“The world we planned for was 10x annual growth. It turned out we saw 80x annualized growth in the first quarter. That’s why we ran into compute problems.”

Anthropic’s founders, siblings Dario and Daniela Amodei, just held a fireside chat at the Code with Claude 2026 conference.

In the conversation, Dario shared a number: in Q1 2026, Anthropic’s revenue and usage saw roughly 80x annualized growth.

Not 80%—80x.

Then he added, half-jokingly: “I hope this doesn’t continue, because this kind of hypergrowth is ‘too hard to handle.’ Maybe ‘just’ 10x growth would be better.”

▋An era where supply can’t keep up with demand

The importance of that 80x figure is that it reveals the AI industry’s most central structural contradiction right now: demand is growing far faster than infrastructure can expand.

On the very same day as the fireside chat, Anthropic announced a major compute deal, taking over 300 megawatts of power capacity from SpaceX’s Colossus One data center.

Because Claude Code (their coding agent tool) is simply too popular—users don’t have enough time available each day—so they had to urgently expand capacity.

Anthropic planned for 10x growth; what happened was 80x. That means that even teams at the cutting edge of AI badly underestimated how quickly the market would absorb “truly useful AI tools.”

If even they were that far off, how credible are everyone else’s forecasts in the market?

▋Building for an exponential future

Another core message in the conversation was what Dario repeatedly emphasized: “Build for the Exponential.”

What he means is: don’t design products only for today’s model capabilities. Design for the models six to eighteen months from now.

It sounds like a cliché, but he offered a concrete framework for thinking about it:

“Some products can’t be done with today’s models, but can be done with future models. That makes internal experimentation extremely important. You have to keep testing with the latest models to see what works.”

He also mentioned a trajectory of product evolution:

- Phase 1: Code assistance

- Phase 2: Software engineering (larger projects, architecture design, debugging)

- Phase 3: Helping build and operate an entire company

Right now we’re roughly between Phase 1 and Phase 2. But given the pace of improvement in model capabilities, Phase 3 may arrive sooner than most people expect.

Dario often cites a “chessboard paradox” in other settings: if you place one grain of rice on the first square of a chessboard, two on the second, four on the third, and so on, how much rice will there be on the 64th square?

The answer is an unimaginably astronomical number. And the key is that growth in the first half looks very gentle; the real explosion happens in the second half.

He believes we’re around square 40 right now—we’ve already entered the steep part of the exponential curve, but most people’s intuition is still stuck in linear thinking from the first half.

▋The saturation of chatbots and the rise of agents

“Products will saturate when the models get too good. The way we make models smarter today is more apparent in agents and Claude Code workflows than in chatbots. You have to think about what the next big trend is.”

Chatbots are approaching their capability boundary—not because the models aren’t good enough, but because the chat interaction format itself has natural limitations. You ask one question, the model gives one answer, then you ask the next. It’s a human-driven, linear workflow.

Agents, by contrast, are completely different. You give them a high-level goal, and they autonomously plan, execute multiple steps, use tools, handle errors, and ultimately deliver a result. It’s an AI-driven, autonomous workflow.

This kind of autonomous workflow is an order-of-magnitude leap in the release of productivity (and in compute consumption).

This also explains why 80% of Anthropic’s revenue comes from enterprise customers: companies aren’t just looking for an AI to chat with anymore—they want a digital employee that can work autonomously.

▋Imagination is the only limit

When even the people at the very forefront underestimated things by eightfold, when even the teams building these models are being swamped by the demand they themselves created, we should understand this: trying to make sense of this revolution with linear thinking is destined to be wildly wrong.

The truth is, nobody knows what models will be able to do six months from now; nobody knows what demand will be in eighteen months; nobody can draw the endpoint of this curve.

Our intuition completely fails here, because the human brain evolved for a linear world.

So rather than trying to predict, it’s better to accept a simpler premise: the ceiling is much higher than you imagine—and it may not exist at all.

The only certainty is that measuring tomorrow with yesterday’s yardstick is a mistake.

- KP

p.s. After talking semiconductors for several weeks straight, everyone who was going to make money probably has already made it. Those who are still wavering will probably keep being afraid of heights. It’s time to shift our focus to what, after this war, has been structurally changed—and is worth studying.

In my latest in-depth analysis, I explored the logic behind four major themes: LNG, oilfield services, refining, and fertilizers. I recommend reading the full piece.