Listen to this episode on Spotify or watch it on YouTube.
Welcome back to The Cost of Intelligence, our comedy news show about the money behind AI. Each episode, Ben Sand reads the industry’s invoice and asks one question: are you getting more intelligence per dollar?
This week: Nvidia says Jensen Huang has resigned as CEO, effective immediately, and its finance chief, Colette Kress, is now interim CEO. Bernstein counts about 586 gigawatts of North American data centres in the pipeline, and only about a third look credible. The rest is a groundbreaking ceremony still looking for ground. OpenAI brought in about $10 billion last quarter and posted an adjusted operating loss of almost $11 billion. Big quarter, slightly bigger bill. And the Department of Token Efficiency honours Cartage, a freight startup that swapped Claude Sonnet for its own model and cut its bill about 75%. The AI got moved to logistics. Best performance review it ever had. Here’s the written version, with every source linked.
The tokens
AI reads and writes in tokens, chunks of words. Like Scrabble, but every tile costs money.
The world runs about 466 trillion tokens a day, on tokensperday.com’s live model, as we ran it on 12 November. That’s up about a trillion since Wednesday’s show. Compound interest, but for tokens.
The landlord’s rate card
“Good news. Your rent is trending.”
On the Ornn Compute Price Index, which tracks actual rentals:
NVIDIA H100 settled at $3.25 an hour on the 11 November print, up about 5% from the $3.10 on Wednesday’s show and about 22% in a month (Ornn H100).
B200 settled at about $9.84 an hour, flat since Wednesday and up about a third in a month (Ornn B200).
Both chips got a raise. The new one just negotiated harder.
Power and the work
Power. AI data centres worldwide draw about 30 gigawatts, cooling and all. That’s our estimate. Epoch AI’s data-centre tracker counts 19.6 gigawatts of facility capacity at the sites it follows, which Epoch reckons cover about 44% of the world’s AI compute. Scaled up, that’s roughly 44 gigawatts of AI data-centre capacity worldwide, and we assume it runs at 60–80% of that. The North American pipeline we get to below is about 13 times all that worldwide capacity, by our maths, and even its credible third is more than four times. (Against the 30 gigawatts actually being drawn, it’s closer to 20 times.) Using IEA figures, Our World in Data reckons AI used about half a percent of the world’s electricity in 2025 (Our World in Data). Half a percent of the grid, and a hundred percent of the earnings calls.
The work. Our name for it is SI: Spend-capped Intelligence. Intelligence on a monthly allowance.
On our bet, SI now does about 11.6% of the hours we work, about 40 minutes a day for everyone still clocking in. The share is up about three-quarters of a point in a month. All of these are our estimates. Forty minutes: one long meeting SI sat through so you didn’t have to.
And the allowance is literal. JPMorgan caps some engineers at $2,000 a month on Claude Code, per Business Insider. And it still can’t expense a sandwich.
These numbers are a work in progress. Got better data? Step boldly into our DMs.
The news: Nvidia changes CEO
On Thursday morning, Nvidia announced that Jensen Huang has resigned as president, CEO and board member, effective immediately (NVIDIA newsroom).
Colette Kress, the chief financial officer, is interim CEO. She keeps the CFO job too.
The board is searching inside and outside the company for a permanent CEO.
Harvey Jones, the former CEO of Synopsys, becomes the board’s first-ever chair.
The release gives no reason and doesn’t quote Huang.
The context, as reported: on Sunday, Huang admitted an affair with Alice Huang (no relation), an executive at the cloud company Megaspeed. The Financial Times reports, per CNBC, that US intelligence officials believe she is a spy for China. Separately, Bloomberg reported that Nvidia’s count of chips sold to Megaspeed, as given to US officials, came up about 50,000 chips short. Nvidia says it later checked more warehouses holding the missing hardware, and Megaspeed denies diverting chips (The Next Web). On Wednesday, The New York Times reported a second relationship, with an Nvidia executive who reported to Huang (NBC News). A person familiar with the matter told CNBC the board made its decision on Wednesday night (CNBC live updates). Nvidia has not linked the resignation to either report.
Nvidia stock fell more than 13% at the open, per CNBC, and as much as 16% during the day, before closing down about 4.3% (NBC News). Even after the bounce, that’s nearly a third below its 6 October closing high of $239.24, not even two months ago; CNBC put that at about $2 trillion of market value gone. And Bernstein cut its price target in half, from $400 a share to $200 (CNBC).
Earnings stay on schedule for next Tuesday, 17 November (NVIDIA newsroom).
Nvidia sets the price of the chips almost everyone else rents, so whoever runs it touches every compute bill on this show. Although Anthropic says its Mythos model got Claude running on AMD chips in about three days, per its IPO roadshow; the footnote narrows that to prompt processing for one model on AMD’s MI355X (Simon Willison’s notes; Anthropic’s roadshow). If that holds up, chips may finally be commoditised.
Per-dollar verdict: the board’s new chair says he expects the roadmap to continue without interruption. Tuesday’s earnings call is the first test.
Counting the gigawatts
Bernstein counted the data centres North America says it’s going to build (via ZeroHedge and Tiger Brokers).
The forward pipeline is about 586 gigawatts. Forward pipeline: like a forward-looking statement, but with parking.
About 199 of those look credible. Two-thirds of the pipeline is still at the “we should get lunch” stage.
Announced projects are the shakiest: only about a quarter look credible (123 of 497).
Under construction, it’s much better: about 75 of 87 gigawatts look credible. Once there’s a hole in the ground, it stops being hypothetical.
ZeroHedge called the gap “bragawatts.” The unit of measure for a PowerPoint.
Amazon leads the credible list at about 33 gigawatts, then Google at 20, Meta at 15 and Microsoft at 10. Amazon’s credible pile beats Meta and Microsoft combined. The bookstore did fine.
Elon’s SpaceXSI clusters aren’t on the lists we can see. Neither is Oracle; it leases a lot of its capacity, so its sites likely sit under developers like Crusoe, Vantage and Stack. You don’t count the renter when you’re counting houses.
Then Elon replied on X:
He said his company “has a chance of going about half the GW next year of actually usable compute,” and that his team gets a large cluster working reliably in about two months; for others, he guesses, it’s closer to six. He didn’t say half of what. Two months versus six: that’s every kitchen remodel quote we’ve ever had.
For scale, he’s talked about adding about 10 gigawatts next year (SpaceX Q3 call), and SemiAnalysis expects about 51 new gigawatts of data centres next year, outside China (TechNews; The Neuron). Ten out of fifty-one isn’t half. Maybe this is where the “actually usable” comes in.
Per-dollar verdict: the compute that’s actually getting built sets the price. The rest sets the loss.
OpenAI’s third quarter
The Wall Street Journal has seen OpenAI’s third-quarter numbers (via Investing.com).
Revenue: about $10 billion, up about 50% on the quarter before.
Adjusted operating loss: about $10.9 billion. OpenAI’s “adjusted” measure has left out stock pay in earlier quarters (The Decoder). Like saying the bar tab was fine, if you don’t count the drinks.
So it lost about $1.10 for every dollar it brought in. Buy a dollar of AI, get a dollar ten free.
The loss grew more slowly than revenue: the quarter before, OpenAI lost about $1.50 for every dollar (an adjusted operating margin of -150%, against -109% now). Losing less per dollar, more in total. A net negative net negative is a positive.
Every CFO reading this has seen this movie, usually in a board deck called “path to profitability.”
Per-dollar verdict: the cheapest intelligence is the one someone else is paying for.
The tier list
How many dollars of API usage you get per dollar of subscription, on our 12 November run of Adam Holter’s method against real API pricing. It ranks value, not usage. Same board as Sunday. The top three have started leaving their mugs here.
Claude Max tops it, at about $77 of API value per plan dollar.
Then Claude Ultra, about $66, and Claude Pro, about $64 (Claude pricing).
Below that, the Haiku plans and Google AI Ultra on Gemini Flash, from about $55 down to about $46.
Anthropic owns the podium. Everyone else is fighting over the snack table.
And Gemini Argon’s API price doubles on 1 December, from $2 to $4 per million input tokens and $10 to $20 per million output (Gemini API pricing). Nineteen days left at the old price. Price hikes: the one AI launch that never slips.
Department of Token Efficiency
Somebody out there cut their AI bill. Let’s honour them the only way we know how. Safety goggles on, CFOs.
Today’s cost destroyer is Cartage, a freight startup. The only AI startup where “shipping” means trucks (TrainLoop case study).
It ran its agent on Claude Sonnet, then trained its own 27-billion-parameter model, just for freight. Turns out you don’t need a philosopher to book a flatbed.
The bill went from about $60,000 a month to about $15,000. Seventy-five percent off. Black Friday came early for freight.
Over ten weeks, traffic grew about eight times, and the model bill only roughly doubled. Eight times the work, twice the bill. Your landlord has never offered that deal.
It now runs about 3.3 billion tokens a day, about seven in every million in tokensperday’s world count, by our maths. A rounding error with a fleet of trucks.
“Dedicated GPUs, darling? I still get paid when they’re idle.”
The fine print: the case study is from TrainLoop, which sells the training platform. The quality score is Cartage’s own test. Training cost: undisclosed. Side effects may include a model that only talks about trucks.
Big win, if your work is narrow and your traffic is steady.
Still: chainsaw approved.
A word from ourselves
Your bench press for the bragawatts. Your receipt for the other dollar ten. Your remuneration range for the next CEO. The people who keep the landlord honest.
Strong Compute plugs into 27 compute clouds, from Nebius to Lambda to RunPod, and 10 AI vendors, from OpenAI to Amazon Bedrock to Deepgram, + many more. Bring your own keys and see exactly where the GPU hours and tokens go.
Because somebody has to read the bill.
Try it at strongcompute.com
Keep your context short and your invoices shorter. See you next episode.