SpaceX becomes Google's compute landlord, and the AI buildout hits the grid

Information Technology 5 min 72 sources

A $30 billion deal to rent computing power, a $30 billion data-center push in India, a Texas grid buckling under demand, and AI named the top reason for May's job cuts — the AI boom is colliding with cost, power and people all at once.

Key takeaways

The AI compute land-grab went into overdrive: SpaceX will rent Google computing power for about $30 billion, and AirTrunk is spending $30 billion on data centers in India.
The running costs are biting — companies now route each task to the cheapest capable model to curb "token" bills — while the Texas grid buckles, making electricity, not chips, the real limit.
Washington is moving in, weighing a government stake in OpenAI and readying an Anthropic model for NSA cyber work, even as AI becomes the top reason US employers gave for May's 97,000 job cuts.

Everyone is racing to build somewhere to run AI

SpaceX, Elon Musk’s rocket company, has signed a deal to rent computing power to Google — about $30 billion over 32 months, or $920 million a month [8][11]. The capacity sits in data centers run by xAI, Musk’s AI firm, and the deal lands days before SpaceX’s planned stock-market debut [8]. In May, SpaceX struck a similar arrangement with Anthropic [8]. A rocket maker is now, in effect, a computing landlord — and Google, which builds vast data centers of its own, is still renting more.

It is not alone. AirTrunk, a Blackstone-backed data-center operator, said it will invest $30 billion to build 5 gigawatts of AI data centers in India by 2030 [5]. Five gigawatts is power-plant scale — enough electricity for several million homes. Meta is reported to be weighing tens of billions in fresh fundraising, and Google parent Alphabet is seeking new capital too [3][9].

Why the frenzy? AI runs on “compute” — the raw processing power of chips packed into data centers. Training and running models eats it in enormous quantities, and capacity is the scarce thing everyone needs. So the spending has turned into a land-grab. For anyone in the field, the lesson is plain: compute is the new oil, and the contest is over who controls the supply.

The token bill comes due

The buildout races ahead. The running cost is catching up to it.

Inside big companies, finance chiefs are cracking down on runaway AI spending, and a new discipline is taking hold: “model routing” [7]. Instead of sending every request to the single most powerful — and most expensive — model, companies match each task to the cheapest model that can do it. A simple question goes to a small, cheap model; a hard one goes to the heavyweight.

The cost they are managing is measured in “tokens” — the chunks of text an AI reads and writes, billed by the million. Every answer a model gives is “inference,” and inference is metered, like electricity. As usage exploded, enterprises blew past their AI budgets, and the token bill came due [2][7].

There is a sting in the tail. The sky-high valuations of OpenAI and Anthropic assume companies will keep paying premium prices for top models [7]. If routing spreads and most tasks quietly shift to cheaper models, that assumption weakens. If you build on AI, your token bill is now an architecture decision — not a footnote.

The real ceiling is electricity

All that compute needs power, and the grid is starting to say no. In Texas, the grid operator flagged that several large data centers and crypto sites failed reliability tests ahead of the summer peak, when demand is highest [20]. The regulator, ERCOT, is reviewing the failures and drawing up fixes [20]. One of the campuses straining the system is an OpenAI data center in Abilene, built across eight planned buildings [20].

The industry is already hunting for its own electricity. A startup called Antares got a small modular nuclear reactor to “criticality” — a self-sustaining nuclear reaction — in a first US test [59]. These reactors are compact and factory-built, meant to deploy faster than a giant plant. It cannot generate power yet [59]. But the direction is clear: the AI story is turning into an energy story, and megawatts, not chips, are becoming the bottleneck.

The state moves into AI

As the money and power pile up, so does the government’s interest. The White House and OpenAI are in talks about the US taking a stake in the company, a source confirmed [1]. OpenAI has floated donating equity to seed something it calls a “Public Wealth Fund,” an idea its chief executive Sam Altman first raised with the administration in 2025 [1]. President Trump has said his team will “look into” the government taking stakes in AI firms [46].

At the same time, Washington is pulling AI toward security work. The US says it will speed the development of AI for national security [24]. And the National Security Agency — the US signals-intelligence service — is reportedly preparing Anthropic’s “Mythos” model for cyber operations [33].

The contrast is sharp. In the same week, Anthropic publicly urged a worldwide “temporary pause” on AI development to weigh the dangers [13]. It did so inside a post boasting that its Claude model is progressing toward “recursive self-improvement” — the ability to build better versions of itself [13]. A company warns of the risks, markets its capabilities, and lines a model up for spy work, all at once. The line between AI lab, defense contractor and state asset is blurring fast.

The threat that gets hired

The week’s sharpest security warning is about people, not code. Google’s threat researchers and the FBI said a ransomware gang known as Silent Ransom Group has begun sending fake IT workers in person into law firms [14]. Once inside, the imposters plug USB drives into office computers to steal data directly, or open a door for others to connect remotely [14].

A second story points the same way. A former IBM security executive alleges, in a lawsuit unsealed this week, that Chinese hackers breached IBM’s core network between 2013 and 2016 [15]. The company, he says, covered it up and never disclosed it [15]. IBM has not confirmed the claims.

The thread is uncomfortable. The weak point is often not the firewall but the org chart — who gets hired, who gets trusted, what gets disclosed. The thing worth auditing is the front door, and the people you let walk through it.

The bill nobody puts on the slide

Under all the megadeals, the cost is landing on workers. US employers announced just over 97,000 job cuts in May — the highest figure for any May since 2020, and the third straight monthly rise [50]. The count comes from the outplacement firm Challenger, Gray & Christmas [50].

For the first time, AI was the leading stated reason. Companies blamed it for almost 40% of May’s announced cuts — up from 7% in January [50]. The same technology pulling in $30 billion data-center deals is now the most common reason given for letting people go. Both numbers are true at once, and they are worth holding together.

02 · Lesson · why it matters

Follow the power cord, not the press release

When something new is booming, the real limit is rarely the magic everyone is discussing. It's the dull, physical thing the magic quietly runs on.

The weightless thing that keeps hitting walls

We are sold artificial intelligence as something close to weightless — intelligence in the cloud, a mind you summon through a screen. Yet this week it kept crashing into the most physical things imaginable.

A rocket company signed a $30 billion deal to rent out computing power. A grid operator in Texas warned that data centers were failing its electricity tests. A startup fired up a nuclear reactor. Companies cut 97,000 jobs. The most abstract technology of our age spent the week colliding with concrete, copper, megawatts and people.

That is not a coincidence. It is the shape of every boom.

Everything floats on something heavy

Behind every abstraction sits a physical base holding it up. “The cloud” is not a cloud. It is warehouses full of chips drinking electricity and water in the Texas heat. An AI “answer” feels free and instant, but it is “inference” — a real computation, metered in tokens, burning real power in a real building someone had to wire.

The base has a name in every field: the substrate. The thing the shiny layer runs on. And here is the rule worth keeping: a boom always borrows against its substrate. It races ahead as if the foundation were infinite.

Why the bill always comes due downstairs

During the loud part of a boom, the substrate looks endless. Money is cheap, so spending feels free. The grid has slack, so power feels unlimited. Talent is plentiful, so nobody counts it. The boom behaves as if the floor isn’t there.

Then the floor speaks. The token bill comes due, and companies scramble to route cheap tasks to cheap models. The grid says no, and a data center fails its test before summer even starts. The constraint was there the whole time. The excitement simply hid it. What looked like a story about intelligence turns out to be a story about electricity.

The substrate is more than wires

It is not only energy. The floor under this boom is also money, trust, and people — and each one is reasserting itself.

Money: the towering valuations of the top AI labs assume everyone keeps paying premium prices forever. The moment companies start routing around that, the assumption cracks. Trust: the sharpest security warning of the week wasn’t about code at all, but about a gang sending fake IT workers through the front door. The substrate there is human — who you hire, who you believe. And people: the 97,000 jobs cut, with AI named as the leading reason, are part of the foundation too. The “efficiency” everyone celebrates has a body count, and it shows up in someone’s monthly budget.

How to read a boom

So here is the practical way to size up any boom, in any field. Don’t measure it by the capability everyone is celebrating. Measure it by what it quietly consumes that cannot scale as fast — power, materials, money, trust, goodwill.

The ceiling is always in the boring layer. And there is a second clue hidden in that fact: whoever controls the scarce substrate ends up holding the real power. That is why a rocket company becomes a computing landlord, and why a government starts asking for a stake. They can all see where the chokehold is. It isn’t the model. It’s the megawatts, the chips, the cash.

The whole was always one machine

Here is the part to carry past today. The press release talks about the magic. The truth lives in the power cord, the balance sheet, the payroll and the front door.

Treating the dazzling new thing as if it floats free of its physical base is exactly how you get blindsided. The bill, the blackout, the breach, the layoff you didn’t think was connected — it was all one machine the whole time. Intelligence and electricity. Model and money. The automation, and the person whose job it quietly replaced. Learning to see the heavy thing under the weightless one is most of what it takes to understand a boom — and to not be swept up in it.

03 · Lab · your turn

What the Boom Runs On

Spread a fixed prep budget across power, compute cost, security and people, then scale an AI product and feel that the ceiling is whichever boring foundation you under-built.