6.36B annually, all ten original xAI co-founders have departed, and Grok is falling behind Claude and GPT in engagement.

SpaceX's bull case at the IPO is priced for AI revenue to hit $322B by 2030, according to Goldman Sachs and some analysts at Evercore project $755B by 2031.

Those projections require SpaceX to not just be the infrastructure provider, but to actually win the AI model race with Grok.

Right now, they're doing the first thing and actively losing the second.

The biggest risk is that SpaceX becomes the landlord to the AI companies outperforming it until those tenants build their own infrastructure and no longer need to rent.

The bull case

Here's what the bears are missing.

SpaceX doesn't need Grok to win.

The entire neocloud thesis works whether or not Grok ever challenges Claude or GPT-5.

As long as AI compute demand continues to outrun supply, SpaceX has a product that every major lab needs.

And the supply problem isn't getting easier, power shortages could constrain 40% of AI data centers by 2027, while grid connection timelines are measured in years, not quarters.

SpaceX has something nobody else can quickly replicate: operational, at-scale data centers that exist today, with launch infrastructure that could eventually deliver orbital compute at costs no terrestrial builder can match.

And SpaceX explicitly said in its S-1: 'We anticipate entering into additional similar services contracts.'

And the other person getting rich off all of this is Google.

While everyone is focused on SpaceX's IPO, the quietest winner in this story might be Sundar Pichai.

Google put $900M into SpaceX back in 2015.

At the IPO's $1.75T valuation, that stake, roughly 6 to 7 percent of the company, is worth somewhere between $100 and $126B.

That's a 130x return on an investment that most people forgot existed.

And separately, Google has invested $13B into Anthropic for a 14 percent stake.

At Anthropic's current $960B valuation, Google's 14% stake would be worth roughly $134.4B, more than 10x its original $13B investment.

Add those up, and Google could be sitting on roughly $234–260B in unrealized gains from its SpaceX and Anthropic investments alone, two bets that barely showed up on earnings slides until recently.

It gets better: Google is simultaneously renting compute from SpaceX while owning 7 percent of SpaceX, and every dollar Google pays SpaceX partly flows back to Google through its equity stake.

They are paying for their own investment to solve a problem their investment created.

This is either the most elegant corporate maneuver in recent memory or the most elaborate way to launder a balance sheet ever attempted. Possibly both.

Alright, that’s it for this edition, but we want to hear from you.