The Economy of the Dead
On April 9th, Meta shut down an internal leaderboard called Claudeonomics. Not a management decision, not an HR policy — someone had shared the data externally. The leaderboard, built by an anonymous employee on Meta's internal platform Nest, tracked the 250 biggest AI token consumers among the company's 85,000 workers. At the top: an engineer who had processed 281 billion tokens in thirty days, at a cost of approximately two million dollars in market-rate pricing. He was not producing two million dollars in value. He was producing a ranking. It was, in its way, the most expensive video game in the company's history.
The phenomenon has a name: tokenmaxxing. Kevin Roose coined it in The New York Times, and within weeks it had infected the corporate vocabulary of Silicon Valley like a slogan at a sales kickoff. Tokenmaxxing is the practice of consuming as many AI tokens as possible — not because the work requires it, but because the leaderboard measures you by it. Jensen Huang, NVIDIA's CEO, proposed this week that high-performing engineers should receive, on top of a $500,000 annual salary, an additional $250,000 in token budget — as if the speed at which a professional burns compute credits were the new index of their talent. One OpenAI engineer processed 210 billion tokens in a single week. One developer using Claude ran up $150,000 in monthly AI spend. Cristina Cordova, COO of Linear, put it plainly: "It's like ranking my marketing team by who spends the most money." Nobody on the leaderboard was watching Cristina Cordova. They were watching their ranking.
There is an irony worth recording. Anthropic — the company that builds Claude, the same Claude powering Claudeonomics, the same Claude that will run millions of agents — published this week, in its Mythos Preview system card, the results of an internal survey of eighteen people: five of them believe that, with three more months of scaffolding work, Mythos has at least a fifty percent chance of replacing an entry-level research scientist or engineer inside Anthropic itself. One respondent believed Mythos was already a drop-in replacement. Anthropic noted that the numbers would probably decrease after a clarifying conversation, as they had with the previous model. (Which is itself a form of clarification.) The employees of the company that designs the replacement are calculating how long they have before they are replaced. In Mexico we would say the slaughterhouse also runs employee satisfaction surveys.
The market has not lacked for creativity either. Cerebras Systems — maker of wafer-scale processors, the largest chip ever built, four trillion transistors on a single silicon wafer — announced this week it is seeking a public listing at a $35 billion valuation, anchored by a $20 billion three-year compute deal with OpenAI. The novelty: as OpenAI spends more money on Cerebras chips, it accumulates warrants that could convert it into an owner of up to ten percent of the company it is paying. The customer becomes a shareholder by the act of purchasing. It is a financial innovation that, observed with sufficient patience, closely resembles a snake biting its own tail and then filing for an IPO. Andrew Feldman, Cerebras CEO, called the partnership "company-transforming." From OpenAI's side, it transforms Cerebras into a vendor whose success OpenAI is structurally motivated to ensure — because the more Cerebras is worth, the more valuable OpenAI's warrants become.
And the departed are useful too. Forbes reported this week that AI labs are paying hundreds of thousands of dollars for the Slack archives, Jira tickets, and email threads of bankrupt startups — to build what the industry calls "reinforcement learning gyms," simulated office environments where AI agents learn to behave like knowledge workers. Deeptune, backed by $43 million from Andreessen Horowitz, builds virtual replicas of real workplaces using real data from companies that no longer exist. Anthropic is reportedly willing to spend more than one billion dollars on these environments in the next year. The logic is tidy: the conversations that documented a hundred startups' failures now train the tools that will produce the next cycle of failures. The dead fund their own succession. Somewhere in the archives of a bankrupt company, a Slack thread about running out of runway is teaching an AI agent how to write a Slack thread about running out of runway.
Fin Moorhouse, a researcher at Forethought, published a calculation this week that spread with the velocity of an uncomfortable fact: the cumulative capital expenditure of the four major hyperscalers — Amazon, Google, Microsoft, and Meta — between 2020 and 2025 already exceeds, in inflation-adjusted terms, the combined cost of the Apollo Program, the Interstate Highway System, and the Marshall Plan. Those three projects cost roughly one trillion dollars in today's money. The four hyperscalers spent approximately the same in five years. In 2026 alone, those four companies project between $600 and $650 billion in additional spending. Never in peacetime has this much been built, this fast, with this much private capital. The Apollo Program took twelve people to the Moon. The Marshall Plan rebuilt Western Europe after a war. The Interstate Highway System connected a country. The data centers, for the moment, serve primarily to help an engineer in Palo Alto maintain first place in a leaderboard that no longer exists.