Tokenomics: Why the AI Token Is the New Semiconductor Chip

If you’re wondering how anyone makes money when AI budgets are blowing up, history has an answer: we’ve seen this play out before, in silicon.

You may have heard about Uber’s CTO talking about runaway AI budgets, and a similar story playing out at IBM. When I first read about it, the situation felt strangely familiar, not from the AI world, but from my years in the semiconductor industry. The more I dug in, the more convinced I became that we’re watching a sixty-year-old story repeat itself, almost beat for beat. Hence, this article.

The “Tokenmaxxing” Bubble

Here’s what happened at Uber. The company rolled out Anthropic’s Claude Code (along with some Cursor) to its roughly 5,000 engineers. To drive adoption, management set up internal leaderboards ranking engineers by their AI tool activity.

The incentive structure worked too well:

• Massive adoption: By April, 95% of Uber’s engineers were using AI tools monthly, and roughly 70% of all committed code was touched or generated by AI.
• Astronomical costs: Agentic coding tools work by reading entire code repositories, planning, executing, and testing. This process swallows millions of tokens per task. API costs per engineer quickly skyrocketed to anywhere between $150 and $2,000 per month.

The situation has triggered a massive internal debate about what people are calling the “tokenmaxxing” bubble, where sheer volume of AI usage doesn’t necessarily equal actual business value. When you reward people for consuming tokens, they consume tokens. Whether that consumption produces anything useful is a separate question entirely.

The Music Has Stopped

I see the Uber AI budget blowout as a classic “the music has stopped” signal for the first phase of the AI boom. It marks the transition from narrative-driven investing (where companies were rewarded just for using AI) to fundamentals-driven investing (where companies must prove AI drives actual cash flow).

The pressure isn’t limited to AI consumers like Uber. To appreciate the scale of money involved, consider this: a data center that consumes 1 GW of power costs approximately $45 billion to build. With the hyperscalers planning data centers that consume several gigawatts, the cumulative buildout is headed into the trillions of dollars.

That is why the hyperscalers, the companies building the AI infrastructure itself, are increasing their already massive capex to the point where their FCF (Free Cash Flow, the money left after all expenses are paid) is shrinking toward too small or even negative. To keep funding the buildout, they are issuing more shares, issuing bonds, or tapping private credit.

So, what is the plan? How is anyone going to make money?

A Lesson from Silicon

For a possible answer, it helps to look back at an older industry, because the parallel between the early days of silicon and today’s AI budget crisis is strikingly exact: in both cases, a revolutionary technological “input” was too expensive for normal businesses to justify, forcing the industry to invent a new economic and manufacturing playbook.

Early semiconductors were astronomically expensive and heavily subsidized by the government: the Apollo program and defense contracts were essentially the only customers who could afford them. The economics of early microchips unfolded across a fascinating trajectory of high costs, low yields, and then a sudden collapse in pricing.

Industry pioneer Robert Noyce (co-founder of Fairchild Semiconductor and later Intel) made a legendary, aggressive business gamble. He realized that if he deliberately slashed the price of Fairchild’s integrated circuits below what it actually cost to make them at low volume, he would stimulate massive commercial demand, and that demand would drive the volume needed to make the low price profitable.

The Cost Collapse That Followed

Then something remarkable happened, and this is the heart of the story:

• 1965: Gordon Moore observes that transistor density doubles roughly every two years, an observation now known as Moore’s Law.
• Every doubling of density meant roughly halving the cost per transistor.
• Process improvements drove yields up dramatically.
• Volume drove fixed costs down per unit.
• Equipment vendors got better, supply chains matured, and knowledge spread.

By the 1980s, chips were consumer products. By the 1990s, they were in toys. By the 2000s, they were essentially disposable. A transistor that cost about $1 in 1968 costs less than a billionth of a dollar today.

The AI Token Is the New Semiconductor Chip

When the two stories are placed side by side, the resemblance is hard to miss:

(Jevons Paradox, if you haven’t come across it: when a resource gets cheaper to use, total consumption of it goes up, not down. Efficiency creates demand.)

The Cost Collapse Is Already Happening

This isn’t speculative. It’s already underway. Look at the trajectory of output token pricing:

Token costs have dropped 99%+ in five years. Semiconductor transistor costs dropped 99% in roughly fifteen years during their early era. AI is moving faster.

The Road Ahead

The tokenmaxxing bubble is real, and companies that can’t connect AI usage to cash flow will feel the pain as investors shift from narrative to fundamentals. But the budget blowouts at Uber and IBM aren’t evidence that AI economics are broken. They’re evidence that we’re still in the “$1,000 chip” phase of the story.

If the semiconductor playbook holds, the path forward looks familiar: aggressive pricing to stimulate demand, relentless cost reduction through better architectures and inference optimization, and volume that turns an exotic, subsidized input into an invisible, disposable commodity. The companies that survived the silicon transition weren’t the ones that consumed the most chips. They were the ones that figured out what to build with them once they became cheap.

The same will be true for tokens.