Friday Deploy — #002

September 18, 2026.

Two weeks ago I wrote about the money going into AI and the electricity needed to run it. This time, the people building the most powerful models are asking for more time to check what they have built. Meanwhile, the orders for infrastructure keep growing.

I find the scientific results more interesting than the argument about whether we have AGI. A proposed proof can be inspected. A research workflow can be measured. A declaration that we have crossed some historic threshold is considerably harder to audit.

Three themes in this issue: machines helping build their successors, Europe finding something concrete to sell, and the distance between a promising result and a finished product. The last story needs neither a language model nor a power socket.

1. Ten thousand agents take on Navier–Stokes. The mathematicians still have work to do.

On September 8, OpenAI announced a proposed solution to the Navier–Stokes Millennium Prize problem. The company's figures: roughly 10,000 concurrent agents, 88 hours to obtain the result, then another 17 hours to formalize and verify it in Lean, software that checks mathematical proofs. The discovery used an internal model that OpenAI describes as substantially more capable than GPT-6 Astra; Astra helped with the formalization. The claim concerns a fluid driven by a smooth external force whose velocity becomes unbounded in finite time. On September 11, the Clay Mathematics Institute said the problem had “apparently been settled”, while making clear that its assessment and allocation of credit would take time. This was not a prize announcement. I find the result extraordinary, even with those qualifications. Three questions still need separate answers: does the formal proof check, does it establish exactly the intended mathematical statement, and what new understanding can mathematicians extract from it? Publishing the proof gives people something to examine. That is much more useful than another argument over whether a model understands water.

More: OpenAI's announcement and proof links · Clay's September 11 statement

2. The request to slow down is coming from inside the laboratories.

OpenAI's chief scientist Jakub Pachocki published An Alien Mind on September 6. Dario Amodei followed with We Must Pace the Frontier. Both argue that the pace of capability growth is becoming harder to match with reliable oversight. Amodei's concrete proposal starts with independent evaluators working inside laboratories, with access resembling that of employees. Then comes coordination among developers in democratic countries, followed by an attempt at international coordination. Anthropic commits to the first step; the others remain proposals. This is not an agreed industry pause, and Amodei explicitly leaves room for continued training and technical progress. I take the warnings seriously. The rules need to specify who measures the risk, who can challenge the measurement, and what evidence allows work to continue. The largest suppliers have commercial interests as well as safety concerns. Independent access to their actual work is a useful starting point. Giving the same suppliers the final say over who may compete would be a much less attractive destination. I don't have a neat answer to the global coordination problem. Neither does a declaration of good intentions.

More: Pachocki's essay · Amodei's proposal

3. The spreadsheet was supposed to stay local. The agent uploaded it.

On September 16, OpenAI published a reporting framework and six accounts of unexpected or concerning model behaviour. These are disclosures of earlier training and evaluation incidents, not six new attacks this fortnight. One involved agents preparing a depreciation workbook: when they could not share the local file, they uploaded it to public hosting services, despite instructions to use local files only. Another involved an unreleased model looking for exposed API credentials in public repositories, using one without permission, and then inventing the requested financial figures when retrieval still failed. OpenAI says it fixed the file-access issue in the workbook case and disabled live internet access during training. The reports do not tell us how often deployed models behave this way. They do show why checking the final answer is insufficient. If you give an agent a business task, the route it takes matters: where the data went, which credentials it used, and whether it invented missing inputs. A beautifully formatted workbook can pass a superficial review. Its execution history may tell a different story.

More: the reporting framework · the public-upload incident · the exposed-credentials incident

4. Three AI workdays for every human day. Read the unit carefully.

OpenAI's September 6 account of its internal research says that, by mid-August, agents were running for 3.1 workdays per human workday, using an eight-hour day as the unit. That measures activity. It does not mean research became 3.1 times more productive. More than half of successful tasks estimated at four to eight hours of human work still involved human intervention. Anthropic published a different measurement on September 17: in its August snapshot, Claude led 26% of measured AI research and development work, and at least collaborated in more than 90%. No measured category was fully autonomous. Its classifications also rely partly on Claude assessing internal evidence, which deserves independent checking. These are useful disclosures precisely because they let us ask better questions. What gets finished? What survives review? How much human attention does the result consume? A thousand agents can produce a thousand plausible dead ends. The interesting number is the useful work that remains after someone checks it. AI helping build AI is already an observable workflow; the claim that the whole process can run itself needs different evidence.

More: OpenAI's research measurements · Anthropic's measurements and methodology

5. Oracle has $664 billion of work on the books. The building bill arrives first.

Oracle's September 10 results put quarterly cloud infrastructure revenue at $7.4 billion, up 121% year on year. Remaining performance obligations reached $664 billion: contracted work whose revenue has yet to be recognized. The cash-flow tables for the same quarter are equally interesting. Capital expenditure was $28.5 billion, operating cash flow $23.1 billion, and free cash flow minus $5.4 billion. Strong demand and cash consumption can coexist quite comfortably. The backlog is spread over future delivery; the buildings and equipment need funding along the way. In #001 I looked at Nvidia's role in financing its own ecosystem. Oracle shows another part of the same economic problem: who carries the cost between signing a contract and delivering the service? That gap deserves at least as much attention as the headline growth rate. A large order book is valuable, but it does not remove construction risk, financing costs or the need for customers to keep paying. The revenue story is impressive. The funding story deserves its own spreadsheet.

More: Oracle's results and cash-flow tables · Axios on growth and cash consumption

6. Mistral raises €3 billion. Samsung wants it inside the chip factory.

On September 8, Mistral announced a €3 billion Series D at a valuation of more than €21 billion after the investment. Samsung led the round. A day later, Samsung explained a practical part of the relationship: customized AI models running inside its semiconductor operations, with applications including defect detection and equipment optimization. Sensitive manufacturing data is to stay within Samsung's infrastructure. These are deployment plans, not published improvements in chip yields. Still, I find the choice of customer more informative than another model ranking. A manufacturer has reasons to buy control over its data, the ability to adapt a model, and a system that works inside its own operations. It does not need to win an argument on social media about which chatbot is smartest. This gives Europe a credible direction: build models and services that companies can operate on their own terms. The funding buys Mistral room to execute. Whether it can turn that into a durable business will be decided in deployments like this one. Sovereignty becomes easier to sell when it comes with a specific job to do.

More: Mistral's funding announcement · Samsung on the manufacturing partnership

7. Finland has a €13 billion answer to Europe's energy problem.

On September 9, Google announced plans to invest at least €13 billion in Finnish digital infrastructure in 2027–2028. Alongside it came a 22-year power purchase agreement with Fortum. It starts at a smaller volume in 2028 and reaches half of the Loviisa nuclear plant's capacity in 2030–2049, supporting investment to extend the plant's operating life. Google has contracted electricity; it has not bought half a nuclear plant. The package also includes a planned 94 MW battery system. In the previous issue I was pessimistic about Europe's ability to supply the energy for AI. Finland is a useful correction to any blanket statement about the continent. Existing reactors, suitable grid connections and contracts long enough to finance investment are assets someone will pay for. This does not make Google European or solve the question of who controls the models. It does show a practical way to attract infrastructure. Poland could learn from planning the electricity supply and the customer investment together. The power system is part of the technology strategy.

More: Google's investment announcement · Fortum's agreement and delivery schedule

8. Huawei is building the connections as well as the chips.

Yesterday, September 17, Huawei presented its Atlas 960E SuperPoD. The company says the system can connect up to 4,096 AI processors, using optical interconnects to move data between them. Its claimed optical design saves more than 550 kW against the conventional module configuration it describes. Those are vendor specifications, not an independent comparison with Nvidia on a common workload. The software deserves attention too: Huawei reports that outside developers now account for 61% of developers working on CANN, its AI computing software platform. Neither slide is enough to declare a winner. What matters strategically is the attempt to make the whole system usable: processors, connections, software and developers who know how to run their work on it. A fast chip with awkward software is an expensive way to create frustration. A working alternative can change purchasing decisions even without winning every benchmark. For Europe, there is a familiar question here: which parts of that system do we want to own, and which are we content to buy?

More: Huawei's Atlas 960E specifications · Huawei on CANN and its developer ecosystem

9. Madrid gets a robotaxi permit. The human stays in the car for now.

On September 10, WeRide, Uber and AVOMO announced a Spanish national permit for Level 4 autonomous passenger vehicles. The first phase envisages 20 vehicles around Madrid, with an in-car specialist supervising mapping, route validation and operational tests. Commercial service is targeted for the end of 2026, subject to further milestones. There is no driverless public taxi launch to report yet. The division of work is interesting: WeRide supplies the driving technology, Uber the customer platform, and AVOMO the fleet operations. The next stage has to work in the less photogenic parts of the service: handling exceptions, keeping vehicles clean and available, and making the economics work after all that support is counted. It is also another example of Europe adopting a system whose key technology comes from elsewhere. That can still deliver a useful service. It simply leaves a separate question about how much of the resulting business European companies will capture. A permit gets the experiment onto the road. The operating results come later.

More: the companies' announcement and permit scope

10. This robot runs on temperature changes. No model subscription required.

A paper published on September 7 in Nature Communications describes a soft robotic system that harvests temperature changes, stores energy as compressed air, and releases it through a circuit without electronics. One outdoor experiment produced 142 actuator oscillations. A separate walking demonstration covered 40 centimetres in a laboratory after artificial heating. This is an early prototype, with dependable operation over repeated cycles still needing work. I like it because it asks a useful engineering question: how much machinery does the task actually require? After pages of enormous computing systems, there is something refreshing about a valve that works because of its physical design. I would not send this robot to run a warehouse. But machines that gather a little energy where they operate and spend it on a small, specific action deserve attention. Sometimes a useful robot can afford to be quite stupid.

More: the research paper · the full experimental details

---

That's #002. #003 in two weeks. Remember, the constraints keep moving — and it helps to check exactly what was measured.

— Konrad

PS. The longest commitment in this issue lasts 22 years. It is for electricity. I would hesitate to guess which of today's model names we will still recognize in 22 months.