AI Companies Are Running Out of Compute—and Money

Here's something worth paying attention to: when you use AI right now, you're probably not paying what it actually costs. Even if you're on one of those $200-a-month plans, the company serving you that AI is likely losing money on the transaction. And that's about to change.

The shift isn't abstract. On Monday, Microsoft's GitHub announced it's moving Copilot to consumption-based pricing. The multipliers tell the story: Claude Opus 4.7 went from a 7.5x pricing multiplier to 27x. Gemini 3.1 Pro and GPT 5.3 Codex both jumped from 1x to 6x. Across the board, you're looking at roughly a 6x price increase for frontier coding models.

Developer Peter Deneen put it bluntly: "These new Copilot multipliers starting June 1st are absolutely ridiculous. I can only imagine this pricing is going to force users to lock in with a single foundation model vendor just to manage costs. Honestly, it would be hard to find a clearer indicator that the subsidy era is over than Microsoft literally revealing how deep their subsidies had been with this massive price hike."

The Physics Problem

The immediate cause is simple physics: there isn't enough compute to service all the AI demand that exists. The longer explanation involves what happens when AI stops being a chatbot and becomes an agent.

When AI agents write code, they don't just answer a question—they run multi-hour autonomous sessions, iterating across entire repositories. GitHub's chief product officer Mario Rodriguez explained the shift: "Copilot is not the same product it was a year ago. It has evolved from an in-editor assistant into an agentic platform capable of running long, multi-step coding sessions using the latest models and iterating across entire repositories. Agentic usage is becoming the default and it brings significantly higher compute and inference demands."

The numbers are striking. One power user reported consuming a billion tokens last month—the equivalent of about 7,500 books worth of words. Multiply that across millions of users and you start to understand why companies are scrambling.

Between November and April, the percentage of people who had never used AI dropped from 26% to 17%. At the same time, heavy users grew from 17% to 24%. More people using AI, using it more intensively, consuming exponentially more tokens as agents replace simple queries.

Anthropic's Uncomfortable Position

Anthropic is currently the most visible casualty of this compute crunch, though calling it a casualty misses the point—they're straining under the weight of their own success.

The company has been plagued by outages. They've started metering computing supply during peak hours. They forced some Claude users onto the API. They ran a test removing Claude Code from pro subscriptions. Most tellingly, they've withheld release of their most capable model.

Technology author Tae Kim observed: "It's obvious Anthropic vastly underestimated compute growth needs, which is expanding much faster than expected. Dario is on the record multiple times describing OpenAI as YOLO, recklessly buying too much capacity, but now it looks like Sam Altman was right all along."

OpenAI, predictably, has seized on this narrative shift. After releasing GPT 5.5, Sam Altman wrote: "Really excellent work by the inference team to serve this model so efficiently. To a significant degree, we have become an AI inference company now." For those watching the compute wars, he's basically saying: we can deliver, they can't.

But even OpenAI doesn't have infinite compute. No one does. The difference is in how close each company is cutting it to the bone.

What Wall Street Misses

The market commentary around all this has been predictable and mostly wrong. Some analysts are reviving the AI bubble narrative, just with new reasoning. Author Derek Thompson summarized the shift: "The AI bubble argument has meaningfully shifted from 'the revenue growth curve looks weak' to 'well sure the revenue growth curve is ferocious, but it's being subsidized by below market token pricing and corporate AI FOMO.'"

The problem with this analysis is timing. Wall Street is reacting to data from 2-6 months ago in an industry that operates on what amounts to dog years. OpenAI's Codex app users grew 20x this year—from 200,000 on January 1st to 4 million by late April. Unless you're tracking week-to-week numbers, you're already obsolete.

The Job Displacement Calculus Nobody's Doing

Here's what's more interesting than market dynamics: what happens to the job displacement narrative when AI costs real money?

Most discourse about AI replacing workers assumes AI will be radically cheaper than human labor. But companies are starting to report something different. Hedgeye Markets noted that "Goldman Sachs reports that companies are blowing past their AI inference budgets by orders of magnitude, with inference costs and engineering now approaching 10% of total headcount costs and potentially reaching parity with salaries within several quarters."

Abacus AI's Bindu Reddy wrote: "Our AI bill will overtake payroll in 6 months. We now have limits on how much employees can use our product on a daily basis."

If AI agents cost roughly the same as human workers—at least for now—the calculus changes completely. Survey data from the AI Daily Brief showed that between January and March, cost savings cratered from 19.7% to 12.7% as the primary benefit of AI. Meanwhile, new capabilities jumped from 21.9% to 29.3%.

People aren't using AI because it's cheap. They're using it because it can do things they couldn't do before.

The Accidental Pause

There's an irony here worth savoring. While activists signed open letters calling for AI development to slow down, and protestors demanded a pause, the thing that might actually moderate the pace of AI adoption is much more mundane: not enough electricity, not enough chips, not enough data centers, not enough compute.

Jamie Dimon articulated a version of AI concern at Davos that didn't center on existential risk but on pace: he wasn't worried about long-term adaptation, just whether changes might happen "too fast for society."

If it turns out that physics and economics impose their own speed limits—that running an AI agent costs enough to make organizations think carefully about when and how to deploy them—that might give us exactly the breathing room Dimon was worried about losing.

The subsidy era created a distorted picture of what AI deployment would look like at scale. Now we're finding out what it actually costs. That's not necessarily a crisis. It's just reality catching up.

Marcus Chen-Ramirez is a senior technology correspondent for Buzzrag.