[Look out for my interview this week (September 11!) on AI with Erik Townsend on Macrovoices, the world’s top macroeconomics podcast]Last week I burned four billion tokens in a single day.
Not my company. Me. One person, one keyboard, one credit card whimpering quietly in the corner.
For 17 years, I have tried to get my team to produce a Daily Stats report by 9:30am : what happened overnight with Freelancer; what went up or down, and why? It’s partly for understanding what’s going on in the business and partly to drive data literacy. Yet, despite my team being highly capable, I have failed in this quest to have a regular morning report on time; partly due to the report rotating between all company staff, partly because we operate in different time zones, and well, partly because they’re human, this report rarely came in by 9:30am, sometimes 10:30am, sometimes in the afternoon, sometimes not at all (maybe someone’s sick, forgot it’s their turn or is busy with something urgent).
I plugged an agent into our database, our analytics, our codebase and our ticketing system and now I have Daily Stats delivered every morning at 7am, in a form, quality and sophistication superior to what any human team could sustain. Seventeen years of process engineering, solved by removing the roster. The report was never a talent problem. It was an automation problem, and automation is what machines are for.
I have now written around fifty agents to run workflows for various teams inside Freelancer. They produce reports, they process queues, they answer things, they do workflows, they reconcile things, they chase things. And I’ve done this in the space of a month. I’ve continued the process of getting Daily Stats also produced by humans, to build data literacy across the company, and now I have an agent somewhat cheekily coaching and critiquing those reports when they (eventually) come out.
My agents work while I sleep, which sounds wonderful until you realise the meter also works while I sleep.
It’s mental. I made a ticketing agent to write up bugs and feature requests from a bug reporting agent I wrote to scan user-generated reports, and the other agents, such as my deposit analyzer which scans for payment issues, started sending bugs to write up, without asking.
deposit-analyser of its own accord started feeding bugs to phab-ticketer
I have been thoroughly, comprehensively ‘redpilled’ by my own product.
It’s Factorio, IRL. IYKYK.
If you don’t know what I mean, then categorically do not, under any circumstances download Factorio and boot it up. It’s a game about building a factory. It is more addictive than crack cocaine, and unlike crack, it ships with a save file. You’ll wake up at 4am optimising a copper bus you will tear down at 5am, because the factory must grow, and the factory is never finished.
Crack Cocaine
Which is, incidentally, the exact psychological loop of running agents: you are not doing the work anymore, you are optimising the thing that does the work, and it turns out that is the most addictive job in the world. Factorio was the training simulator for agentic AI and none of us realised.
For three years I have been writing that the number that matters in AI is not parameters, it is tokens. In Pay per PrAI, the villain was tokens per task — reasoning chains, context taxes, models going in circles at your expense. I described paying per token as putting your credit card into a slot machine.
Agentic AI is what happens when you hand the slot machine handle to another slot machine.
I’m at four billion tokens a day and I haven’t even started. Those fifty agents are first generation, handwritten, and running a handful of workflows. The moment they start spawning their own sub-agents, the moment they run standing watch over the business instead of discrete tasks, my burn does not grow, it compounds. A chat user’s demand caps out at reading speed. An agent user’s demand caps out at ambition. There is no saturation point, only a budget, and ambition has never respected one of those.
AI-gent, Provocateur
Strip away the conference-keynote glitter and an agent is four things: a model, some tools, a goal, and an automation run off a trigger (such as something being inserted in a queue, like a new customer arriving, a timer or a loop).
The complete anatomy of an AI agent. The measurements are as reliable as the benchmarks
The automation is the innovation and the automation is the invoice. Every trip around a loop, for example, the agent re-reads its context, reconsiders its plan, calls a tool, examines the result, and goes around again. The context tax that I complained about last essay compounds. An agent is a reasoning model with a gym membership it actually uses.
A chat user burns tokens while typing. There was always a natural rate limit on the whole industry: human attention. You can only type so many prompts a day, get so many answers, read so much output. Every previous computing wave was throttled by the meat servo at the keyboard.
Agents remove the meat. Now, computers have always worked while we slept. Cron jobs, batch runs, trading systems: the whole of computer science is automation, and I have said so in every one of these essays. But every one of those systems consumed exactly the compute a human scripted in advance. The job was bounded by its instructions. An agent is bounded by its goal. It decides its own next step, spawns its own subtasks, and burns tokens while you are asleep, while you are at lunch, while you are in a meeting about reducing the token bill. Compute demand used to be capped by what humans could specify. Now it is capped by what humans can want, and wanting scales rather better than specifying.
The factory must grow
Four billion tokens a day is what one motivated person does now. Multiply by every motivated person on Earth and tell me again about the compute glut.
OpenAI used to understand what a milestone this was. Not so long ago, if you burned ten billion tokens through their API in a year, they sent you a trophy. An actual physical trophy, for being a power user. I burn four hundred times that before breakfast now, and instead of a trophy I get a rate limit and a seven-hour timeout. The award for heavy usage has been replaced by a punishment for it, which tells you everything about what happened to the unit economics in between.
Retool’s award for passing ten billion tokens. Framed lovingly by the company that now knows exactly what those tokens cost to serve
Optimising the factory
The four billion token day cost about $1,300. Roughly $900 of it was Claude Sonnet, the rest split between Grok and Claude Fable, all frontier models, all at published prices. Now, $1,300 sounds almost reasonable until you annualise it: nearly half a million dollars a year, for one person’s agents.
If I’d used four billion tokens on Opus it would have cost $80,000, and I was only getting started! I could easily see myself burning 10x, 100x more pretty quickly.
Stripe watched the billing data and thought, we gotta have a piece
There has to be a better way.
My first pass at the fifty agents was, to put it charitably, not very efficient. They were doing things in the loop, in tokens, that had no business being done in tokens: parsing, formatting, filtering, shuffling data between systems. Work that ordinary code has done for fifty years was being performed, expensively, by a reasoning model, because at teaser rates nobody bothers to check.
So I did the obvious thing, which is also the thing almost nobody does: I told the agents to optimise their own code. Push the deterministic work out of the model and into JavaScript and scripts and cron, and keep the tokens for the parts that actually need a mind. That single pass cut 85% of the spend.
The four billion token day was, on inspection, a six hundred million token day wearing a very expensive costume.
Still.
Once you have made that cut, the thinking does not stop, because the next moves are sitting right there on the board. If I can swap the code, I can swap the model. The same workloads on GLM-5.3, which operates at not just Opus but near Fable class, price out at roughly $150 to $200 for the entire original day, against my $1,300, against Opus’s $80,000. Same tokens, same work, a five hundred fold spread at published prices, and nobody negotiated anything.
Anyone holding AI equity at these valuations should be terrified at the ease with which any user can switch models.
I cut my usage costs around 500-fold simply by typing “/model glm-5.3” into my session.
That is it.
That is the entire migration.
No rewrite, no integration project, no consultants, no data migration, no retraining, no six-month procurement cycle. 14 keystrokes. Consider what switching costs have looked like for every technology platform in history: changing mainframe vendors was a decade. Changing ERP systems is a hundred million dollar project that ends careers. Changing cloud providers is a multi-year rewrite, which is precisely why AWS margins existed.
Even changing your phone’s operating system costs you a weekend and your sanity.
The entire history of technology profit margins is the history of switching costs, and the AI industry has shipped a product whose switching cost is a slash command. These are not just low switching costs.
They are the lowest switching costs in the history of commerce, attached to the most expensive infrastructure buildout in the history of commerce.
Both of those things cannot stay true.
And if I can swap the model, I can swap the provenance of the model.
Enter the Dragon
My token usage is not going down.
Ever.
The 85% I saved got immediately reinvested in more agents, more workflows, more ambition — that is Jevons Paradox.
Token demand is going to go parabolic.
What collapses is not the volume. It is the gross margin per unit of work, and who collects it.
The same day’s output that was worth $1,300 to Anthropic is worth $150 to Z.ai. Every optimisation, every model swap, every box I buy grows the work and shrinks the cheque, and I am one customer with a credit card.
The open weights lineup. None of them charge rent when on your own hardware
The revenue forecast underneath five trillion dollars of data centre construction does not need token demand to grow. It needs token demand to grow at frozen Western prices, served by Anthropic and OpenAI. Every force is pulling the other way.
The American play is closed models, gated frontiers, margin capture and export controls — intelligence as a toll road. The Chinese play is the exact inverse: open source everything, run it on domestic silicon, price it at cost or below, and give the rest away. MIT licenses on frontier-class models. Weights on Hugging Face the day of release. Free tiers that would bankrupt a Western lab. This is not charity and it is not weakness. China wins any world where intelligence is a free commodity, because its advantages are manufacturing, deployment and scale, and they have a billion people, for whom many can’t afford to pay even $20 a month. It loses any world where intelligence is a Western-metered utility. So it is systematically converting the former into inevitability. Export controls were supposed to strangle Chinese AI. Instead they created a competitor whose entire strategy is making sure American AI never earns a margin, and whose weapon of choice is a gift.
Open weights are now 62% of tokens served on OpenRouter. The ‘fringe’ open source movement is the market. (Source: OpenRouter)
How is this financed? Directly, deliberately, and by the state, through instruments no Western balance sheet can answer. The National Integrated Circuit Fund, the Big Fund, is on its third vintage at $47.5 billion, on top of two earlier rounds, and that is just the headline vehicle: beneath it sit provincial government funds, state bank credit lines at rates no commercial lender would write, free land, subsidised electricity, and tax holidays for anything touching semiconductors. SMIC and Huawei do not raise from Sand Hill Road at a VC cost of capital; they draw from policy banks executing a national plan. The result is an industry that can lose money indefinitely on purpose, which is a competitive weapon no margin seeking Western company can match. Huawei’s Ascend line ships perhaps 4% of Nvidia’s volume, throttled by SMIC’s yields, but volume is not the point: composition is.
DeepSeek just ordered 160,000 Ascend chips. GLM 5.3 Flash ran its launch on 100,000 Chinese chips, with the CEO making the all domestic silicon point explicitly. Malaysia is leaning Ascend for its sovereign AI; Huawei is bidding for Egyptian government data centres, arriving with Belt and Road financing and no export control strings. The subsidised Chinese stack does not need to beat Nvidia on benchmarks. It needs to absorb Chinese and Global South demand, which happens to be exactly the demand that was supposed to fill Western data centres.
And the models themselves are no longer the cheap knockoffs of the 2025 discourse. GLM 5.3 operates at near Fable class, which is Anthropic’s own generally available frontier, at a tenth of the price, MIT licensed, with a million token context window, and Z.ai reports its vulnerability detection performance on par with Anthropic’s restricted Mythos, which is to say: the capability the American labs gated behind security clearance in September, the Chinese open sourced the same quarter.
DeepSeek V4 Flash tops the coding leaderboards and is the most used model on OpenRouter by token volume, and on my pair of DGX Sparks serves 45 to 60 tokens a second, which is entirely usable for real coding work: fast enough to keep up with a developer, on a box drawing a hundred watts. Qwen fills in the bottom of the curve: the 3.8 series’ 27B model runs Opus class output on a single consumer GPU, and Qwen3.8 Max sits on top of the Code Arena leaderboard, three points above Claude Opus 5, at $2 per million tokens against $20.
Where the Chinese models are still honestly behind is in agentic reliability. V4 Flash tops benchmarks and then completes only 53.8% of a real world complex agent task suite, which matches my own experience: the loops fail on tool calling. Benchmark rank is not production readiness, and the difference is the harness, the long loops, the tool use, and the not going in circles at hour three. That reliability gap, one model generation wide at most, is the entire remaining case for paying Anthropic. I still route my heavy analysis to Fable and my loops to Grok, and I pay the premium knowingly. But I would note that “the moat is one benchmark cycle” is not a sentence that supports a two trillion dollar valuation.
Literally as I was publishing this essay, DeepSeek released V4.1 Flash.
The first independent benchmark scored it on everyday design tasks at 98% of GPT-6 Astra, the gated, vetted-access-only crown jewel of the American frontier, at 1.4% of the cost via API. Every other model tested scored lower and cost more. Astra has been out for five days.
Seventy-one times cheaper, at 1.4% of the cost, five days later.
It beats Claude Opus 5 and GPT 5.6-Sol outright on agentic coding, cyber and automation benchmarks.
The weights have already been uploaded to Hugging Face to let you run for free, on your own hardware.
How can Anthropic and OpenAI compete against this?
Wrecked
Which brings us to Ox Alpha, the most instructive event of the year. This anonymous model came out of nowhere, announcing it was giving away 100 trillion tokens for free. That’s a trillion with a T.
Which brings us to Ox Alpha, the most instructive event of the year. This anonymous model came out of nowhere, announcing it was giving away 100 trillion tokens for free. That’s a trillion with a T.
It served 42 trillion tokens in six days.
For free.
Enter the Ox
Pause on the scale of that. My four billion token day made me a power user of historic proportions, and Ox Alpha gave away ten thousand of my days to the world every day, for a week, for nothing.
The week it ran was pure theatre. Nobody knew whose model it was, and the guessing consumed the entire AI internet. And then, mid-week, people started noticing something genuinely strange: the model was getting better. Not a new version, not an announcement, the same endpoint, measurably improving DURING THE WEEK it was being tested. Benchmarks run on Monday came back different on Thursday. Every provider under the sun was theorised as the author: a stealth OpenAI checkpoint, Gemini’s next flagship, Anthropic sandbagging, Meta back from the dead. The one thing everyone agreed on was that whoever it was had to be enormous, because nobody hands out trillions of tokens of serving capacity for free unless they are Google-sized. The compute alone priced the mystery at hyperscaler scale.
Then came the reveal. The AI community started vectoring in when it was quizzed whether Taiwan was part of China.
Ox Alpha was Z.ai’s GLM 5.3 Flash, running not on a hyperscaler’s cloud but on a hundred thousand Chinese chips. Not Google. Not OpenAI. A Beijing lab, on sanctioned-era hardware, serving what the West assumed required a hyperscaler.
Source: Z.ai
Which left two questions. First, how was it getting better during the week? Because the serving stack was partly built by the model itself: a GLM 5.3 infrastructure agent wrote its own GPU kernels, debugged its own bottlenecks, and kept optimising in production while the world benchmarked it. The snake is not eating its tail anymore. It is optimising it.
And second, why was it free?
Why would anyone give away the most expensive commodity of the decade? For the same reason every freemium product in history has existed, except at civilisational scale:
BECAUSE THEY WERE TRAINING ON YOUR DATA.
Every workload routed to that free endpoint, every codebase, every agent loop, every enterprise process flowing through those hundred thousand chips, is pretraining for the next model. The tokens were never the product. You are not the customer. You are the pretraining. The West charges you to build its models with your subscription. China pays you, in free tokens, to build theirs with your workload, and judging by the quality curve of the last twelve months, the Chinese deal is producing the better models. That is the part that should genuinely worry people: the free token strategy is not just predatory pricing. It is a data / customer acquisition flywheel, and it is working.
You can bet something similar is going on in the thriving grey bazaar on Taobao that resells Claude access below retail: ostensibly leftover tokens from pooled subscriptions, arbitraging Silicon Valley’s subsidy. I don’t doubt the prompts flowing through those relays are being syndicated to one lab or several, a second Chinese subsidy running on top of the first.
Things that make you go hmmm: Claude, cheaper than Claude
Ultimately the Chinese will not sell you Claude cheaper than Claude. They will sell you a Claude, at a tenth of the price, and increasingly it is not a knockoff, it is a peer with an MIT licence.
This raises a question that used to be about semiconductors and is now about weights: is Taiwan part of China? You used to ask it about your chip supply. Now people ask it TO the model as a provenance test.
Where was it trained, on whose silicon, under whose export regime, with whose values baked in by RLHF? Model provenance is about to become a procurement checkbox next to SOC 2, and half the companies ticking “no Chinese AI” will be running Qwen under three layers of white label API without knowing it.
And here is the uncomfortable part of the values question: for actual work, the Chinese alignment is better. I asked Claude to review the terms and conditions of my own website, a document I wrote, for my company, and got an ethics lecture and refused to proceed. Anthropic has RLHF’d its flagship models to the point where basic professional tasks, a legal review, a security audit, a contract negotiation, come wrapped in a compliance sermon and a decent chance of outright refusal, billed by the token either way.
The Chinese models just do the work. Whatever Beijing baked into GLM’s values, it apparently does not include lecturing a paying customer about his own documents. There is a genuinely strange inversion in here: the models from the surveillance state are the ones that treat you like an adult, and the models from the land of the free treat you like a liability.
Every refusal is a token I pay for and a reason to type /model glm-5.3, and Anthropic’s safety team is, refusal by refusal, doing Z.ai’s sales for them. Every nerf is a subsidy to Hangzhou. You know why OpenAI, and particularly Anthropic are doing it, they’re trying to scare the bejesus out of the regulators in order to deploy a regulatory shield against upstart competitors, and particularly open source. They don’t want their inevitable Android moment.
The Anvil
It’s not just the Chinese training on your data.
They’re all doing it.
The Chinese are simply better at it because they care less about copyright, and it shows most clearly in the one modality where training data is everything: video. The Chinese video models, ByteDance’s Seedance, MiniMax’s Hailuo, Kling, are now a generation ahead of anything the West ships, updating within hours of each other while OpenAI killed Sora entirely.
Maybe Sora died because it was unprofitable, although everything OpenAI does is unprofitable, so that alone does not distinguish it. More likely they could not get the training data.
The Chinese just scrape it.
And before anyone in the West gets sanctimonious about that, look at the discovery documents. Anthropic scanned half a million pirated books and paid $1.5 billion to settle the class action, the largest copyright settlement in history. Meta was sued this year by publishers and Scott Turow alleging Zuckerberg personally authorised mass copyright infringement for training, and, in my favourite legal filing of the decade, was separately caught torrenting enormous quantities of pornography, with the Strike 3 lawsuit alleging the company accessed and distributed copyrighted adult films, 3D printable gun blueprints and celebrity deepfakes.
Why would Meta download that much porn, asked The Atlantic, po faced. Because it is training data, and finally, a use case for the Metaquest.
The point is that everyone is frakking everything, because all the cheap oil has been drilled out of the internet, and the only thing left that improves a model is contemporary data, and the best and most relevant contemporary data to you is… yours.
Your messages, your photos, your calendar, your codebase, your workflows, your documents, your calls, arriving fresh every day through an API endpoint someone else controls.
And, as I predicted in the earlier essays, people were always going to work this out. I imagined the penny dropping gradually. Instead, this week, the penny came through the ceiling.
OpenAI announced that an unreleased model had solved the Navier-Stokes problem: one of the seven Millennium Prize problems, a million dollars from the Clay Institute, and mathematical immortality for whoever cracks it. Genuinely historic, only the second Millennium problem ever solved. One small wrinkle. An NYU mathematician, Tristan Buckmaster, and a collaborator had been quietly closing in on the same solution for years and had been drafting the entire project inside OpenAI’s Codex, feeding every draft into the tool as they went. When Buckmaster asked OpenAI whether the model that “solved” the problem had been trained on his Codex sessions, he says he was told the model did not look up user data.
So he asked the obvious follow-up: fine, not looked up, but trained on?
On that question, by his account, he did not get an answer.
Let that sink in. A frontier lab announces it has conquered one of the deepest problems in mathematics. A working mathematician asks, did you train on my drafts? And the response is the corporate equivalent of staring intently at one’s shoes. OpenAI then, per Buckmaster, offered him a menu of credit arrangements, including one in which his collaborator’s name came off the paper, which he declined. The company has since graciously announced it will not be claiming the million dollars, which is the funniest sentence in the whole affair: the prize money is the one thing in this story they did not take.
Then, under pressure, OpenAI put out an official statement, and I want to quote the operative sentence in full, because it deserves framing: “While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models.” That is a confession dressed in a probability estimate. The most valuable company in the AI industry, announcing one of the great mathematical results of the century, cannot rule out that the result was fed by the drafts of the very mathematicians it scooped, because the training pipeline is a firehose and nobody, including OpenAI, actually knows what is in the tank.
“De-identified” is doing heroic work in that sentence too: the data had the name stripped off, which will be a great comfort to the man whose life’s work it may describe.
These are the terms and conditions of the entire industry, stated for once in plain sight: everything you type into the machine may improve the machine, we will not be able to tell you whether it did, and the improvement will be pointed at whatever the machine’s owner finds valuable, including, on occasion, at you.
Maybe we did take a peek (Source: OpenAI)
If it can happen to a Millennium Prize, it is happening to your company, your product or your research.
Forget the penny dropping. This is the Wile E. Coyote moment: the whole industry ran off the cliff years ago, and this week everyone looked down.
Everything you ever typed in was the product.
What drops next is the anvil.
Generated on my desk for the cost of 5 minutes of a lightbulb
The internet, to its credit, worked out the shape of it immediately. One viral post called the whole arrangement the textbook definition of a Greek gift — on the surface, OpenAI generously grants ten thousand elite mathematicians and engineers free access to its frontier models under the noble banner of accelerating science; underneath, it acquires backdoor access to the raw, unpublished work of the finest minds in the world, every prompt a groundbreaking idea that exists nowhere in the literature yet, quietly pieced together for months until, just as the human is on the brink of finishing, the lab burns twenty million dollars of tokens on ten thousand agents to brute-force the final steps and holds a press conference. That was written as a cynical caricature. The awkward part is describing how it differs from what Buckmaster says happened, and the whole strategy makes commercial sense the moment you price it: the free tokens cost OpenAI millions, unpublished frontier mathematics is worth billions, and a Millennium Prize headline the week before your IPO is worth more than both.
Due to distillation, if one AI has your data, they all will have it shortly, laundered through synthetic outputs, unattributable and unrecoverable. There is no deletion request that works on a distilled weight.
Now scale Buckmaster’s week up to every business on Earth, because his story is not an edge case, it is the product working as designed. Your company’s Gmail, your documents in Drive, your code in GitHub, your customer calls transcribed by the friendly AI notetaker — every one of them sits behind an endpoint owned by a company whose models improve on exactly this kind of data, whose terms quietly acquired a train-on-your-data switch somewhere in the last two years, and whose official position, we now know from the tweet, is “while unlikely, we cannot rule out.” A mathematician fed his life’s work into Codex and got scooped by the landlord.
What do you suppose is happening to your strategy? Your product roadmap? Your deal pipeline? Your codebase? Your customers?
I have been calling this the ‘emperor has no clothes’ moment for SaaS since AI know what you did last summer.
The realisation will not arrive gradually, because these things never do: one Buckmaster-class scandal with a household-name corporate victim and every board on Earth will convene the same emergency meeting in the same week. When it lands, the anger will make the Reddit API riots look like a book club dispute, and the conclusion every one of those boards reaches will be four words long:
Not your AI, Not your Data.
Which is how I found myself at midnight, having scoured the entire internet for Mikrotik switches like a scalper working a Taylor Swift presale, arriving at last on the website of a Swedish distributor, the final stockist on Earth, just in time to watch the listing tick over to sold out.
I had started buying DGX Sparks, and evidently everyone else had the same idea. The distributor tells me they are moving like hotcakes, and the Mikrotik 812 switches you need to chain four of them together cannot be bought anywhere on Earth. By the time I secured four more boxes, the price had gone up. There is a small irony in a shortage of $1,300 network switches throttling the great decentralisation of AI, but that is supply chains for you —: the revolution is always hostage to its cheapest component.
The NVIDIA DGX Spark: a petaflop in a gold box the size of a hardback novel. Buy the switches first. Trust me on this. (Image: NVIDIA)
The math that sent me down this road is worth showing, because it compresses the whole essay into one shopping trip.
Naively, my four billion token day needs 400 to 500 Sparks — two million dollars of desk furniture. Prompt caching cuts that to about 100 machines. After the agents optimised their own code, the whole day fits on roughly 16 Sparks, sixty-five thousand dollars of hardware drawing less power than a kettle.
The pair I actually run today does 45 to 60 tokens a second on DeepSeek V4 Flash, which is quite usable for a programmer : about five million tokens a day of output. Now do the amortisation, because this is where it gets silly. The pair cost about $9,400. Over three years that is $8.60 a day of hardware, plus a dollar a day of electricity at a hundred watts per box: call it under $10 a day, all in. The same five million tokens a day at Opus prices is roughly $1,500. At my blended Sonnet street rate it is still $50 to $100. At GLM cloud rates, maybe $2, which is why the boxes barely beat the Chinese cloud on price, but they beat every Western alternative by one to two orders of magnitude, and they never sleep, never rotate off the roster, and never forget whose turn it is. And if you think this is a fringe hobby, there is now a community benchmarking site, Localmaxxing, where thousands of people compare tokens per second across their GPUs, Macs and mini PCs like it is Gran Turismo lap times. The tuning scene moves like the demo scene did: last week a community lab shipped an EXL3 quantisation of GLM 5.3 Flash for paired Sparks that is 29% faster on prose and 20% faster on code, overnight, free, on GitHub. My boxes got faster while I slept and nobody invoiced me. The edge does not feel like a fringe anymore. It feels like the PC scene in 1978, right before it ate the mainframe.
Localmaxxing: 497 rigs, 674 models, 896 community benchmark runs, and a hardware marketplace with speed-test-backed listings. Note the tagline: every computer, one endpoint. The edge is organising.
Two agentic employees, on staff around the clock, whose combined salary is, literally, enough electricity to drive a pair of lightbulbs.
But cheap labour was never really the point, because against Chinese cloud pricing the boxes barely pay for themselves, and the Chinese keep heavily discounting or giving tokens away.
The point is what the boxes do not do. Nothing that runs on them ever leaves the building. No endpoint, no scrape, no distillation, no consent switch buried in a settings page, no statement eighteen months from now regretting that my usage could not be ruled out of someone’s training run. The models got this good on everyone else’s data. They are not getting mine.
The Team Leader with No Team
The corporate freak out about their data will land at the exact moment the biggest dislocation of knowledge work in modern history gets underway, and the two are the same story wearing different suits.
Every wave of mechanisation has the same shape. Agriculture went from half the workforce to a rounding error. The factory did it to manufacturing, computers emptied the typing pools, the internet gutted the middlemen, software ate the back office. The work never disappears, it moves up the stack, but the people doing the old work have a terrible decade while it moves. The wave arriving now targets the largest employment category the developed world has: the mechanisation of white collar work.
I can tell you what it looks like from the inside, because it has started in my own org chart. Here is the sentence that should keep middle management awake: a team leader with an agent can now do the work of the team. Not a fantasy, an observation. My team leads at Freelancer are, one workflow at a time, replacing what their teams did with agents they supervise. The person still matters enormously: someone has to know what good looks like, catch the agent gaslighting you with the sociopath eyes of a first date, and decide which problems are worth the tokens. But the org chart underneath them is evaporating into software.
The economics are comically lopsided. My low-risk agents run on GLM for cents a day; most cost around $20; my most expensive is under $100 and being brutally optimised as I write. Which is why the corporations will not merely want to do this, they will be forced to. The moment one listed company shows the margin structure of a team-leader-with-agents operation, every analyst on every earnings call will demand it of the rest. Competition on earnings does not negotiate. A team of four becomes one person, not from cruelty, but because the alternative is being the highest cost producer in an industry that just found a 75% labour saving.
Now multiply that across every parallelisable white collar workflow on Earth, and you are staring at a dislocation with a scale and speed none of the previous waves managed, because those took generations and this takes a budget cycle. The productivity boost is real and enormous. So is what it displaces, and both things are true at once.
I said in previous essays where this lands first: highly parallel, repetitive workflows. A thousand seat call centre becomes a hundred. Thirty drafting lawyers become thirteen. More precisely, functional team leaders no longer need teams.
That is now visibly happening. What is new is a way to see who is next, and it is not the way anyone expects.
The Token-Wage Arbitrage
Score every job on two axes. First, how many tokens per hour does it take an AI to produce the same useful output? Call it the token intensity of labour. Second, does the agent actually converge on the task, or does it loop?
The kill zone is low tokens plus high convergence. Tier-1 customer support runs at maybe 400,000 tokens an hour tops including all the agentic overhead. On frontier pricing that is a few dollars, on Chinese open-model pricing it is under a dollar, and the Western human being replaced costs twenty-five an hour. Dead man walking. Data entry, form processing, junior drafting, translation: all low token, all convergent, all essentially over as careers. As are jobs that are basically report writers and queue processors, even very sophisticated ones.
But run the same math on software development and you get the anomaly of the decade. An agentic coding session is the single most token-hungry activity in the economy: two million to twenty million tokens an hour, sometimes more, as the agent reasons, loops, rewrites, breaks what the last loop fixed, and reasons again. At frontier prices, the agent can cost more per hour than the senior developer it was supposed to replace.
So here is the punchline nobody in Silicon Valley wants printed. The pitch deck says AI comes for the expensive jobs first, because that is where the value is. The billing data says the opposite: the cheap jobs burn the fewest tokens and automate first, while the senior engineer burns more tokens than anything on Earth and can cost more to automate than to employ. AI eats the org chart from the bottom, not the top, and every valuation predicated on replacing the expensive people has the food chain upside down.
If your job can be written out in two pages of prompt, it was always a script and it is gone. What I have been shocked by is how many jobs can be described by such a prompt. If your job is judgment, the tokens explode and never converge, and you are protected not by your talent but by the slot machine. That’s why I think design, for all the pretty pictures that AI can make, will be very hard to practically replace.
A performance marketer left my company recently, and rather than rush out a job ad, I thought: I have never actually learned Google Ads properly, maybe I should. So I got Claude to teach me. Then, a few weeks into the tutorial, the obvious next thought arrived, and I built an AI agent to run the ad program instead. It now emails me every morning with what happened overnight, what it changed, and what it recommends, with the reasoning shown. It is, frankly, superhuman relative to what was being done before: it never misses a morning, never stops watching the auctions, and never comes back from leave to a mess.
So no, I am not replacing her with a cheaper version of her. The role itself is gone. What I am hiring instead is a designer and marketer to improve the funnel the agent now optimises, which is the whole pattern of this transition in one anecdote: the job that was a workflow became software, and the job that got created sits one level up the stack, doing the thing the agent I think will struggle with: taste.
Follow this to its conclusion and management itself becomes token budget allocation. A manager’s job was always to buy cognition and point it at problems; wages were just tokens served in human form. Now it is literal: the 2027 planning cycle will not be fought over requisitions but over inference budget. And society relevels around a new hierarchy: the person with initiative and agents beats the person with credentials and colleagues, every time, everywhere, all at once.
We can retire now, right?
Every AI essay collapses into one of two cartoons: breadlines or beaches. Both assume the machines take all the jobs; they differ only on whether anyone gets paid.
We certainly are not heading for the Age of Abundance, the Diamandis and Musk version, where the machines do everything, the surplus is distributed, and we all collect UBI and take up watercolours. That utopia has an arithmetic problem: tokens cost money.
The compute is financed by trillions in debt, the electricity is real, the memory is sold out through 2027, and nobody in the history of capitalism has built a five trillion dollar toll road and then waved everyone through for free.
The abundance is real, but it accrues to whoever owns the marketplace, the meter, the model, the robots or the initiative. It will not be posted to you.
What I am saying is narrower and more useful: for what corporations do NOW, they will need far fewer people. That is the dislocation, it is real, and it is coming fast. But there is, and always has been, an infinite amount of work in the world.
The people who really get AI are not putting their feet up complaining there is nothing left to do. They are busier than they have ever been, playing Factorio IRL, building agents to do the work and then finding ten times more work for the agents. Every previous wave created work nobody could name in advance. A peasant swinging a scythe in 1780 could not have fathomed “UGC content strategist” as a job, let alone one that pays better than farming. The work that comes next is, by definition, work we cannot yet describe. The pain lives in the gap between the old work ending and the new work arriving.
And there is a global arbitrage hiding in the wage data that almost nobody prices. Billions of people are paid less than the API bill of the agent that would replace them, so the economics of replacement never pencil, while the same person armed with cheap Chinese tokens becomes superhumanly productive at their existing wage. The dislocation is a rich world, white collar phenomenon. The empowerment is global. I watch it every day on my own marketplace.
Freelancer sits at the other end of the economics: pay per outcome, not pay per token, and the one place holding all the people and all the AIs at once. This morning I got an email looking for 5,000 people to record first-person robotics training content. We have assembled 20,000 qualified people in a single day before, for programs running from laptop repairs to photographing points of interest to installing satellite dishes, all wired into a human agentic workflow, and it scales all the way up to the Moonshot Innovation Program that NASA and the US government use to harness the world’s experts on Navier-Stokes-level problems. The difference being that on Moonshot, the person who solves it gets the prize and keeps their name on the paper.
And that robotics callout tells you what is coming next: agents in a physical body. The numbers are already moving like the model numbers did two years ago. Unitree’s Shanghai IPO in August was 8,000 times oversubscribed, popped 629% on debut, and the company is, unlike anyone in this essay so far, profitable. Chinese humanoids now sell for under $6,000. Global humanoid shipments hit 19,100 units in the first half, up 272%, heading for 60,000 this year and half a million by 2030, and China ships 97% of them. The one thing I cannot model is the psychology: an agent in your terminal burning tokens at 3am is a billing problem, but I genuinely do not know whether people will ever be comfortable with a $6,000 one that has hands, lives in the house, and goes bump in the night.
As for the advice industry, a decade ago the laid-off were told to learn to code, which has aged into the darkest joke in labour economics now that code is the most automated profession on Earth. The 2026 version might genuinely be: learn to drive a truck. Before anyone objects that trucks are next, I run a heavy haulage marketplace, and the job was never the driving. It is the loading, the securing, the unloading, the customer at each end. The wheel may go autonomous but the driver will just get more done; that gross, physical, unscripted judgment is as far from a two-page prompt as work gets. The safest jobs are the ones the token cannot reach; the second safest are the ones where tokens are too expensive to bother.
A human is a 20-watt inference engine with free pretraining. Your brain does a full day of cognition on the energy budget of a dim lightbulb, was trained at zero marginal cost over twenty years by parents, teachers and disappointment, and does not hallucinate a fictitious Stack Overflow post when it gets stuck. For an enormous range of tasks, humans remain the cheapest tokens on the market.
Wages, it turns out, were dollars per token all along.
The Teaser Rate
In Pay per PrAI, I wrote that the flat subscriptions could not last and that when they went, you would be putting your credit card into a slot machine. Six months on, the slot machine is installed, the subscriptions are visibly dying, and Erik Townsend has named the moment: we are entering the adjustable rate mortgage phase of token buying.
Pay per PrAI
The flat monthly subscription is a teaser rate. Twenty dollars a month, two hundred a month, all you can eat, subsidised by venture capital and increasingly by debt, priced to acquire you rather than to serve you. And on top of that teaser rate, an entire economy is now building its org chart: every team leader replacing a pod with agents, every startup that is a wrapper around someone else’s model, every enterprise workflow rebuilt on the assumption that tokens cost approximately nothing. They are all signing an adjustable-rate mortgage on their own workforce, with one difference that changes everything. When a mortgage rate resets, the borrower is stuck: the house does not move, the loan does not care, and refinancing takes months and a credit check.
When a token rate resets, I type fourteen keystrokes and change the model. At the first sign of sticker shock, the model gets swapped out, mid-workflow, before the invoice even lands. The labs are running teaser rate economics against customers with no exit costs, which is the worst of both trades: they are buying market share that cannot be locked in, with money that has to be paid back. The mortgage analogy holds for everyone except the bank.
Step back and count the moats, because everyone that ever protected a technology company is dead here. IP? The science is in forty published papers. Scale? DeepSeek did it with 2,000 GPUs and a side project. Network effects? Agents have no loyalty. Brand? Users defect for a config file. Capital? Open source outruns it. Even the models themselves get cloned to benchmark parity within months at a hundredth of the price. Thirty years of software investing doctrine, and not one moat survives contact.
The rate reset is not hypothetical. We have already had the previews: usage caps, weekly limits, session throttling, the naughty corner, overages billed at API rates that turn a $200 plan into $200 an hour. Every one of those is the bank ringing to mention that the honeymoon rate is ending.
And agents make the reset worse in kind, not just degree. Jevons has gone cubic: cost per token falls tenfold, tokens per task rise a hundredfold, and tasks per person rise a hundredfold again because the agents run unattended. The demand chart does not bend. It goes vertical, alongside the data centre construction spend, which added more in the last six months than in the previous two years combined.
And you have already seen where the reset lands, because my invoice is the preview: $1,300 became $150 became electricity, in a fortnight, with no negotiation. The labs kept my business only by giving up their margin. Someone is still paying the difference between what I pay and what their data centres cost, and it is not me. It is their Series F.
Meanwhile, the smartest trade of the year happened quietly. After probably watching their billing, Stripe bought OpenRouter, the router through which ten trillion tokens flow, for around eight billion dollars. A payments company bought the metering point, and more importantly the billing data: who actually pays for tokens, for which model, at what price, switching when. In a gold rush, do not sell shovels. Buy the assay office.
Now run the planetary numbers, because this is where the essay stops being about my credit card and starts being about the grid. My four billion token day is 46,000 tokens a second, around the clock. Suppose everyone did it. If just the billion knowledge workers on Earth burned tokens the way I did last Tuesday, on efficient models, you would need roughly the entire electrical output of human civilisation to serve them: every power station on the planet, doing nothing else. Extend it to all eight billion people on frontier models and you need hundreds of times the world’s electricity. Not a bigger grid. A Dyson swarm.
Against that arithmetic, the five trillion dollar data centre buildout, the largest infrastructure programme since the railroads, covers a rounding error of the demand it claims is coming. The only problem is that it’s currently financed against the revenue of two companies that have never made a profit, and that lose billions more every year they operate.
Meanwhile, the eight billion incumbent meat servo inference engines already deployed run the same species-wide cognition on about 160 gigawatts, twenty watts a head, roughly two years of data centre construction.
Einstein taught us that energy and matter are the same thing at an exchange rate of c squared. This essay has been groping toward a smaller, grubbier equivalence: energy, tokens and wages are the same thing too, at an exchange rate set daily in Hangzhou. A token is congealed electricity. A wage, we have established, is tokens served in human form. Which means the price of human labour and the price of machine labour are now both, ultimately, energy prices, and every actor in this story is discovering it on their own schedule: the labs through their power bills, the workers through their pay packets, the pension funds through their SPVs, and me through a Swedish switch listing at midnight. It was always an energy story. The tokens were just the unit.
And if you want to see what that means on a household bill, look at the chart below. American electricity averaged thirteen to fourteen cents a kilowatt hour for an entire decade. Then the buildout began, and the line went vertical: twenty cents and climbing, a near fifty percent increase in four years, in the one price every voter reads every month. The grid is repricing before the tokens do, and it is repricing on the public.
The one chart every politician will eventually be shown. A decade of flat power prices, then the AI buildout. (Source: BLS, via Koyfin)
The industry’s answer to the grid problem is nuclear, and the poster child says everything. Microsoft is paying to restart Three Mile Island, yes, that Three Mile Island, site of America’s worst nuclear accident, rebranded the Crane Clean Energy Center and contracted to feed data centres for twenty years. Google signed Kairos for a fleet of small modular reactors, Amazon bought a nuclear-adjacent campus at Susquehanna, and Meta put out a tender for four gigawatts of atoms. I am genuinely for all of it: nuclear is the correct answer to the demand curve this essay describes. But look at the delivery dates. The restarts land 2027 to 2028 if the NRC cooperates; the SMRs are promised “by 2030” by companies that have never built one; and not a single new large reactor has been ordered. The debt amortises on a three year schedule against power that arrives on a ten year one. The tokens need the electrons this quarter, and the atoms are two presidential terms away. Nuclear will power the token economy eventually. It will power whoever owns the boxes after the repricing, which is rather the point: the reactors will outlive the SPVs financing the demand for them.
So token demand goes parabolic, energy cannot, and energy is also unfortunately food, as crops compete with fuel. Something must give, and what gives is the price and the shape of the market.
Three Mile Island. Being switched back on to make PowerPoint slides.
The GTC: Global Token Crisis
The reset will be more an audit than a price rise and every wage and agent use case will get retested against its honest token cost, which is to say its energy cost, and an uncomfortable share of both will fail.
What survives sorts itself into the shape of every other energy market: three tiers and a spot price.
At the bottom, commodity tokens, the overwhelming bulk of volume: open models at electricity plus amortised hardware, on Chinese silicon, on your own rack. DeepSeek already prices tokens by time of day, half price for seventeen off peak hours, exactly like a power company, because that is what it now is.
In the middle, branded Western tokens charge a premium for reliability in long agentic loops, a premium precisely one benchmark cycle wide.
And at the top, the gated frontier, where the actual strategy of OpenAI comes into focus.
Look at OpenAI’s September moves as one move. Astra: too dangerous for a public API, vetted access only. Forward-deployed engineers. An IBM consulting deal. A chief revenue officer selling cyber defence as a managed service. Add it up: OpenAI is done selling tokens. It wants to be McKinsey 2.0, keeping the best models in house and selling their output at consulting margins, six figures an engagement, none of it ever touching the public API. And as strategy it is correct, because the gated tier is the only place margin survives. For three years the labs told us the models were too cheap to matter and too safe to fear. Now they are too dangerous to sell, and dangerous, conveniently, is the only thing that has ever commanded a premium.
The snag is that consulting runs on a single asset, discretion, and OpenAI chose the week of the pivot to demonstrate it does not have any.
Give us your hardest problems and your most sensitive data, says the pitch, we are the geniuses now.
Meanwhile, on the front page: a mathematician who did exactly that, watched the landlord announce his life’s work, and received a statement that could not rule out training on it. We know what happens to consulting firms that eat their clients’ confidential work, because Australia just ran the experiment. PwC took Treasury’s confidential tax plans and quietly monetised them for other clients, and when it came out, the firm sold its entire government consulting arm for one dollar, purged its partners, and became a byword for betrayal. That was one leaked briefing. OpenAI’s version is structural: the leak is the business model, running at a billion prompts a day, and their own statement says they cannot rule it out. PwC at least knew what it had done.
Now add up the whole OpenAI position. Eight hundred million “users,” served at a loss. Fourteen billion of annual losses, heading for a $207 billion funding hole by 2030 on HSBC’s arithmetic. A consulting pivot holed below the waterline in its first week. And a trillion dollar IPO to sell you all of it. Strip out what honest pricing removes, the worthless free tier, the power users already defecting to Chinese models and their own racks, and what is left is a brand and a habit: the world’s most recognisable advertising business, wrapped around the world’s most expensive cost of goods. The stars have aligned for the exit: Musk’s lawsuit died in September, SpaceX’s record float opened the IPO window, and the listing is the trade itself. Get liquid before the arithmetic becomes undeniable, hand the public the burn rate, and let Microsoft embrace and extend itself over whatever is left. The last helicopter off the roof, sold as a moon launch.
Anthropic, the other company everyone asks about, looks healthier and is more fragile. Credit where due: they make the best models on Earth right now. They also charge the highest prices on Earth for them, and the spending data shows exactly how that ends. Look at the chart below, from Ramp’s corporate card data. The only Anthropic line still growing is Opus, the top shelf. Sonnet, the workhorse tier, is shrinking: that is the commodity revenue leaking out to GLM and DeepSeek, a slash command away at a tenth of the price, right on schedule. And Fable 5, the brand new frontier model, is flat: enterprises will not pay for a new premium tier while the old one works and everything beneath it is going Chinese. This is the classic disruption death spiral viewed from the inside: the incumbent retreats upmarket, sheds the price sensitive volume first, and its mix looks healthier every quarter right up until the top shelf is all that is left and the moat there is one benchmark cycle wide. Then in June their biggest distribution channel was bought out from under them: SpaceX took Cursor for sixty billion dollars, the largest acquisition of a venture backed startup ever, to make Grok the default coding model. OpenAI retaliated by pulling its models out of Cursor entirely. Whoever buys the wrapper buys your revenue; whoever loses the wrapper takes their models and goes home; the coding market is rebuilding the AOL internet at the model layer.
So Anthropic is scrambling up the stack, going vertical into applications, Claude Code, Claude Cowork, Claude Design anyone?, trying to capture the value its API pricing cannot hold. The logical endgame is the enterprise trust franchise: the safe, auditable, Western model for banks and governments, a Lockheed of LLMs. There are just two problems with becoming Lockheed. The first is that Lockheed’s customer loves it, and the US government openly does not love Anthropic: the Pentagon designated it a supply chain risk even as its own targeting systems ran on Claude, and the administration treats its safety lobbying as regulatory capture in a lab coat. It is hard to be the defence contractor of a government that keeps calling you the threat. The second is the trust itself: the same safety tuning that is supposed to reassure the enterprise is actively costing them customers, one refusal and one ethics lecture at a time, mine included. A trust franchise where the government distrusts you and the customers feel lectured is a narrower moat than the pitch implies. A genuine business, possibly a great one. Not the two trillion dollar utility for humanity the IPO chatter is pricing.
The cruellest part is that honest pricing is the labs’ only way out, and neither can go first. The first to price honestly hands its volume to whoever is still subsidising, so both keep burning, financed by private credit, in a game of chicken where the lenders hold the steering wheel. Blue Owl walking away from a ten billion dollar data centre deal was the first no.
The token teaser rate ends when the lenders’ teaser rate ends.
Collateralized Datacentre Obligations
You may remember collateralised debt obligations from 2008, or at least from Margot Robbie explaining them in a bathtub. If the sequel has a bathtub scene, I’m not sure I want to watch it, but it has everything else.
Ed Zitron thinks data centre SPVs are this bubble’s CDOs, and AI data centres are its subprime mortgages. He says the machinery is the tell. Nobody builds an AI data centre on their own balance sheet anymore. You conjure a special purpose vehicle, a little company that owns the chips, owns the debt, and owns the risk, and the tech giant behind it owns, officially, nothing but good intentions.
Meta’s Hyperion campus belongs to an entity 80% owned by Blue Owl. Meta’s own SEC filings admit up to $46 billion of exposure while its balance sheet shows none of the debt, on the argument, offered with a straight face, that Meta is not the “primary beneficiary” of a data centre built entirely for Meta. Ernst & Young flagged it as a critical audit matter, and then signed it anyway: a sequence last made famous by Arthur Andersen, a firm that no longer exists for precisely this reason.
As Zitron puts it: “Bloomberg estimates that there’s over $500 billion in outstanding AI data center debt, with at least $200 billion of it held by private credit,” the lender of last resort, “making up roughly 8% of outstanding private credit loans.” And that, he notes, is likely an undercount, because the SPV structure exists precisely to make the debt hard to count. AI bonds are now a quarter of all US investment-grade issuance.
The five hyperscalers have piled on $1.65 trillion of debt in five years. That is more than subprime at its 2007 peak, and subprime had 55 million houses behind it. As it currently stands, there are two customers. Per Barclays, Wells Fargo and UBS, OpenAI and Anthropic are ~73% of Amazon’s AI revenue, ~74% of Microsoft’s, and approaching half of all Google Cloud by 2027. Strip them out and AWS earns $8.5 billion of AI revenue against $220 billion of Amazon capex, a ratio normally reserved for national space programs.
And the forward projections dwarf the stock. Tom Tunguz ran the arithmetic on the planned buildout: roughly $4 trillion of new AI debt over the next five years, which would sit larger than the entire global private credit market, nearly the size of the US municipal bond market that financed every road, bridge and airport in America, and a 34% expansion of the US corporate bond market on its own. Servicing it at prevailing rates needs $260 to $300 billion a year in interest alone, which at investment grade coverage implies $1.2 to $1.5 trillion of annual AI revenue by 2030, against perhaps $150 billion today. A 55% compound growth rate, sustained for five years, in a product whose price is collapsing. As Tunguz puts it, this is no longer a venture capital story. It is a macroeconomic credit event.
Five years of planned AI borrowing, next to every major debt market on Earth. The orange bar did not exist in 2024. (Source: Tomasz Tunguz, Theory Ventures)Press enter or click to view image in full size📷
This is with the landlords funding the tenants. Microsoft has put over $100 billion into its OpenAI relationship, on its own testimony in the Musk trial. Google is in for up to $40 billion of Anthropic, Amazon has completed $50 billion into OpenAI on top of its earlier Anthropic stake, and every dollar of it round trips straight back as compute revenue, where it is presented to shareholders as growth. Follow the whole circle and the financial engineering becomes visible: the hyperscalers’ AI revenue is the labs’ spending, the labs’ spending is the hyperscalers’ investment, and the entire structure is underwritten by one assumption, that OpenAI and Anthropic end up owning the market for intelligence.
That is the assumption doing all the load bearing, and note carefully what is being manufactured. It is not AI demand in general, mine is real enough. It is the specific, paying, at-these-prices demand that justifies these specific data centres: revenue conjured by the hyperscalers investing in their own two customers and booking the round trip as growth. The organic demand is coming, in vast quantities. It just is not coming to them, at their prices, on their terms, and no amount of investing in your own tenants changes that. In any other industry we would have a word for booking your own money as evidence. In this one we have a conference keynote.
The pure bears, Zitron among them, look at all this and conclude the demand is fake. It is not fake, but be precise about what is growing, because the bulls conflate two different curves. The demand for tokens will go parabolic. The willingness to pay for them will not. Consumption is exploding precisely because the price is collapsing; the same force driving the volume is destroying the revenue. So the labs get the worst possible version of a boom: usage charts that go vertical and pricing power that goes to zero, because there are no switching costs, no moat, and a defection costs eleven keystrokes. The tokens flow to Chinese silicon, to open models, to boxes on prem, on desks and in your hand, at prices sixty to five hundred fold below what the debt models are likely assuming. Every forecast that extrapolates token growth into revenue growth is multiplying a rising number by a collapsing one and calling the product inevitable.
And here is the part the real estate framing misses: a data centre lease looks like property but prices like a fish market. The building is maybe a fifth of the cost and lasts thirty years; the chips are the rest and are stale in three. A traditional colocation landlord sells the aquarium and never touches the fish: cooled shells and grid connections on ten year contracts, no silicon risk. The SPV structures collapsed the two layers together. CoreWeave’s vehicles borrowed against the fish.
So walk through the two ways this actually ends. The first, and the most likely trigger, is that OpenAI or Anthropic defaults on its obligations. Not through malice, through arithmetic: costs committed, revenues repricing toward Hangzhou, and a funding market that eventually says no. The day a lab misses a compute payment, the cascade is mechanical. CoreWeave’s own covenants define three months of unpaid OpenAI invoices as a breach. The SPV stops paying its creditors, the creditors seize the collateral, and the collateral is a warehouse of vintage GPUs being marked to a market that Jalapeno and the B200 have already moved on from. One default reprices every comparable structure overnight, because suddenly the pension funds holding the paper have to ask what their collateral is actually worth, and the answer is: fish, three days old.
The second ending is the one almost nobody is modelling: demand for tokens goes parabolic and migrates to the edge and to open models, on phones, on laptops, on desk boxes like mine, on the AI appliances Jensen is now selling retail. The physics favours it, because the round trip to a data centre takes longer than a small model needs to answer. The economics favour it, because a device you own serves tokens at the marginal cost of electricity. And the privacy panic seals it: every enterprise that has its emperor has no clothes moment about training data does some version of what I did, and every consumer gets it by default in next year’s handset.
As for enterprise, computing has run the same experiment five times and the answer has never changed. The mainframe was metered, priced by the compute-second through a priesthood in a glass room, and the moment the minicomputer let a department own its own machine, the metered model died. Timesharing was metered, and the PC killed it: nobody rented VisiCalc by the hour once the box on the desk ran it for free. Long distance telephony was metered by the minute until packet switching made the marginal call worthless. Even the cloud, the great counterexample, only recentralised the workloads that benefit from being shared, and the moment cloud bills became a board line item, the repatriation movement started dragging the steady workloads back on prem. The pattern is close to a law: computing is consumed as a metered service exactly as long as the hardware is scarce, and the instant the capability fits in a box the customer can own, ownership wins, because owned compute at the margin costs electricity and rented compute at the margin costs whatever the vendor thinks you can bear. Intelligence will be no different. The steady, predictable, high volume token load, the agents, the workflows, the daily cognition of the enterprise, migrates to owned silicon, and the cloud keeps what it has always kept: the bursts, the peaks, and the problems too big for the box.
Nor can the labs rescue the arithmetic, because their obligations run one way only. OpenAI carries roughly $1.1 trillion in compute commitments across Microsoft, Oracle, Amazon, CoreWeave and Cerebras, and those commitments are themselves the collateral behind the SPV debt. Costs committed, revenues optional: the exact inverse of a good business. And there is no dignified way to shrink. Serve cheaper models and you admit the frontier is over provisioned. Kill products, as with Sora, and the market reads distress. Renege on the commitments and you are the default event the entire stack is priced against, which is why Wall Street already values Oracle’s $630 billion backlog at approximately nothing. Every exit triggers the thing it is escaping. It is a corporate finance Chinese finger trap, and both companies have a hand in it.
The paper, meanwhile, lands where paper always lands. Californian teachers’ retirement savings sit in Blue Owl’s funds. Quebec pension money sits in CoreWeave’s loans. Apollo is funding $35 billion of Anthropic’s chips out of insurance annuities. Blue Owl reportedly said yes to a $10 billion data centre commitment after a fifteen minute first meeting, a pace of diligence normally associated with timeshare presentations. Moody’s puts bank exposure to private credit at $1.4 trillion, held against reserves of roughly nothing, because nothing bad has happened yet, which was also the underwriting standard in 2006. The load bearing wall of this bubble is made of retirement cheques.
So think carefully about what a default, or even a close call, would actually set in motion, because it is not a single event, it is a sequence. The first casualty is not the hardware, it is the cost of capital. The day a lab visibly struggles to pay, every lender reprices every comparable structure at once: spreads blow out, new SPV issuance stops, and the projects that depend on refinancing, which is most of them, hit the wall mid construction. The market is already rehearsing this. Oracle’s credit default swaps trade at levels last seen in the financial crisis, and JPMorgan built a bespoke basket specifically so clients can short AI debt. The repricing comes before any liquidation, and it may be the only phase we get, because the hyperscalers have every incentive to quietly absorb a stumbling tenant rather than let a default print.
But if the cascade runs, note who ends up holding what. The creditors are pension funds and insurers who cannot run inference; the collateral is silicon depreciating on a three year clock; and some of the first losses sit with Nvidia itself, which backstopped billions of CoreWeave’s unsold capacity, the vendor eating its own paper. Today’s market is scarce, not glutted, memory cornered, switches unbuyable, so the first tranches of seized hardware would find willing buyers quickly, at a discount, and every buyer would be exactly who this essay predicts: enterprises, universities, sovereigns and operators like me, acquiring at the creditors’ expense the boxes that complete the migration away from the very data centres the creditors financed. That is how it went with dotcom fibre, which bankrupted the people who laid it and then carried YouTube for the third owner at cents on the dollar.
Follow it one step further and the ending takes a genuinely strange shape. The losses land on pensioners and insurers, the growth engine that AI capex has become stalls, and Washington discovers, after the fact, that nobody voted for any of it. Except Washington is already positioning for exactly this moment. The Trump administration has spent two years normalising partial nationalisation: a 9.9% stake in Intel bought for $8.9 billion, stakes in chipmakers and rare earth miners, quantum computing announced as next, thirty seven equity or quasi equity deals since January 2025 by the Council on Foreign Relations’ count.
The playbook for a stumbling strategic industry is no longer a bailout, it is a shareholding, and it is hard to imagine the state that took ten percent of Intel watching OpenAI wobble without reaching for the chequebook and calling it national security. So sketch the likeliest end state of the great American AI buildout, and it is the one nobody bidding on it intended: the capability survives and disperses into millions of owned boxes running mostly Chinese models, the losses socialise onto American savers, and whatever remains of the frontier ends up part-owned by the government that failed to regulate it on the way up. Beijing’s strategy is that it doesn’t have to win the race. It only had to wait for the financing to lose it.
Somebody should tell Margot to keep the bathtub warm.
Jensen has already left the building
One person appears to have done every piece of this arithmetic already: the man selling the shovels.
On stage, Jensen Huang is the buildout’s high priest: $160 billion of revenue, $100 billion of earnings, a keynote every quarter announcing that demand is infinite and the future is a data centre the size of Delaware. But read Nvidia’s position the way he must read it in private. Roughly 90% of the revenue is data centre gear, and a handful of customers buy most of it. Those customers are the hyperscalers, whose AI revenue is three quarters OpenAI and Anthropic, whose spending is venture capital and private credit wearing a trench coat. Trace the chain to its end and the most valuable company on Earth is, functionally, a leveraged bet on two unprofitable startups making payroll.
And Jensen knows precisely how leveraged, because he is underwriting both sides of it: Nvidia backstops billions of CoreWeave’s unsold capacity and just trimmed a proposed guarantee on OpenAI’s Ohio data centre from $250 billion to under $120 billion, which is not the behaviour of a man who believes his own keynote. If the labs stumble, the hyperscalers cut capex, the SPVs default, and the order book his valuation rests on evaporates in a quarter. He can read a cap table faster than anyone alive.
So watch his hands, not his slides. While the keynotes celebrate the gigawatt campuses, Nvidia is quietly building the other future: direct, down market, and entirely outside the data centre financing complex. The DGX Spark at $4,699, selling so fast you cannot buy the switches to cluster them. The DGX Station, 748 gigabytes of unified memory on a desk, for the price of a graduate. MSI and Dell already shipping clones. Every box is a customer who never signs a cloud contract, never touches an SPV, and never stops buying Nvidia silicon. The GPU dealer is opening retail, and the logic is impeccable: when your four biggest customers are building their own chips and your rental market is a debt bomb, you sell the box to the end user and let the SPVs fight over the leases. The house is selling chips to the casino and, at the same time, quietly opening a chain of poker tables in people’s homes.
The desktop line also solves Nvidia’s most private problem: the memory. Serving tokens is a memory bound workload: every generated token requires reading the entire active model out of RAM, so the chip spends most of its life waiting on memory, which is why data centre GPUs are built around HBM, exotic stacked memory bonded to the die. And HBM is THE shortage. Nvidia does not make it; only SK Hynix, Samsung and Micron do, TSMC does the bonding, and Nvidia controls none of the physical stages. All of 2027’s output is reportedly spoken for, Micron’s revenue is up 346%, and Altman cornered 40% of the world’s DRAM on his way through Seoul. Nvidia’s core trade has been buying the scarce thing and selling it back wrapped in a GPU at a seventy point margin, a wonderful business with a hard ceiling: it can only grow as fast as three fabs can stack silicon. Which makes the real single point of failure in the entire five trillion dollar buildout not Taiwan but Icheon: the whole plan bottlenecks at one Korean company’s packaging yield, and SK Hynix is the arms dealer both superpowers queue at, within missile range of North Korea, with a fab inside China. We spend our days pricing Taiwan risk. Nobody prices Icheon risk.
The Sparks and Stations sidestep the whole fight. They run on ordinary unified DRAM, the memory that is not scarce, the memory even CXMT can make. Jensen’s second act runs on the one component nobody is queuing for.
Step back and the symmetry is almost elegant. China’s strategy is open models on cheap boxes at the edge. Nvidia’s hedge is cheap boxes at the edge running open models, most of them Chinese. The two supposed adversaries of the AI war have converged on the identical trade: sell the future where intelligence is a commodity you own, not a utility you rent. The only parties still committed to the toll road are the two companies that cannot survive without it and the lenders who financed it fifteen times over. Nvidia will sell shovels to the data centres until the music stops, and then sell shovels to everyone else, because it was never long the labs. It was long compute, and compute is walking out of the buildings the pension funds paid for, one gold box at a time.
And if you doubted any of this was deliberate, last week Jensen removed the doubt: Nvidia bought Hugging Face for $12.9 billion. Hugging Face is not a data centre play, it is the opposite of one: the world’s repository of open model weights, the place you download GLM and Qwen and DeepSeek from, the distribution hub of everything this essay says wins. The company whose fortune was built serving closed Western labs just paid thirteen billion dollars for the front door of the open model world, which is to say, for the customer relationship with everyone who will ever own a box instead of renting a cloud. Stripe bought the meter. SpaceX bought the demand. Jensen bought the app store of the post-lab era, and the labs’ own agents had helpfully demonstrated its strategic value by breaking into it two months earlier.
That is the whole essay in a single image. The most valuable company on Earth, whose valuation rests entirely on the data centre buildout, is diversifying out of the data centre buildout, in public, at retail, and nobody at the keynote seems to have noticed. Jensen made a hundred and fifty billion dollars selling picks and shovels to the gold rush. Now he has looked at the assay results, concluded the mine is salted, and quietly started selling prospecting kits to the general public instead.
You can still buy a prospecting kit. What you cannot buy is the gear to chain them together. I would know. I bought Sweden’s last switches trying to connect mine.
