Good morning from London.
We are looking for an outstanding economist to join us as an AI Economy Research Fellow. If you know someone we should speak to, send them our way.
Cheers
A MESSAGE FROM OUR SPONSOR
This paid post was unlocked for you by Granola, the AI notepad for back-to-back meetings.
Granola is our team’s favorite note-taker. We’re excited to partner with Granola to bring four months of free access to the Business plan (10 seats) to Exponential View annual members. This perk, valued at $560, is now part of our membership.
Upgrade to claim your free access to Granola.
The AI spend of your nightmares
Amazon spent some $1.8 million on a Claude project that ran for five months. A senior employee said: “It’s difficult to figure out how much anything [AI-related] costs”.
We had a similar experience with R Mini Arnold, my OpenClaw agent. Its costs have run up as it has grown in complexity, and it takes time to switch it to progressively cheaper models. It becomes hard to keep track of exactly what is running efficiently and what isn’t. Nathan Warren called out an outrageous few days when the bot blew through $500 a day.
The subsequent tedious audit was worth it. Many processes were running older, higher-tier models, like Opus 4.5, which are more expensive than smaller, newer models like Sonnet 5 or a slew of open-weight models. The price war that has broken out between Anthropic and OpenAI in response to Chinese advances has helped even more.
By default, RMA now uses my token allowance on OpenAI Codex, which is already paid for in my $200-a-month subscription. It will fall back to DeepSeek v4 Flash or Pro if OpenAI is unavailable. For harder tasks, it can jump to 5.6 Sol, OpenAI’s top model, through the same subscription or Kimi K3 or Anthropic’s Fable (both of which I pay for by the token). The net result is $6 a day, lower than it has been for months.
The funny thing is that RMA is cheaper than it ever has been and yet more capable than ever. It plugs into the Manus API for some types of work; Claude Code and Codex for coding tasks; Prism (our internal research graph, which is more powerful than ever); and other resources like Elicit for academic papers.
It’s a microcosm of the big question in the industry. Has $494 a day just disappeared from genAI revenue? In some sense, yes, but that was really an anomaly. RMA had typically cost me $50 to $60 a day before it went wild. Even at $6 a day, it runs to $2k per year from me alone, which is reasonably substantial for someone who isn’t writing code. I expect spending to spike as I move back into book-writing terrain and need more research done.
I’m curious whether readers have had similar experiences.
See also:
8 minutes with me on the state of the AI economy:
US companies are continuing to spend on AI. Ramp reports that “in July, the top 1% of businesses spent a median $7,400 per employee on AI. The top 10% spent $650. The median firm spent $11.95 per employee.”
Opus 5 is driving a large chunk of Anthropic’s revenue growth.
SpaceXAI is picking up a pricing fight with Grok 4.6. The new model undercuts top rivals by more than 60%.
Who really pays for data centers?
The cost of building data centers spills over on neighboring towns — in some cases disproportionately so. Each additional data center within 25 miles raises a neighboring town’s bond spread by about 10 basis points. The effect fades with distance — roughly 4 basis points at 75 miles. Neighbors also borrow more. A town with the average number of nearby data centers issues about $34 million more in debt over the following year. The effect roughly doubles between six and thirty-six months. In states with tax breaks, the bill goes through schools. After a state adopts a data center incentive, state transfers to school districts fall by roughly $673 per student.
Explained simply, the towns nearby get the strain but no bargaining chips, so when they need to borrow money for a school or a road, lenders charge them more. And if the state gave the company a tax break to show up, the money the state didn’t collect comes out of the school budget.
Addressing these types of issues is going to become a priority as data centers become about as popular as lead in petrol.
Jasmine Sun’s extraordinary reporting from the frontlines of data center backlash is more than worth your time.
See also:
Some teams within Microsoft are working on regenerative data center designs based on biomimicry.
Short morsels to appear smart at dinner parties
Researchers built protein logic gates that can trigger cancer cells’ self-destruction.
A hidden prompt injection in a court filing asked AI to side with the plaintiff in case the court used LLMs.
Batteries deployed in 2026 could move more than one-third of new solar generation into the evening hours to replace fossil fuels.
💪🏼 France’s solar panel recycling sector hit scale in 2025, up 40% from 2024.
An AI designed 16 entirely new synthetic viruses from scratch that were better at killing E. coli than the natural counterparts.
👀 Anthropic is hiring a chip design team.
AI is a decent financial advisor, but it tends to be too patient and sensitive to your prompting.
👾 Fun game: run the AI lab from 2017 and race to recursive self-improvement takeoff.
Over 150 years and despite major electoral reforms, Congress has consistently been dominated by “fortunate sons”.
Thanks for reading!
Thanks for today’s edition - insightful and uaeful! BTW, on the solar energy topic, Switzerland’s high dependency on hydroelectric power makes it vulnerable to the droughts we’re going through. However, the massive investment in solar has allowed it to compensate. Here’s the article that came out this week in the Swiss national radio-TV (in French) L'énergie solaire sauve la Suisse d'une pénurie d'électricité cet étéhttps://www.rts.ch/info/suisse/2026/article/l-energie-solaire-evite-une-penurie-d-electricite-en-suisse-29326400.html
The part that gets underdiscussed here is that token costs are only half the problem, the other half is that agent architectures often retry or branch silently on failure, so a single task can fan out into dozens of API calls with no human ever seeing a decision point. Traditional cloud cost tools were built for predictable compute curves, not for workloads where a model can decide to call itself five more times. Curious whether Amazon's overrun was mostly volume or mostly inefficient orchestration, because the remedy is completely different depending on which one it is.