Meta Started Selling Its Spare Compute — What the $10 Billion Anthropic Talks Really Mean

A $1.6 trillion company quietly started a cloud business

Meta's AI infrastructure spending is moving from the cost column into the revenue column. In my analysis, that's the single least-priced-in change at the company right now.

The stock is down 11% over the past year while most of big tech kept climbing. But reading that decline as 'the business is broken' leads you to entirely the wrong conclusion. When I went through the numbers, what I found wasn't a broken business — it was a spending figure large enough to scare people. Those are very different problems.

In the short run the market is a voting machine, and the votes are currently going against Meta. The gap between those votes and the underlying fundamentals is exactly the kind of situation worth doing real work on.

First, the number that scared everyone

Meta raised its 2026 capital expenditure guidance from $125 billion to $145 billion. That's up to $145 billion in a single year.

Most of it goes into data centers, same as every other hyperscaler, with AI chips and compute infrastructure layered on top. Watching that much cash walk out the door is unnerving, and it's a completely reasonable reaction. It also produces one obvious question: when does this spending turn into profit?

So far the company's answer has effectively been 'soon.' That is the least comfortable answer in all of investing. However long-term you intend to be, it's hard to sit still when you can't see light at the end of the tunnel.

Meanwhile, the business wasn't declining

Here is what the company actually delivered over the same stretch.

• Earnings per share of $7.31 last quarter, beating Wall Street
• More than 3.5 billion people reached daily across Facebook, Instagram and WhatsApp
• Operating margin above 40% — top-tier for a company this size
• Revenue growth above 30%

A company approaching multi-trillion-dollar scale is growing revenue more than 30%. I don't think that sentence should slide past you. That combination is rare in mega caps.

For context, consensus for the July 29 second-quarter print was around $7.13 per share. Nowhere in that setup is there a signal that earnings are collapsing. What collapsed was the stock. What got complicated was expectations.

The turning point: deciding to sell the leftovers

This is where the story changes direction.

For years Meta poured tens of billions into AI infrastructure — data centers, chips, servers — to run its own apps. Plenty of investors looked at that and said money was simply leaking out. Mark Zuckerberg looked at the same assets and reached a different conclusion: the company had more compute than it needed internally, so it should sell the excess.

The reason I find that decision sharp is simple. It extracts additional revenue from spending that has already happened. It isn't new money going out; it's one more path for money already spent to come back.

What the Anthropic talks tell us

Reports indicate Meta is in late-stage discussions with Anthropic on a compute agreement worth up to $10 billion over two years.

The number matters less than its character. Infrastructure that looked like a drag on the balance sheet starts getting reclassified as a revenue-producing asset the moment a counterparty signs. The depreciation schedule doesn't change, but now there's revenue running against it.

I pay particular attention to the cost structure here. A company building a cloud business from scratch has to secure land, sign power contracts, procure servers and staff operations — all of it new. Meta already finished that process for its own products. Same revenue, different marginal cost.

That's the basis for the bull argument that Zuckerberg is effectively building a competitor to AWS and Azure. I think that framing is early, but directionally it isn't wrong. I laid out the broader capex debate in Meta: The Ad Machine and New Cloud Business Hiding Behind Capex Fear, and the Anthropic negotiation is the first piece of physical evidence attached to that thesis.

The ad engine is still home base

If compute leasing is the new story, advertising remains the machine that actually prints the cash.

Meta's family of apps sees more than 3.5 billion daily active users. Outside of Google and YouTube under Alphabet, no advertising platform on earth has that reach.

On top of that, generative AI is now built directly into the ad platform. Advertisers can auto-generate and customize video ads, still images and messaging, with everything optimized in real time toward clicks and conversions. Ads are getting smarter, cheaper to produce, and better performing all at once.

That structure is why operating margins clear 40%. Gross margin runs at 82%, meaning 82 cents of every advertising dollar survives as profit before overhead and taxes. Companies with that cost structure convert revenue growth into earnings growth much faster than average.

One more thing worth noting: revenue per user in the US and Canada is far higher than in international markets. As those markets get wealthier, the gap has room to close, and I read that as the remaining upside in the advertising line.

What I'm actually watching from here

Three items are on my checklist for the next several quarters.

First, does compute leasing revenue start getting broken out? Once it's material, the company has little choice but to disclose it separately. That's the first real test of whether this business is genuine.

Second, do customers beyond Anthropic sign on? One contract is an experiment. Three or four is a business.

Third, does capex keep growing faster than leasing revenue? As long as that gap widens, free cash flow stays under pressure. I covered the industry-wide spending race in Half a Trillion Dollars a Year: The AI Infrastructure Arms Race Built by Four Companies, and within that group Meta is the one most aggressively building new recovery paths for the money.

The stock fell. The company kept growing. Whether that gap closes, or whether the market turns out to have been right, comes down to those three numbers.