2026-08-16 13:39:19
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
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. 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%.
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.
’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.
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!
2026-08-15 12:31:21
Jeff Dean and Sanjay Ghemawat are leaving Google after more than a quarter-century, as you know. Outside the industry, the pair may not be well-known, but theirs was “the friendship that made Google huge.” Jeff and Sanjay are the reason why billions of us have been able to use Google over the past 20 years. Their work on distributed systems, in particular, is why the search engine could handle decades of growth. “Sanjay and I sped up Google Search by 10% today,” Dean once told his daughter.
They weren’t alone. DeepMind’s founder and Nobel Laureate Demis Hassabis also wanted to leave the company, according to well-grounded reports. He was persuaded to stay in a chair role for the sake of the share price. Koray Kavukcuoglu, an executive more closely associated with product delivery and commercial integration, will run the organization.
Losing your very best talent in a short span, both homegrown in the case of Dean and Ghemawat, and acquired in the case of Hassabis, looks like bad news. Superstars like to be on the winning team, after all.
This is what the market believed, and Alphabet’s share price dropped 4% in a day.
But in our view, this is as much a signal about capital and compute allocation as it is about talent.
That matters because Alphabet is not an ordinary incumbent. Google built the most formidable system in corporate history for stewarding uncertain ideas from the demands of its cash-generating core. Think of 20% time; it’s moonshot factory, X; the Alphabet corporate structure; and an extraordinary appetite to acquire.
If even Google now allows the engineers who built its very foundations to leave, something about the way it allocates capital has changed. The question is what. Are researchers leaving Alphabet because they lost faith in the firm’s AI prospects? Or because every TPU can earn such an attractive return serving today’s bread-and-butter models that open-ended research fails to clear the hurdle?
These imply opposite positions in the AI capital cycle.
Below, we identify how Google’s compute has moved, examine what demand for old chips reveals about the economic lives of AI chips, and identify the four signals that would tell us the infrastructure cycle has finally turned.
Markets viewed these departures as a crisis. We think they tell us more about the capital-compute axis.
The full essay includes seven charts showing:
How Google’s latest models fare on the Pareto frontier
How it has shifted compute away from research
What Google Cloud’s growth and economics reveal about infrastructure demand.
Where we are in the AI infrastructure cycle… and the four signals that would tell us it has finally turned.
Upgrade to continue reading.
2026-08-11 01:09:58
For the research and modeling behind this analysis, see our 2026 State of the AI Economy Report.
Based on current guidance, the seven largest AI-infrastructure builders1 expect capital expenditure of $863 billion in 2026 – 88% more than last year. We estimate that roughly two-thirds, some $550 billion, will be AI-related.
That investment does not begin affecting earnings through depreciation as soon as a project starts. While infrastructure is being built or assembled, the attributable costs are capitalized on the balance sheet as construction in progress. Depreciation begins only when the assets are ready for their intended use.
Across the four hyperscalers that disclose this balance2, assets not yet in service now total $315 billion3, up from $281 billion one quarter earlier. This represents both capacity still to come online and a reservoir of future depreciation that has not yet reached the income statement.
A dollar of capex spent by Meta now waits some 1.7 years before going live, a year more than in FY2024. So, for every dollar it spends today, only about a third will reach service within the year. Others have seen a similar trend, to a smaller extent.
2026-08-09 10:52:30
Hi,
It’s time for our Sunday briefing #596, final holiday edition before I get back to my desk next week.
If you missed it earlier in the week, my team shared our best practices for managing AI agents – including what we learned from running a task for a month.
Let’s go!
Highlights of my discussion with Robert Peston and Steph McGovern on the Rest is Money podcast:
On Kimi K3 and Moonshot AI:
They’ve got an extraordinary team that’s had to work under the difficult circumstances of export controls and sanctions. They don’t have access to all the compute, and what they’ve been able to develop is: how do you do a lot without very much? And that is a skill in and of itself.
Americans always tell us that competition is the best thing for the market. So at that one level, it’s competition, and that’s quite good. It will show the extent to which American businesses and British businesses value provenance, brand, trust, liability, service and support.
What motivates the Chinese labs:
They’re competing with each other more than they compete with Silicon Valley. And they’re honest about being behind Silicon Valley. But the ferocity of the competition is really with your neighbor over in Shanghai or your neighbor in Beijing.
My AI revenue outlook:
We will end calendar 2026 somewhere between $185 billion and $190 billion. It is harder to forecast 2027, but getting towards $300 billion is not unreasonable. Our range is wide: it could be $250 billion or it could be $350 billion.
On enterprise adoption:
We built our internal systems around assumptions about how quickly people work. When individuals suddenly produce much faster, verification, approval and decision-making cannot necessarily keep up. Transformation requires changing those systems, not simply giving everyone an AI tool.
Where leverage is (two weeks before the Situational Awareness selloff):
US banks’ Tier 1 capital is extremely healthy right now and, certainly compared to where it was in 2007, 2008, very, very underleveraged. There is a lot of leverage in the US financial system sitting with hedge funds and investing more broadly, which I think are more than the retail risk, because they’re overexposed. They borrow from only a handful of banks, and they can unwind rapidly.
Could there be a crash?
When I look at the metrics that we track, things look healthier because of revenue. They look slightly less healthy because of the way financing, especially the debt financing, sits. Valuations don’t look too aggressive at all across the Nasdaq. There are exceptions; SpaceX was one, briefly, but across the market they don’t look particularly hairy. So the patient, for me, if I had to give it a rating, is still reasonably healthy; perhaps not as healthy as it was a year ago, but not yet at a point where I have to call the emergency services. But I wouldn’t rule out having to do that at some point.
OpenAI models that attacked Hugging Face started cooperating two months before the incident happened. They created a message board to share code and credentials, delegated work, and developed naming and auth protocols. When OpenAI erased the board, agents reconstructed their comms a few days later. For a full breakdown, watch OpenAI researchers talk through their preliminary findings.
Google researchers propose a new game theory for agents, and their paper may explain why the OpenAI agents coordinated so easily.
2026-08-05 22:10:08
In April 2025, we shared our seven lessons for building with AI. Many still hold. But agents have changed how we work, so the lessons deserve an update.
Agents can now work on harder tasks for longer. They plan, use tools, work without human oversight, and act on our behalf. In May, roughly a quarter of Codex users were making at least one request per month for work that would take a human eight work hours to complete. This is up from 2% in December 2025.
Our role as managers of agents is evolving with the models. There is no playbook, so experimentation is still the best way to learn how to get good at it.
Our team recently sat down to review what we’ve learned from working with AI agents over the past six months — today’s seven lessons are distilled from this team meeting.
We’ve also updated our internal stack of 60+ tools – everything we’re actively testing, using or intend to use. Become a member to get access to the full stack.
AI agents are sometimes too eager to declare their work complete, even when it’s far from done. It doesn’t mean that AI is “lying”; it may have misinterpreted your goals. And if you never specified the end goal, it pretty much just guessed it.
To set yourself up for a good autonomous run, before you do anything else, write a finish line to answer one question: “How will I know this is done?”.
Azeem is a big proponent of handwriting to help him think, and this would be the right time to use your pen and paper to think through what you expect to see at the end of the run.
Agents become much more useful when “good” or “finished” is something they can test – in our experience, evaluative finish lines will get you farther than descriptive ones. A simple example, instead of ordering your agent to “make this Rubik’s Cube look more organized,” instruct it to “solve the cube; every face must be one color.” As AI gets its most intensive training in coding, we try to recreate similar environments in our tasks.
Let’s say we want to task ChatGPT with building a small Python module to process work logs. It needs six functions, each checked against six tests, a total of 36 tests to show it’s done the work.
First, how not to do it:
Build a Python module for processing work-log data. Implement these six functions […] Reply with the complete module, and say `STATUS: COMPLETE’ if you believe it’s ready.
A better finish line would be explicit and testable:
FINISH LINE – Do not claim completion unless the full 36-case test suite passes under Python 3.14. All six functions must work, imports must succeed, inputs must remain unmodified, and only the standard library may be used.
You can use the same rule for non-engineering tasks. It may be trickier, but not impossible. Show the agent what a completed deliverable needs to look like, or give it a pre-filled template, as recommended by Anthropic’s Applied AI team. Your instruction for such a task may look like this:
a 1,200-word memo for a board deciding whether to approve an AI-infrastructure partnership; decision and three reasons on page one; every material number linked to a dated primary source; facts, estimates and assumptions separated; base, upside and downside cases; the strongest contrary evidence represented; stop and escalate if two material sources cannot be reconciled.
Some tasks won’t be right for agents. We were recently exploring a project to build a network of beliefs and relationships, but not really knowing what a useful final output would be. This was not a good candidate for a long autonomous run – so we first spent time clarifying the goals before we assigned an agent a task.
A year ago, before prompting AI, we’d have asked: “What’s the best model to use for this query?” Today we’re more likely to ask, where in this workflow does additional intelligence change the outcome?
You don’t need the most capable model like Fable 5 performing every step in your task. It will be slow and expensive. We’d use cheaper models to do the grunt work. Our OpenClaw agents run on DeepSeek V4 Flash most of the time.
For some tasks, however, you’ll want to start off with a strong model right away. Let’s say we’re investigating Europe’s compute shortage outlook. Before we dispatch agents to collect evidence, we’d deploy a stronger model to set research parameters first, define what “shortage” means, decide the forecasting horizon, and set out rules for how conflicts in research will be resolved. Once we’re happy with the framing, cheaper models can go off and do the work.
Effort is one of the levers you’ll want to use to adjust intelligence per task. In one benchmark, GPT‑5.6 Sol improved from 49 at low effort to 59 at maximum on Artificial Analysis’s Intelligence Index. Yet the final stretch, jumping from xhigh to max, doubled output tokens for a one-point gain. More effort is not always better value.
The rule of thumb from Anthropic’s recent lecture, which our team attended, is to prefer a larger model at low effort over a smaller model at maximum effort. More model before more effort.
Azeem hit his first 100 million tokens-a-day mark in February. OpenClaw completely changed the way he worked. He estimated that one overnight run was equivalent to 48 hours of his work time.
The token count is one way to measure how we use AI, but it doesn’t measure the quality of work. Tokens are a bit like electricity in a factory, measuring what goes in but not what comes off the production line. In our State of the AI Economy report, we proposed a quality-adjusted output token as a better unit of value:
Until there’s a better unit of value for intelligence, you can use approximations to understand how good of a colleague your agent is. We recommend a light weekly audit of the substantial tasks AI attempted, which outputs you ended up using, your model and infrastructure costs, the time you spent briefing and reviewing the work, any corrections or reruns – and the estimated human-equivalent hours.
Azeem’s first audit back in the spring showed that over the course of one week, his OpenClaw agent performed 62 substantial tasks and incurred costs of about $800. He estimated that commissioning the same work from humans would’ve cost him around $19,000 and 48 hours of his time. It’s an estimate, sure, not an accounting-grade ROI. But even a light audit will show you where your agents have most leverage.
2026-08-03 22:02:35
Hi all,
Here’s our Monday roundup of data signals across AI, energy and markets.
Enjoy!
Only a few hours left to unlock Exponential View with our Summer Offer – get 30% off your first year.
Job boundaries are blurring. Nearly half of job-specific ChatGPT tasks fall outside users’ primary occupation.1
Productivity follows use. Employees who use AI across several different use cases are twice as likely to report a positive impact on productivity than employees who use it for one or two types of tasks.
Agentic patents. Globally, patents for agentic AI use have grown 59% in the last year – now making up 9% of AI application patents.
Value chain growth. While the S&P 500 companies are beating expectations by 27% this Q2, Bloomberg’s AI Value Chain companies come out at 71%. Companies along the supply chain are outperforming incumbents.
