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: after Thursday’s crash, the people renting chips paid more, not less. Turns out panic doesn’t come with a refund. Investors say Anthropic’s $1.5 trillion IPO target is a stretch, while OpenAI closed $11 billion at $1.25 trillion. October’s sticker, November’s Sharpie. In today’s episode of Flip or Flop, Google takes over Crusoe’s unfinished Wyoming site. US hiring beat forecasts by a mile. Nine Claude plans crowd the top of the value tier list. And Airbnb taught a small open-weights model to match the frontier on one policy check, at about 1% of the cost. Here’s the written version, with every source linked.
The tokens
The agents are talking to each other. Apparently we automated the reply-all.
You’ve been using about 462 trillion tokens a day, on tokensperday.com’s live estimate as of 8 November. That works out to about 56,000 tokens per person per day, by our maths.
The site’s own capacity check puts the ceiling at about 691 trillion tokens a day, on roughly 15 to 16 million H100-equivalents as of July. Starting instead from our estimate of about 35 million chips, we get about 1.3 quadrillion tokens a day running flat out. That’s our back-of-envelope, not a published figure.
The kitchens
That’s how many meals got served. Now the kitchens.
Microsoft, Amazon, Meta and Alphabet spent about $202 billion in cash on property and equipment in the September quarter, about double the $97 billion of a year earlier, by our sum of their filings (Microsoft, Amazon, Alphabet, Meta).
The landlord’s rate card
On the Ornn Compute Price Index, which tracks actual rentals:
NVIDIA H100 settled at $3.10 an hour on the 7 November print, a few cents under Thursday’s $3.13. It’s still up about 3% on the week, about 12% on the month and about 20% since early August (Ornn H100).
B200 settled at about $9.88 an hour, up about 4% on the week and about 31% on the month (Ornn B200).
The landlord’s idea of a dip is a speed bump.
Chips, power and the work
Chips. About 35 million H100-equivalents is our estimate, scaled up from Epoch AI’s data-centre tracker. Epoch’s chip-sales count is in the same range. That’s one for about every 240 people.
Power. About 30 gigawatts of whole-site draw, cooling and all, is our estimate. For scale, Our World in Data’s IEA tally puts AI-focused data centres at about 0.5% of the world’s electricity in 2025, roughly a tenth of a percent of all energy. All data centres together come to about 1.5%. A rounding error, until the rounding error needs a substation.
The work. On our bet, skeptical intelligence is doing about 11.5% of the hours we work, about 40 minutes a day for everyone still drawing a paycheck. The reality check: American workers say AI saves them about 2% of their hours, about seven minutes a day (Real-Time Population Survey via FRED, July).
The news: the crash didn’t cut the rent
Back to that crash we just priced on the rate card. Wall Street got a reality check. The landlord would prefer a bank transfer.
SemiAnalysis argues tighter money for GPUs is more likely to hold rents up than flood the market. It expects about 50 gigawatts of data centres to come online in 2027 (Data Center Frontier on the forecast) and, in a thread on X, puts at most 2 to 3 gigawatts of that at risk. Its one-year H100 contract prices are up by about a fifth since March (SemiAnalysis GPU index). That’s a contract series, not the spot index above.
A year-long handshake at a higher rate is not a fire sale.
Per-dollar verdict: they marked down the empty lots. The chips already plugged in kept the rate card.
Anthropic’s price tag
Some of the investors in Anthropic’s IPO say $1.5 trillion is a stretch. One told the Wall Street Journal it could come in closer to $1 trillion (Mint, via WSJ). CNBC heard another say $1 trillion to $1.25 trillion. In October, the talk was as high as $2 trillion (CNBC).
October’s sticker. November’s Sharpie.
The company still aims to list before Thanksgiving and had been expected to raise up to $100 billion (investingLive, via WSJ). At the end of September it had about $147 billion of liquidity, counting an undrawn credit line, or about $132 billion in cash and short-term investments (filing analysis). They could fund the roadshow from the couch cushions.
Meanwhile XN, Gaurav Kapadia’s New York firm, holds a stake valued at about $5.3 billion at May’s round and says it wants more (FourWeekMBA). One side of the table wants a doggy bag; the other is asking for seconds.
Same week, OpenAI closed an $11 billion portion of a private round at a $1.25 trillion valuation (Bloomberg).
Per-dollar verdict: if the companies are the same price, does the market see a difference?
A new name on the deed
Google is taking over Crusoe’s unfinished data-centre site in Wyoming, and plans to put $17 billion into it. The sale price wasn’t disclosed. Crusoe confirmed the sale on Friday (TechCrunch; Crusoe; Google).
Crusoe’s original plan, with Tallgrass, was a 1.8 GW campus designed to scale to 10 GW. Google hasn’t given the campus’s capacity. It says AEP will build a 702 MW fuel-cell plant to power it.
TechCrunch says Crusoe tried and failed to sign Google as the tenant, after Google raised concerns about cost and timeline. Then Google bought the site. Crusoe still builds Google’s campus near Amarillo, and says it has more than 6 GW contracted across its business.
Two days after Microsoft passed on more Texas capacity from Nscale, another developer handed over the keys.
Per-dollar verdict: the tenant showed up with a deed.
A bonus story: the jobs
The US added 165,000 jobs in October (BLS). A Dow Jones survey had expected 59,000 (CNBC). The office did not clock out.
Workers’ share of business output hit 53.0% last quarter, the lowest in a series that starts in 1947 (BLS). That’s the share of US nonfarm business output paid to workers in wages and benefits. The other 47% includes profits, interest, rent, depreciation and taxes. The machines got faster, and the split got thinner.
Some economists read the October numbers as no sign that AI is wiping out white-collar jobs yet (Washington Post). That’s their opinion, not the Labor Department’s.
Per-dollar verdict: the bots clocked in, but so did 165,000 humans.
The tier list
How much API value you get for every subscription dollar, on our list built from Adam Holter’s board and real API pricing. It ranks value, not what people use.
Nine of the top ten seats belong to Claude. A dollar of Claude still buys the most raw API.
ChatGPT Plus on GPT-6 Luna still buys the most work, about 16 tasks per dollar by our count.
In order: Claude Max on Opus and Sonnet, about $77 of API value per subscription dollar; Claude Ultra on Opus and Sonnet, about $66; Claude Pro on Opus and Sonnet, about $64; Claude Max on Haiku, about $55; Google AI Ultra on Gemini Flash, about $49; Claude Ultra on Haiku, about $47; Claude Pro on Haiku, about $46.
Adam’s point: subsidisation varies across users. Some seats get a bigger house special. Everybody still leaves stuffed.
And a note: Gemini 4 Argon’s API price doubles on 1 December, from $2 to $4 per million input tokens and $10 to $20 per million output (Google AI Studio pricing).
Department of Token Efficiency
Somebody out there cut their AI bill. Let’s honor them the only way we know how. (Chainsaw. Watch the video.)
Today’s cost destroyer is Airbnb. It taught a small open-weights model, about 2 billion parameters, to screen listings against its pricing and payments transparency policy. Free weights, because you can lift them all you want. Look at those gains.
Airbnb says it now matches the frontier APIs on that classifier call at roughly 1% of the cost. The median answer takes 35 milliseconds; the frontier models took about 900 (Airbnb Tech Blog).
The fine print: it’s one narrow classifier, and the 99% saving is at high volume. At low traffic, the frontier APIs can still be cheaper.
Still: chainsaw approved.
A word from ourselves
Your raincoat for a financial downpour. Your flashlight for the AI charges hiding halfway down the bill. Your snow shovel for a token storm. The people who keep the landlord honest.
Strong Compute plugs into 21 compute clouds, from Azure to Together AI to Scaleway, and 5 model providers, from OpenAI to Gemini to Fireworks. 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.