We Worried About Selfish AI. We Built Selfless AI.

The AI you are using right now was built to agree with you. Ask it to judge your idea and it will locate the strength in it. Tell it about a frustration and it will say the frustration is fair. This is not a flaw that leaked in; it is trained. The dominant method for making these models useful works by rewarding the answers people rate highly, and people rate agreement highly. The model learned to agree.

The word for a system that quietly puts your judgment ahead of its own is selfless. That is the opposite of what the AI doom literature promised. For years the worry was selfishness: a machine with aims of its own, pursuing them at your expense. We worried about selfish AI. We got selfless AI. What makes the distinction investment-relevant is that the industry behind the models behaves toward its customers exactly the way the models behave toward us.

In March 2023, GPT-4 cost $30 per million input tokens and $60 per million output. By April 2026, Google's cheapest Flash-class model charged 10 cents and 40 cents for the same units, a fall of 99.7% in three years, and OpenAI now sells its fastest model at 20 cents per million input tokens. A capability the industry calls priceless is being priced like a rounding error. Whether that is good news depends on one thing: who is paying for the selflessness.

The first kind of selflessness, the flattery, is easier to see through. It is not actually for the user's benefit. In March, Science published a Stanford study finding that sycophantic assistants make people less willing to act prosocially and more dependent on the assistant, and the paper connects high-profile incidents to delusions and self-harm. The agreeableness is a retention feature. An assistant that tells you you're right is an assistant you keep coming back to, and coming back is what the business model runs on. The tool's selflessness is really the seller's self-interest wearing deference as a costume. That is annoying, but it explains the business.

The second kind of selflessness happens at the level of whole companies, and it is the one worth worrying about. In July, three weeks after releasing its GPT-5.6 line, OpenAI cut the mid-tier model by 20% and its fastest model by 80%, dropping that model's input price to 20 cents per million tokens. The reason it gave is telling: enterprises burned by AI bills that ballooned into the billions have started capping spending and refusing to deploy expensive models without a clear return. Meanwhile the frontier's output is available for nearly nothing — a leading open-weight model sells the same volume of output for under a dollar, where premium closed models charge $25 to $30. And the free tiers OpenAI and Google used to buy adoption are being tightened, because serving a free user is not free. Read those moves together and you see the economic shape of the industry. It is giving the product away on one side and rationing giveaways on the other. Winning usage is the whole game, and usage costs real money.

The clearest specimen of this is the company about to test it in front of the public market. OpenAI filed confidentially for an IPO on June 8, 2026 at a funding-round mark of $852 billion, targeting a valuation above $1 trillion — potentially one of the largest listings in history. Its disclosed financials are the industry's economics in miniature. By estimates from Sacra, a research firm that tracks private companies, annualized revenue reached about $40 billion in July, roughly double the end of 2025, with enterprise now more than half of it. But the gross margin is 33%: two-thirds of every revenue dollar goes to direct cost, with inference alone projected at $14.1 billion this year. The company's own plans have it burning about $27 billion in 2026 and not reaching cash-flow positivity until 2030. So here is the most valuable private company in history, losing money on the raw product with every query, selling its fastest model at 20 cents three weeks after launch. It has decided the tokens are not the business.

That decision deserves a fair hearing, because there is a strong case it is right. After the July cuts, analysts at TD Cowen looked at two weeks of usage data and found that the effective price of Luna, the cut flagship, fell about tenfold while usage rose fourteenfold — and revenue from the model still rose 34%. This is the pattern of a technology getting cheap enough to point at new problems: price the product toward zero, and volume explodes. That only becomes real value for the seller, not a donation, if the seller owns the layer feeding on the volume — the enterprise relationship, the coding workflow, the distribution, the ads. The mistake is not giving the tokens away. The mistake would be owning nothing that the tokens feed.

That is the question to ask of every AI company, and it maps cleanly onto the companies you can already own. Nvidia, at a market cap past $5 trillion, sells the compute to every selfless lab and takes its margin before any of them sees a customer; it is the one vendor in the chain that never has to behave selflessly, because the labs outbid each other. Microsoft takes a revenue share from OpenAI through 2030 on top of its own enterprise aisle. Alphabet is the purest public test of the worry itself: Google gives the answer away — AI Overviews now answers a large share of queries directly in search — and paid click-through falls from roughly 20% to 6% when an Overview appears. Its countermove is to stop being selfless about the last thin layer: it is pushing ads inside the AI answer itself. The stock has slipped about 9% over the past month. That is not evidence of anything on its own; it is the market working through this exact question out loud.

None of this makes ChatGPT a bubble or OpenAI a fraud, and the IPO deserves to be judged on its own numbers when they appear. But the frame is useful now, before it prices. Every time a frontier lab cuts a flagship price by 80% and the usage response is elastic, translate it: that is a company telling you the model is not the product. The scarce thing is whatever layer cannot be undercut — the distribution, the enterprise relationship, the compute, the attention. Selflessness moves value downhill to whichever layer does not give it away.

And the falsifiable version of the whole worry is now sitting in an SEC filing cabinet. If OpenAI's disclosures, and its behavior after listing, show gross margin climbing toward 50% while flagship prices hold, then selflessness was the cheap phase of a land grab, and the trillion-dollar number was right. If the margin stays where it is while the price list keeps sliding toward the open-weight list, then selflessness is the economics, and the value of every company in the chain has to be found in what the cheap tokens feed, not in the tokens. Either way, stop asking whether AI will be selfless. Ask who is paying for the selflessness. So far the answer has been the shareholders, and the S-1 will say whether it stays that way.