Meta's $145B Spend Only Makes Sense If You Believe in Robots

Meta raised its 2026 capital expenditure forecast to $125–145 billion last week. The stock fell 9.4% in five days. Investors saw the number and panicked.

Two days later, Meta quietly acquired Assured Robot Intelligence, a 20-person startup building AI for humanoid robots. Most coverage treated it as a footnote.

That framing has it backwards. The acquisition is the receipt that explains the capex. And once you connect the two, the spending starts to look less like a panic move and more like the most rational bet Meta has made in a decade.

Here's the case.

What actually happened

On April 29, Meta reported Q1 2026 earnings and raised its full-year capex guidance from $115–135 billion to $125–145 billion. The company attributed the bump to higher component prices, particularly memory, and additional data center costs.

For context, Meta spent $72.2 billion on capex in 2025. The 2026 guidance is nearly double that, and more than 2025 and 2024 combined. According to a Bank of America tally, Meta's planned spend now sits alongside Microsoft at $190B, Amazon at $200B, and Alphabet at $185B. The four companies will spend a combined $725 billion on capex in 2026, a 77% jump over last year's record.

Investors responded with skepticism. Meta shares closed the week at $608.75, down 9.4% from the earnings release.

Then on May 1, Meta announced the acquisition of Assured Robot Intelligence (ARI). Co-founders Lerrel Pinto and Xiaolong Wang join Meta Superintelligence Labs to work on whole-body humanoid control. Financial terms undisclosed.

The two stories ran on parallel tracks in most coverage. They shouldn't have.

The Reality Labs comparison everyone is making is wrong

The default reading is that Meta is doing it again. Zuckerberg burned $50 billion on Reality Labs and the Metaverse with one genuine hit (the Ray-Ban smart glasses). Now he's pivoting to robots and asking investors to fund another moonshot. The capex jump is read as evidence that nothing has been learned.

This reading misses a structural difference between the two bets.

Reality Labs was a vertically integrated hardware play. Meta designed the headsets, manufactured them, sold them, and built the software. When Quest underperformed, Meta absorbed every layer of the loss.

The robotics strategy is the opposite. Meta CTO Andrew Bosworth has said explicitly that the goal is to build a software platform that other companies can license, like Google did with Android. The ARI acquisition fits this thesis: Meta isn't buying a robot, it's buying the AI models that let robots understand and adapt to human behavior. Meta plans to develop sensors, software, and AI models for robots and make them available to the rest of the industry.

This is a lower-capital, higher-leverage strategy than Reality Labs. Meta doesn't need factories. It doesn't need to win on hardware margins. It needs to be the intelligence layer running inside someone else's machine.

The capex jump still has to be funded. But the path to recouping it looks fundamentally different from the Metaverse playbook.

Why the math actually requires robots

Here's the part most analysts are skipping. Meta's core advertising business is profitable but not growing fast enough to justify $145 billion in annual capex. Even generous assumptions about AI-enhanced ad targeting don't close the gap.

The math only works if Meta is building toward a market that's bigger than the one it currently serves.

Look at where the spending is going. Meta attributed the increase to "higher component pricing this year, particularly memory." Data centers now consume 70% of the world's memory output. Microsoft alone attributed $25 billion of its 2026 capex to memory cost inflation. This is infrastructure that scales with model size and inference demand, not with social media users.

You don't build that kind of infrastructure to show better Reels ads. You build it because you think the next computing platform runs AI models continuously, in physical environments, on hardware you don't manufacture.

That platform is robotics. And the timing tracks. Tesla is pushing Optimus toward a target of one million units a year. Amazon acquired Fauna Robotics in March. Figure AI has raised over $2.6 billion. China's State Grid Corporation has allocated approximately $1 billion to deploy 8,500 robots in 2026 for power grid inspection alone. Government-scale buyers and industrial customers represent infrastructure contracts worth more than any single hardware sale.

Meta's $145 billion capex starts to look less like spending and more like positioning. The data centers being built now are what trains the models that run on robots later. The components being stockpiled are what powers the AI brain inside humanoids built by manufacturers Meta doesn't own.

The bet that's actually being made

There's a useful comparison hiding in Meta's own history. The LLaMA strategy.

When Meta open-sourced LLaMA, the move was widely misread as charity or strategic confusion. The actual logic was sharper. By giving away the foundation model, Meta commoditized OpenAI's core asset and ensured that a Meta-built stack ran underneath the broader AI ecosystem. The capture wasn't direct revenue. It was distribution, data flywheel, and the option to monetize later.

The robotics platform play is the same logic applied to physical AI. Don't build the robots. Build the layer every robot needs. Let the hardware become the commodity. Capture value through the data flywheel that emerges when your models run on millions of machines learning from real-world physical interaction.

This is why the ARI acquisition matters more than its size suggests. Pinto and Wang aren't building robots. They're building foundation models for whole-body humanoid control: the part of the stack that's hardest to replicate and most valuable when it runs at scale.

If this strategy works, the $145 billion capex isn't a bet on robots being big in 2026. It's a bet that by 2030, every humanoid manufacturer needs a software stack, and Meta has spent five years making sure theirs is the obvious choice.

What investors are missing and what they're getting right

The 9.4% stock drop reflects two real concerns. The first is that capex this large compresses near-term free cash flow, which matters for valuation regardless of long-term thesis. The second is that Meta's track record on hardware-adjacent bets is genuinely poor.

Both concerns are valid. Neither addresses the strategic logic of the robotics platform.

What investors appear to be missing is that the robotics strategy isn't a hardware bet at all. It's an AI platform bet that uses hardware partnerships as the distribution layer. The capital intensity is in compute and data centers (Meta's strength), not in manufacturing (Meta's weakness).

What investors are getting right is that the strategy depends on conditions that don't yet exist. The humanoid market currently has no Samsung, no Xiaomi, no Oppo. The leading companies (Tesla, Figure, 1X, Boston Dynamics) are vertically integrated and building their own AI stacks. For Meta's platform play to work, the market needs to fragment into hardware specialists who outsource intelligence. That hasn't happened yet, and it might not happen at all.

The honest assessment: this is a real bet with real downside. It's just not the same bet as Reality Labs, and pricing it like a Metaverse repeat is probably wrong.

What to watch

Three signals will tell you whether the strategy is working.

First, manufacturer partnerships. If Meta announces formal integrations with humanoid manufacturers, particularly Chinese or industrial-focused companies that lack their own AI capabilities, that confirms the platform thesis. The smartphone equivalent would be the Open Handset Alliance that Google built around Android in 2007.

Second, talent flow into Meta Superintelligence Labs. Aggressive hiring from Boston Dynamics, Tesla Autopilot, Nvidia's robotics division, or DeepMind Robotics would signal that Meta is building the engineering depth required to execute this at speed. Hiring at that level is also expensive and visible, which makes it a useful tell.

Third, the open-source release pattern. If Meta starts releasing robotics models under licenses similar to LLaMA, the strategy is clear. If it doesn't, the platform play is probably stalled.

By the end of 2026, the answer will be obvious. Either Meta has positioned itself as the foundational layer for the next decade of physical AI, or the $145 billion capex looks like another Reality Labs in slow motion.

The bottom line

Meta's 2026 capex is the largest single-year infrastructure bet in the company's history, and the market is reading it as another Zuckerberg moonshot. The ARI acquisition complicates that reading.

The robotics strategy isn't asking Meta to win at hardware, where it has lost before. It's asking Meta to win at AI platform distribution, where it has actually succeeded with LLaMA. The capex makes sense if you accept that the data centers being built now are training infrastructure for models that will run on robots Meta doesn't manufacture.

That's a real bet with real risk. But it's a more rational one than the Metaverse ever was, and it's the only framing under which $145 billion is a number that adds up.

Watch the partnerships. Watch the hiring. Watch the open-source releases. By December, you'll know.