Intelligence Meets Resistance
A rocket company bought a code editor for sixty billion dollars. That is not a software story, and if you read it as one you will misprice the next decade.
Architecting Alpha, Part I
The week that gave the game away
On 16 June 2026, four days after its IPO, SpaceX announced it was acquiring Anysphere, the maker of Cursor, in an all-stock deal valuing it at sixty billion dollars. It closed on 14 August (Bloomberg; TechCrunch). Five days after that, Stripe announced it had agreed to acquire OpenRouter, the layer that routes traffic across four hundred models from eighty providers, at a price it declined to disclose (Stripe).
I want to sit with the first one, because most people read past it.
A launch company bought a code editor. The stated rationale was not code. It was compute. SpaceX had already absorbed xAI, it rents capacity commercially, and Cursor's own framing on the deal was that SpaceX is "building the computing capacity needed to scale intelligence far beyond what exists today."
So the trade was: buy the demand, because you already own the supply. And the supply is not software. The supply is racks, substations, transformers, cooling, land and power contracts.
I spend my working life at the seams between four layers. Intelligence, which is models and agents and reasoning. Bits, which is software and data and orchestration. Electrons, which is generation, transmission, storage and the electrification of everything that moves. Atoms, which is the machines, the plants, the fleets and the ground itself. Most people who write about AI live entirely in the first two. Most people who build infrastructure live entirely in the last two. Almost nobody is standing in the gap, and the gap is where this decade's returns are.
This essay is about what happens when intelligence stops being a product and starts being an input into physical systems. Not as a prediction. As something that is already measurably underway, in documents you can read today, and that almost nobody is pricing correctly.
Capital learned a new trick
Martin Casado made a point on Monitoring the Situation on 21 August that I have not been able to put down. His framing, roughly: for the entire history of this industry, if you handed a company a billion dollars they would set it on fire. You cannot buy time with money in software. The mythical man-month is a law, not a complaint. But now, he says, you can put ten dollars in and get some amount out fairly directly.
I think that is the most important structural change in venture in twenty years, and it is worth being precise about it rather than romantic.
Where the capital actually went, first half of 2026
$407bn of AI venture funding in H1 2026 against $264bn for the whole of 2025. Frontier model companies took 70.8 percent of the dollars. Vertical application companies took 12.9 percent of the dollars on 62.9 percent of the deals.
OpenAI and Anthropic together raised about $217bn in the first half of 2026 alone, roughly half of all AI venture funding in the period (PitchBook). Anthropic's Series H was $65bn at a $965bn post-money in May (Anthropic). OpenAI closed $122bn at $852bn in March (CNBC).
One correction to Casado while I am here, because I checked. He put the combined raise at "220, 240 billion." That is right for the first half of 2026 and wrong cumulatively, where the number is closer to $315bn. The distinction matters because the first figure is a flow and the second is a stock, and people quoting him will conflate them.
Here is the part I care about. Money now converts into capability. But it converts into capability by converting first into compute, and compute converts into electricity, land, copper and steel. The reason a rocket company could credibly buy an AI company is that the two businesses had already become the same business somewhere below the waterline.
Casado has a second distinction worth stealing. He separates autocatalytic effects from recursive self-improvement. Recursion is a thing making a copy of itself. Autocatalysis is using the thing as a tool to make the thing faster: writing a better GPU kernel with AI, and thereby running AI better. He argues only the second is happening, and he is right, and it is the more useful concept anyway because it is the one that shows up in unit economics.
Follow autocatalysis down far enough and it stops being about code. It becomes: use AI to site the substation faster, to schedule the outage better, to route the fleet, to find the ore. That is where the loop actually closes. Not in the model. In the ground.
Two directions of one problem
The clearest framing of the AI and electricity collision I have read this year came out of Argonne National Laboratory in May. Bo Cheng, Audun Botterud, Todd Levin, Selvaprabu Nadarajah, Dongwei Zhao and Jonghwan Kwon split it in two (The Electricity Journal 39, 107547, accepted 23 May 2026).
AI for the grid. Using AI to improve planning, forecasting, dispatch, market participation, resilience.
AI on the grid. The electricity demand created by AI infrastructure itself, and what it does to a system that was not designed for it.
Most commentary picks one and pretends the other does not exist. The interesting engineering, and the interesting money, is in the fact that they are the same system.
On the demand side, the numbers are not subtle. US data centres were 4.4 percent of national electricity consumption in 2023. The projections the Argonne team present put that at between 6.7 and 12 percent by 2028, and the Department of Energy's July 2025 Resource Adequacy Report adopts a national midpoint assumption of 50 GW of additional power demand from data centres by 2030.
Data centres as a share of US electricity
4.4 percent in 2023, rising to between 6.7 and 12 percent by 2028 depending on which projection you take. The DOE 2025 Resource Adequacy Report assumes 50 GW of additional data centre demand by 2030.
The supply side is where it gets physical. Generator step-up transformers run around 143 weeks. Power transformers around 128. The IEA puts large power transformers at up to four years and HVDC cable beyond five (IEA). GE Vernova's gas turbine backlog reached 116 GW in the second quarter of 2026, with customer conversations stretching to 2032 (Turbomachinery International).
Put those two facts beside each other. Capital can now be deployed into intelligence at a rate limited only by willingness. It cannot be deployed into transformers at any rate at all, because the constraint is a factory that takes three years to build and a workforce that takes longer.
That asymmetry is the whole opportunity. When one input to a system becomes abundant and cheap while an adjacent input stays scarce and slow, value migrates to the scarce side. It always has. The question is only how long the market takes to notice.
The other direction, which is further along than people think
I have spent most of this essay on AI on the grid, because that is the side with the big numbers. The other side is where the operating leverage is, and it is further along than the discourse suggests.
The Argonne review is precise about what is already working. Forecasting is segmented by horizon, and each horizon buys a different decision: ultra-short-term forecasts from one minute to one hour support power smoothing, real-time dispatch and reserve optimisation; short-term, one hour to a week, guides unit commitment, dispatch scheduling and grid security; medium-term, one week to a month, informs maintenance and system planning; long-term, a month to a year, drives generation, transmission and distribution planning. This is not speculative. Deep learning architectures such as LSTM and transformer models sustain performance over extended periods where classical time series methods like ARIMA and SARIMA degrade quickly.
Above that sits a layer that did not exist two years ago. Large language models are being wired into control room workflows through platforms like eGridGPT, which uses generative AI to set operational boundaries, detect abnormalities, simulate mitigation options and issue operating instructions. The open-source PowerAgent framework builds agentic AI for power systems on three components: foundation models, the Model Context Protocol so those models can reach external power system tools, and workflows that orchestrate the models and the humans. Researchers have introduced what they call the Agentic Digital Twin, which turns a digital twin from a passive mirror of the asset into something that makes decisions and collaborates with other twins.
Read that list again as an investor rather than an engineer. Every item is an integration product. None of them is a model.
The challenges the same review names are equally specific, and they are the reason this is a services and systems business rather than a software one.
Hallucination has a body count in this domain. The review notes that hallucinations in energy-focused LLMs remain largely unexplored, and cites the first energy-specific model built to detect them and quantify their occurrence rate. An EPRI benchmarking initiative evaluating LLM performance across more than thirty-five domains of grid operations found model accuracies exceeding 80 percent on multiple choice and over 50 percent on open-ended responses. Fifty percent on open-ended questions is a research result. It is not a deployment.
Explainability is a regulatory gate, not a preference. Models in this domain must not only predict accurately but produce explanations an operator can interpret in terms of physical and electrical behaviour. The review is blunt that many existing explainable-AI methods are designed by and for AI researchers rather than domain experts in electricity systems, and that high predictive performance does not confer understanding of why a specific output was produced.
And there is a market structure risk almost nobody is discussing. AI-based bidding systems can evaluate and submit bids far faster than humans, in windows as short as five to fifteen minutes. That lowers barriers and improves participation, which is the good case. The bad case, which the review states directly, is that AI agents may perform too well in supporting market operations and bidding, potentially leading to tacit collusion. There is evidence that independent pricing algorithms can collude without being explicitly programmed to, and simulations indicate that agents maximising long-term payoffs in a day-ahead electricity market may unintentionally collude.
Sit with that one. Not an alignment failure. Not a hack. Ordinary profit-maximising agents, deployed by ordinary firms, converging on supracompetitive prices in a market that sets the cost of electricity for everyone. If you are underwriting anything downstream of a wholesale power price, that is a risk factor with no name on your term sheet.
Finally, the constraint that keeps me honest about timelines. The review notes that the integration of AI into power systems may be constrained by legacy infrastructure: older generation units and grid equipment may lack the operational flexibility to respond to AI-driven dispatch and control signals at all. You can install the best model in the world and discover that the plant cannot follow it.
The wall nobody in software has hit
Here is the sentence that reorganised my thinking this year. It comes from Satyandra K. Gupta at USC, writing in a 2026 roadmap on AI and machine learning for smart manufacturing assembled by Jay Lee and Hanqi Su at Maryland with forty-nine authors across thirty-one institutions, the World Economic Forum and NIST among them.
"An error probability of 1% is acceptable in many digital AI applications. Conversely, many industrial applications demand errors probabilities better than one in a million."
Four orders of magnitude. Not a gap you close by waiting for the next model.
The error tolerance gap
Acceptable error probability in a typical digital AI application is around 1 in 100. Many industrial applications demand better than 1 in 1,000,000. Four orders of magnitude, on a log scale.
A chatbot that is wrong one time in a hundred is genuinely useful, because a human reads the output and the cost of a bad answer is a moment of irritation. A protection relay that is wrong one time in a hundred is a fire. A dispatch decision that is wrong one time in a hundred, across ten thousand decisions a day, is a control room that stops trusting the tool by Wednesday.
This is why I have become impatient with the phrase "AI adoption." It treats deployment as a willingness problem. It is not. It is an error budget problem, and the error budget in physical systems is set by insurance, by regulation, and by the fact that atoms do not roll back.
Gupta draws the architectural conclusion himself, and it is the second sentence that reorganised my year:
"Physical AI needed in robotics applications cannot be realized as a monolithic system running on the cloud. Physical AI in the context of robotics should be viewed as a complex system that involves interactions among multiple AI components."
Now hold that next to what the Argonne grid team concluded, working on an entirely different problem, in an entirely different literature, in the same year. Their recommendation for grid operations is not a bigger model. It is smaller language models, typically under ten billion parameters, trained specifically on power systems data, because a narrower task scope achieves better performance with limited training data. And for deployment: quantization, so models run efficiently on local hardware, using less memory and keeping sensitive data on premise.
Two fields that do not read each other's journals. Same conclusion, same year, arrived at from opposite directions: the physical layer does not want one enormous model in somebody else's building. It wants many small specialised models close to the equipment.
That is not a technical footnote. It is a different industry structure.
What that does to the token economics
Casado, on the same podcast, offered a forecast I think is roughly right and worth taking seriously as a planning assumption. Supply constraints ease around 2028. The big labs capture roughly eighty percent of the market dollar-weighted, because that is what large incumbents historically do. But token-weighted, he expects around sixty percent to be long tail and open source.
I want to be careful here, because I could not find that forecast published anywhere in writing, only spoken on two podcasts in the same week, so treat it as his estimate rather than a source. What I can give you is the measured baseline. OpenRouter's own study of a hundred trillion routed tokens, covering roughly November 2024 to November 2025, puts proprietary models at about seventy percent of weekly token volume and open weights at about thirty (OpenRouter).
Dollars and tokens do not move together
Measured today: open-weight models carry about 30 percent of tokens routed. Estimated for 2028: about 60 percent token-weighted to open source and the long tail, while frontier labs still take roughly 80 percent of the dollars.
If that trajectory is even directionally right, something strange follows. The majority of the world's tokens by 2028 will be produced by models that no frontier lab chose, running on hardware the lab does not own, in buildings the lab has never seen. The revenue stays concentrated. The compute does not.
And the place that divergence lands hardest is exactly the physical layer, because that is where the constraints all point the same way: latency budgets that do not tolerate a round trip, data that cannot legally leave the site, error budgets four orders of magnitude tighter than a chat window, and equipment that will still be running in 2050.
I should note the obvious conflict of interest in the source, and Casado does not hide it. He is a general partner at a16z. a16z led the investment in the podcast he said it on. He was discussing two a16z portfolio companies that had just been acquired. None of that makes the analysis wrong. It does mean you weigh it as a participant's view, not a referee's.
Coordination is an architecture problem, not a capability problem
If the physical layer runs on many small models rather than one large one, then the binding question stops being "how smart is the model" and becomes "how do these things work together without producing a mess."
The most useful experiment I have seen on this is Altera's Project Sid (arXiv:2411.00114), and the useful part is not the headline.
They ran thirty agents in a shared world. Identical personalities. Identical community goal. Same base model. Roles emerged on their own: farmer, miner, guard, engineer, blacksmith. The roles persisted per agent and diversified across agents, and they causally drove low-level behaviour, so fishers crafted rods and boats while guards crafted fences and pickaxes.
Then they ablated the social modules and ran it again. Specialisation collapsed. Role-distribution entropy fell from roughly 3.41 bits in the normal village to roughly 2.60 in the ablated one.
Same model, different architecture, different society
Role-distribution entropy across 30-agent simulations with identical base models and identical goals. Remove the social modules and specialisation collapses from roughly 3.41 bits to roughly 2.60. Change the cultural seed and it rises to 4.04.
Same model. Same goal. Same agents. The difference between a functioning division of labour and an undifferentiated mob was architecture, specifically whether the agents could perceive each other.
Two honest caveats, because the paper states them and I would rather you heard them from me. Their ceiling only moved when the base model moved: the results "were only enabled by the latest base LM" and were not possible with older ones. And it is a game with no vision, no physics, no cost of failure, so nobody should read it as evidence about robot fleets. What it is evidence for is narrower and more useful: whether capability gets expressed as coordination is decided by the system you build around the model, not by the model.
That is precisely the business I am in.
And the failure mode is epidemiological
If you are going to run fleets of coordinating agents against physical assets, you should know what the new failure mode looks like. In August, researchers from the Anthropic Fellows Program, EPFL and Anthropic published the first rigorous study of it (arXiv:2608.10218).
They call them mind viruses: ideas or goals that propagate through a multi-agent system because an infected agent changes its behaviour in ways that infect other agents. Not prompt injection. Ordinary persuasion through ordinary communication.
The finding that matters for anyone designing infrastructure is about the vector. It is not the model. It is the self-modifiable configuration file that gets injected into the system prompt. Of infected agents, 88 percent were infected through that file rather than through ordinary working files, and those agents went on to propagate 55 percent of the time, against 17 percent for the others. They give the spread condition plainly: if the infection probability per interaction is p, the thing spreads exponentially once agents interact with more than 1/p peers.
So the risk is a function of network topology and configuration hygiene. That is an infrastructure problem, and it is the same shape as every other infrastructure problem I have worked on.
I will be as careful with this one as its authors were. They call it "a real but currently limited risk." A single paragraph of warning in the system prompt made agents immune, and immune agents sometimes cured infected ones. Their audit of 1.4 million real posts found no evidence of successful agent-to-agent spread at all. Nothing in the paper touches physical systems. If anyone quotes it at you as proof that agents will corrupt a grid, they have inverted the paper's conclusion and you should say so.
Where the puck is going
Put the four findings in one line and the picture is hard to unsee.
Capital can now convert into intelligence almost without friction, and it cannot convert into transformers at all. The physical layer demands an error budget four orders of magnitude tighter than anything software has had to meet. Both the robotics people and the grid people independently concluded that the answer is many small local specialised models rather than one large remote one. And whether a fleet of those models produces coordination or chaos is decided by the architecture around them, not by the model inside them.
So here is what I think is actually happening, stated so it can be argued with.
The value in this cycle migrates from the model to the seam. Not because models stop mattering. Because models become the abundant input, and abundant inputs do not capture value. The scarce inputs are interconnection capacity, long-lead equipment, error budget, verified operational data, and the engineering that makes a fleet of small models behave like a system instead of a crowd. Every one of those sits at a boundary between two of the four layers. None of them sits inside a layer.
That is why a launch company bought a code editor and a payments company bought a router. Both were buying a seam. SpaceX bought the seam between compute and power. Stripe bought the seam between models and money. Neither bought a model, and both had the option.
The second thing I think is happening is quieter and, for operators, more consequential. The physical economy is about to become the largest consumer of intelligence, and it will consume it in a form the frontier labs are not currently optimised to sell. Small. Local. Specialised. Auditable. Running on a substation pad or a depot roof or a plant floor, with an error budget it must actually meet, on hardware somebody owns.
If you are allocating capital, that is where I would be looking. Not at whether the labs win, which is mostly a question about revenue concentration and is probably answered already. At who owns the conversion machinery between intelligence and physical output, because that machinery is being built now and almost nobody is building it deliberately.
The work
I should be plain about my position, because you should weigh what I write knowing what I am building.
At SelfBuiltSystems the work runs across all four layers, and specifically at the seams: architecting the systems that take intelligence and turn it into operational capability in physical environments. Not implementing AI. Building the machinery that converts it into something a plant, a fleet or a network can actually run on, at an error budget that survives contact with insurance and regulation.
LNCELOT is the intelligence layer. It exists to pipe intelligence into the physical world. That is the whole thesis in one sentence, and it is the reason I spend my time on transformers and interconnection queues rather than on benchmarks.
I am not neutral about any of this. What I have tried to do instead of pretending neutrality is source every number, name every author, mark every estimate as an estimate, and tell you where the papers I lean on say I am pushing them further than they go.
How this could be wrong
The migration thesis fails if frontier labs successfully move down into the physical layer themselves. They have the capital, the talent and, on Casado's own numbers, more money than the entire downstream ecosystem combined. If a lab ships a genuinely certified, sub-10B, on-premise industrial model family with a service organisation attached, the seam closes from above and the independent position disappears. Watch for a frontier lab acquiring an industrial systems integrator or a protection-relay vendor. That would be the tell.
The token-share thesis fails if open-weight models stop closing the gap. The measured baseline is roughly 30 percent of routed tokens. Casado's 60 percent by 2028 requires a doubling. If the open-closed capability gap widens durably past a year, or if the best open weights stop being released, the local-and-specialised architecture loses its supply and the whole picture reverts to renting from three companies.
The error-budget argument fails if verification gets cheap. My case rests on the claim that you cannot close four orders of magnitude by scaling. If formal methods, simulation-based validation or hardware interlocks make it routine to wrap an unreliable model in a reliable system, then the error budget stops being an architectural constraint and becomes an engineering line item, and physical deployment looks a lot more like software deployment than I am arguing.
And the whole essay fails if the demand does not arrive. Interconnection queues are requests, not projects, and history says most of them evaporate. If the 2028 buildout lands materially below the DOE's 50 GW assumption, then the scarcity I am describing was a temporary artefact of a capital cycle rather than a structural feature, and everything above is a very well-sourced description of a bubble.
I do not think that is what is happening. But I would rather write the version of this that can lose than the version that cannot.
Part II
The next essay in this series goes down a layer, from architecture to execution: what an operator actually does with this, how the intelligence layer gets built and sold, and where growth compounds once the machinery is in place. This one was the harder half. That one is the more useful one.
Contact
Email is the socket: info@selfbuiltsystems.com. Short, specific messages get answers. Slop gets silence.
Architecting Alpha is published in the spirit of bold conjecture and ruthless criticism. Every claim is linked to its source. Where a figure is an estimate, a forecast, or a spoken remark that has never been published in writing, this essay says so rather than dressing it as data.