Who Made This a Race?

Anthropic is reportedly backing a consultancy in Europe. OpenAI already sends engineers into businesses to help them put AI to work. Both are building increasingly autonomous agents, while the wider industry expands into cybersecurity, drug discovery, specialized engineering tools, consumer devices and its own infrastructure.

Most of these moves make sense to me. What interests me is what they suggest about the business of selling increasingly intelligent models. How much more intelligence do customers actually need, and how much additional revenue can the labs extract from providing it?

I use AI extensively. I am a software engineer; give me a better model and I will probably find something interesting to do with it. But I am less certain that the bulk of the population has a comparable appetite.

Enough for what?

What does a baker need from a model that current ones cannot provide? Or a barista, a gardener, someone running a small shop?

There is plenty they could use help with: accounts, supplier comparisons, correspondence, estimates, schedules, promotion. Some of it could save considerable time. But how much additional intelligence does it require?

The remaining difficulty may be getting the right information into the system, connecting it to the software, and making it reliable enough that using it does not become another job. The baker needs the order to agree with the invoice. They have little reason to care that the same system can discuss advanced mathematics if they still have to supervise it through that.

For a substantial share of ordinary tasks, I suspect the models are already good enough, while the products and workflows remain inadequate. Better judgment and fewer mistakes would still matter. So would a lower price, a useful integration, or simply less attention required from the person supposedly being helped.

Those improvements may be commercially more important than another spectacular demonstration of reasoning. A customer has a finite task and a budget. Making the model more intelligent does not automatically enlarge either.

That makes the move into consulting quite understandable. Consultants can turn a model’s capabilities into something a particular business can use. Specialized tools, integrations and access to useful data could also give the labs advantages more durable than a temporary lead in model intelligence. Devices can provide distribution; custom chips can improve costs. There is considerably more to building these businesses than making the model smarter.

Local inference could further reduce what customers are willing to pay for access to a model. Apple already provides developers with access to on-device models, but the milestone I am interested in is practical substitution: models good enough to comfortably handle a meaningful share of what people currently send to a server. Once a task can be done adequately on hardware the customer already owns, the superior remote model needs to offer a reason to pay for it.

There could be considerably more AI use alongside less willingness to pay a premium for intelligence itself.

At the demanding end, the calculation is different. Scientific research, advanced engineering and robotics offer problems where additional capability could be enormously valuable. But that does not establish that everyone needs those capabilities, or that the people who benefit from them will become direct customers of the labs.

The baker may eventually use equipment or materials improved through AI-assisted research. That benefit does not tell us how much intelligence the bakery will purchase, or how much of the value the model developer can capture.

The fascination and the expense

Building artificial general intelligence is a fascinating engineering ambition. I would like to see how far it can go. But I can also see how easily people like me might pursue it with the intensity of an engineer’s fever dream, mistaking our interest in its existence for a society-wide demand for its use.

OpenAI’s announced Navier–Stokes result made me think about this. The company reported that its system had produced a proof addressing a longstanding question about whether the equations describing fluid motion can develop a mathematical singularity. The reported customer-equivalent inference cost was around $15 million (an estimate of what such a run would cost a customer, rather than a disclosed internal bill).

That is a remarkable scale of computation to apply to a mathematical problem. I would like to know what a strong researcher might accomplish with a salary and a much smaller inference budget, and when removing that researcher improves the result enough to justify the difference.

The $1 million Millennium Prize attached to the problem is hardly a measure of what the knowledge is worth, and OpenAI said it did not intend to claim it. Research can be worthwhile at a cost far beyond whatever prize someone has attached to the result.

Human judgment may eliminate enormous amounts of unproductive search. An autonomous system may explore possibilities a person would miss, or continue through work they could not complete in a reasonable time. Different problems will reward different combinations. I am interested in where those combinations become economical, and less persuaded that an expensive demonstration of autonomy tells us how much of it anyone will want to buy.

The money changes the question

This is where I place a considerable share of the blame on investor culture.

Coming out of the crypto and NFT fiasco, one might have expected more restraint, or at least some humility about technologies the people allocating capital did not necessarily understand. Instead, the next opportunity arrived and much of the same behavior followed it: spectacular certainty, spectacular rounds, and a willingness to declare entire professions and business models obsolete before establishing what would replace them.

The technology being more useful this time makes the consequences more substantial.

At the late stage, companies such as OpenAI and Anthropic offered something investors were eager to buy: considerable traction, recognizable products, extraordinary visibility, and a plausible claim to technological importance. An enormous check could secure exposure to the category everyone was discussing.

But an important company can still be a bad investment at a particular price, under particular terms, for a particular fund. Capital-intensive research with an uncertain commercial timetable may suit patient or strategic investors. It does not become appropriate for every venture fund because missing the round would be embarrassing to explain to its limited partners.

The unusual governance of the labs should have encouraged careful thought about those expectations too. OpenAI’s nonprofit control and earlier capped-profit structure, and Anthropic’s public-benefit purpose and trust arrangements, were attempts to accommodate obligations beyond conventional shareholder returns. Their existence did not make commercial investment impossible. They did make it reasonable to ask whether the capital being offered was compatible with the undertaking.

At the early stage, the enthusiasm acquired a different appearance: elite pedigree, a provocative launch, impressive-looking traction, and the suggestion that asking about durability meant failing to understand the moment.

Cluely is a particularly unpleasant example. Its original appeal included helping people cheat without being detected. Andreessen Horowitz’s investment announcement treated its founder’s controversial behavior and viral stunts as evidence of deliberate strategy. The willingness to undermine trust was visible from the beginning.

I find it difficult to take the subsequent concern about AI’s reception seriously when something like that was celebrated as entrepreneurial promise.

Selling the excitement

Investors also complain about inaccessible valuations and impossible hiring markets, then contribute to the same atmosphere by circulating the next astonishing figure. Harry Stebbings is a particularly irritating example. In an Instagram reel, he claimed that freshly graduated students from top universities were receiving offers worth $20 million a year, according to his founder friends.

Which companies? What roles? How many people? What does the compensation consist of, and how is it valued? Cash or equity? Annual compensation or a package spread over several years? A new undergraduate, or a researcher finishing a doctorate after years of important work? There was not enough information to assess the claim, but plenty to make the number sound exciting.

This is “trust me, bro” presented with the authority of privileged access. Perhaps an offer exists. That does not make the impression created by its presentation representative, or even particularly informative.

The stories influence the market being described. Founders revise hiring expectations. Employees reconsider what they should be earning. Investors acquire another explanation for why the company needs so much money. An anecdote does not have to be entirely invented to give a thoroughly misleading impression of the market.

What irritates me is the ease with which Stebbings moves between complaining about inflated expectations and helping inflate them. His platform participates in the market he describes. He should take some responsibility for the impressions he helps create.

How the race is made

The pressure does not stop with the companies receiving the money.

Fund a competitor lavishly enough and you change the choices available to everyone around it. Someone who might have built a useful, reasonably paced business now faces subsidized prices, expensive distribution and hiring ahead of revenue. Customers begin expecting the subsidized price. Investors begin expecting the growth.

A valuation then establishes demands of its own. A company that could have become an excellent business at one scale has to become something much larger to justify what has already been paid for it. Another round can raise that threshold again.

At some point, the pace is being set partly by obligations the financing created.

That is why I find the language of an inevitable AI race so irritating. The competition is real. But give several competitors extraordinary amounts of money, attach extraordinary expectations to each, and their inability to pause is hardly an independent discovery about how technology must develop.

Room for prudence

The infrastructure makes those commitments harder to escape. Data centers, chips, power and computing capacity are being secured against anticipated demand. Some will be necessary. Some may prove premature or excessive. The expenditure itself does not establish which.

Anthropic publicly filed for its IPO on October 26. Its revenue for the first half of 2026 reached approximately $16.3 billion, compared with $1.2 billion a year earlier. That is extraordinary growth. The question is how much further it must grow to support the valuation investors expect and the infrastructure being secured around it.

Long-term computing agreements can require payment regardless of how much capacity is ultimately used. Once those commitments are made, the company has considerably less freedom to respond if demand develops more slowly than expected. That is part of what makes the decision about how much to secure so consequential.

Dario Amodei’s stated approach to that decision is a green flag in my opinion. He has explicitly accepted the possibility of falling short of demand rather than committing to so much computing capacity that a delay in the expected growth could bankrupt the company. I appreciate hearing someone acknowledge that even an exceptionally promising business can destroy itself by committing too much, too early.

Anthropic’s performance makes that position particularly interesting. Its warnings about advanced AI have often been dismissed as doomerism. It has defended restrictions on mass domestic surveillance and fully autonomous weapons despite pressure from the US government, and chosen not to develop its own image and video generators. These are the kinds of reservations and omissions that the race narrative encourages us to regard as disadvantages. Yet the company is performing spectacularly.

I find that a rather persuasive argument in favor of prudence. Anthropic has managed to become a leading company while declining some opportunities and defending limits on others. I would like to see more attention paid to that, especially by people who speak as though restraint were something a serious competitor could scarcely afford.

Who ends up exposed?

Circular financing can make the underlying demand harder to assess. A supplier can invest in a company that commits to buying its infrastructure. The money becomes revenue elsewhere in the same network, helping substantiate the excitement around it. The transactions can be real while leaving a difficult question about how much independent customer demand will eventually support the arrangement.

Even the profits require reading. Extend the estimated useful life of equipment and its annual depreciation expense falls. That can be reasonable. But I want to understand the assumption when the same industry emphasizes how quickly each generation of hardware is superseded.

Older equipment can continue doing valuable work. What work, at what cost, for how long? Those details matter more than a confident projection whose profitability depends on the answer.

The financing is also being repackaged. Data-center debt is entering asset-backed and commercial mortgage-backed securities, with different classes receiving different repayment priorities. Behind the resulting investments are properties, leases, equipment, and assumptions about the revenue all of them will produce.

The resemblance to the repackaging involved in the 2008 financial crisis deserves to be stated plainly. Financial engineering can allow the system to take on more risk than it can comfortably absorb, while making that risk appear sufficiently divided, distributed, and insured. Several apparently separate investments may still depend on the same customers, the same demand forecasts, and equipment whose future value is far from assured. If demand falls short, customers struggle to pay, and equipment loses value, several layers of the financing can suffer together.

Meanwhile, the exposure reaches banks, insurers, pension funds, and people who never deliberately chose to finance a speculative technology buildout. Those institutions have different obligations from a venture fund. Their ability to buy a financial product does not settle whether they understand its underlying concentration, can withstand its losses, or should be exposing their beneficiaries to it.

That is what concerns me about the increasing complexity: how much exposure it permits, how difficult the connections become to follow, and who eventually bears the consequences.

Parts of the credit market have already become considerably less comfortable with the borrowing. The cost of buying protection against default has risen sharply for several AI-linked companies. These contracts concern whether companies will meet their debt obligations, which is a rather more consequential question than whether their shares look expensive.

There is an uncomfortable question about incentives here too. An institution can help finance the expansion while protecting itself against the consequences. Depending on its overall position, that protection may simply limit its losses, or it may allow it to profit from deterioration. The distinction matters, particularly if the institution is also encouraging others to invest.

The increased demand for default protection does not tell us which of those positions particular institutions hold. But we have already seen what undisclosed opposing interests can look like in mortgage finance. I would like to know how much exposure the people arranging and promoting these investments actually retain, how much they have passed on, and whether their own financial interests resemble the confidence they are selling.

Ordinary savings

Public equity markets provide another route into ordinary savings. Nasdaq introduced accelerated entry for qualifying large new listings; S&P retained its waiting period and financial requirements for the S&P 500. Those choices determine how quickly a newly public company can enter portfolios whose owners bought a broad index, rather than deliberately backing its particular promises. Going public gives the expectations accumulated in private markets a much larger population of potential investors.

South Korea offers a less abstract example of how policy can help turn market excitement into household exposure. In April, authorities relaxed diversification rules to allow domestic single-stock funds with up to twice the exposure to their underlying shares. The products launched in May. By July, regulators were suspending new listings and banning advertising, concerned about rapidly growing bets on memory-chip companies and the losses investors could suffer.

This was happening in a market the government had encouraged people to participate in. When the boom reversed, retail investors faced severe losses, including people who had borrowed to invest or committed money intended for retirement. The authorities could change the rules again; the people who had already lost their savings could not undo their participation.

That is where I draw the line. Encouraging citizens into increasingly risky financial products carries an administrative responsibility that cannot be discharged by reminding them afterward that investments carry risk. A government helps shape what appears legitimate, accessible, and reasonably safe. It should take that influence seriously before relaxing the safeguards.

Widespread household losses can also damage much more than investment accounts. People cut spending, struggle with debts, postpone plans, and lose trust in the institutions that encouraged their participation. At sufficient scale, those consequences affect businesses and households that never joined the speculation. That possibility belongs in the decision to expand access, rather than becoming a concern only once the losses arrive.

Adoption as an obligation

Inside businesses, the pressure to deliver the expected transformation affects how people are asked to work.

SaaS was finished. Software engineers were about to become unnecessary. Companies that failed to rebuild themselves around agents would be left behind. Adoption became a demonstration of seriousness, sometimes before anyone had established what successful adoption should accomplish.

There is a considerable difference between making a useful tool available and asking employees to prove they have used it.

When AI use enters performance reviews and requests for additional staff, it acquires a purpose beyond completing the work. Employees have reasons to demonstrate adoption; managers have reasons to present a convincing account of transformation. The amount produced is easier to observe than whether anyone needed it.

Software makes the distinction between output and useful work particularly clear. Producing code, deciding what should exist, understanding its consequences and maintaining the resulting system remain different activities. Progress in one can increase the demands on another.

A company can need fewer people while still depending heavily on the judgment of those who remain. It can generate more software while becoming less capable of understanding everything it now operates. The cost of an agent’s work includes what everyone else has to do with it afterward.

Tobi Lütke’s discussion of “slop grenades” interests me in that context. He describes a real problem: someone generates unchecked work and hands the burden of understanding and correcting it to a colleague. He is also the CEO who made AI use a baseline expectation and required teams to explain why AI could not meet their needs before requesting more headcount.

That does not mean he asked people to submit bad work. It means the incentives established by management belong in the explanation. Employees should remain responsible for what they send; executives should remain responsible for what they reward.

I have a similar reaction when not reading code is presented as a professional achievement. There are ways to establish confidence in software beyond inspecting every line. If someone has a better method, I want to hear it. But generating a change from a Slack conversation tells me little about whether the behavior being introduced is understood, or whether the change should enter production.

The infrastructure has to accommodate this activity too. GitHub itself identifies AI-assisted and agentic development as a major source of its traffic growth, while its incident reports describe services exhausting their available headroom.

An outage need not originate in AI-written code for the pace of adoption to be relevant. Capacity, recovery and operational competence are requirements too. A financial commitment can be made much faster than the systems and expertise needed to support it can be developed.

I see little patience for that distinction when acknowledging it would complicate the growth story.

The customer’s willingness

The capacity of society to absorb AI cannot be separated from how people feel about it. Trust affects what they buy, what information they share, what they tolerate at work and what restrictions they support. People can continue using a technology while becoming deeply hostile to the industry behind it.

Some uses can be imposed. That does not make the hostility inconsequential.

A persistent assistant asking to read messages, remember private circumstances and act on someone’s behalf depends on a degree of trust that the provider has to deserve.

Unwanted features, contempt for workers and creators, political interference, and the rather aberrant entitlement of some exceptionally wealthy individuals influence whether people want that relationship. Those objections do not disappear because the next model performs better.

Nor do the material effects remain inside the companies. Allocating more manufacturing capacity to AI infrastructure can leave other customers paying more for ordinary memory. Someone who never requested an AI product can encounter the boom through the cost of maintaining their computer.

For others, it arrives through a job they no longer have, a service that has become harder to use, or another demand to reorganize their working life around tools whose limitations they are expected to absorb.

The technology could have been welcomed as useful assistance, new possibilities and relief from tedious work. Some of it still is. But much of the surrounding culture has introduced it as an ultimatum: adopt it, compete with it, invest in it, or accept that you are about to become irrelevant.

I do not find the resulting resentment mysterious.

I can imagine AI becoming considerably more capable and widely used while many of the financial expectations attached to it turn out to have been preposterous. Expensive systems could do exceptional research. Cheaper models could handle ordinary tasks. A substantial amount of work could remain in making either useful to particular people.

I would be pleased to live with that outcome. It would offer plenty for someone like me to build, use and be fascinated by.

What concerns me is how much has already been committed on the assumption that something far larger must arrive, and how many people are being required to accommodate the attempt. The people allocating that capital helped determine the pace. They should also bear responsibility for what it costs everyone else to keep up.