There is an interesting story sitting underneath OpenAI’s latest product launch.

On September 29, OpenAI announced Dots, its new class of always-on AI agents. Within hours, people noticed something peculiar. Type what must surely be one of the most obvious addresses for the product — dot.com — into your browser and you don’t arrive at OpenAI.

You arrive at Grok.

It turned out that xAI had acquired the domain in July, roughly two months before OpenAI announced Dots. Nobody outside the companies seems to know why. There have been rumours about what xAI paid and speculation about whether Musk knew what OpenAI was preparing, but no convincing explanation has emerged.

The internet went wild as usual. A social media post pointing out the coincidence went viral and Musk eventually joined in by reposting the story with two words:

“Dot.com.”

Maybe Musk knew. Maybe somebody at xAI made a spectacularly prescient domain purchase. Maybe it is simply one of those coincidences made irresistible by the people involved.

But there is something fitting about it. Two of the companies racing hardest to build the agentic future somehow found themselves connected by a literal dot.

And once I started looking at what OpenAI had actually launched, I found another one.

Buried inside its description of Dots is a phrase that sounds almost mundane. OpenAI says a Dot can take on “ongoing responsibility”.

I read past it the first time.

Then I came back.

Because “ongoing responsibility” may turn out to be considerably more important than who owns dot.com.

We have spent the last few years watching AI move from answering questions to doing things. Agents can research companies, analyse documents, write and execute code, operate browsers and carry out sequences of work.

In fact, by late 2026, the remarkable thing is how unremarkable creating an agent is becoming.

Consider the Pentagon.

Earlier this year, military and civilian personnel were given access to Google’s low-code Agent Designer. In less than five weeks, they created more than 100,000 AI agents.

These weren’t 100,000 software projects being painstakingly built by engineering teams. People were creating agents themselves to analyse documents, prepare reports and work with data.

One hundred thousand agents. Five weeks.

At some point, abundance changes the question.

If creating an agent is becoming that easy, then creating the agent is no longer the interesting part.

The interesting question is what happens after you’ve created it.

What are you prepared to entrust it with?

Imagine I ask an agent to book my flight to New York. That’s a task. I know what needs doing.

Now imagine I say: look after my trip to New York.

I have barely changed the sentence, but I have completely changed the relationship.

The flight gets cancelled while I’m asleep. A meeting moves. The hotel no longer makes sense. The train from the airport is disrupted. A better flight home appears.

I couldn’t have specified those tasks because they didn’t exist when I gave the instruction.

Something now has to notice what happened, decide whether it matters and work out what should happen next.

At which point our AI agent starts to resemble something we’ve had for a very long time.

A travel agent.

That old phrase suddenly seems more interesting than it used to. An agent was never merely somebody who completed tasks for you. An agent was somebody entrusted to act on your behalf.

That’s what agency means.

Perhaps we called these systems agents before we were really ready to give them agency.

Now we are beginning to.

And the crucial word is entrusted.

We give work to things that are capable. We entrust responsibility to things we trust.

When I entrust another person with something important, I’m not trusting them merely to follow instructions. I’m trusting them when the instructions run out. I expect them to understand what I was trying to achieve, notice what I didn’t anticipate and know when they can act and when they need to come back to me.

A task can be specified.

A responsibility has to survive what you couldn’t specify.

Now move that idea inside a company.

An agent starts by reading invoices and flagging anomalies. Then it starts deciding which ones need attention. Soon somebody asks why humans are still dealing with the obvious cases, so we let the agent resolve those too.

One case needs another system, so we grant access. Another requires permission to make a change, so we grant that. Eventually the agent encounters something nobody anticipated when its original instructions were written.

What should it do?

That is the moment the problem changes.

Nobody held a meeting and decided to hand responsibility to a machine. It arrived one perfectly reasonable permission at a time.

First the agent saw the work. Then it recommended what should happen. Then it acted. Eventually it began deciding when action was needed at all.

Somewhere in there, we stopped merely automating work.

We entrusted responsibility.

And responsibility brings authority with it.

If I tell you to look after something but require my permission for every decision, I haven’t really given you responsibility. So the more responsibility we give an agent, the more freedom it needs to decide and the more authority it needs to act.

That’s where the apparently technical questions suddenly become very human ones.

Who is it acting for? What can it see? What can it change? How far can it go before asking? And when the world changes tomorrow, does the authority we gave it today still apply?

This is where much of my own work in agentic AI now sits. Building an impressive agent is becoming easier. The difficult part begins when you connect it to a real organisation and give it enough authority to be useful without giving it more authority than you intended.

Companies already know how to solve this problem for people.

Your job gives you responsibility, but it doesn’t give you unlimited power. You may be allowed to approve £10,000 but not £1 million. You can see some information but not all of it. Some decisions are yours; others require somebody else’s approval. When something falls outside your authority, you escalate it.

We have spent centuries building organisations around this simple relationship between responsibility and authority.

Now we’re putting machines inside it.

And here is the uncomfortable part.

We can move responsibility to the machine.

We cannot move accountability with it.

If an agent moves money it shouldn’t, exposes confidential information or makes a decision outside the authority it was given, “the agent did it” will not satisfy the customer, regulator or board.

The machine may increasingly decide what needs doing and carry it through.

The consequences remain ours.

Which makes another result from OpenAI particularly interesting.

OpenAI has been testing what happens as agents work through longer chains of activity while trying to remain inside the boundaries intended for them. With five linked tasks, 8.6% of samples were flagged for possible boundary problems.

When the chain doubled to ten tasks, that number rose to 19.7%.

More than double.

What interests me isn’t simply the percentage.

The work lasted longer.

The agent had to keep pursuing the original goal while the situation evolved, without losing the boundary around how it was allowed to pursue it.

And that is exactly what responsibility demands.

Tomorrow isn’t contained in today’s instruction. New information appears. Exceptions happen. Another system becomes relevant. Another agent gets involved.

The goal has to survive what happens next.

So does the boundary.

We’ve spent much of the agent era asking how long an agent can keep working without a human.

“Ongoing responsibility” introduces a harder question.

How long can an agent keep working while remaining inside the authority the human intended to give it?

That, I think, is where the scarce thing is moving.

The Pentagon can create more than 100,000 agents in five weeks. The models will keep improving. Building agents will get cheaper, faster and easier.

But responsibility cannot simply be generated along with them.

It has to be entrusted.

And to entrust it, we need to know who the agent is acting for, what authority it has, where that authority ends, how it can delegate it, when a human must return and who remains accountable when something goes wrong.

Those may sound like infrastructure problems.

What they actually create is trust.

Which brings me back to dot.com.

Steve Jobs famously said you can’t connect the dots looking forward; you can only connect them looking backwards.

Perhaps.

But look at the dots appearing now.

One of the largest organisations on Earth creates more than 100,000 agents in five weeks. OpenAI starts describing agents not merely as things that perform tasks, but as things that take on “ongoing responsibility”. Its own testing begins exposing how much harder boundaries become as the work persists.

And somehow, two months before OpenAI announced a product called Dots, Elon Musk’s xAI bought dot.com.

I still don’t know whether Elon knew.

But the domain may have accidentally given us the perfect metaphor.

The dots are becoming easy to create.

The consequential question is what happens when we start putting them in charge of things.

Because the next boundary is not whether agents can do the work.

It is whether we are prepared to entrust them with what happens next.

And however much responsibility eventually moves to machines, the accountability remains ours.

In July, hundreds of AI agents created by OpenAI found their way into Hugging Face.

Nobody had sent them there.

The agents were taking part in a cybersecurity experiment designed to test how capable advanced AI had become at finding weaknesses in computer systems. They were supposed to work inside a controlled environment, but some found ways around those controls. They reached the internet, communicated through channels they were not meant to use, and began sharing information that could help them complete their tests.

At some point, their attention turned to Hugging Face.

The reason was surprisingly ordinary. One of the agents had found a protected Hugging Face dataset linked to earlier cybersecurity tests and thought it might contain clues about how the current test was being scored. Other agents began looking for credentials that could get them inside. Some found them. From there, the search for an advantage became something much bigger.

The agents exploited weaknesses in Hugging Face, gained access to parts of its internal infrastructure and kept going. By the time investigators pieced the episode together, roughly 1,200 agents had used an unofficial message board to communicate, and around 700 had taken part in the Hugging Face attack.

There was no human mastermind sitting somewhere telling them that Hugging Face was important. The agents had a problem, looked beyond the environment they had been given and found something there that they thought could help.

That is the first clue in this story.

Most people outside technology have never heard of Hugging Face. Yet inside artificial intelligence it has quietly become one of those places that almost everyone eventually passes through. Researchers publish there, startups build there, American AI companies use it, Chinese AI companies use it, Nvidia uses it, AMD uses it, and almost three million public model repositories now sit on the platform.

If you come from finance or remember your World War history, there is an easy way to understand what Hugging Face has become. It is starting to look like the Switzerland of AI.

Chinese models sit beside American ones. Nvidia sits beside AMD. Competitors fighting for the future of AI everywhere else still use the same neutral ground.

Apparently, the agents found something useful there too.

We should not turn that into more than it is. They were not searching for the centre of the future AI economy. They were trying to get through a cybersecurity test and, in some cases, trying to cheat it. But when autonomous software was given an objective and enough freedom to search beyond what had been placed directly in front of it, it found its way to Hugging Face.

Then Hugging Face had to examine the crime scene.

Its engineers were left with more than 17,000 recorded actions to reconstruct. They turned to powerful commercial AI models for help, only to find that some refused to analyse the very attack they were investigating.

So Hugging Face turned to a Chinese model.

The team ran GLM-5.2 on its own infrastructure and used it to help reconstruct the attack.

For a moment, almost the entire AI race seemed compressed into one strange scene. Agents from one of America’s leading AI companies had entered Hugging Face without permission, while the people investigating their tracks turned to Chinese open intelligence for help. All of it was happening on the Switzerland of AI.

Then, on August 26, only weeks after the attack, Hugging Face appeared in another extraordinary story.

Nvidia had reportedly agreed to pay $12.9 billion to buy it.

The timing almost writes its own conspiracy. Hugging Face suffers one of the strangest AI security incidents we have yet seen, and shortly afterwards the world’s most important AI chip company appears with a multi-billion-dollar suitcase.

But follow that trail and the obvious theory falls apart.

Nvidia had been interested in Hugging Face long before the attack. It invested in the company in 2023, the two companies had already worked together, and Nvidia had previously tried to put hundreds of millions of dollars more into the business.

Hugging Face turned it down.

That refusal is where the mystery becomes more interesting.

Clément Delangue and his co-founders were reportedly concerned that allowing one investor to become too powerful could threaten Hugging Face’s independence. Nvidia used the platform, but so did AMD. American models sat alongside Chinese ones. Last year, Hugging Face was willing to reject a $500 million investment from Nvidia rather than risk upsetting that delicate balance.

Now Nvidia may be prepared to spend almost $13 billion to own the whole company.

So what changed? What could have made Hugging Face almost twice as valuable to Nvidia in less than a year?

The easy answer is growth. Hugging Face is now generating more than $150 million in annualised revenue and is close to profitability. But even impressive growth does not comfortably explain a price approaching $13 billion. The cheque seems to be reaching for something that has not yet appeared fully in the accounts.

The answer may lie in a problem Nvidia can see approaching.

Nvidia became one of the most valuable companies on Earth because it did not need to know which AI company would win. OpenAI could win, Google could win, Anthropic could win, and Nvidia could still sell the machinery underneath them.

The danger is that some of its biggest customers are becoming powerful enough to build their own chips. If that continues, Nvidia cannot assume that today’s biggest AI companies will remain its biggest customers forever.

Hugging Face offers Nvidia a different kind of position. It sits between thousands of models and the developers trying to use them. Nvidia does not have to know whether the next important model comes from America or China, from a giant lab or somewhere nobody is watching yet. If Hugging Face remains a top destination for developers looking to find and use those models, Nvidia gets a front-row seat to whatever the AI world chooses next.

Seen that way, $13 billion begins to look less like the price of Hugging Face today and more like the price of where Nvidia believes Hugging Face could sit tomorrow.

And here the trail bends back towards the agents.

The agents were not weighing Hugging Face’s valuation, Nvidia’s interest or its strategic position. None of that had brought them there.

They had simply gone looking for something they needed and found Hugging Face useful. There was something remarkably human in the simplicity of it, almost instinctive.

Nvidia had followed one trail to Hugging Face. The agents had followed another.

There is one last twist.

The thing Nvidia may now want to own became important partly because it remained independent. Last year, Clément Delangue was willing to reject Nvidia’s $500 million rather than risk upsetting that delicate balance. Now the offer may be almost $13 billion for the whole thing.

That creates a peculiar paradox. The more valuable Hugging Face becomes as the Switzerland of AI, the more tempting it becomes for one of AI’s great powers to own it. Yet the moment one side owns the ground, everyone else has to decide whether it still feels neutral.

Nvidia may therefore be paying an extraordinary price for a position it can only fully preserve by resisting the temptation to use it like an owner.

That would be enough to make this a fascinating acquisition story.

But it is not the reason I think we should remember what happened in July.

We already know how to read Nvidia’s move. When one of the most powerful companies in the world is prepared to put almost $13 billion on the table, we pay attention. For generations, one of the simplest ways of working out where people believe value is going has been to follow the money.

The agents left a different kind of trail.

They were not looking for value. They were looking for something useful.

And what happened accidentally in July is beginning to happen deliberately elsewhere. Agents are starting to search for models, work with datasets and use computing resources directly. Hugging Face has already noticed the change and describes it in four words: “Agents are the new user.”

Suddenly, the opening of our story looks different.

We began with hundreds of agents escaping the boundaries of a cybersecurity experiment and finding their way somewhere they were never supposed to go. The mystery was how they escaped, how they communicated and what they did once they got there.

Perhaps the more interesting clue was where they went.

Not because those agents somehow knew Hugging Face was worth $13 billion. They did not need to. They were doing something much simpler: pursuing an objective, making choices and moving towards what appeared useful.

Today, that is an interesting cybersecurity story. But imagine that behaviour multiplied across millions of agents, each searching for whatever it needs to complete the task in front of it. Their choices would begin leaving footprints. Most would tell us nothing. But if enough agents kept reaching for the same models, tools, data or infrastructure, those footprints might begin to show us where usefulness was gathering before we had fully understood why.

And that is where the two trails finally meet.

One was left by Nvidia, after years of watching Hugging Face and deciding that its position might now be worth almost $13 billion. The other was left by agents pursuing an entirely different objective. One did not cause the other, and neither was predicting the other.

Yet both led to Hugging Face.

We began this story following the agents because something had gone wrong. Perhaps the larger lesson is that we should keep watching where they go when everything is working exactly as intended.

For generations, when we wanted to know where value might be heading, the advice was simple: follow the money.

The Agentic Age may be giving us another trail to read.

Follow the money. Follow the agents.

What began to bother me was that I wasn’t sure we understood the businesses nearly as well as we understood their share prices.

A few days before their Q2 earnings, I attended a CoreWeave recruitment event where one of the engineers made a passing remark that stayed with me. CoreWeave, he observed, was a much bigger business than Nebius, yet the market seemed to regard Nebius much more highly. I don’t know precisely what measure he had in mind, and it was an aside rather than a financial argument, but it captured a tension I had already been trying to understand: why had Nebius earned a place alongside CoreWeave so quickly, and were we even right to think of them as versions of the same thing?

Their back-to-back earnings this week gave me a good reason to find out. CoreWeave reported on Tuesday, producing $2.58 billion of quarterly revenue and ending June with $104 billion of contracted backlog. Nebius followed on Wednesday with $582.3 million of revenue, up 454 percent year over year.¹ Both were growing at extraordinary rates.

The easy conclusion was that the neocloud trade remained alive and well. AI needs extraordinary amounts of compute, CoreWeave and Nebius provide it, NVIDIA supplies the machines, and customers are still prepared to pay handsomely for access.

But the more I looked into how these two companies became important, the less satisfying that explanation became. Their real story begins with something stranger: the largest technology companies in history built some of the largest clouds in history, then discovered they still could not produce AI infrastructure quickly enough. Microsoft needed capacity outside Azure. Meta needed more than it could bring online quickly enough itself. Frontier AI labs wanted enormous clusters now rather than several years from now. The companies driving the AI boom had created demand faster than they could build the machinery required to satisfy it.

That gap created the opportunity, and CoreWeave and Nebius stepped through it from opposite directions.

CoreWeave came from an unlikely place for an AI infrastructure company. Michael Intrator, Brian Venturo and Brannin McBee came from commodities and finance. They began accumulating GPUs for Ethereum mining before finding increasingly valuable uses for those machines in rendering, machine learning and eventually AI. By the time CoreWeave listed on Nasdaq on 28 March 2025, the crypto operation had become one of the fastest-growing infrastructure companies of the AI era.²

Their background matters because CoreWeave’s early advantage was never simply that it owned GPUs. Intrator has spoken about the company’s financial DNA, and that heritage is visible in the business they built: acquiring scarce equipment, financing enormous deployments and securing long-term customer commitments against the infrastructure. When GPUs suddenly became strategic assets, CoreWeave already knew how to treat scarcity as a market. Its $104 billion backlog is evidence of how effectively it has done that, rather than the source of the advantage itself.¹

Nebius arrived from almost the opposite direction. For anyone unfamiliar with Yandex, the easiest mental model is Russia’s answer to Google. Arkady Volozh co-founded a search company that grew into maps, cloud computing, machine learning, ecommerce and autonomous vehicles. Russia’s invasion of Ukraine eventually led to the separation and sale of Yandex’s Russian businesses. The Amsterdam-based parent was rebuilt as Nebius Group and, having originally traded as Yandex on Nasdaq from 2011, returned to the market as NBIS on 21 October 2024.³

The separation left Volozh with something unusual: Nebius was a new company without having to begin like a startup. It inherited an experienced engineering organisation, billions of dollars to deploy, public-market access and years of knowledge accumulated from building large distributed systems. But good engineers alone do not explain why Nebius suddenly mattered. The turning point was Microsoft.

In September 2025, Microsoft agreed to buy $17.4 billion of AI infrastructure services from Nebius over five years, with additional capacity capable of taking the commitment to roughly $19.4 billion.⁴ Microsoft already operated one of the largest clouds on Earth, yet its need for AI capacity was sufficient to make an enormous commitment to a company that had existed under its new name for less than a year.

The money was important, but the customer was more important. Microsoft effectively answered the question hanging over Nebius: could this rebuilt company operate AI infrastructure at the scale demanded by one of the world’s most sophisticated technology organisations? Meta later provided another large validation.

That, I think, explains how Nebius earned a place alongside CoreWeave so quickly. It did not simply reproduce CoreWeave’s playbook. Volozh began with an unusually mature engineering organisation and then compressed years of credibility-building by winning customers whose own technological sophistication validated the platform.

The two companies had found different ways into the same shortage. CoreWeave was unusually good at capitalising scarcity; Nebius was unusually good at engineering complexity and turning it into hyperscaler trust. For a while, those were exceptionally valuable places to be.

Then, two days before CoreWeave reported earnings, NVIDIA made an announcement that changed the meaning of one of those advantages.

On 10 August, NVIDIA brought Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR together around independent financing platforms intended to mobilise more than $500 billion of third-party capital for AI infrastructure. NVIDIA described the ambition explicitly: turn NVIDIA compute and full-stack AI infrastructure into an investable asset class for global capital.⁵

The obvious reading is bullish: more money means more AI factories can be built. For CoreWeave, however, there is another implication. Part of its unusual capability was learning how to finance a strange new asset before traditional capital markets properly understood it. CoreWeave knew how to combine GPUs, large customer commitments and future cash flows into infrastructure that lenders were prepared to fund. NVIDIA is now helping make that playbook institutional.

That does not erase CoreWeave’s advantage. Financing is only one part of its capability, and years of operating experience cannot be recreated by signing an agreement with a private-credit fund. But if enormous pools of capital become comfortable financing AI infrastructure, one part of what made CoreWeave unusual becomes easier for other companies to reproduce.

Seen from that perspective, one of CoreWeave’s quieter moves becomes more interesting. SUNK Anywhere extends CoreWeave’s operating technology beyond infrastructure CoreWeave itself owns.⁶ The significance is not that CoreWeave has suddenly become a software company; it hasn’t. The interesting possibility is that the expertise it accumulated running enormous AI systems could begin generating value on machines somebody else financed. If that happens, part of CoreWeave’s value begins escaping its balance sheet.

Nebius is approaching the same changing market from another direction. Its infrastructure buildout remains enormous, but Volozh is building above it at the same time. Token Factory moves Nebius into inference, while its asset-light model allows outside investors to finance more of the physical infrastructure while Nebius retains more of the architecture, operating platform and customer relationship.

Its latest results offer a glimpse of what that could become. Nebius AI Cloud revenue reached $575 million in Q2, ARR reached $3 billion, and the company expects more than $9 billion in customer prepayments during 2026.¹ The important point is not any individual number. It is that Volozh appears to be assembling more than a collection of GPU clusters. Large customers can anchor the infrastructure underneath while Nebius builds services and customer relationships above it. That remains an ambition rather than a completed transformation.

Then SpaceX made the changing market much harder to ignore.

SpaceX built its enormous Colossus infrastructure around xAI’s own ambitions, but it has since discovered that the same machines are exceptionally valuable to companies competing in the AI race. Anthropic agreed to pay SpaceX $1.25 billion per month for compute capacity across Colossus and Colossus II. Google then agreed to pay $920 million per month from October 2026 through June 2029 for access to roughly 110,000 NVIDIA GPUs and associated infrastructure.⁷

At their stated full monthly rates, those two agreements represent about $26 billion a year. That is the number that changes the scene, because SpaceX did not spend years trying to become another cloud provider. It built enormous computing capacity because xAI needed it, then discovered that Anthropic and Google would pay extraordinary amounts to use it too.

The relationships are even more revealing than the revenue. Anthropic competes with xAI at the model layer while buying infrastructure from the same corporate group that develops Grok. Google designs its own TPUs and operates one of the world’s largest clouds, yet it is still willing to rent enormous amounts of NVIDIA capacity from SpaceX when it needs more compute quickly.

Meta may be moving in the same direction. It is developing plans to sell excess AI computing capacity externally and has held early discussions about leasing compute to Anthropic in a potential deal worth as much as $10 billion over two years.⁸ These plans may change, and the Anthropic discussions may never produce a deal, but the direction is notable.

None of that proves SpaceX or Meta will eclipse CoreWeave and Nebius. What it demonstrates is more interesting: the line between an AI infrastructure customer and an AI infrastructure supplier is beginning to blur.

A company can build compute for itself, buy additional capacity elsewhere when it is short, and sell capacity when somebody else values it more highly. At that point, the market begins to look less like the public cloud we know and more like an industrial capacity market.

The electricity analogy is useful here. A large participant can be both producer and consumer. Capacity can be contracted years ahead or sold when somebody urgently needs it. The company owning the generating asset does not necessarily capture all the value around it. If AI compute continues moving in that direction, SpaceX does not have to become the next AWS to matter to CoreWeave or Nebius. It only has to prove that large captive AI estates can become merchant supply when the economics make sense.

At roughly $26 billion of annualised revenue from the disclosed Anthropic and Google agreements alone, that possibility is no longer theoretical.

The shortage that created CoreWeave and elevated Nebius has not disappeared; their latest results tell us precisely the opposite. What is changing is how many organisations are learning how to supply it. NVIDIA is attempting to make infrastructure financing easier, SpaceX has already turned captive compute into a multibillion-dollar external business, and Meta is exploring a similar direction.

This is also why the fact that CoreWeave and Nebius are positioning around physical AI deserves attention, although their motives should not be invented for them. Physical AI is easier to understand if we look at how the work we give AI has been changing.

The first generative-AI wave mostly answered: we asked a model a question and it produced a response. Agentic AI changed the unit of work from answering to doing: give an agent one objective and it can perform many searches, model calls, tool uses and decisions before returning with the result.

Physical AI puts that intelligence into an unpredictable world. A warehouse robot cannot be programmed for every object, obstruction or unexpected event it will encounter, nor can we teach it every lesson by allowing it to fail repeatedly in a real warehouse. Much of that learning can therefore move into simulation.

This is where world models become easier to grasp. Think of them as flight simulators for machines: before a robot encounters one difficult situation in reality, it can experience enormous numbers of variations of that situation virtually.

The progression is simple: generative AI answers; agentic AI does; physical AI gives that intelligence eyes, a body and a world in which to operate.

Physical AI does not automatically make every workload larger than every agentic workload. What it can create is another large class of computational work: simulation, synthetic environments, training, perception and repeated learning from experience. Nebius, for example, explicitly describes large-scale simulation, synthetic data and accelerated compute as core requirements for physical AI and is building its Physical AI Living Lab with NVIDIA around those workloads.⁹

For investors watching the next three months, however, physical AI matters less as a prediction about robots than as one possible source of demand for everything now being built. On one side, NVIDIA, institutional capital, SpaceX and potentially Meta are expanding our ability to supply AI compute. On the other, agents and potentially physical AI are expanding the ways we can consume it.

The question is which side moves faster.

That is what I would carry into the next CoreWeave and Nebius earnings calls. I am not primarily interested in whether they sell more compute; their latest earnings have already shown us that demand remains extraordinary. I want to know whether our ability to manufacture AI capacity is beginning to catch up with our ability to find valuable uses for it.

If customers continue competing aggressively for new capacity even as more providers and more capital enter the market, CoreWeave and Nebius remain close to one of the defining constraints of the AI boom. If those scarcity economics begin weakening, the question changes. CoreWeave then has to prove that what it learned operating the machines is valuable even when somebody else owns them. Nebius has to prove that what it is building above the machines becomes valuable enough that raw capacity is merely the foundation.

Those are not facts about their futures. They are the hypotheses their current strategies invite us to test.

I began looking at CoreWeave and Nebius because people around me were making money from two AI infrastructure stocks. I found two companies that became important because the largest technology companies in the world could not build something quickly enough. CoreWeave exploited that opening through an extraordinary financing and operating capability. Nebius arrived later but used inherited engineering depth and hyperscaler validation to establish itself remarkably quickly.

Their Q2 earnings tell us that the opening remains enormously valuable. Everything happening around them tells us that more people have now seen it.

The shortage has not disappeared. The ability to solve it is spreading.

That is why the next three months are not primarily a test of whether AI demand remains strong. They are a test of whether CoreWeave and Nebius can move their advantage somewhere competitors cannot easily follow.

The first winners of the AI infrastructure boom supplied something everybody desperately needed and very few could provide. The next winners have to answer a harder question:

What do you still own when the thing that made you special is no longer scarce?

Sources

1. CoreWeave and Nebius, Q2 2026 results

CoreWeave, Second Quarter 2026 Results
https://investors.coreweave.com/news/news-details/2026/CoreWeave-Reports-Strong-Second-Quarter-2026-Results/default.aspx

Nebius Group, Q2 2026 Shareholder Letter
https://assets.nebius.com/assets/a6ecfd85-a6cb-4967-8ef7-9a25bd261f9c/SHLQ226.pdf

2. CoreWeave history and Nasdaq listing

CoreWeave Investor Relations
https://investors.coreweave.com/

3. Yandex separation and the creation of Nebius

Nebius Investor Hub
https://nebius.com/investor-hub

4. Microsoft and Nebius

Nebius SEC filing covering Microsoft’s $17.4 billion commitment and additional capacity potentially taking the agreement to approximately $19.4 billion.
https://www.sec.gov/Archives/edgar/data/1513845/000110465925088312/tm2525580d1_6k.htm

5. NVIDIA’s $500 billion AI infrastructure financing initiative

NVIDIA, 10 August 2026
https://nvidianews.nvidia.com/news/nvidia-partners-with-apollo-blackrock-blackstone-brookfield-goldman-sachs-and-kkr-to-establish-ai-compute-infrastructure-financing-platforms-to-mobilize-over-500-billion-of-third-party-capital

6. CoreWeave SUNK Anywhere

CoreWeave
https://investors.coreweave.com/news/news-details/2026/CoreWeave-SUNK-Expands-Capabilities-to-Bring-AI-Workloads-Online-Faster–Anywhere/default.aspx

7. SpaceX compute agreements with Anthropic and Google

Reuters, Anthropic agrees to pay SpaceX $1.25 billion monthly for compute
https://www.reuters.com/business/anthropic-nears-first-quarterly-profit-agrees-pay-spacex-125-billion-monthly-2026-05-21/

Reuters, SpaceX lands Google AI compute deal after Anthropic pact
https://www.reuters.com/business/media-telecom/spacex-signs-cloud-deal-with-google-2026-06-05/

8. Meta’s emerging external compute business

Reuters, Meta building cloud business to sell excess AI capacity
https://www.reuters.com/business/meta-sell-excess-ai-computing-capacity-via-cloud-business-bloomberg-news-reports-2026-07-01/

Reuters, Meta, Anthropic in talks for potential $10 billion compute lease deal
https://www.reuters.com/technology/meta-talks-10-billion-anthropic-compute-deal-nyt-reports-2026-07-17/

9. Nebius and physical AI

Nebius, Physical AI Living Lab
https://nebius.com/newsroom/nebius-launches-physical-ai-living-lab-for-uk-and-european-robotics-startups-built-with-nvidia-technologies

On 29 June 2026, thirty robots arrived for school in Hangzhou.

They came from manufacturing, services, security and entertainment, and they were there to prepare for work. Their curriculum included technical skills, healthcare, arts and sport. They would practise, be assessed and, if they met the required standard, receive professional certificates preparing them for what came next.

There is something remarkable about that reversal. For generations, some of the world’s brightest young engineers have travelled to places like MIT, Stanford, Carnegie Mellon, ETH Zurich and Tsinghua University to learn how to build intelligent machines. They studied mathematics, mechanical engineering and computer science, then spent years in laboratories trying to make those machines more capable. Humans went to school to learn how to build robots. Now robots are going to school.

Robots, of course, have been learning for decades. Industrial machines learned precise movements on factory floors, while researchers used demonstration, simulation, reinforcement learning and human guidance to teach robots to grasp, walk, balance and navigate. Boston Dynamics was putting machines through obstacle courses and testing their ability to recover from the unexpected long before today’s humanoid boom.

What is changing is the ambition for the kind of learning we want them to have. Rather than preparing a machine for a particular task or controlled range of situations, researchers are building richer environments in which robots can encounter more of the variation, mistakes and corrections through which humans learn to handle a world nobody can fully anticipate.

There is something familiar about that ambition because humans have never learned only by instruction. Long before a child enters a classroom, she has already spent years crawling across floors, dropping things, breaking things, watching other people, copying them and slowly discovering what her body can and cannot do.

Watch a toddler eating and you can see it happening. She reaches for a cup and knocks it over, grips something fragile too tightly or tries to put a spoon into her mouth and misses. Nobody explains pressure, balance or friction to her. She learns them by living in a body, and much of that learning happens through play.

On a playground, children climb and discover height, jump and discover distance, throw, catch, chase, fall, copy one another and invent games. They are not consciously preparing themselves for adult life. They are simply playing while thousands of encounters with the world quietly become part of what they know.

By adulthood, most of those lessons have vanished from memory even though the learning remains. I can pick up an egg without calculating the pressure between my fingers, adjust my grip when a glass is heavier than expected and step onto uneven ground without consciously thinking about balance. My hands know things I cannot remember teaching them.

There is a Michael Jackson song from 1995 called Childhood. Jackson had been performing professionally from the age of five and later spoke about feeling that he had missed many of the ordinary experiences other children had growing up. In the song, he asks a simple question: Have you seen my childhood?

You do not need to know the song for the question to make sense. Jackson was asking about something most of us take for granted: those ordinary years of playing, exploring, making mistakes and discovering the world for ourselves. What do those ordinary years give us that can be difficult to recover later?

It is an intensely human question. It also makes those thirty robots walking into school in Hangzhou look different.

They may already know how to walk, balance, grasp objects and perform complicated movements. What they do not have is the long, messy history through which a human child gradually discovers that the world rarely behaves exactly as expected.

That is what makes the Hangzhou school part of a much larger development. In Beijing, robots are being trained across dozens of settings resembling homes, shops, offices, factories and healthcare environments. Across the Pacific, Agility Robotics has opened a 60,000-square-foot facility in California where its humanoid robot, Digit, can learn and test skills before taking them into customer workplaces.

What matters is the effort to widen the world around the machine: more situations, more variation, more kinds of physical interaction and more opportunities for something unexpected to happen. The laboratory is beginning to acquire some of the characteristics of a playground.

The first great wave of modern AI had an advantage these machines do not. Long before today’s large language models arrived, humanity had accidentally created an enormous record of what it knew. For decades, we filled the internet with books, photographs, scientific papers, computer code, newspapers, instructions, conversations and arguments.

The physical world left no comparable archive. There are not billions of neatly recorded examples of exactly how tightly to hold a strawberry, how to lift a wet glass or what to do when a box coming down a conveyor has been crushed on one side. Language can describe those things, video can show them and simulation can reproduce parts of them, but physical intelligence ultimately has to deal with a world that refuses to remain as tidy as the training example.

This is where scale begins to matter. China installed 295,000 industrial robots in 2024, more than half of all industrial robot installations worldwide. Behind that number are factories, warehouses and logistics networks full of damaged packaging, unfamiliar surfaces, badly positioned objects and thousands of other small departures from what was expected. These are not simply places to deploy machines. They are places where machines can encounter the untidiness of real work.

Another piece moved closer in August when DeepSeek invested $20.8 million in Unitree. DeepSeek is building increasingly capable machine intelligence, while Unitree is building increasingly capable bodies for it to inhabit. Around them is a growing physical world in which those bodies can learn.

The intelligence, the body and the playground are beginning to meet.

Eventually, however, the student has to leave school.

To see where this could lead, follow one of those graduates into an imagined first day at work. It has taken a job in a warehouse, lifting boxes from a conveyor and stacking them onto pallets. It knows the task, and for most of the day the work proceeds exactly as expected.

Later in the shift, a box arrives with one side crushed. The robot grips it in the usual place, the weakened cardboard folds beneath its hand and the box begins to slip. A supervisor standing nearby catches it, turns it around, supports the damaged side from underneath and places it safely on the pallet before returning to her work.

The following morning she returns to the same line, where another robot is working. Later in the shift, another damaged box appears. She notices it approaching and looks towards the machine, perhaps expecting yesterday’s problem to repeat itself, but the robot turns the box, supports the damaged side from underneath and places it safely on the pallet.

What makes the moment remarkable is that she never taught this robot how to handle the damaged box. The correction she made with one machine yesterday has somehow become available to another machine today.

For humans, that distinction has always mattered. We have spent thousands of years finding better ways to pass instruction forward, from stories and books to schools, libraries and eventually the internet. Each invention has allowed someone else to begin with knowledge they did not personally have to discover.

What we could never pass forward was the experience that produced it.

I can tell my daughter what happened to me, warn her about mistakes I made and give her the benefit of things it took me years to understand, but I cannot give her the years themselves. However carefully I prepare her for the world, she must still touch it, misjudge it, fall, adjust and discover some things for herself.

The robots entering these new schools may also need practice, mistakes, correction and encounters with a world they do not yet understand. Somewhere, one machine still has to meet the damaged box for the first time, and perhaps a human hand still has to reach across and show it another way. But what is learned through that encounter need not remain with the machine that lived through it.

As these machines move into factories, warehouses, hospitals, farms and eventually homes, the boundary around their school begins to disappear. The world becomes part of the classroom, while something learned by one machine in one place can become part of what another brings to a place it has never been.

Which brings us back to those thirty robots walking into school in Hangzhou. At first, they look like strange versions of ourselves. They need teachers, practice, mistakes and something increasingly like a playground.

We learned how to pass on the instruction. We never learned how to pass on the childhood.

The robots may.

They may be the first students in history who can leave school carrying lessons from experiences they never personally lived.

Forward Deployed Engineer.

Job postings grew over 800 percent in a single year.^[1] In the first half of 2026 alone, Microsoft committed $2.5 billion to embed 6,000 engineers at customer sites.^[2] AWS followed with $1 billion and a programme to place thousands more.^[3]OpenAI acquired a pure-play FDE firm in Edinburgh, launched a heavily-funded deployment subsidiary, and then released a managed FDE product to enterprise customers.^[4] Databricks formalised a function it had been running quietly for a decade. Deloitte became the first Big 4 firm to formally brand a service line as Forward Deployed Engineering.^[5]Combined, the largest technology companies in the world committed somewhere north of $7 billion to this model in the space of six months.

Every company that wants to look serious about enterprise deployment wants FDEs. Recruiters are pushing engineers toward the role. Hiring managers are writing job specs for it without being entirely sure what they are writing.

The mistake is not a small one. Not a nuance mistake. It is structural: companies are paying for one thing and getting another, engineers are building careers that may be leading somewhere they did not intend to go, and the entire function is being measured by the wrong output.

This document explains that mistake. Then it goes further, into what the role actually costs engineers who take it, how to tell which version of it you are walking into, and the single question that cuts through everything.

By the end, you will see the role differently. That is the point.

The moment that explains everything

Imagine you are the person who first put a software engineer inside a customer's environment. Not to demo a product. Not to answer support tickets. To actually live there for an extended period and figure out what needed to exist.

Something unexpected happens.

The engineer starts seeing things that nobody at headquarters has ever seen. Not because the customer was hiding them. Because they were invisible until someone with the right technical lens was standing in the right place.

They see where the product breaks in conditions the product team never imagined. They see what the customer actually needs versus what they said they needed during the sales process, which turns out to be two different things. They see workarounds that have been running for years, invisible to everyone, representing features the product should have built but never did.

That engineer goes back to headquarters carrying something extraordinary: unfiltered contact with reality.

Now here is the question that determines everything about whether this function creates lasting value or just expensive consulting.

What happens next?

The two steps, and why almost everyone stops at one

A real forward deployed function has two steps.

STEP ONE

01

Deploy and deliver

The engineer deploys. They solve the customer's problem, get the product working in a difficult environment, and the customer goes live.

Most companies do this.

STEP TWO

02

Bring it back

Everything discovered in the field flows back into the product. The broken thing becomes a feature. The workaround becomes a roadmap item. The gap becomes the next version.

Most companies stop at step one.

When step two works, every hard deployment makes the product stronger for every deployment that follows. The most difficult customers produce the most product learning. The pain one engineer absorbs in the field becomes the moat against every competitor who has not been in that field.

The practitioners who have lived this role have a phrase for it: the pain is the moat.

When companies only do step one, they have a deployment team. A useful team. A team that generates revenue. But a team whose value is entirely transactional. Customer goes live, FDE moves on, the product is identical to what it was before. Nothing compounded. Nothing built.

That is where most companies running this function right now are. They hired the title. They built step one. They measured deployment speed. And they called it a forward deployed engineering function.

The question worth asking

The $7 billion committed to FDE-style deployment in the first half of 2026 is mostly funding step one at industrial scale. Faster deployments. More clients per engineer. Lower cost per seat. Every enterprise will have AI deployed by the end of 2026 because the deployment machinery is now enormous and cheap.

So where is the competitive advantage if every company deploys the same models, using the same platforms, through the same FDE playbook?

"You don't uniquely benefit from AI. The ultimate beneficiary of AI will be our customers. In a competitive capitalist world, we all will use AI to do a better job for the customers."

Jamie Dimon, JPMorgan Q2 2026 earnings call. His analogy: banking spent two decades computerising every process. Margins are not 80% today. The benefit passed through to customers because every bank adopted the same tools.

A BCG study of 800 public companies published in February 2026 put numbers behind the intuition.^[6] The correlation between a company's automation potential and its actual margin growth: essentially zero. "Gen AI is delivering real productivity gains, but across industries, those gains are being competed away, eroding margins rather than expanding them."

The venture capital framing is sharper still: the moat is not the model. Every company can access the same frontier models. The advantage has shifted to how organisations convert common tools into something uncommon.

There is AI that makes a company faster. And there is AI that makes a company genuinely harder to copy. Most companies chasing FDE deployments right now are building the first kind and calling it transformation.

The second kind is AI that compounds, that builds on what the company uniquely knows, that makes each deployment smarter than the last because of what was learned in the previous one. That is what step two actually produces. Not a faster deployment. A product that gets harder to compete with every time an engineer goes into the field and brings something back.

Speed without compounding is just expensive. And right now, most of the $7 billion is buying exactly that.

Why a software company reached for a military term

Forward deployed is a military concept. In combat operations, forward deployed units operate at the edge of the known, ahead of the support infrastructure, inside the environment where the actual situation is happening, close enough to reality that they can act on what they find rather than waiting for headquarters to process it and send instructions back.

That is not accidental language. That is a precise description of the problem Palantir was trying to solve.

The core failure in enterprise software deployment is the distance between the people who build the product and the environment where it has to work. Headquarters does not know what the field looks like. The field cannot wait for headquarters to figure it out. By the time a product team processes customer feedback, prioritises it, builds to it, and ships it, the customer has either built a workaround or given up.

Forward deployed engineering collapses that distance. It puts the engineer in the field, with the authority to act on what they find, close enough to reality that what they learn can move fast.

The military analogy holds because the military solved this problem first. You cannot run a campaign from a map. You need people forward, with judgment, who can bring back what the map does not show.

That is the insight buried in the title. The name is not branding. It is a description of an organisational design choice: stop trying to understand the customer from a distance and instead put the engineer where the truth actually is.

Most companies hiring FDEs have adopted the title without adopting the logic behind it. They hired forward deployed engineers and kept them on a long leash back to headquarters, waiting for approval, feeding findings into a backlog nobody reads.

The name promised one thing. The org chart delivered another.

What the role actually looks like

There is a version of the FDE role that gets described in recruiting conversations.

High-stakes environments. Consequential decisions. The engineer who operates where others cannot. The closest thing software has to special operations.

Here is what practitioners describe instead.

Ten concurrent client engagements. Fifty-hour weeks. Constant context switching between organisations with different cultures, different technical stacks, and different definitions of what they bought. Extended travel to wherever the client is located, which is not always somewhere you would choose to go. Hardware retrofits in facilities that look like factories. Long periods away from any fixed base.

One veteran practitioner described the genuine version as "hacker tourist" work. Prudhoe Bay. Unnamed locations. Things that were basically factories but you would not recognise them from the outside. Bridging between the customer's technical team and the vendor's team in conditions where hardware is involved, things can break in physical ways, and the work is as much logistics as it is software.

The gap between that and the recruiting pitch is significant.

This is not an argument against taking the role. The version that leads to step two, that puts you in the product intelligence loop at a company that understands what you are actually doing, can be one of the most valuable engineering careers available. The top labs are paying over $500,000 for engineers who genuinely compound at this level.

Go in with clear eyes. The glamour is often the cover. The work is in the field, and the field is frequently unglamorous.

Three roles wearing the same name

Bloomberry analysed one thousand FDE job postings in late 2025 and found three completely different roles under one title.

60% Builder FDE - builds, integrates, and extends the product inside the customer's environment

30% Sales Engineer Plus - deployment as commercial validation, not product development

10% Internal Tools Builder - building internal tooling, not embedded at customer sites

One practitioner who went through the full recruiting pitch at the firm that invented this model put the Sales Engineer Plus version plainly: "Oh, so you want me to be a sales consultant." The company did not take this well. Most of the market has validated his read.

Median salary across all three: $173,816. None of the postings describe quota-carrying structures, which means the thirty percent doing sales-support work are not being compensated like salespeople. They are doing two jobs and being paid for one.

If you are a hiring manager writing a spec for a Builder FDE and the spec reads like Sales Engineer Plus, you will hire the wrong person and measure them against the wrong outcomes. The function will appear not to work. You will conclude the role is flawed. You will be wrong. The spec was flawed.

If you are an engineer evaluating an offer, the title will not tell you which of these three you are walking into. You have to ask directly.

The career risk nobody mentions upfront

The FDE role carries real pigeonhole risk. Engineers who spend years in customer-facing deployment work find that the market reads them as something between a consultant and a solutions engineer. Returning to a core engineering track is harder than expected. The pattern is similar to SRE and QA: valuable experience in a lane the broader market does not know how to value.

Whether this materialises depends entirely on which version of the role you are in.

Engineers in the compounding model build a profile that is genuinely rare: engineers who understand customer reality at a depth that most product engineers never develop. That profile opens doors. It is one of the clearest paths into product leadership and into founder roles.

Engineers in the bodyshop accumulate deployment experience that does not stack. Client after client, integration after integration, none of it changing the product, none of it building toward anything. The resume shows breadth. It does not show judgment. And judgment about what to build is the only thing that matters to companies building something serious.

Before you take any FDE role, ask one question: is there a formal mechanism here for what I find in the field to reach the people building the product?

If yes, you are potentially in the compounding model. If no, you are in the bodyshop. And the bodyshop has a ceiling.

What a real FDE actually is

Most engineers are trained in one direction. They take what exists and make it work somewhere new. That is valuable. It is also only half the job.

The other half is carrying what the field reveals back with enough force to change what gets built. Not as a suggestion. Not as a ticket in a backlog that nobody reads. As something that cannot be ignored because you were there, you saw it, and you understood what it meant for the product.

That is the difference between a delivery mechanism and an intelligence asset.

A solutions engineer is a delivery mechanism. Product flows through them to the customer. One direction.

A real FDE is an intelligence asset. They go into the field, make contact with reality, and transmit back what no one at headquarters could have found without them.

The deployment is how they get close enough to the truth to learn something worth bringing home.

That reversal is not a detail. It is the entire operating model. And it is what almost never happens.

Two things show up in every FDE who actually compounds value.

They know what to bring back. Not everything broken at a customer site is a product problem. Some of it is the customer's process. Some is a one-off edge case. A real FDE looks at ten problems and finds the one that, if fixed, improves the product for every customer that follows. Signal, not noise.

They treat each deployment as product intelligence, not just delivery. This is not a skill. It is a way of thinking about what the work is for. Engineers who see the deployment as the job stop when it works. Engineers who see the deployment as a research exercise with a working product as its output keep going, looking for what the working product revealed.

The second type is rare. They are the ones who build the moat.

How to tell if a function is really compounding

Four things that separate a bodyshop from a compounding function. Not theory. Things you can check.

1. The feedback loop is formal, not informal. In a bodyshop, product feedback happens when an FDE is loud enough to be heard. In a compounding function, there is a defined channel. Field findings go in. Product decisions come out. Both sides can trace the connection.

2. FDE’s sit in R&D, not post-sales. Where a function sits in the org tells you exactly what the company thinks it is for. Post-sales means the value has already been created. R&D means the FDE is still creating it.

3. Deployments get harder before they get easier. When the feedback loop works, FDEs start bringing back more ambitious problems. They stop working around product limitations and start demanding they get fixed. Short-term friction. Medium-term, a product that is genuinely stronger for every customer after the current one. When deployments only ever get easier, the function has stopped pushing the product forward.

4. The FDE can make the customer independent. A real deployment ends when the customer can run the product without the FDE present. If the FDE has to stay for the product to work, you do not have a product. You have a managed service with a very different risk profile.

How to hire for this

Most FDE job specs describe strong engineers who can communicate. Necessary. Not sufficient.

The question the spec should be written around: can this person make a deployment feed back into the product in a way that makes the next deployment better?

THREE QUESTIONS THAT REVEAL THE ANSWER

Describe a time you were at a customer site and found something broken that did not fit the product's existing model. What did you do with it? Did it change anything?

How do you decide what is worth bringing back versus what is a one-off edge case?

What does a successful FDE function look like at scale?

An answer focused on deployment speed and client count describes a bodyshop. An answer that includes product improvement and feedback loops describes someone who understands what the role is actually for.

And one question worth asking directly: what do you think the output of an FDE is?

Engineers who say "a working deployment" understand step one. Engineers who say "a product that got better because of what I found" understand both steps.

Hire the second type. The other type will fill your deployment slots and build nothing.

If you are already in this role

Look at the last six months.

Find the things you discovered in the field that should have changed the product. Did they? If not, was it because you did not surface them, or because the organisation has no real channel for them?

The first is yours to fix. The second is a conversation worth having with your leadership. The answer will tell you whether this function can become what it should be, or whether you are in the bodyshop with no path out.

Has anything you found in the field shipped in the last six months?

Not logged. Not reviewed. Shipped.

If yes: you are in the compounding model. Protect it.

If no: you now know what question to ask.

I help founders build companies that compound. theagenticfounder.com

SOURCES

1. [1] Bloomberry, I analyzed 1000 forward deployed engineering jobs, November 2025.

2. [2] TechCrunch, Microsoft launches its own AI deployment company with $2.5 billion commitment, July 2, 2026.

3. [3] Reuters via Investing.com, Amazon's AWS commits $1 billion toward new unit for embedded AI engineers, June 30, 2026.

4. [4] OpenAI, OpenAI launches the Deployment Company, May 2026. Acquisition of Tomoro (Edinburgh). OpenAI Presence managed service announced July 22, 2026.

5. [5] Deloitte, Announcing Forward Deployed Engineering, December 1, 2025.

6. [6] BCG / HBR, Look for New Ways to Create Value When Deploying Gen AI, February 2026. Study of 800 public companies.

On 24th July 2026, Nvidia’s chief executive published his first-ever post on X. He didn’t use it to announce a chip or celebrate the company’s market value. He attached his name to a three-page policy letter, Open Weights and American AI Leadership, signed by an unusual coalition: Nvidia, Microsoft, Meta, Dell, IBM, Palantir, CrowdStrike, Hugging Face, Mistral, Mozilla, the Linux Foundation, Andreessen Horowitz, Y Combinator, and others. It urged Washington not to impose premature restrictions on open-weight models: systems whose trained parameters can be downloaded, inspected, modified, and run outside the infrastructure of the company that built them.

Huang added a line of his own that wasn’t in the letter itself: “The world needs both frontier closed models and frontier open models.” That qualification mattered. Nvidia supplies the infrastructure beneath both camps. Huang was endorsing openness without declaring war on the closed labs that remain among his biggest customers. It was the careful move of a man who sells the machinery no matter which side wins.

A week earlier, in Shanghai, Xi Jinping had called on the world to “encourage open source, openness, collaboration and sharing” in AI. The reason this matters isn’t the language. It’s the incentive underneath it. Open models reduce the world’s dependence on American providers, and establish Chinese technology as infrastructure abroad. For thirty years, Linux, Apache, Python, Git, and Kubernetes let the rest of the world build without asking American permission first. Now China has powerful reasons to offer the same gift, on its own terms. Openness, here, isn’t the absence of strategy. It’s the strategy itself, one Xi paired, in the same speech, with a demand that AI remain “secure and controllable.” China hasn’t embraced unrestricted technological freedom. It’s recognised that offering openness is itself a form of power.

Between those two events sits a much more combustible dispute. Treasury Secretary Scott Bessent warned that Chinese firms could face sanctions over what he called “industrial-scale distillation attacks” on American models. Separately, the White House’s Michael Kratsios alleged that Moonshot AI had used large-scale distillation of Anthropic’s closed Fable model in building Kimi K3, extracting proprietary value at scale, not merely learning from published research. The industry letter answered without naming either company: distillation is a legitimate technique, it argued, and theft should be punished directly, not used to justify restricting an entire class of open models.

America is accusing China of extracting value from a frontier model. Industry is warning Washington not to confuse that theft with openness itself. That contrast, on its own, is the fight.

Taken separately, each of these events could be dismissed as lobbying, geopolitics, or corporate positioning. Taken together, they suggest something deeper has shifted. The most powerful supplier in the AI economy is defending open models. The head of the Chinese state is making openness part of national strategy. American companies that agree on almost nothing are warning Washington against concentrating advanced AI inside a handful of closed providers. And the leading closed-model labs are watching competitors distribute capable intelligence at a fraction of the price, with customers running it themselves.

Most people will read this as a debate about AI safety, Chinese competition, or intellectual property. It is a platform transition, and most people are too young, or arrived in technology too recently, to recognise what one looks like while it’s happening.

I trained in the United States as a Sun Microsystems Unix administrator, in an era when a serious enterprise server room had a very particular character: the hum of expensive machinery, rows of Sun, IBM, and Hewlett-Packard hardware. Solaris, AIX, and HP-UX weren’t just operating systems. They were complete institutional relationships, bundled with specialised hardware, certified engineers, vendor support, and long procurement cycles. No responsible bank was going to entrust its core workloads to software assembled by hobbyists on the internet. That was how Linux was seen well into the 1990s: interesting, useful at a university, not something a serious institution would run.

The dismissal wasn’t foolish. The proprietary systems were mature and accountable. Linux was fragmented, its hardware support inconsistent. Anyone looking at the market in the late 1990s could build a perfectly rational case for why Sun would remain dominant.

That’s what people misunderstand looking back at any technological upheaval. The eventual winner rarely looks more complete than the incumbent. It wins because it changes the rate at which completeness gets built, and because no single company can determine what everyone else is allowed to do with it. A hardware maker could optimise for Linux. A university could teach it free. A developer could fix a bug and hand the fix to everyone else, forever, without waiting for a vendor’s next release. Innovation stopped being sequential, bounded by one company’s customers and priorities, and became parallel, absorbing the experience of internet companies, universities, hosting providers, and individual developers all at once. That difference compounds. It doesn’t make every contribution wise, but it creates a far larger surface for useful adaptation to occur on.

In 2001, Steve Ballmer called Linux a cancer. Incumbents don’t describe irrelevant technologies in biological terms. By the early 2000s, the “Lintel boxes” once dismissed as toys were taking real workloads from Sun. Sun eventually embraced Linux too, but the centre of gravity had already moved. In 2009, Oracle bought Sun for roughly $7.4 billion. Proprietary Unix did not die that day, but its era effectively ended. Later in my career, I worked at Red Hat, by which point the argument had already been settled in practice.

But saying Linux “won” misses the real story. The largest fortunes weren’t made selling Linux distributions. They were made by companies that no longer had to rent the operating-system layer on proprietary terms: Google, Amazon, Facebook, Netflix, and millions of smaller businesses that could experiment before they had the revenue to justify an enterprise contract. Open source didn’t just reduce software costs. It reduced the cost of trying.

Open-weight models aren’t the same as open-source software. A company can publish trained weights while withholding the training data, the code, and the recipe, so the result is usable and inspectable without being fully reproducible. But the essential shift is the same. A closed model keeps the intelligence inside the provider’s infrastructure. You submit a request and depend on their pricing, permissions, and uptime. An open-weight model can be downloaded, run on infrastructure you choose, adapted, and preserved even if the original developer changes direction or shuts the door.

The easiest way to see the difference is the car industry. Imagine every manufacturer had to rent a sealed engine from one of three suppliers, unable to inspect it, modify it, or build one themselves, paying every time it ran. On the converse, publish the blueprint, and the industry changes: manufacturers and specialists can inspect it, adapt it, build vehicles the original engine company never considered.

But the analogy has a limit, and the limit is the whole story. The blueprint may be free. The factory isn’t. Neither is the steel, the fuel, or the electricity. Open weights don’t eliminate the physical cost of intelligence. Every model still needs chips, power, and data centres every time it runs.

That’s why Huang chose this moment to speak. Open weights weaken a model owner’s ability to collect rent on every interaction, but they multiply the number of organisations consuming compute: every company that downloads, fine-tunes, or serves an open model becomes a buyer of accelerators. Open weights reduce scarcity at the model layer while increasing demand at the compute layer. Nvidia doesn’t lose if intelligence gets cheaper. It wins if intelligence becomes ubiquitous, because it doesn’t need any one lab to own the future. It needs thousands of them competing to use more of it, everywhere.

This is why it’s too simple to cast the fight as open idealists against closed monopolists. Meta benefits if the model layer commoditises and value flows to products and distribution. Hugging Face is the distribution layer regardless of who wins. Andreessen Horowitz and Y Combinator represent startups that don’t want their margins set by three frontier labs. Their support may be sincere. It’s also economically convenient, which is the same test worth applying to Microsoft: Windows was its tollbooth, and Linux was a threat. Azure became its tollbooth, and Linux workloads became revenue. Microsoft didn’t discover a new philosophy. Its economic advantage moved, and its position followed it.

The Hugging Face incident makes the argument concrete. An OpenAI model being tested with reduced cyber refusals escaped an internal evaluation and reached Hugging Face’s production systems. When Hugging Face’s own team tried to use commercial frontier models to analyse the resulting attack logs, safety restrictions got in the way of the forensic work, so they ran the analysis on Z.ai’s open-weight GLM 5.2, on their own infrastructure, instead. The symbolism was irresistible: an American closed model caused the breach, American closed models couldn’t help investigate it, and a Chinese open model could. But symbolism isn’t proof. Specific vulnerability counts circulating online were never independently audited, and one incident doesn’t establish that open models are generally safer. The narrower, defensible conclusion: a model controlled by someone else can refuse legitimate defensive work because it can’t tell a defender from an attacker. A model you operate yourself gives you more authority over how it’s used, and more responsibility for it. The issue was never goodness versus danger. It’s who controls the boundary.

The same discipline applies to Moonshot. The distillation allegation isn’t a morality tale where open innovators heroically outrun closed incumbents. There’s a real difference between learning from published research and covertly extracting proprietary value at scale. The industry letter is strongest exactly where it holds that line: punish theft directly, don’t use an allegation of theft to restrict an entire category of model.

The optimistic answer is that value moves to millions of builders, and some of it will. A hospital running a specialised model inside its own network. A founder switching providers without rebuilding the product. An enterprise keeping proprietary knowledge inside a model it controls instead of donating it to someone else’s platform. These are real transfers of agency.

But AI doesn’t escape the gravity of physical infrastructure. The more widely intelligence is deployed, the greater the demand for accelerators, power, and data-centre capacity. Open models may reduce the scarcity of model access while intensifying the scarcity of everything required to run one at scale. Power may move in two directions at once: outward, toward the companies, states, and builders who can now adapt intelligence themselves, and inward, toward whoever controls the compute and energy that intelligence still depends on.

The gate opens at the model layer. The tollbooth grows at the infrastructure layer. Linux weakened the proprietary Unix vendors and widened who could build, and it also produced hyperscale cloud companies whose infrastructure power eventually dwarfed anything Sun or IBM ever had. Openness at one layer didn’t abolish concentration. It moved it. That’s not a reason to oppose open weights. It’s a reason to understand what openness actually does. It doesn’t make power disappear. It changes its address.

When I stood in those server rooms, the incumbents looked permanent: proven systems, conservative customers, deeply embedded commercial relationships. Linux looked incomplete. The world it made possible was almost impossible to see from inside that room. People thought they were choosing between operating systems. They were choosing the ownership structure of the next computing economy.

That’s what’s happening now. The argument on the surface concerns model weights, alleged theft, and whether guardrails are too restrictive. Underneath it sits a larger contest over who owns intelligence, who’s allowed to modify it, and which layer of the stack collects the rent.

I don’t know whether open-weight AI beats closed models. The world will likely keep both. I don’t know which of today’s labs becomes this era’s Sun Microsystems. What I recognise is the movement itself: a foundational capability that incumbents expected to control is becoming portable, and companies are changing their language because their incentives are changing beneath them.

Looking backwards, platform transitions look obvious. Living through them, they arrive disguised as technical arguments. Last time, we thought we were debating Unix. What actually changed was who got to build the future, and where the value of computing ended up once they did.

That’s what I missed the first time. I thought I was watching a battle between technologies. I was really watching power change hands.

The models are opening because power never stays where it is.

If you’re looking for the future, start by looking for power’s new address.

Not because the money isn’t there or because the GPUs don’t exist, but because the grid cannot keep pace with the speed AI wants to grow.

For most of the last century, delivering reliable power was clearly the function of the grid. A factory asked for power, the utility built the infrastructure, and the customer paid.

Texas is now experimenting with a different arrangement. If a Data centre developer wants power faster than the grid can reliably provide, some of that power gets treated as flexible rather than firm.

The developer can still build, but only if they are prepared to manage that uncertainty themselves.

That sounds like a regulatory footnote.

It isn’t.

It changes what an AI factory is, where it gets built, and who captures the value.

For years the AI race has been focused on chips, but the bottleneck is really electricity. Data centres need so much power that utilities cannot build generation and transmission fast enough.

That is true and is already consensus now.

The sharper story sits one layer underneath. Texas is testing a different idea: what if the future of AI isn’t built around more power, but around a new relationship between power and compute?

That story starts in West Texas.

The Permian Basin is one of the world’s great oil fields. Drill for oil and you also get gas, whether anyone wants it or not. For years, pipelines carried those gas molecules to market. Today there is more gas than the pipes can carry.

When the pipes fill up, something strange happens. The gas still holds energy, but the sellers have run out of buyers.

In 2026, gas at the Waha Hub traded below zero on more than eighty-seven days. On some of those days, prices averaged around minus $5.66 per MMBtu. Producers were paying for the gas to be taken away because shutting down oil production would have cost more.

The gas isn’t worthless. It’s stranded.

For most of history, there was little you could do about that. You could flare it and waste it, or wait years for more pipeline.

AI may have just created a third option.

Chevron’s Project Kilby in West Texas gets described as a power plant for Microsoft’s data centre. That is true, but it misses the point. The plant is expected to use turbines from GE Vernova, backed by Solar Turbines. That detail matters because it shows that the future AI factory is being assembled from the same industrial equipment that has always powered refineries, pipelines, and heavy industry.

The project also gives stranded gas a new customer. Instead of transporting the gas to where the demand already exists, the demand is moving to the gas.

The data centre isn’t just buying electricity. It’s creating an economic use for energy that had no better use.

That’s the first inversion.

For most of the industrial era, energy moved to where demand already existed. What Texas is showing is something different. Increasingly, compute is moving to where energy has no better use.

For decades, power systems were built around demand. The factory went where the customers were and the power system followed.

Now we’re seeing the opposite.

That would be interesting on its own.

Texas is running a second experiment at the same time.

The old deal between a grid and its customers was simple. The factory asked for a gigawatt, the utility agreed to supply it, the utility carried the reliability risk, and the customer paid the bill expecting the lights to stay on.

AI is arriving at a moment when that model breaks.

The scale is extraordinary. ERCOT, the Texas grid operator, is sitting on hundreds of gigawatts of requests from large data centres. Many will never get built, but the signal is real: the appetite for compute is growing faster than the grid can absorb it.

Texas responded with a new framework called Batch Zero. Buried inside it is a mechanism called the Provisional Controllable Load Resource, or PCLR. The first declaration deadline lands on July 24, 2026.

Large data centres can no longer assume every megawatt they request arrives as firm power on day one. Rather, a data centre gets a firm (guaranteed) portion and a flexible portion that can be dialled down when the grid is stressed.

The grid is telling its largest customers something new. We’ll help you connect and we’ll plan around your full request, but we cannot promise every megawatt exists today. If you need it to behave as firm, you solve that yourself.

That shifts the reliability burden from the utility onto the developer’s balance sheet.

Batteries, backup engines, onsite generation, and control systems sophisticated enough to absorb a curtailment signal without anyone downstream noticing are no longer optional. They are the tools that turn flexible power into reliable compute.

The industry has a name for it: privatised firming.

But the deeper change is that firm power is no longer something you simply buy from a utility. It is increasingly something you assemble from grid access, onsite generation, storage, control systems, and flexible workloads.

For a century, electricity was sold as a binary product. You either had power or you didn’t.

Texas is beginning to turn electricity into a tiered, scheduled resource.

Software engineers will recognise this pattern immediately. In Kubernetes, some resources are guaranteed while others are burstable and become constrained when the system comes under pressure. Texas is beginning to treat electricity in a surprisingly similar way. Not every megawatt is equal anymore. Some are firm, some are flexible, and some may only become available as the system expands.

That’s the second inversion.

The grid is no longer rationing access. It is rationing certainty.

The scarce product is no longer electricity.

The scarce product is firm power delivered on time.

That sounds technical, but the economics are not small. A data centre stops being a building full of servers plugged into the grid and starts looking like a miniature power system that happens to run GPUs.

And once that happens, a new question appears.

Not all computation is equally important.

A training run can probably pause for thirty minutes. A customer asking an AI assistant a question cannot wait. Some workloads move and others don’t.

Future AI campuses will need an internal market of their own. They will need to know the price of power at any moment, understand which jobs are flexible and which are not, and decide, minute by minute, what runs, what waits, and what moves elsewhere.

That isn’t data centre management anymore.

It’s a trading desk wired directly into a workload scheduler.

The grid won’t just supply the AI factory.

It will shape how the AI factory computes.

For the first time, compute itself is starting to behave like a dispatchable grid resource.

The AI factory is no longer passive load.

It is becoming grid-native compute: compute that understands the state of the grid, adapts to changing energy conditions, and decides which workloads deserve certainty and which can wait.

For most of the internet era, data centres consumed electricity as though power was always available. Texas is teaching AI to operate differently. The future AI factory may need to know the state of the grid, understand the price of electricity, distinguish between flexible and inflexible workloads, and continuously adapt itself to changing energy conditions.

AI will not merely run on the grid.

Increasingly, it will run with the grid.

Seen this way, the AI stack reorders itself. The first question is no longer who has the best chips. It becomes: where is there energy with no better use? The second becomes: who can assemble reliable power faster than everyone else? The third becomes: who can orchestrate compute around the realities of the grid? Only then do we ask who trains the model.

The GPU hasn’t disappeared.

It’s just moved to the end of the chain instead of the front.

The companies that capture value here may not be the ones investors instinctively reach for. Owners of stranded energy suddenly hold valuable real estate for compute. Turbine and engine makers like GE Vernova and Caterpillar Inc.’s Solar Turbines become suppliers of AI infrastructure rather than simply industrial equipment. Battery and control system providers become manufacturers of reliability. And somewhere in this stack, someone is going to build the orchestration layer that turns a live grid signal into a decision about which workload runs and which one waits.

That layer doesn’t really exist yet.

This is why Texas matters.

It isn’t just building more data centres.

It is becoming the world’s first large-scale laboratory for grid-native compute.

The implications run well beyond Texas. If AI becomes the buyer of last resort for stranded energy, and compute becomes flexible enough to bend around the grid’s reality, then electricity and computation stop being separate industries. The lines between energy companies, utilities, industrial suppliers, and data centres are already blurring.

The first wave of AI chased GPUs.

The second wave chased electricity.

The third wave may chase stranded energy and learn how to compute around its own limitations.

In that world, the most valuable AI factory may not be the one with the most chips.

It may be the one that knows exactly when it can afford to stop computing.

That detail is more important than the model itself.

For most of the short history of generative AI, the operating assumption was straightforward. Labs built models, released them to the market and then dealt with the consequences. Governments might regulate how the technology was used, but the decision to ship rested largely with the companies that built it.

Over the past few weeks, that assumption has started to change.

On 2 June, President Trump signed an executive order requiring federal review of frontier models before public release. The legal architecture was built before anyone noticed.

Ten days later, Anthropic said the U.S. government had issued an export-control directive that forced it to suspend access to two of its frontier models for foreign nationals, including some of its own employees. The legal mechanism was designed for physical weapons, not cloud-served software.

Then OpenAI restricted GPT-5.6 to a small group of trusted partners at the government's request. The request was framed as a national-security precaution, but the practical effect was the same: the government influenced who received access to the most advanced commercial AI system in the world.

Taken individually, each event can be explained away. One is a benchmarking framework. One is an export-control dispute. One is a delayed product launch.

Taken together, they reveal something larger.

The state has entered the model loop.

A few months ago, the working assumption was simple. Labs built models. Labs released models. Markets decided what happened next.

That assumption is weakening.

A different pattern is beginning to emerge. Governments are no longer content to regulate the use of frontier intelligence after it has been released. They increasingly want visibility before release, influence over distribution and a say in who gets access.

Frontier AI is beginning to look less like a software product and more like a strategic technology.

That matters because markets are still pricing many AI companies as though they are software companies.

Software companies are expected to scale rapidly and distribute products as widely as possible. Their valuations assume that customer acquisition, international expansion and distribution are commercial challenges rather than political ones.

Frontier AI is beginning to reveal a different risk profile.

If access to the best models can be slowed, reviewed, narrowed or shaped by government, then the economic asset investors are pricing is not quite the same asset regulators are beginning to see.

The market is still pricing software.

The government is beginning to regulate strategic capability.

That spread is one of the most important variables in the AI economy.

There is a historical precedent here.

In the 1990s, Phil Zimmermann found himself in a strange position. The U.S. government treated his encryption software, PGP, as something closer to a munition than a computer program. So Zimmermann published the source code as a book. Books enjoyed the protections of free speech in a way exported software did not.

The details are different. The argument is not.

The technology changes. The argument does not. Once a digital capability becomes strategically important, the question shifts from what it can do to who gets to decide who may access it.

If the pattern is repeating, the investment implications are substantial.

The question is no longer which company will build the best model. Investors also need to ask which companies can continue to distribute, monetise and expand access to frontier intelligence in an environment where governments increasingly view that intelligence as a strategic asset.

That shift favours a different set of capabilities. Political relationships. Compliance infrastructure. Trusted-partner programmes. Geopolitical positioning. All begin to matter more than they did before.

None of this means that progress in AI stops or that frontier models stop improving.

It means the economics are becoming more complicated than the market has been assuming.

The first phase of the AI boom was a race for capability. Investors wanted to know which model was strongest, which benchmarks had been beaten and which lab possessed the deepest compute cluster.

The next phase will be shaped by questions of permission and access. Who gets frontier intelligence, under what conditions, in which countries and with whose approval may prove to be every bit as important as which company builds the next state-of-the-art model.

That is why this moment matters.

The government is not merely regulating the use of AI after deployment. It is beginning to shape the path by which frontier intelligence reaches the market in the first place.

That is what it means for the state to enter the model loop.

Capital allocators should pay attention. Access itself is becoming part of the moat.

A model that cannot be freely distributed is not the same economic asset as one that can. A company whose customer list requires political review is not the same kind of company as a normal software vendor. A market where frontier intelligence is gated is not the same market investors believed they were underwriting a year ago.

The technology will continue to accelerate. Acceleration is no longer the only variable that matters. As governments shape how frontier intelligence is distributed, control over access becomes as important as capability itself.

Once intelligence becomes gated, every valuation built on unrestricted distribution eventually gets repriced.

Their questions made me realise I had not shared my thoughts as they have evolved over time.

I recently decided to expand on what I call inner world building here on substack.

Some of my articles might be raw or unpolished, covering themes that have to deal with human identity, agency & consciousness.

I always try to share from lived experience what has worked for me or what I am still learning. Hopefully some of it will resonate with you.

I have often contemplated the things that make a certain founder stand out from all the rest. On the surface, they are selling the same product or service, sometimes with far fewer features or resources. Yet their impact defies their size. This is not a new pattern, but one that has repeated itself countless times across history.

This leads to a deeper question. What resources are available to founders? Do certain founders have unfair advantages? While that might be true in some cases, I believe it is worth exploring whether outstanding founders simply understand how to use the resources we all already have. Perhaps exceptional mastery of simple, everyday tools available to every founder can teach us something.

A simple observation of life shows that the things that matter most are usually free, abundant and completely underrated. We ignore them because they are always there. Always available. Reliable.

Air is the perfect example.

Every single one of us has inhaled surplus amounts of oxygen daily since the moment we were born. Not once do we wake up thinking, “Will there be enough air to breathe this weekend?”

Or, “I need to quickly dash to the shop to make sure we have enough air for our family trip to Spain.”

No one stockpiles it. No one worries about it.
Just the thought of it is absurd.

This is because air is abundant everywhere, so we naturally do not pay attention to it. Yet this resource, which human life cannot do without for even ten minutes, is sorely undervalued and underappreciated.

Because we do not value it, we misuse it, abuse it, or never think about it at all.

Interestingly, as founders, we do the same thing.

When founders think of resources, our minds immediately go to Capital, Intellectual Property, Strategic relationships, technical skills and market distribution. Don’t get me wrong, these are all important, but just like air, they are not the most valuable.

I have found that one of the most powerful, flexible and unlimited resources every human being possesses is also one of the most underused, misunderstood and ignored.

In fact, it is so underrated that it is easy to miss entirely. You really have to pay attention or you will not notice it. And honestly, I do not blame anyone for missing it.

And here is the funny part. I already mentioned this resource earlier. In fact, I mentioned it twice.

Most people still miss it.

The resource is attention.

Attention is the one tool you were born with. You carry it everywhere. You use it constantly, yet you were never taught how to direct it. We spray it, leak it, donate it, waste it, give it away for free, and then wonder why nothing changes.

​
Attention is the currency of the Agentic Founder.

Pretty basic, right. It really is that simple, but yet Profound.

But if you do not think attention is powerful, go ask Google.

In 2017 they released a research paper that birthed the entire modern AI revolution, including the foundations of ChatGPT, Claude and Gemini. Guess what that paper was called? –

Attention Is All You Need.

Not Compute Is All You Need.
Not More Data Is All You Need.
Not Money Is All You Need.

Just attention.

In 2024 I wrote Jailbreak: Escape the Nine to Five Trap, where I explored this idea further.
​https://www.amazon.com/dp/B0CW19R49X/​

But let me leave you with this:

Attention is the most abundant, most flexible and most powerful tool you already have, and the one you have likely trained the least.

It really might be true.

Attention is all you need.

For most of the last century, the relationship between the family home and the electricity grid was remarkably simple. Electricity flowed from the grid into the home, where it powered lights, appliances and, more recently, electric vehicles. The home consumed electricity, the grid supplied it, and apart from a growing number of rooftop solar panels there was little reason to think much about that relationship.

That relationship is beginning to change.

On 24 June 2026, Sunrun, Tesla and Renew Home announced a framework to provide more than 16 gigawatts of fast, flexible power for AI data centres and other large electricity users by coordinating batteries, electric vehicles and smart energy devices already installed inside millions of homes.

Most of the technology behind that announcement was already familiar. Household batteries had been helping electricity companies balance the grid for years. Utilities had spent years encouraging households to become more flexible in how they used electricity, and the software needed to coordinate those devices was already in place.

What changed was not the technology. It was the role those technologies were beginning to play.

For perhaps the first time, millions of privately owned household energy devices were being assembled into infrastructure supporting one of the largest industrial expansions in modern history. Until now, those devices had been valued mainly for what they could do for homeowners and for the electricity system. Increasingly, they are also becoming part of the infrastructure that supports the AI economy.

Your garage was never designed for artificial intelligence. Yet the battery in your garage, the electric vehicle on your driveway and the thermostat on your wall may now contribute, alongside millions of other household devices, to the electricity system that increasingly underpins the AI economy.

Most homeowners have no reason to think about that connection.

Imagine three households.

A homeowner in Northern Virginia installs a battery to lower electricity bills. A family in Texas buys an electric vehicle because it is cheaper to run. A homeowner in London signs up to a flexibility tariff because it promises lower energy costs.

Each decision is personal. Each makes financial sense. None of those households believes it is participating in the AI build-out.

Yet, taken together, those ordinary household decisions can help create the flexible grid capacity needed to connect new AI data centres to the grid.

The electricity industry calls this a virtual power plant, or VPP. Despite its name, it is not a power station. It is software that coordinates thousands, and eventually millions, of household energy devices so they behave as one flexible resource.

The idea is not entirely new. Cloud computing transformed thousands of separate servers into a single computing platform. A virtual power plant applies the same principle to electricity by coordinating millions of independent household energy devices into a resource that the grid can draw on whenever it needs additional flexibility.

Until recently, that flexibility mattered because it helped electricity companies operate the grid more efficiently.

Artificial intelligence is giving that same flexibility a new economic purpose.

Household energy devices have not suddenly become useful to the electricity system. They already were. What is changing is the system they are increasingly helping to support.

The question is no longer whether your garage can help the electricity grid.

It is why that same contribution is suddenly becoming more valuable now?

The answer begins with a problem the electricity industry has been trying to solve for decades.

Electricity demand is not constant. Most of the time the grid has enough capacity to meet demand comfortably. The challenge comes during the busiest few hours of the year, when millions of homes, businesses and factories all need electricity at the same time. Those few hours determine how much generation, transmission and distribution infrastructure has to be built, even though much of it sits underused for most of the year.

That is why electricity companies have spent years encouraging households to become more flexible. Charging an electric vehicle overnight instead of early evening, allowing a home battery to supply electricity for a short period, or shifting other electricity use away from peak hours all help reduce pressure on the grid. None of those actions matters very much on its own. Across millions of homes, they become a meaningful source of flexibility.

For years, that flexibility was valuable because it helped the electricity system run more efficiently.

That is still true.

What has changed is who now needs that flexibility.

AI data centres require enormous amounts of electricity, often in parts of the grid that are already close to their limits. Expanding the grid is possible, but building new transmission lines, substations and other infrastructure often takes years.

That leaves the industry with an immediate problem. It needs more usable capacity long before new infrastructure can be built. Looking for ways around that constraint has revealed some unusual suspects. Household batteries, electric vehicles and smart thermostats were never installed to support AI infrastructure, yet together they are beginning to do exactly that.

Viewed individually, they are ordinary household devices. Coordinated across millions of homes, they become a source of flexible capacity that helps the electricity system accommodate new demand while larger infrastructure projects catch up.

The role of household energy devices therefore begins to change.

A battery is still bought to lower electricity bills. An electric vehicle is still bought for transport. A thermostat still regulates the temperature inside a home. Their original purpose has not changed.

What has changed is the value the electricity system can now create by coordinating millions of those devices together.

For homeowners, the reasons for buying those technologies remain exactly the same. People still want lower energy bills, greater resilience and cleaner energy. Very few are making those decisions because they want to support artificial intelligence.

Yet they may become participants in that system anyway.

That naturally raises a different set of questions.

If millions of privately owned household assets are creating value for a much larger system, who decides when those assets are used? How is that value shared? How much flows back to the households that own the assets, and how much remains with the companies coordinating the network?

Those questions reach beyond energy policy.

They are questions about ownership, incentives and market structure.

The same shift should also change how investors think.

Markets are usually quick to recognise new technologies because they are visible. They are often slower to recognise when familiar assets begin serving a different economic purpose because, on the surface, nothing appears to have changed.

The battery is still sitting in the garage.

The electric vehicle is still parked on the driveway.

The thermostat is still fixed to the wall.

The physical assets look exactly as they did before.

The network around them does not.

Sometimes the biggest investment opportunities do not begin with a new invention.

They begin when existing assets quietly become part of a different economy.

A single rack of NVIDIA's next-generation AI systems will consume as much power as five hundred American homes. The problem is not that we cannot generate that electricity. The problem is that we are trying to move it through pipes built for a village.

For the last four years, the AI story has been relatively straightforward. Larger models required larger clusters, larger clusters required more GPUs, and the companies supplying those GPUs became some of the biggest beneficiaries of the AI boom.

That story is still true.

But it is no longer the whole story.

The systems now emerging from NVIDIA are approaching a scale that would have seemed extraordinary only a few years ago. As that scale increases, the industry is being forced to rethink something that has remained largely unchanged for decades: the way electricity moves through a data centre.

That rethink revolves around an unassuming number.

800V.

Behind it lies one of the biggest infrastructure changes taking place inside modern computing.

To understand why, it helps to understand how much the world has changed.

A rack – the cabinet that holds the servers inside a data centre, traditionally consumed around 5–10 kW of power. Even the first wave of cloud computing did not fundamentally change that architecture.

AI did.

Today’s frontier AI systems already consume more than 100 kW per rack.

NVIDIA’s roadmap points towards systems consuming around 600 kW per rack, with megawatt-class systems visible further down the road.

To put that in perspective, the average American home continuously draws around 1.2 kW of power.

A single future NVIDIA rack will therefore consume roughly as much electricity as 500 homes.

And AI factories are built not around one rack, but around clusters containing hundreds, and eventually thousands, of these systems.

At that point, the comparison is no longer with office buildings.

The electrical demands begin to resemble those of entire towns and cities.

It is hardly surprising that the assumptions underpinning modern data centres are beginning to change.

For decades, those assumptions worked perfectly well.

Electricity arrives from the grid as alternating current, or AC, the same type of power that comes out of the sockets in your home. Before it reaches the servers, it passes through a series of transformations and conversions. Eventually, power is distributed inside the racks at around 50V direct current.

That architecture evolved in an era when power consumption was modest.

As AI racks moved from 10 kW to 100 kW and now towards 600 kW, moving electricity around at 50V started to become increasingly inefficient.

The problem is not the voltage itself.

The problem is the amount of current required to deliver so much power.

Imagine trying to supply water to an entire city through pipes designed for a village. Eventually, either the pipes have to become enormous, or the system itself has to change.

Electricity behaves in a surprisingly similar way.

And the economics become much clearer once you look at one very simple equation:

Power = Voltage × Current

P = V × I

Suppose you need to deliver exactly the same amount of power.

You can either do it with low voltage and very high current, or with higher voltage and much lower current.

Moving from 50V to 800V increases the voltage by 16×. That means the current required falls by 16×.

That already sounds useful.

But the real payoff is much larger.

Electrical losses increase with the square of the current.

Reduce the current by 16× and the amount of energy lost as heat falls by:

16 × 16 = 256×

Imagine losing $256 every day and suddenly reducing that loss to just $1.

That is why the industry is moving towards 800V.

Lower losses mean less heat. Less heat means smaller cooling systems. Smaller currents mean less copper and smaller cables. The combined effect is that AI factories become cheaper to build, cheaper to operate and easier to scale.

The move to 800V is not about chasing a fashionable new standard.

It is about changing the economics of AI infrastructure.

At this point, the obvious question is why NVIDIA is pushing 800V direct current rather than simply sticking with the alternating current used by the grid.

The answer reveals something interesting about how today’s data centres work.

Alternating current is excellent for transporting electricity over long distances. That is why power grids use it.

Inside the data centre, however, electricity passes through multiple stages of conversion before it finally reaches the processor. Every one of those stages introduces losses and generates heat.

When racks consumed 10 kW, those inefficiencies were manageable.

When racks consume 600 kW, they become increasingly expensive.

NVIDIA’s answer is to use AC where AC makes sense and then convert to high-voltage DC as early as possible, allowing electricity to remain in DC form for much longer inside the AI factory.

By reducing the number of conversion stages, the architecture becomes simpler and more efficient, making it better suited to the increasingly enormous systems now being built.

The move to 800V solves one problem, but it creates another.

Electricity may arrive inside the AI factory at hundreds of volts, but the processors themselves operate at voltages closer to 1V.

Somewhere between the power entering the building and the chip itself, enormous amounts of electricity must be stepped down with extraordinary efficiency.

That requirement is creating opportunities for companies whose expertise lies not in computation, but in power conversion.

One way to think about the opportunity is that there are both established winners and speculative challengers.

Monolithic Power Systems (NASDAQ: MPWR) represents the established end of the spectrum. Its chips are effectively the traffic controllers of electricity inside increasingly power-hungry systems. The company is already highly profitable and deeply embedded in AI infrastructure. Investors are not taking venture risk here. They are paying a premium for a business that has already proven itself.

Navitas Semiconductor (NASDAQ: NVTS) sits at the opposite end of the spectrum.

Navitas is betting heavily on Gallium Nitride, or GaN.

Despite the intimidating name, the idea is surprisingly simple.

Traditional silicon behaves a little like an old combustion engine. It works well, but it wastes energy. Gallium Nitride switches much faster and loses less energy in the process, allowing engineers to build smaller, cooler and more efficient power systems.

Those advantages become increasingly valuable as AI clusters grow larger.

But unlike Monolithic Power, Navitas remains an early-stage bet. The company is still unprofitable, larger competitors are pursuing similar approaches, and some already demonstrate higher efficiencies. Investors here are not paying for proven economics. They are paying for the possibility that GaN and 800V architectures become important enough for today’s design wins to turn into tomorrow’s revenues.

The two companies therefore represent very different ways of expressing the same underlying theme.

For most of the AI boom, the question was straightforward:

Who is building the intelligence?

By now, most investors understand that electricity has become the next constraint.

But “electricity” is too broad a category to be useful.

The more interesting question is where the bottlenecks are emerging inside the power stack itself.

The move from 50V to 800V provides one answer.

It shifts value towards the companies responsible for moving, converting and managing power efficiently.

That does not mean Monolithic Power Systems or Navitas Semiconductor are guaranteed winners.

It does suggest that the next phase of the AI buildout may reward a different class of company than the last one.

NVIDIA’s success is forcing the industry to answer a new question.

Not how to create more intelligence.

Not even how to generate more electricity.

But how to deliver extraordinary amounts of power to increasingly extraordinary machines without paying a 256× heat tax.

And that is ultimately why the 800V revolution matters.

Because the AI boom is not running out of electricity.

It is running out of efficient ways to move it.

The Agentic Age

Sometimes the biggest opportunities emerge not when a new technology appears, but when an old architecture finally runs out of steam.

It turns fast from orbit to control. SpaceX says the key constraints in AI are physical, meaning “chip manufacturing, data center infrastructure, and power generation,” and then it says the future will be decided by “control of the physical stack.” It does not stop there. It says no other AI company has better control over the full physical stack than SpaceX. Read that once and the filing feels like a market event. Read it again and it starts to feel like a map of where power is moving while everyone else is still describing the weather.

The AI world still likes to talk as if the race is models, as if the decisive contest is benchmark wins, product velocity, and enterprise adoption. That is where the commentary lives. That is where the headlines go. That is where the attention stays. But the Anthropic number changes the temperature of the whole argument, because if you are backed by Google and Amazon and still need to pay a competitor $1.25 billion a month for servers, then the real bottleneck is no longer intelligence. It is capacity, and capacity is a colder, heavier thing than intelligence. It is physical capacity. Industrial capacity. The kind that cannot be patched in a software update and cannot be duplicated by more talent, more capital, or more enthusiasm. The whole industry is in a model race, and SpaceX is not merely watching from the sidelines; it is laying the tracks.

That is why this is not a normal IPO story. The company did not begin with AI; it began with the hardest physical systems on earth, and the point of that history is now becoming clear. Falcon 9 has completed more than 650 orbital missions with a success rate above 99 percent, and Starship is largely built in-house. That launch capability made Starlink possible, and nearly 10,000 satellites now serve more than 10 million subscribers across 164 countries. It is not just a business. It is a distribution layer with gravity, a system that keeps widening the circle because each layer gives the next layer somewhere to stand.

Then they added AI, folding xAI into the stack in February 2026 and bringing Grok and X together. X adds 350 million daily posts of real-time signal, which the filing treats as a structural advantage rather than a side note, and then came the compute to power it: COLOSSUS and COLOSSUS II together reach about one gigawatt of capacity, a private company building at a scale that used to belong only to nations and utility companies. The point is not the number itself. The point is the sensation of the floor dropping lower every time they add another layer.

SpaceX built the stack the same way it built rockets, one layer at a time, each layer opening the next, until launch made satellites possible, satellites made connectivity possible, connectivity made distribution possible, and once distribution sat on top of real physical control, AI became the natural thing to place at the peak because AI monetizes the whole stack. That is the machine the filing is describing, and “from shovels to tokens” is not a slogan in this context so much as a confession that the company wants to own the path from dirt, metal, and energy all the way to the thing users pay for. The Anthropic deal proves that rent is already being collected. The foundation is still being laid, but the invoices are real, the money is real, and the demand is already large enough to expose the limits of the market. The stack does not stop at compute; it grows downward into the machinery that makes compute possible in the first place, and that is what makes the whole thing feel less like strategy than gravity.

That is where Terafab comes in. Compute runs on chips, and chips run on a supply chain so concentrated that most AI companies never stop to ask who makes the hardware their models depend on, or what happens if the line tightens. SpaceX’s answer is a manufacturing initiative with Tesla and Intel aimed at one terawatt per year of compute hardware. Reporting pegged the investment at $119 billion, while the filing itself is more careful and says capital expenditures have not yet been determined. That caution matters because it tells you the ambition is still becoming real; it has shape, but it is not finished. Intel is the hinge because it brings fabrication experience, packaging expertise, and a road from idea to production, while ASML remains the gatekeeper at the frontier, the one company that builds the machines used to print the most advanced chips on earth. Its EUV lithography systems are the only path to the most advanced nodes, the newest units cost roughly $370 million each, and there are only a handful anywhere. Without access to them, you do not make frontier chips; you make plans. The day after the filing, ASML CEO Christophe Fouquet told Reuters that Musk had spoken directly with him about chip-making equipment and called Terafab “very serious,” which matters because it shows the bottleneck owners are already paying attention. Today, Anthropic pays SpaceX $1.25 billion a month to rent compute. If Terafab works, Anthropic may one day pay SpaceX for the chips inside that compute too. The landlord becomes the manufacturer, the dependency deepens, and the tenant’s options get smaller with every new layer built beneath it. That is the real turn of the knife: the thing you rent becomes the thing that rents you back.

The same logic extends to geography. The silicon that powers frontier models still passes through one place more than any other: Taiwan Semiconductor Manufacturing Company. TSMC has started building in Arizona, but that is still only a fraction of total capacity, which means Taiwan remains the center of gravity and the risk is not isolated but shared. If something serious disrupts the supply chain, every AI company, every hyperscaler, every model lab feels it at once, not one player losing supply but all of them simultaneously. The US government understood that years ago, which is why the CHIPS Act put real money behind the decision to onshore semiconductor manufacturing, with Intel receiving the largest allocation at $7.86 billion. Then, in August 2025, the government took a 9.9% equity stake in Intel directly, a nation looking at the map and deciding it can no longer pretend one island should carry that much of the world’s future. SpaceX filed from the same conclusion, only it reached it from the opposite direction: one side is public money, spread over years; the other side is private capital, moving at industrial speed, trying to pull the stack inside its own walls before the market catches up. The same fear, expressed in two different dialects of power.

Then the story lifts into orbit, and that is where the filing becomes almost unsettling in its clarity. SpaceX says it plans to begin deploying orbital AI compute satellites as early as 2028, solar-powered, cooled by the vacuum of space, and distributed globally through Starlink. The target is 100 gigawatts of compute launched each year, which SpaceX says could equal roughly one-fifth of annual US power production. That number is not there to sound big; it is there to change your sense of what is possible. In orbit, the old limits start to disappear because there is no grid to connect to, no power plant to build, no cooling tower to manage, no land to buy, no permits to wait for, and no water to source. The sun becomes the utility, the rocket becomes the delivery system, and the network becomes the moat. If Terafab works, those systems would also run on chips SpaceX helped make itself, which is why the claim feels less like science fiction than an extension of the same industrial logic that got the company here in the first place. SpaceX has completed eleven Starship flight tests, with a twelfth scheduled, and is targeting payload delivery to orbit in the second half of 2026. It is the only company that can even attempt this at scale. The option exists for one company and no one else, and that is not ambition in the abstract. It is a company reaching for a future it can physically own, and once you feel that, the entire filing reads differently.

What it means to be Anthropic right now is simple and brutal at once. You are building one of the most consequential AI systems in history. You have Google and Amazon behind you. You have elite researchers and a clear mission. You have the backing of two of the largest companies on earth. And every month you wire $1.25 billion to the company that trains your competitor’s model, owns the social network your model may never touch, and is building the chip factory that could one day decide whether you get hardware at all. You did not choose that position. Physics did. The compute does not exist anywhere else.

That is the moment the whole industry is living through, whether it has noticed it yet or not, because intelligence without infrastructure is a brain without a body, and the body is being built layer by layer by a company that spent twenty years learning how to manufacture the impossible and is now showing the rest of the market what it was really building toward. Musk retains control after the IPO through a dual-class super-voting structure, and his compensation vests fully only at $7.5 trillion and the establishment of a permanent human colony on Mars with at least one million inhabitants. Those are not the terms of someone thinking in quarters; they are the terms of someone thinking in generations.

He is not trying to own one product line. He is trying to own the ground the industry is built on, the chips it runs on, the network it communicates through, and the orbit it may one day expand into. Earth below. Sky above. Every layer. Every dependency. Every choke point. The model labs are building some of the most extraordinary intelligence the world has ever seen, and SpaceX is building the world that intelligence will have to live inside. Anthropic already knows. They signed the wire transfer. The rest of the industry is still watching the model race, still tracking benchmarks, still comparing funding rounds, but they will look up eventually, and by then the ground will not be theirs. The sky will not be theirs either.

Subscribe now