For the last three years, the AI industry had a simple mantra: more is more. More data, more compute, more tokens. OpenAI and Anthropic raced to build ever-larger models, fueled by venture capital and the promise of AGI. But the music has stopped. The check writers are now asking for returns, and the hangover is brutal.
Companies are tightening their AI budgets. The shift from tokenmaxxing—the practice of throwing massive compute at problems—to efficiency isn't just a trend. It's a survival reflex. And it could dampen growth rates at the two most prominent AI labs on the planet.
The Party's Over
Walk into any Silicon Valley boardroom and you'll hear the same question: "What's the ROI on our AI spend?" Three years ago, that question was laughed off. Today, it's a threat. CFOs have seen the bills for GPT-4 and Claude deployments—millions in API calls, fine-tuning, and inference—and they want proof that this translates to revenue.
At OpenAI, the pressure is acute. The company has reportedly burned through billions in operating costs, with revenue growth failing to keep pace. Anthropic, more cautious but still capital-intensive, faces similar headwinds. The era of building without a business model is ending.
"The low-hanging fruit is gone. Now we have to sweat the assets." — former OpenAI engineer
From Tokenmaxxing to Pragmatism
The technical shift is just as stark. Tokenmaxxing—cranking out massive responses with minimal filtering—is giving way to efficiency engineering. Companies are now pruning models, quantizing weights, and using sparse attention mechanisms. The result? Smaller models that do 80% of the work for 20% of the cost.
This is bad news for labs that sell access to large models. If a custom-tuned 7B parameter model can handle customer support as well as GPT-4, why pay the premium? The market for giant, general-purpose models is commoditizing.
Who Wins?
Not everyone loses. Nvidia might see a dip in demand for its H100s as companies stop building new clusters. But the real winners are the lean startups that never bought into the tokenmaxxing hype. They built efficient pipelines from day one, and now they're eating the lunch of the incumbents.
OpenAI and Anthropic have a choice: adapt or get disrupted. OpenAI is already hedging, rolling out smaller, cheaper models and pushing enterprise contracts. Anthropic is doubling down on safety as a differentiator, but safety doesn't pay the bills unless clients buy it.
The next 12 months will determine whether these labs are the next Google or the next Netscape—disrupted by a market that stopped caring about size and started caring about results.
