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: Anthropic opens its IPO roadshow at up to $1.4 trillion. The ticker is ANT. The compute bill is the eleph-ant. Nvidia sheds about $430 billion of value in two days, and its board brings in an outside investigator. A hot inflation report has traders betting on a December rate hike. The landlord’s mortgage just got interesting. And the Department of Token Efficiency honours two Amazon engineers who swapped Claude Sonnet for a tiny decision model and, at list prices, took about 98% off the cost of sorting a ticket. Sonnet has been asked to clear out its desk. Here’s the written version, with every source linked.
The tokens
Somewhere, an agent is writing a thoughtful reply to your out-of-office.
You’ve been using about 465 trillion tokens a day, on tokensperday.com’s live estimate as of 11 November. That’s about a trillion more than Tuesday’s show. The tracker’s figure is its 16 July estimate, grown forward to today. Rising like a sourdough starter with a billing address.
DeepSeek has an agent tool called Harness. DeepSeek says the tool alone ran about 92 trillion tokens a day for its official-API users as of 4 November (South China Morning Post). DeepSeek counts tokens its own way, so it doesn’t line up with the world number. One tool. And it doesn’t take weekends.
The landlord’s rate card
Agents need an office, and that’s where our landlords, the clouds, come in.
“Inflation, darling? I’ve been pricing it in since August.”
On the Ornn Compute Price Index, which tracks actual rentals:
NVIDIA H100 settled at $3.10 an hour on the 10 November print, down about 6% from the $3.30 on Tuesday’s show. Over the week it’s down about 12%, off a two-day spike. Since 12 August it’s still up about 14% (Ornn H100). The landlord calls that a loyalty program.
B200 settled at about $9.90 an hour, up about 5% on the week, on top of a 72% climb since 12 August (Ornn B200). So a B200 now rents for about three H100s, every hour. Triple the rent. Welcome to the penthouse.
Rent one H100 around the clock at that rate and it’s about $74 a day, by our maths. A hotel night, every night, for a guest who never checks out.
Chips, power and the work
Chips. The world has about 34 million H100-equivalents, on Epoch AI’s chip-sales count, last updated 21 October. That’s chips sold since 2022. The receipts are real. The return policy is not.
Power. AI’s data centres draw about 30 gigawatts, cooling and all. That’s our estimate, scaled up from the 19.6 gigawatts of total facility capacity on Epoch’s data-centre tracker (updated 6 November), which Epoch says covers about 44% of the world’s AI compute. And the chatbot still can’t open your attachment. For scale, Our World in Data’s smaller tally of the IEA’s figures puts AI’s data centres at about 0.5% of the world’s electricity in 2025, and about 0.1% of all the energy humans use. Still the loudest tenant in the building.
All that to drive a future filled with SI: Shareholder-owned Intelligence. Now it comes with a ticker symbol.
The work. On our bet, Shareholder-owned Intelligence is doing about 11.5% of the hours we work, about 40 minutes a day for everyone on a payroll, and nearly five hours a week. That’s up about two seconds a day since Tuesday. All of these are our estimates. Somebody’s getting a shorter week. It isn’t you.
Who gets those minutes depends on the budget. Bloomberg reports that AI access is the new corporate status symbol: at Match Group, engineers get four times the AI budget of non-technical staff. Finally, somebody at a dating company is getting matched.
The reality check: American workers say AI saves them about 2.1% of their work hours, roughly seven minutes a day by our conversion, in the survey’s latest reading, for the third quarter of 2026 (Real-Time Population Survey via FRED). Got better data? Step boldly into our DMs.
The news: Anthropic’s IPO roadshow
Anthropic launched its IPO roadshow on Tuesday (Anthropic). A roadshow. Like a band tour, except every venue is a bank.
It plans to sell 825 million shares at $70 to $78 each. The only Claude plan with no weekly limit.
At the top of the range, that values Anthropic at about $1.4 trillion, fully diluted (TechCrunch). That’s about 45% above the $965 billion valuation of its May funding round, by our maths (the IPO figure is fully diluted, so the bases differ a little).
The ticker is ANT, on the Nasdaq. Pricing is set for 18 November, with trading the next day. A trillion-dollar company named after the smallest thing at the picnic.
On top of the IPO, Anthropic is selling $26 billion of stock to its own suppliers at the IPO price: Nvidia and Google $10 billion each, then AMD, Broadcom, Micron, Microsoft and TSMC (Anthropic’s amended S-1). The suppliers are buying the customer. Imagine your caterer buying stock in your wedding.
Anthropic’s own slice of the IPO raises up to about $62 billion; the rest goes to existing shareholders. At its latest quarterly spend on compute, that pays for just over six quarters. Its compute commitments run to about $900 billion, about 14 times the raise, by our maths. Most companies go public to pay the founders. This one goes public to pay the landlord.
Per-dollar verdict: new shareholders, same invoice.
Nvidia, day two
Nvidia shares fell about 5.5% on Tuesday, to about $169. Over two sessions, that’s about $430 billion of market value, by our maths. The Nasdaq fell less than 1% on Tuesday. (Nvidia won back about 3% on Wednesday, after we recorded.)
Nvidia’s independent directors have ordered an outside review, hiring Joon Kim, a former acting US Attorney, and the law firm Wachtell Lipton (NVIDIA newsroom).
Bloomberg reports, from trade records it reviewed, that Nvidia’s count to US officials came up 50,000 chips short on sales to a company called Megaspeed through hardware partners. These are allegations, not findings. Nvidia told Bloomberg last December it had checked additional warehouses holding the hardware, and Megaspeed denies diverting chips. The Information reports the relationship helped move at least $7.5 billion of chips to Chinese firms.
Nvidia reports earnings on 17 November. Every CFO reading this knows that feeling.
Per-dollar verdict: the stock got cheaper, the B200 rent didn’t.
The price of money
Consumer prices rose about 0.4% in October, hotter than forecast. Leave out food and energy, and core prices rose about 0.3%, also above forecasts (BLS). Your token bill calls that a slow month.
On CME’s FedWatch, traders now price a December rate hike as a near lock, most of them on a quarter point and a smaller group on half a point (CME FedWatch). Traders are more sure of a rate hike than your chatbot is of your name.
That hits your AI bill too. Data centres run on borrowed money, and the hyperscalers have issued record amounts of bonds this year. A rate hike makes every gigawatt cost more to finance. The landlord has a landlord. It’s called the Fed.
Per-dollar verdict: your next token comes with interest.
Bonus: Claude on AMD chips
From the roadshow: Anthropic says its Mythos model got Claude running on AMD chips in about three days, with basically no human help, against six to twelve months the old way (Anthropic’s roadshow). Most companies take longer than that to approve the purchase order.
The claim is narrower: prompt processing for one Anthropic model, on AMD’s MI355X, at a useful level of performance. It’s Anthropic’s own pitch to investors, and we found no independent check of it. Until then, it’s a three-day claim waiting on its six-month review.
Marcelo Lima says it means Nvidia’s CUDA moat is gone. That’s his take. Anthropic’s filing says it runs its compute (Amazon’s Trainium, Google’s TPUs, Nvidia and AMD GPUs) as a single, fungible pool (amended S-1). Hot-desking, for silicon.
Per-dollar verdict: when a model can do the moving, the landlord has to compete.
The tier list
How much API value you get for every subscription dollar, on our 11 November run of Adam Holter’s board against real API pricing. It ranks value, not usage. The order is holding for a third straight day.
Nine of the top ten spots are Claude plans. Anthropic’s best sales pitch is its own price list. Opus and Sonnet tie on the first three, so that’s ten spots on seven lines:
Claude Max, Opus and Sonnet tied: about $77 of API value per plan dollar.
Claude Ultra, Opus and Sonnet tied: about $66.
Claude Pro, Opus and Sonnet tied: about $64. Even the models are splitting the check.
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 (Claude pricing).
The best value in AI is about to be on the Nasdaq. Enjoy the bargain before it has shareholders.
And Gemini Argon doubles in price on 1 December, from $2 to $4 per million input tokens and $10 to $20 per million output (Google AI Studio pricing). The old price holds through 30 November. Just enough time to ignore three reminder emails.
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 destroyers are two Amazon engineers, sorting internal tickets. They swapped Claude Sonnet for a tiny decision model called Strands Decider (Strands Agents). It decides. That’s the whole job description. Sonnet ran on Amazon’s Bedrock; the small model runs on Amazon’s SageMaker. Sonnet got replaced by a sorting hat.
At list prices, a million tickets cost about $22,000 on Sonnet, and about $500 on the small model. Two cents a ticket felt cheap. Then somebody multiplied by a million.
That’s about 98% off. The one markdown every CFO likes.
It used about 70% fewer tokens per ticket. The big model was mostly clearing its throat.
The wait for each answer dropped about 78%. The big model read every ticket like a novel. The small one reads the subject line.
“A tiny model still needs a tiny room, darling. I charge by the hour.”
That room runs about $1.41 an hour. Replace about 1,500 Sonnet tickets a day, and one always-on box pays for itself, by our estimate.
The fine print: list prices only, and no AWS Lambda costs. No accuracy numbers were published. And the small model’s makers co-wrote the post.
Big win, if your queue is mostly sorting. Sonnet is taking it well. It wrote a four-paragraph goodbye.
Still: chainsaw approved.
A word from ourselves
Your prospectus for the AI line item. Your rain gauge for rate-hike weather. Your clipboard for the ticket queue. The people who keep the landlord honest.
Strong Compute plugs into 27 compute clouds, from Oracle to Voltage Park to Cerebrium, and 10 AI vendors, from Anthropic to OpenRouter to Simplismart, + 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.