Friday charts: the tokenmaxxing economy

Will tokens be more productive than people?

That’s the bet Mark Zuckerberg is making. Announcing plans to cut 10% of Meta’s staff, he explained that lower headcount would “make the company more efficient while freeing up resources to fund its other investments.”

In other words, he’s getting AI to do more of what people do now, and reinvesting the payroll savings to make it do even more of it later.

You only do that if you believe paying for tokens — the billable unit of AI work — is more cost-effective than paying for people. 

Many do.

SemiAnalysis founder Dylan Patel says token costs are already 25% of his payroll — and on track to be 100% by year-end. “Which is a bit terrifying,” he adds.

Indeed. Few companies can afford to double their payroll costs in the course of a year. So tokens will eventually have to pay for themselves — either by replacing labor or generating revenue.

That might already be happening. But if so, CEOs seem to be keeping it a secret. A report from Goldman Sachs says that S&P 500 companies have reported only $300 million of productivity gains from AI so far.

$300 million is roughly what Anthropic earns in a day and a half selling tokens.

It’s unclear exactly what to make of that, because the incredible demand for Anthropic’s tokens may not be a measure of how useful they are.

Employees at Meta and Microsoft have reportedly been burning tokens as fast as they can, just for the sake of burning tokens.

These “tokenmaxxers” look industrious, but it’s not clear what they’re industrious about.

Bill Gurley notes as well that, because VCs like him have been so keen to subsidize tokens (by investing billions in OpenAI and Anthropic), we don’t yet know what the real cost of producing them is. He suggests it’s much higher than what they cost now, citing the example of ride-sharing, which got a lot more expensive after VCs stopped subsidizing it.

Alternatively, the cost of providing tokens could go down. But there’s not much sign of that. So far, the biggest users of tokens always want to use the latest, most expensive models.

Patel thinks the new models will be more than worth it: “The economic value that the best model can deliver is growing faster than our ability to serve those tokens to people.”

But that, too, suggests that tokens could soon get a lot more expensive.

If so, will companies be able to afford all the AI agents they’re currently employing?

If not, could the agents soon be facing layoffs?

It’s possible.

Humans may be more cost-effective than we think.

Let’s check the charts.

Makes sense:

An Anthropic study finds that job fears are pretty well correlated to exposure to AI — ie, people in jobs more exposed to AI are more concerned about losing their job to AI. This entirely sensible finding suggests, to me at least, that job fears are not irrational.

When code writes code:

Google reports that 75% of its code is now written by code (AI). That does not, however, mean they can fire 75% of their software engineers. Someone has to tell the AI what code to write. And the new code creates new work for engineers: someone has to review it, integrate it, maintain it…

More work?

Ed Yardeni finds that, on Indeed, postings for open software developer jobs (in red) are rising while total job postings are flat. Instead of doing more with less, companies might be doing more with more.

Tokenmaxxing:

Coatue charts data from Google showing incredible growth in token usage: tokens per minute served by Google’s API were up 43% in Q4 (the start of agentic AI) and 60% in Q1. Incredible. 

GPU availability: 

The dotted lines show what percentage of the time different generations of GPUs are available for AI compute. Thanks mostly to agentic coding, it’s trending toward 0%.

Why the market is up:

It’s strange to see the stock market completely ignore war in the Middle East. But the above is why: thanks to the demand for all things AI, earnings estimates are going up at a time of year they are almost always going down. Roughly 45% of the S&P 500 is now AI-related.

Higher, but cheaper:

Earnings estimates are up so much that, despite making new all-time highs, valuation looks pretty cheap relative to history. (This data is to April 17, so it’s gotten slightly less cheap since.)

EPS vs. FCF:

Via Vincent Deluard: S&P 500 companies are reporting high profit margins but low free-cash flow ones. They’ll have to convert the earnings to cash at some point. If not, the market’s not as cheap as it looks.

Why the market is not down:

The US economy is far less exposed to oil-price shocks than it once was. The volume of US petroleum products supplied to the US as a percentage of real GDP has been trending lower since the early 1990s. “The economy now requires roughly half as many petroleum inputs to generate the same unit of output as it did in the early 1990s,” Ed Yardeni writes.

The new oil?

Coatue estimates that demand for DRAM (memory semiconductors) could 5x over just the next five years. This is likely to cause problems, because DRAM suppliers (like Micron) won’t be able to increase capacity that fast. Hopefully, some coders will still remember how to code when the token shortage hits.

Early returns?

Morgan Stanley notes that non-farm labor productivity has turned higher. If that really is attributable to AI, it could have much higher to go.

Wall Street v. Main Street:

This is a chart crime: a cumulative measure like the S&P 500 should never be charted against an oscillating survey like the University of Michigan Consumer Sentiment Index. And yet, it’s probably the chart that most captures the moment. Thanks to AI, business is booming for US tech companies, so the stock market is up. Also thanks to AI, people are worried about their jobs, so consumer sentiment is down.

I suspect that could change — because AI might soon price itself out of the job market.

Have a great weekend, tokenmaxxing readers.

— Byron Gilliam

Brought to you by:

Zcash Q1 2026: the privacy-focused blockchain had a pivotal quarter.

Nearly a third of all ZEC is now held privately, the network's mining power hit an all time high, and the SEC closed its investigation without action. A leadership transition also brought a new development team and a $25M+ raise.

Get the full Q1 report from Blockworks.