Tokenmaxxing
In which very large numbers do peculiar things
As a thanks to my paid subscribers they received this article yesterday. After a day, it opens up for everybody else.
There is, traditionally, a school of capitalist thought which holds that being banned by the Pentagon as a national security risk is bad for the share price. Anthropic, with quite remarkable patience, has spent the past four months testing this proposition.
It seems to have gone rather well for them.
Last Thursday, in what we are reliably informed is the largest private capital raise in the recorded history of this universe and possibly several smaller ones besides, the company closed a $65 billion Series H at a $965 billion post-money valuation.
That makes it, by about $113 billion, larger than the company that received the Pentagon contract instead.
By Friday morning, Bloomberg was reporting that Apollo and Blackstone were quietly arranging a separate $36 billion debt deal, run through a special-purpose vehicle that would buy Google’s TPUs and lease them back to Anthropic. One of the largest private credit transactions ever assembled.
$101 billion of fresh capital across twenty-four hours, for one company. Against, and I want to give them their friendliest possible number, a “run-rate revenue” of $47 billion - which means take the best recent month, multiply by twelve, squint. So in the course of a business day, they signed up for more financing than their bestest extrapolated annual revenue number.
This was widely received as good news.
Everyone looked at the headlights headlines. No1 looked at the structure.
First thing I noticed is that this deal was so big that Anthropic itself - now a trillion-dollar company - isn’t creditworthy enough to support the debt.
The shortfall has been solved with characteristic ingenuity… Broadcom - the company that manufactures the TPUs that Google sells, that Apollo is now financing the purchase of, and that Anthropic will lease - is backstopping the largest portion of the loan. (gasp for breath)
Long story short: Broadcom is on the hook to pay the noteholders if Anthropic defaults on the lease and the resold chips don’t cover what’s owed.
Google is on multiple sides. Amazon, too. They dropped another $5 billion into Anthropic earlier this quarter, on top of the $13 billion already in. In return they got a hundred-billion-dollar Anthropic commitment to spend it on AWS.
Funny how modern finance works, right? That same dollar that gets counted as Anthropic equity, gets re-counted as AWS revenue, as Anthropic capex commitment, as Broadcom backstop, as an Apollo asset, and now also as TPU sales.
The technical term for this in modern finance is “synergy”.
Not going to dive into that here, but if you want to get up to speed, this is my latest on that particular topic:
While they were negotiating this SPV, the spot rental price of the H200 - Nvidia’s last-gen chip, the sort of asset the SPV is implicitly collateralised against - dropped 40% in three weeks.
From $7 an hour to $4.
Bulls were all like “Blackwell’s rotating, not weakening demand”. 🤦♂️
My read? Those chips, those things behind the loan? They lost half their resale value. And it didn’t even roll out of the garage yet! Apollo should have done their due diligence. They knew. Or they’re really bad at their job. And I don’t believe the latter for one second.
You know what else happened this week? The rest of the tech press spent writing about cost discipline. Peculiar little thing that is so 1990s. Companies have to earn money, shouldn’t spend more than they get in… So pre-QE, pre-2008, pre-printing press times. Good (g)old days…
Microsoft yanked Claude Code from thousands of engineers in December on the rather modern grounds that the bills exceeded the productivity.
Uber blew its entire 2026 AI tools budget by _April_.
Amazon, with the kind of corporate self-awareness that gives one some hope for the species, was discovered to have run an internal leaderboard ranking employees by token consumption. The staff verb for the desired behaviour is “tokenmaxxing”. The leaderboard went quietly dark after engineers were caught padding their numbers for performance reviews.
Meta had a near-identical one called “Claudeonomics”, which went equally dark two days after The Information picked it up.
But the most candid contribution came from one Bryan Catanzaro, Vice President of Applied Deep Learning Research at Nvidia, who told Axios on the record that on his team the cost of compute is, and one quotes, “far beyond” the cost of the employees.
Nvidia sells the chips. Its four-trillion-dollar market cap rests entirely on the proposition that the chips are cheap and getting cheaper.
And it adds up… Microsoft’s cancelled licences and Uber’s blown budget are the kind of expense managers argue about at QBRs. Tens of millions, occasionally a hundred.
The financiers saw that and said: “watch me!”…
The reassurance you’ll hear whenever someone notices this discrepancy is the J-curve. Don’t worry about it.
The tokens, you see, are about to multiply. The costs are about to fall. The curves will intersect at some unspecified point in the future, and the unit economics will at that moment become favourable. Please do not press for further detail.
Goldman helpfully supplied the consumption number: enterprise tokens rising 24x by 2030.
Gartner provides the cost number: inference 90% cheaper by 2030.
Both numbers get quoted side by side in every “but the economics will work out eventually“ piece. Reassuring.
Try multiplying them.
Twenty-four times the volume at one-tenth the unit price… the bill goes up by 2.4x.
Pure primary-school math.
And that’s the floor. The “chatbot” we all know. Agentic workflows is another beast altogether. Those multiply tokens like crazy. “One prompt, one answer” becomes fifteen branching reasoning calls with three sub-agents recursively verifying each other. Agents calling agents calling agents. And of course: each one’s billable.
And this is where I had to put my coffee down: Goldman’s own delta-one desk noted the same day that DeepSeek had ALREADY cut its token pricing by 75%. Xiaomi’s MiMo did 99%.
Back to the future I guess? Welcome to the year AD 2030…
And Catanzaro was saying that his bills were crushing him already at the falling prices!
Then there’s the productivity question hiding inside this productivity story.
Demirer, Musolff and Yang published a paper at the NBER last week measuring the gains from agentic AI coding tools at four stages of the software pipeline. In descending order:
Writing individual files: +290%
Bundles of work: +150%
Distinct products: +50%
Released, shipping product: +30%
The further up the abstraction chain you go - from “this AI helped me write a file” to “this AI helped us ship a product” - the more spectacularly the gain collapses.
It works in the small. It fails in the large.
The bills, of course, are sized to the small.
This is the same dynamic I described in “The faster you grow, the faster you go bankrupt”:
Anthropic’s own gross margin is about 40%. Sixty cents of every revenue dollar is used to pay for inference. In December ‘25 they had to cut their own 2025 margin projection because inference costs ran 23% over budget.
The company that everyone is paying, can’t make the numbers work either…
And that’s just one side. The operations.
Somehow - as if that is not fucked up enough - the capex side is worse.
Panmure Liberum did the homework so the rest of us wouldn’t have to.
They took the announced 2025-2030 hyperscaler capex, lined it up against the implied revenue assumptions, and asked the obvious question: what return on capital does this clear?
Then they ran it again. This time assuming the entire build cost nothing.
Free chips. Free electricity. Free land. Free salaries. Free everything.
Only Amazon is slightly positive… and most of that is Anthropic’s hundred-billion AWS commitment. Which’s… ‘hello…’ the thing we just talked about.
None of those multinational companies can make their AI capex pay back. Even if you give them everything for free.
I’m not making this up. The Financial Times printed the chart. So it’s THE truth! (joking joking) In general I hold a distrust for legacy media, but if they print THIS chart… How much worse are the REAL numbers??
So Panmure did more. They calculated that, to justify the $750 billion of planned 2026 capex at any sensible return, that these five hyperscalers between them need to find $2-5 trillion of additional annual revenue by 2030.
Last year, their combined revenue was $1.66 trillion.
Walmart, the biggest company in America by sales, does $725 billion.
So each individual hyperscaler needs to extract Walmart-scale revenue from its datacentres alone. On top of search, social, retail, whatever else they sell. By 2030.
Sure.
I’m not done yet. The valuations are equally ridiculous:
Anthropic at $965 billion. $47 billion revenue (give or take a fairy tale). Somehow still burning $3 billion a year. Blacklisted by the Pentagon (ok ok, temporarily unblacklisted by the courts because it was “political retaliation”), CEO on the record that if they grew faster, they would go bankrupt less slowly.
Now compare that to an ACTUAL business. Walmart. $940 billion valuation. $725 billion of ACTUAL revenue (no fairy tales needed). Profitable. Pays a dividend. Owns about a third of America’s grocery aisles. And it’s been at it for sixty years.
Same valuation. Give or take. Take one guess which I prefer in my portfolio?
None of these AI companies are publicly traded. Yet. Anthropic private. OpenAI private. SpaceX private. So how does retail money actually get into the trade of this century?
Nasdaq quietly rewrote its index methodology in May. Top-40-by-market-cap newly listed companies can now enter the Nasdaq-100 fifteen trading days after IPO instead of three months. The 10% minimum free float requirement, gone. Low-float listings get weighted up to three times their actual tradable share base.
S&P Dow Jones has opened a parallel consultation on cutting its own seasoning window from twelve months to six. And waiving the GAAP profitability requirement for “megacaps”.
For clarity: “megacap” means SpaceX, OpenAI, and Anthropic.
SpaceX is targeting an IPO at $1.75-2 trillion. Floating just 3-5% of shares.
Bloomberg Intelligence reckons the rule changes will mechanically force somewhere between $15 billion and $200 billion of passive index buying - from 401(k)s, pension funds, target-date retirement vehicles, your mum’s IRA - into a single low-float, unprofitable IPO. I counted that at around $4-30 billion in “Honest recount”.
SpaceX, then OpenAI, then Anthropic. In that order.
The rules of passive index investing got rewritten so that retirement money would mechanically buy unprofitable AI-adjacent megacaps. At the price the insiders set on day one.
And meanwhile, in a development that has surprised approximately No1, the chief executives are already walking their predictions back.
→ Sam Altman, last June: “a lot of jobs will go away”.
→ Sam Altman in March: OpenAI “shouldn’t have rushed” the Pentagon contract. “Opportunistic and sloppy”, in his own message to staff.
→ Sam Altman this week, to the CEO of Commonwealth Bank of Australia: he was “pretty wrong” and is “delighted to be wrong” - “I thought there would have been more impact on entry-level white-collar jobs being eliminated by now than has actually happened”.
Same man.
I agree that you should adjust to the facts, but this guy’s looking like a weathervane.
→ Dario Amodei last May: AI will eliminate 50% of entry-level white-collar work within five years, brace for 10-20% unemployment.
→ Dario Amodei this month: actually if you automate 90% of a given job, the remaining 10% scales up and humans get vastly more productive.
Same gentleman. Same technology.
But what I loved the most was the huge boycott of OpenAI in the weeks that followed the Pentagon’s decision to blacklist Anthropic for refusing to drop the autonomous-weapons restriction.
ChatGPT uninstalls surged 295% day-over-day.
Claude topped the App Store.
And Anthropic’s annualised revenue jumped from $14 billion to $19 billion.
It was unorganised, unbranded, and absolutely ferocious.
People can, when shown the choice, tell the difference between a company holding a line and a company scrambling for the next contract.
Some of that $965 billion came out of that.
Not just out of the financial engineering machine.
But you can’t sell a trillion-dollar listing to retail savers if your own pitch deck reads “we will eliminate a hundred million of you from the workforce“. So the message gets polished for public consumption. At the precise time the IPO pipeline lights up. How unexpected.
But the displacement they were warning about hasn’t shown up in the data anyway. Yale Budget Lab’s labour tracker, updated through March, finds no AI-specific signal. The shifts that are visible were already happening before ChatGPT existed.
Altman himself, at the India AI Impact Summit, told the room: “there’s some AI washing where people are blaming AI for layoffs that they would otherwise do”.
He’d know.
Two-thirds of the companies that ran AI-driven layoffs last year are now rehiring. One in three spent more on the rehiring than they saved on the original cuts. Robert Half calls them “AI boomerangs” - which is the name a consultant invents when there’s a market in unwinding what the last consultant sold you.
Different titles, same desks. Often the same humans. Back at higher rates. Because they’re the only ones who understand the systems nobody else does.
And then, yesterday, this appeared on the wires: the Trump administration is, according to NOTUS, in active discussions with major AI labs about taking equity stakes in them.
The firms would “voluntarily cede” shares to the Federal Government - give them away, that is, rather than have the government buy them - and the dividend income would, in due course, be distributed to American households as an “AI dividend”. (I think I heard that about tariff dividend too?)
Anthropic is not in the room.
So now we have ourselves a Funding Pyramid.
Tier One: private equity, augmented with circular debt.
Tier Two: retirement money mechanically channelled into the IPOs.
Tier Three: freshly proposed this week: voluntary government equity. The taxpayer in the role of the Final Bagholder.
The economics still won’t work. The machine needs more. So bring in the taxpayer. The same taxpayer whose retirement account was already drafted at Tier Two. The same one whose taxes fund the data-centre subsidies.
I’m sure it’s a matter of “National Security”…
Don Durrett reposted something last week that’s been nagging at me:
If you actually believe the buildout would happen at the scale UBS describes - $1.3 trillion a year by 2030 - you’d need physical things. Copper for the wiring. Silver for the panels and the contacts. Rare earths for the magnets. Uranium for the reactors. Gas turbines, steel, water rights, transmission, land.
The AI bulls, in aggregate, own none of it.
Look at the positioning. The loudest voices on AI changing everything are long Nvidia, long the hyperscalers, long the model labs, long the index funds whose rules just got rewritten on their behalf. But zero commodities.
Long the narrative, short the inputs.
If they actually believed the buildout would arrive at the scale their own decks describe, they’d be loading up on the stuff you can’t generate from a GPU.
They are not.
Which tells you most of what you need to know about what they actually think happens next.
The juniors who got fired last year will be back next quarter under new titles. “AI Operations Specialist”, or whatever the consultancies bill the rename at. Doing the same job at the same desk. For a higher pay.
While the CEO who fired them rings the IPO bell of a listing whose index inclusion was negotiated separately. Whose profitability requirement was waived in writing. Whose 401(k) buyers had no say in any of it.
The pension fund that owns the index fund that owns the share they were just forced to buy belongs, in the end, to the workers downstairs.
And somebody, somewhere, still has to dig the copper out of the ground.
My other publications:
[Daily Digest] → The news in 5 minutes, without the forty open tabs.
[Portfolio] → What I do with my own money
"If they actually believed the buildout would arrive at the scale their own decks describe, they’d be loading up on the stuff you can’t generate from a GPU.
They are not."
That's the tell, innit?
I'm sure there are plenty of Elon/spacex believers who'll hit back with "we'll just go mine some asteroids."
That'll fix it for sure.
/sarc...
I know it's wobbly, but spin that thing faster, because if it slows down the bomb goes off and we all die... ;-/