Zuck's God Complex: Meta's Superintelligence Reality Check

Remember when Mark Zuckerberg wanted to build the metaverse? Legless avatars, $1,500 face computers, cartoon boardrooms nobody asked for? That was 2021. Adorable era.

Now Zuck wants to build God.

Specifically: artificial general intelligence. Superintelligence. The machine mind that does everything a human brain can do, then keeps going past the event horizon of comprehension. And according to SemiAnalysis's year-one progress report on Meta's superintelligence ambitions, the company has pivoted from "metaverse maximalism" to "AGI or bust" faster than you can say "legs are an upcoming patch."

Here's the thing though: Meta is actually spending money like civilization depends on it. And they might be closer than the haters think.

FOLLOW THE GPUS

Let's talk compute, because in 2024-2025, compute is destiny.

Zuckerberg announced in January 2024 that Meta would possess the equivalent of 600,000 H100 GPUs by end of year. At roughly $25,000-$30,000 per H100, we're talking $15-18 billion in silicon alone. That's before data centers, power, cooling, and the engineers keeping the whole thing from becoming the world's most expensive space heater.

For context: that's more AI compute than most nations can muster. It's a provocation. SemiAnalysis, the semiconductor analysis outfit run by Dylan Patel that's become required reading for anyone tracking the AI arms race, has been dissecting Meta's compute trajectory with forensic intensity. Their read: Meta isn't just buying GPUs — they're constructing a vertically integrated AI war machine.

Custom silicon via MTIA (Meta Training and Inference Accelerator) chips. Massive training clusters consuming hundreds of megawatts. A data center footprint now including a reported $10+ billion facility in Louisiana purpose-built for AI workloads. Nuclear power exploration because — shocker — training superintelligence requires industrial-scale electricity that fossil fuels can't cleanly provide.

Reality Labs burned $18 billion in 2023 on the metaverse. The AI buildout makes that look like a hobby.

THE MODELS: ACTUALLY NOT EMBARRASSING?

Here's where the story gets genuinely interesting instead of just expensive.

Meta's open-source strategy was initially dismissed as cope — "we can't beat OpenAI so we'll give away our seconds for free." Then Llama 3.1 405B dropped in July 2024 and the narrative cracked.

405 billion parameters. Trained on over 15 trillion tokens. At launch, it went benchmark-to-benchmark with GPT-4-class models from OpenAI and Anthropic — and you could download it. Run it locally if you had the hardware. Fine-tune it for your own purposes. This wasn't a toy model being charity-cased to developers. It was a legitimate frontier-tier system being handed to the world.

The Llama ecosystem detonated. By late 2024, Llama downloads crossed hundreds of millions. Fine-tunes, merges, quantized variants, and custom derivatives flooded Hugging Face. Companies that couldn't afford GPT-4 API costs suddenly had a viable alternative. Meta AI — the assistant injected into Facebook, Instagram, WhatsApp, and Messenger — claimed 500 million monthly users. Whether those users represent genuine engagement or accidental activations while scrolling Reels at 2 AM remains an open question, but the distribution is undeniable.

The strategic play is elegant in its cynicism: Meta doesn't need to win the model quality race outright. They need to commoditize it. If open-source approaches closed-source performance, OpenAI's pricing power erodes, Google's differentiation shrinks, and Meta — sitting on a 3-billion-user distribution graph — wins by default. It's not about building the best AI. It's about making sure nobody else can charge rent on it.

THE LECUN PROBLEM

Then there's Yann LeCun. Chief AI Scientist at Meta. Turing Award winner. The man who has spent the last two years appearing on every podcast and conference stage that will have him to declare that large language models — the exact technology powering ChatGPT, Claude, Gemini, and yes, Meta's own Llama — will never achieve genuine intelligence.

LeCun advocates for "world models" — architectures that understand physical reality, causality, and common sense rather than pattern-matching text sequences. His JEPA (Joint Embedding Predictive Architecture) framework is Meta's contrarian bet: a fundamentally different path to AGI that doesn't rely on next-token prediction. V-JEPA for video understanding. I-JEPA for images. Research papers showing theoretical promise, but no consumer product anyone actually uses.

This creates a delicious internal contradiction. Meta's commercial AI stack — the stuff generating headlines and user metrics — runs on the precise transformer architecture LeCun publicly calls a dead end for superintelligence. The company is simultaneously betting billions on and against its own roadmap. It's the R&D equivalent of ordering the keto plate and the chocolate cake.

SemiAnalysis's analysis implies Meta is straddling: pouring resources into the LLM pipeline because it ships products today, while funding LeCun's alternatives because they might matter tomorrow. Hedging at planetary scale.

THE SCORECARD

One year into the superintelligence pivot:

Working: Open-source is genuinely disruptive. Llama changed the industry conversation. Compute capacity is world-class. Distribution through Meta's apps is unmatched.

Not working: No clear bridge from "very good language model" to "superintelligence." LeCun's world models remain academic. Revenue model beyond advertising is unclear. The assistant's 500M users haven't translated to a must-use product the way ChatGPT has.

The existential question: Is Zuckerberg building superintelligence, or building the narrative of building superintelligence? One reshapes civilization. The other moves the stock price. Both cost $100 billion.

But underestimate Zuckerberg at your peril. He out-waited Snapchat. He cloned TikTok with Reels and survived. He burned $50+ billion on Reality Labs and emerged with infrastructure now relevant for spatial AI computing. The metaverse didn't work, but the VR headset line became unexpectedly useful for AI research visualization and developer tools.

Superintelligence might follow the same arc: the destination proves illusory, but the infrastructure built in its pursuit — the GPU clusters, the data centers, the open-source ecosystem, the talent pipeline — accrues real value regardless.

Zuck might not build God. But he's building the cathedral. And either way, Nvidia shareholders are sending thank-you cards.