TL;DR
– Meta’s 2026 AI infrastructure budget sits between $125B and $145B
– Zuckerberg told shareholders on May 27 that selling cloud services is “definitely on the table”
– Outsiders are already knocking on Meta’s door weekly trying to buy GPU access
– A new internal group called Meta Compute launched in early 2026
– Inference pricing could crater if they follow through. Make your workloads portable before it happens
Something shifted at that shareholder meeting on May 27, 2026. Zuckerberg just… said it. Companies come to us basically every week wanting to rent our compute. That’s not a hypothetical.
That’s a man describing inbound demand for a product he hasn’t launched yet.
And here’s what makes it weird.
Meta right now is both a massive customer of AWS and Google Cloud and CoreWeave and Nebius. Spending billions on other people’s servers. While apparently noodling on the idea of becoming a vendor themselves. You don’t do that unless you’ve crunched the numbers and seen something that makes your eyes go wide.
tbh I’ve been thinking about this all week and it keeps getting more interesting the more you sit with it.
Building Too Much on Purpose
The word Zuckerberg used was “overbuilding.” Direct quote: “If we get to a point where we feel that we have overbuilt, then that is an option that we have.” That option being.
Sell the extra capacity. Rent it out. Turn a cost center into a revenue line.
$145 billion in a single year. Sit with that number. It’s not rounded. The range they gave was $125B to $145B and honestly the fact that they’re willing to publicly attach a figure that enormous to AI infrastructure tells you how confident they are. Or how desperate they are to not fall behind. Could be both.
Here’s the thing nobody mentions. Amazon didn’t sit down one morning and say “let’s invent cloud computing.” AWS grew out of Amazon having too much server capacity after the holiday rush and realizing they could sublet it.
The origin story of the most profitable division at one of the world’s largest companies is basically “we had extra stuff lying around.” Meta is staring at the same playbook.
They also cut something like 8,000 jobs recently. Redirected all that headcount spend into data centers and GPUs.
So the money’s already moving.
Side note: I tried to find out exactly how many GPUs Meta operates and the numbers are all over the place.
Some reports say 600,000 H100 equivalents. Others say north of a million by end of 2026. Nobody outside Meta actually knows, which is itself kind of telling.
Why Inference Gets Cheaper First
If you’re running AI agents or automated pipelines or anything that hits an API repeatedly, you’re paying AWS, Azure, or Google Cloud. That’s basically it. Three players. They set the terms. You pay.
Cartels work like that.
Meta entering the room changes the math.
More supply. Pricing has to respond. We watched this exact movie before — Google Cloud showed up and AWS suddenly discovered the concept of competitive pricing. Azure scaled up and it happened again.
But Meta’s play wouldn’t be training. Training needs absurd cluster sizes. Weeks of contiguous runtime. Inference is different. It’s bite-sized. You can shard it across regions, across providers, across whatever. If Meta starts selling inference at volume, cost per token drops. Not just on Meta’s platform. Everywhere. Because AWS can’t charge $0.003 per 1K tokens when someone else is charging $0.0015.
We run client automation pipelines at my shop.
GPU time is the single biggest line item that isn’t salaries. We’re not theorizing about this. We’ve already mapped which workloads move first if Meta pricing hits 30-40% below current spot rates.
Make Your Stuff Portable Before You Need To
Most small teams picked a cloud provider years ago. Optimized around it. That was rational. Switching costs were real and pricing was stable enough that lock-in didn’t hurt much.
That’s about to stop being true.
If Meta launches a cloud product in the next year or two.
And they’ve created a dedicated division for it, so this isn’t just Zuckerberg freestyling. You’ll want the ability to move fast. Not given that you’ll switch on day one. As when your current provider sends the renewal email with a 15% bump, you’ll want options.
Start testing your inference workloads on at least two providers. AWS and CoreWeave. GCP and Lambda. Whatever combo works. Build a cost and latency baseline right now while nothing’s urgent. If Meta enters with a serious price undercut, you want to be able to port in weeks. Not after six months of replatforming.
This isn’t a Meta-specific bet. It’s a bet that the compute market is about to get crowded.
Crowded markets favor buyers.
What to Watch For
Nothing’s guaranteed. Meta hasn’t announced a cloud product. No pricing page, no beta signup, no timeline. Zuckerberg’s managing expectations around the $145B number and keeping his options open.
But you don’t spend $145 billion without a plan to recoup it.
Meta watched Amazon, Microsoft, and Google turn cloud into one of tech’s most profitable businesses. Now they’re publicly saying the quiet part out loud.
Meta Compute is the team to track.
Created early 2026. That’s the group that would run a commercial cloud offering. Watch their job postings. When you see roles for billing infrastructure, customer support, SLA management, that’s when it’s real. Press releases and beta access will follow.
Run a test workload across AWS, GCP, and CoreWeave this month. Document the numbers. Cost per inference call. P95 latency. Cold start time. You’ll thank yourself later.
And keep an eye on spot GPU pricing across all the major providers. When Meta starts filling idle capacity, the first sign won’t be a press release. It’ll be someone’s spot price dropping 20% overnight and a bunch of DevOps people posting screenshots on X.
Amazon built AWS from spare servers. Google built GCP since they needed compute and figured they might as well sell the extras. Meta’s at the same fork in the road. The only real question is timing. And whether you’re positioned to grab the savings when they arrive.
—
Sources:
NDTV. Meta tech layoffs, AI push, Zuckerberg cloud computing | Benzinga — Zuckerberg drops hints on cloud wars | Biggo Finance — Meta cloud signal
