OpenAI kicked off the AI arms race when it launched ChatGPT in late 2022, sparking hundreds of billions in data center investment and sending Nvidia's AI-related sales spiking fourfold in 2023. Global AI computing power has since ballooned to roughly 20 million H100-equivalents worth of operational compute. But here's the plot twist nobody's talking about: the labs that started this frenzy don't actually control most of it.

The Compute Gap Nobody Expected

According to analysis from Epoch AI, OpenAI likely used only 10–15% of the world's operational AI compute as of end-of-2025—roughly 1.7 million H100-equivalents based on their disclosed data center power capacity of 1.9 gigawatts. Add Anthropic and xAI into the mix, and you get fewer than 4 million H100-equivalents combined. Even expanding to include Google DeepMind and Meta Superintelligence Labs, the five most well-resourced frontier developers probably commanded less than half of global AI compute. That's right: somewhere between 50–80% of all deployed AI chips are running workloads that have nothing to do with training the next GPT or Claude.

Where's All That Compute Going?

The remainder isn't sitting idle in warehouses. Second- and third-tier LLM players, open-weight model inference (think DeepSeek), recommendation systems powering your Instagram feed, audio/video generation tools, biology models, robotics applications—all consume significant compute. Google alone owns roughly a quarter of the world's AI chips, but much of that infrastructure supports Google Cloud customers and internal recommender systems rather than frontier research. Meta's situation is similar: their GPUs split time between training next-generation models and ranking content in your feed. The transformer architecture innovations behind GPT-4 also turbocharged these adjacent domains, creating a compute ecosystem far larger than the frontier lab narrative suggests.

Anthropic's Revenue Trajectory Is Absolutely Insane

But this distribution won't last. Anthropic grew its annualized revenue run rate from $9 billion to $30 billion in Q1 2026 alone—an acceleration from an already-extreme 10x growth rate in 2025. OpenAI tripled its data center power capacity in both 2024 and 2025, translating to roughly 4× annual growth in H100-equivalent computing power. Greg Brockman testified that OpenAI would spend $50 billion on compute in 2026, triple what it spent in 2025. If these trajectories hold, the math gets wild fast.

The Consolidation Scenario

If Anthropic and OpenAI continue growing their compute at 33% above industry average—say, quadrupling annually while global compute "merely" triples—their combined share of world AI compute doubles in under three years and hits ~80% within five. That's not science fiction; it's basic exponentials. The author notes this headroom could be consumed even faster if Google DeepMind and Meta Superintelligence Labs also capture significant shares, which seems likely given Mark Zuckerberg's aggressive hiring push and promises of "industry-leading compute" for MSL.

The Infrastructure Ceiling

Here's where it gets interesting. Total AI capex will approach $1 trillion annualized in 2026—nearly 1% of global GDP and 3% of US GDP. Hyperscalers are growing capital expenditure at a steady ~70% annually, which suggests the broader compute buildout can't match frontier lab growth rates forever. The author puts it bluntly: maintaining 4× annual compute growth would require more than doubling capex every year from a $1 trillion starting point in 2027—an unsustainable trajectory unless AI dramatically accelerates economic productivity.

xAI's Wild Card Position

Meanwhile, xAI finds itself in an unexpected role: Anthropic agreed to rent xAI's entire Colossus 1 data center (~300,000 H100e) plus part of Colossus 2 for up to $15 billion per year. This is a fascinating development—xAI built massive infrastructure to compete with OpenAI and Anthropic, then immediately monetized it by renting to its main competitor. xAI retains the compute for its own models but now has a revenue stream while it scales Colossus 2 toward ~1.4 million H100e.

Key Takeaways

  • The five largest frontier labs control less than half of global AI compute despite kicking off the investment boom
  • Anthropic's Q1 2026 revenue run rate hit $30B, up from $9B in late 2025—an unprecedented growth curve
  • If current trajectories hold, OpenAI and Anthropic alone could dominate most operational AI compute by ~2030
  • Maintaining frontier lab compute growth requires accelerating an already $1T/year capex buildout

The Bottom Line

The AI industry's favorite narrative—that a few compute-rich labs will crush everyone else—may be true, but the timeline keeps compressing. What's wild is that we're not even there yet. Today, most GPU cycles run on stuff that has nothing to do with frontier models. Tomorrow, assuming Anthropic's revenue ramp holds and OpenAI hits its 2027 capacity targets, that changes fast. The real question isn't whether compute consolidates at the top—it almost certainly does—but whether the physical infrastructure can keep pace with the exponential ambitions driving it.