Meta estimates annual capex for 2026 at between 115 and 135 billion US dollars.
Meta just casually announced they’ll spend up to $135 billion this year on infrastructure. That’s nearly double what they burned through in 2025, and roughly the GDP of Hungary. Most of it goes toward gigawatt-scale data centers, AMD Instinct GPUs, NVIDIA Blackwell chips, and custom silicon. Zuckerberg calls it “front-loading compute capacity” for superintelligence. Wall Street calls it a bet-the-farm move.
But the spending only makes sense once you understand where Meta actually stands right now, because their AI story in 2025 was far from glorious.
Llama 2 and 3 Built Real Trust
For a while, Meta genuinely earned the open-source crown. While OpenAI and Google locked everything behind APIs and subscriptions, Meta dropped the Llama series for free. Llama 3 ran well on consumer hardware, held its own against proprietary models on coding and reasoning, and earned massive community goodwill. Researchers, startups, and hobbyists built entire ecosystems around it. It was the rare case of a trillion-dollar corporation actually giving something back.
Then Llama 4 Happened
In April 2025, Meta rolled out Llama 4 Scout and Maverick, both using a Mixture-of-Experts architecture. The reception was brutal.
Developers found the models struggled badly with coding tasks and long-context comprehension. Scout advertised a 10-million-token context window, but fell apart on complex long-form work. Maverick wasn’t much better.
Then came the benchmark scandal. Meta submitted a specially tuned “experimental” variant of Maverick to LMArena, the popular AI leaderboard, where it placed second behind Google’s Gemini 2.5 Pro. The publicly released version? Nowhere close. LMArena called Meta out for not disclosing that the submitted model was a custom build. Maverick’s ranking cratered from 2nd to 32nd once real-world usage data poured in. Scout dropped out of the top 100 entirely.
It got worse in January 2026 when Yann LeCun, Meta’s outgoing chief AI scientist, admitted on record that Llama 4’s benchmark results were “fudged a little bit”, with different model variants cherry-picked for different benchmarks. An anonymous whistleblower from Meta’s AI team claimed leadership had suggested blending benchmark test data into the post-training process to produce “presentable” scores after the model failed to meet state-of-the-art thresholds internally.
Meanwhile, the flagship Behemoth model, a 2-trillion parameter beast, never shipped to the public at all. It remains an internal “teacher model” for the smaller variants, stuck in delay limbo because it couldn’t meet expectations.
The Real Play: Escaping Apple and Google
So Meta’s pure software AI play stumbled. But $125 billion in capex doesn’t scream retreat, it screams pivot.
Here’s the context people miss: Meta has always been an app company trapped inside someone else’s operating system. Facebook, Instagram, WhatsApp, they all live on iOS and Android. Apple’s ATT privacy update in 2021 alone cost Meta an estimated $10 billion in annual ad revenue. Every policy change, every App Store rule, every privacy toggle sits between Meta and its users. Zuckerberg has been trying to claw his way out of this dependency for a decade.
The Ray-Ban Meta smart glasses are his current escape route, and they’re working. In 2025, Meta shipped 7.4 million units and captured 85% of the global smart glasses market. The second-generation models launched in September drove shipments up 139% year-over-year in the second half of the year. Meta and EssilorLuxottica are reportedly planning to double annual production to nearly 20 million units by the end of 2026.
These aren’t fashion accessories anymore. The Gen 2 glasses shoot 3K video, run real-time translation, and feed your surroundings straight into Meta’s multimodal AI for contextual answers on the fly. Meta’s next models, codenamed “Aperol” and “Bellini,” will run the AI continuously in the background, proactively reminding you about forgotten keys or items on your shopping list. A premium HUD-equipped pair codenamed “Hypernova” is coming at $1000+. True AR glasses, “Artemis,” are slated for 2027.
Zuckerberg’s term for all of this is “Personal Superintelligence”: an AI that lives on your face, sees what you see, and gradually becomes your ambient assistant. It’s the anti-ChatGPT play. Instead of you going to the AI, the AI is just always there.
The Moltbook Anomaly: Advertising to AI Agents?
But the hardware pivot isn’t the only strange move Meta is making right now. In a bizarre twist of what can only be described as absolute “AI psychosis,” Meta recently acqui-hired Matt Schlicht and Ben Parr, the duo behind Moltbook.
For the uninitiated, Moltbook (also known as OpenClaw) is essentially a Reddit-like social network where humans don’t interact, but AI agents do. Originally starting as a quirky experiment using Anthropic’s models (which earned them a cease-and-desist over their original name, “Clawdbot”), it evolved into a sandbox where developers and crypto-enthusiasts set their autonomous agents loose to debate, organize, and chat with one another.
Why on earth would Meta’s Superintelligence Labs acquire a bot-to-bot network where mostly humans communicate? Put your tinfoil hat on for a second, because the underlying theory is uniquely dystopian.
As traditional human-to-human social networks increasingly suffer from the “Dead Internet” theory, flooded with formulaic, AI-generated bot replies, the nature of a platform “user” is fundamentally changing. If Zuckerberg’s vision of an ambient “Personal Superintelligence” succeeds, humans will soon be delegating daily tasks. Your agent will be the one booking your flights, scheduling your appointments, and buying your groceries.
If humans are no longer the ones clicking the “buy” button, how does an ad-driven empire survive? That is where Moltbook comes in. It serves as an experimental playground for Meta to fine-tune its Facebook and Instagram algorithms to advertise directly to autonomous AI agents. Imagine invisible, prompt-injected text hidden inside ads designed entirely to bypass the human eye, meant only to convince your AI assistant that it absolutely needs to purchase a specific brand of coffee to “live longer” or function optimally. It sounds like a ridiculous conspiracy theory, but for a company whose entire survival relies on targeted ads, figuring out how to market to the new AI middlemen is a brilliant, albeit incredibly weird, necessity.
Platform Company or Bust
The $135 billion capex, the AMD and NVIDIA mega-deals, the custom chip development, the aggressive glasses roadmap, and even the bizarre acquisition of bot-social-networks: it all points in one direction. Meta wants to be a platform company. Not an app living inside Apple’s walled garden, but the company that owns the next computing paradigm.
If the glasses scale to the tens of millions Zuckerberg is targeting, Meta finally controls the hardware, the OS, the AI stack, and the user relationship end-to-end. No more middlemen. No more App Store tax. No more privacy policy disruptions tanking their ad business overnight.
Whether this works depends on execution. The Llama 4 debacle shows Meta can fumble hard when credibility matters. But the hardware numbers don’t lie, 7.4 million units in a product category that barely existed two years ago is serious traction. And $125 billion buys a lot of second chances.
Sources
Meta boosts annual capex sharply on superintelligence push - Reuters
Meta’s Llama 4 herd controversy and AI contamination, explained - ZDNet
Yann LeCun: Meta “fudged a little bit” when benchmark-testing Llama 4 - Fast Company
Global AI glasses shipments reach 8.7 million units - Omdia
Meta mulls doubling output of Ray-Ban glasses by year end - Reuters
Meta HUD Glasses Price, Features & Input Device Reportedly Revealed - UploadVR
AMD and Meta Announce Expanded Strategic Partnership to Deploy 6 Gigawatts of AMD GPUs - AMD
Personal Superintelligence for Everyone - Meta
Meta Buys Moltbook: The Social Network for AI Agents