Salaries to Servers
I want the AI optimists to be right. The numbers I'm watching aren't yet letting me.
Three university commencement speakers were booed this month for talking about AI. Eric Schmidt, the former Google CEO, was one of them. He had to wait out the noise before he could finish his thought.
The reflex move is to call this AI propaganda. Or generational unseriousness. Or foreign influence operations seeded by hostile state actors. The third is not paranoid: state-aligned campaigns out of China and Russia almost certainly are amplifying anti-AI sentiment in the United States, the same way those operations have amplified every other domestic fracture for decades. Gavin Baker raised this possibility on Thursday’s All-In Podcast, and his suspicion is not unreasonable. But pointing to the campaign is not the same as explaining the booing. Information operations amplify what is already moving. They do not invent the underlying signal.
The booing is information.
Where I sit
My primary vocation, for more than twenty-five years, has been pastoral. Hospital visits, funerals, weddings, counseling appointments, the conversation after Sunday service about a job that disappeared and a mortgage that did not. That is the room I write from. I am with real people every day, and what is happening to them is not a hypothetical to me. The AI work came later. I am a PhD student in AI ethics. I run a company doing AI research and development. I use Claude, ChatGPT, Grok, Gemini, and DeepSeek every day, and not only through the chat windows most people see, but inside the systems that sit underneath those chat windows. I want the AI optimists to be right.
That last sentence is not a rhetorical concession. It is the actual position. Mass white-collar displacement is a disaster I would prefer not to live through. Universal abundance is a future I would prefer for my children. When David Sacks says the United States cannot afford to slow down on AI, I find myself nodding more than I would like to admit.
What follows, then, is not the complaint of an outsider. It is what I see from inside the room, written by someone who walked in hoping the optimists would be right and is still waiting for the data to confirm it.
The optimists’ case, steelmanned
The optimists are not making things up.
Goldman Sachs reported this week that the U.S. labor market is healthier now than it was when ChatGPT launched in late 2022. That is not a forecast. That is the most recent reading of the actual hiring data.
Jeff Bezos pointed out, correctly, that every general-purpose technology in history (steam power, electricity, the personal computer, the internet) eventually produced more jobs than it destroyed. The fear of mass unemployment has a long track record of being wrong. Economists call the mistake the lump of labor fallacy: the assumption that there is a fixed amount of work in the world and machines eat it. The Lump of Labor Has a Lump of Labor Problem walked through the argument earlier this month. The fallacy is real, and a century of automation predictions has overshot for exactly the reasons Bezos says it has overshot.
Anthropic, the company that builds Claude, just posted its first profitable quarter at scale. It is now running at roughly $45 billion in annualized revenue, meaning that if its current monthly sales held steady for twelve months, that would be the total. Gavin Baker put it simply on Thursday’s All-In Podcast: if Anthropic and OpenAI together are pulling roughly $100 billion in revenue right now with margins north of 80 percent, “the returns are there.”
The optimists’ best concrete example came from the same episode. Baker described a hedge fund manager he met recently whose daughter was born with a rare genetic condition that should have made an ordinary life impossible. The father refused the diagnosis. He spent months running medical research through large language models, found an existing safe drug, and tested it on his daughter. Her brain function went from 30 or 40 percent to 80 or 90 percent. He now believes he is months away from a complete cure. Not just for his daughter. For everyone with the disease.
That story is real. The drug is real. The daughter is real. If the steelman of the optimists’ case were a single image, it is that father.
What the steelman doesn’t cover
This month was also when Matthew Prince, the CEO of Cloudflare, sent a memo to his company informing them that he was laying off more than 20 percent of the workforce. He cited AI directly. The eliminated group was given a label: “measurers,” meaning the employees whose job was to track and report on data. The memo opened with these words: we posted record revenue growth, have strong free cash flow, and are adding an unprecedented number of customers.
Record revenue. Twenty percent of the workforce eliminated. Same paragraph.
The same window, Mark Zuckerberg laid off 8,000 employees at Meta, about 10 percent of the company. The same week, Meta announced that it would install recording software on every remaining employee’s computer so the company could study how those employees work and use the recordings to train its internal AI models to do the same jobs. One observer on Thursday’s All-In Podcast summarized what the remaining Meta employees walked away believing: the best they could now hope for was to keep the job long enough to train their own replacement.
Microsoft, the largest single investor in OpenAI, canceled the licenses it had been paying for its own engineers to use Claude Code, Anthropic’s coding tool, because the bill, charged in tiny units called tokens the way a phone bill is charged in minutes, had grown faster than the productivity gain it was buying. Uber burned through its entire 2026 AI spending budget in four months.
And Anthropic, the same company posting its first profitable quarter, is now paying SpaceX roughly $15 billion per year to rent the use of a computer farm called Colossus, essentially a warehouse full of stacked-up specialty processors that the AI models live on. Fifteen billion. From one AI company. To another company. To run computers.
The CFO term for what is happening is operating leverage. The plain-language phrase is something else. Companies are moving money from salaries to servers.
That is the line worth pausing on.
Companies posting record profit and record layoffs in the same quarter is historically unusual. The two normally move in opposite directions. When a company is growing, it hires. When a company is contracting, it fires. The combination usually only appears when management discovers it over-hired the year before and is correcting for the mistake. That is not what is happening now. The work itself is moving. Out of the spreadsheets that show payroll. Into the spreadsheets that show cloud infrastructure.
The routine should we hire another analyst, or buy more software? questions every CFO faces every quarter are now favoring the second answer with the kind of regularity that did not exist eighteen months ago. That is not a moral claim. That is an accounting claim. But the accounting one is where the social one begins.
Everybody is trading their own book
The most useful thing said about the AI debate on Thursday’s All-In episode was not said by an optimist or a doomer. Chamath Palihapitiya said it about all of them, simultaneously. “Everybody is trading their own book.” The phrase is a Wall Street idiom for promoting a position you already hold. Pushing Their Book applied it to Anthropic two months ago, when the company’s “accidental” leak of its Claude Mythos model read like the most sophisticated product launch in AI history. What Palihapitiya did on Thursday was extend the same diagnosis to every loud voice in the debate. Each of them has a financial position that depends on the public believing a particular version of the future.
Dario Amodei, the CEO of Anthropic, warns of mass unemployment because regulatory moats around frontier AI models would benefit the few companies large enough to comply with them. Anthropic is one of those few companies. Bezos forecasts a labor shortage produced by abundance because the productivity-abundance scenario benefits the warehouses he owns, the cloud he rents, the equity he holds. David Sacks personally called the President on Thursday to kill a federal AI executive order that would have required pre-release review of frontier models. Sacks holds a portfolio that ships faster when frontier models ship faster.
I should be transparent about my source here. I have listened to the All-In Podcast since Chamath Palihapitiya, David Sacks, David Friedberg, and Jason Calacanis first started recording it during the COVID shutdowns. It is one of the better business and tech conversations on the internet most weeks, and I learn from it. It is also, every week, four extremely successful tech investors whose portfolios benefit when AI keeps moving in the direction they are publicly arguing it will. They are trading their books too. I listen anyway. I am just careful not to forget which seat they are in.
This applies to me as well. I run an AI company. I have a financial interest in AI working out well. Any piece I write about AI’s downside risks is being written by someone who, if AI somehow turned out to be a footnote in the history of technology, would prefer that not happen.
Chamath’s correction is the right one. The people whose books are smallest and whose voices are quietest are the people whose testimony is least trapped by incentive. He quoted Shyam Sankar, the CTO of Palantir, on the same episode: stop asking the model-makers what they think. Ask the truck driver. The package sorter. The ICU nurse who now has more time at the bedside because the AI charted the patient’s vitals while she was in the next room. The writer whose drafts now have to compete with prose generated in seconds by something her own publisher trained on her catalog without paying her.
When you do that, you find both. You find the ICU nurse. You find the hedge fund father who got his daughter back. You also find the Cloudflare “measurer” who, while looking for a new job, gets to explain to every recruiter why his last employer eliminated his entire job category by name in a press release.
The data is not in any of the loud voices. The data is in the spreadsheet.
The PR problem is pattern recognition
The optimists have framed public sentiment as a PR problem. Gavin Baker, on the same episode, floated the possibility that there is a Chinese Communist Party-funded influence campaign behind the protests against new AI data centers. That is not impossible. Influence operations are real. But it does not explain a Harvard senior booing a commencement speaker. It does not explain UK polling published this month showing that 57 percent of British adults believe AI will destroy more jobs than it creates, and 65 percent believe the benefits of AI will flow mainly to wealthy investors and large corporations.
That second number is not a forecast. It is a description. Sixty-five percent of an entire country looked at the actual economic structure of the AI industry, a handful of foundational labs, a handful of hyperscale cloud providers, a small group of large equity holders, and named what they saw.
David Friedberg offered the deeper version of this on the same episode. He pointed out that technology, as a category, has always created leverage for a small group of people, but the time-to-diffusion has historically been long enough that the rest of the population eventually catches up. The internet did not arrive everywhere in 1995. It diffused over twenty years. AI is compressing that diffusion curve in ways that nobody is hiding. The companies are explicit about it. The CFOs are explicit about it. The CEOs are explicit about it in the same memos that announce the layoffs.
The booing at commencement is what pattern recognition looks like before it has a vocabulary. The 22-year-old in cap and gown is almost certainly seeing some of this on a TikTok feed that hostile information operations are working to tune. That is a fair concern and worth taking seriously. But she is also reading the same news the rest of us are reading. She is watching the same memos go up the same week she is being told her degree was the right investment. She is in the group text with classmates and recent graduates who cannot find a first job because the entry-level positions that used to be the open door into a career have been replaced by AI agents that one senior employee can manage instead of training a team of new hires. The amplification is real. The arithmetic underneath it is also real.
The optimists’ PR problem is not a PR problem. It is a math problem. And calling it propaganda will not make the math come out differently.
Why this matters for the optimists themselves
This is where the optimists should be paying attention more than anyone else. Their political position is structurally unstable.
If 65 percent of voters in any country say AI’s benefits are flowing to wealthy investors, the regulatory response arrives before the abundance does. The European AI Act already exists. California’s governor signed an AI labor protection executive order this past week, the same day the federal one was scrubbed at the last minute by a phone call. State legislatures are next. The longer the optimists call public sentiment a propaganda problem, the further the actual regulators move ahead of them, and the harder the eventual reckoning becomes.
The corrective is not better messaging. The corrective is the one Gavin Baker himself half-articulated on Thursday when he said it was incumbent on everyone in technology to be an advocate for the positive possibilities of AI. He is right. It is incumbent. But advocacy from inside a memo that calls 20 percent of your workforce by the name of the function being eliminated does not land as advocacy. It lands as confirmation of what the booing students were already telling you.
And if foreign campaigns really are amplifying anti-AI sentiment on social media, the optimists own the same channels. The platforms are largely American. The capital is largely American. They could counter-program with abundance any time they wanted to. Better still, they could do what the wave of newly rich industrialists did at the turn of the twentieth century. Andrew Carnegie built more than two thousand public libraries. John D. Rockefeller funded the General Education Board and helped build the University of Chicago. J. P. Morgan underwrote hospitals and museums. They were not philanthropists in spite of being rich. They were civic builders because they were rich. They poured the upside of the industrial transition back into the public life of the country that produced them. The current generation of tech billionaires, soon to be tech trillionaires, has the same option on a faster timeline and with a wider reach. NVIDIA announced eighty billion dollars in additional stock buybacks last week. That is great for shareholders. Is it great for society? It is a legal and common choice. But so is the alternative.
The optimists have to choose what they are actually defending. If it is the medical breakthrough the hedge fund father produced, defend that. Tell that story everywhere. Make it the public face of the technology. But if you defend the daughter with one hand while you fire 8,000 people with the other and announce recording software on every laptop in the same week, you have not made the optimists’ case. You have made the doomers’ case for them, and signed it with your own name.
What I am watching that I want to be wrong about
Three patterns. I am writing them down so that, if AI’s first year of compounding really does produce the abundance the optimists promise, I will have a record to return to.
The first is salaries to servers as a recurring quarterly pattern. Cloudflare, Meta, Microsoft, Uber, Anthropic in one cycle is a data point. It is not yet a trend. But it is a data point that, if it appears again in the second quarter, and again in the third, will reshape how every CFO in the country thinks about headcount. There is a number, somewhere between one quarter and four quarters of repetition, where this becomes the new baseline. I would like that number not to exist.
The second is the cohort asymmetry. The previous wave of labor displacement, the offshoring of manufacturing to Mexico and China in the 1990s and early 2000s, hit a population that absorbed it slowly because the people displaced had two things at once. They had a theological vocabulary for absorbing suffering. And they had the social structures (congregations, extended families, civic groups, neighborhoods where the unemployed neighbor was visible and named) that gave the vocabulary somewhere to live. The credentialed white-collar class that is about to absorb the next wave has neither. That deserves its own piece, which is what the next article in this series will do. The Weight of What’s Coming was the early version. The full version is queued up.
The third is the solution that is going to be offered. Elon Musk has been talking for over a year about a “universal high income.” Sam Altman has been running Worldcoin and writing about UBI for longer than that. Both are real proposals from people with the capital to fund them. Both have problems I do not yet hear them addressing. That is the third article in this series.
Close
I am still rooting for the optimists. I run a company that is rooting for them. The future where AI quietly cures diseases, lowers the cost of living, and frees people to do work the previous generation could not have imagined is the future I would pick. Twice.
But rooting for an outcome is not the same as observing one. The price of progress is paid, somewhere, by someone (A Debt, Not an Externality named the principle in April). The optimist case is honest only when it names the bill.
If the structural shift I am watching this quarter is real, then calling the public’s pattern recognition a PR problem is the optimists’ problem, not the public’s. The fix is not better messaging. The fix is either to let the data catch up to the story, or to let the story catch up to the data.
For this week, the question is the simpler one.
Where, exactly, is the money going?
Sources
All-In Podcast, episode 274, “SpaceX’s $2T Case, Nvidia’s Shock Selloff, America Turns on AI, Trump Pulls AI Order, Bond Crisis” (May 22, 2026)
Goldman Sachs labor market commentary (May 21, 2026) via @unusual_whales
Anthropic Q1 / Q2 financials and SpaceX-Colossus deal: Sami Minhas / Dan Nystedt and All-In S-1 teardown (May 22, 2026)
Meta 8,000 layoffs / Zuckerberg recording software framing: @interesting_aIl
Bezos labor-shortage framing (amplified by Bill Ackman and David Sacks): @BillAckman
David Sacks intervention on Trump AI executive order: Sophia Cai
Newsom California AI labor protection executive order: Cointelegraph
Shyam Sankar (Palantir) on listening to frontline workers: quoted on the All-In Podcast (May 22, 2026)
AI disclosure
This piece was drafted with Claude as a writing partner against an outline I built from the underlying source materials (the All-In episode transcript, the past week of Watson daily briefs, my own working notes). The argument, the framing, and the editorial decisions are mine. The voice is mine. Claude assisted with structure, fact-checking against the source documents, and prose tightening.
The next couple of quarters will be very interesting. Great article Pastor Miles.
Meta, Oracle, Cisco, Intuit, HSBC… Drucker’s model in action. Prince just said it out loud.