10→1 vs 1→0
I don't think AI is making enough money yet to justify what's been invested in it. The default AI business model that comes to mind has been B2B SaaS, but from the data I mined, I think investors are scrambling to find what else this AI thing is good for.
I believe the lack of ROI mostly comes from the shape of B2B SaaS. Taking a workflow from 10 steps to 1 (10→1) is capped at the human operator's speed. However, taking the person (your customer) out entirely (1→0) lets you engineer the pipeline to the speed of software. To do that, you have to become the business, and maybe put the AI in a robotic body if the work is physical. You stop being Software-as-a-Service and become the Service.
Summer '25, when I wasn't fucking around Japantown, I interned at Soff, a YC startup that sold AI software to help manufacturers quote orders faster. Berni, the CEO, told me last month that Soff now makes steel. They rebranded to Westgate Supply and are now a real distributor.
The sales reps we worked with spent 3 hours a day building quotes by hand. We eventually got good at extracting info from orders and quoting them, but deals still took days. The rep still had to review the agent's numbers, send it internally, and follow up regularly with the buyer.
That is the ceiling on the whole B2B AI pitch: we make your business' employees faster. It's useful software, I helped build some of it. I’m just unconvinced there’s much money in making a sales person 25% faster or have nicer keybinds/UX. There is now an AI copilot for every business out there (even car washing!), and techy customers can build their own now. In those cases, the vendor selling software to DoorDash competes with Josh down the hall.
Investors know this too. Matter of fact, I think investors have a gut feeling that AI is vaguely useful, but are sweating because it might not be $3 Trillion useful. About $1.5 trillion is going into AI infrastructure this year, Sequoia’s math says the industry needs $3 trillion back, and the two biggest labs together run at under $100 billion. I see two places that investors1 are looking.
One is the physical world. I got all companies that YC lists publicly and tracked industry labels by batch.2 B2B peaked at 69% in 2023 and has fallen about three points a year since. Industrials were 2% of the W23 batch. By S26 they were 23% (Fig. 1). YC’s request for startups says it plainly: “AI is moving into the physical world.” I believe this 1→0 idea is one of the last shots they have at making generational money from AI before the market admits that copilots will never pay for $1.5 trillion of data centers.
The second bet is Soff’s. If your AI can do most of an accountant’s work, the old move was to sell it to accounting firms. The new move is to start the accounting firm. These are called full-stack companies. YC doesn’t label them in their website, but I ran all YC companies through an LLM call (Sonnet 5) explaining what a full-stack company is + company data. They were 1-4% for most batches. YC published a request for them in June 2025. By S26 they were 12% (Fig. 2).
To be clear, nobody is retreating from AI. Spending on chips, data centers, and models is still enormous. But I think founders and investors are scrambling to find good ways to make long-term revenue to justify the trillions invested. The first answer was B2B SaaS, but that's not producing enough.
In almost every operation today, the slowest part is people. Take the last person out and the whole thing runs at machine speed, which is why the money is moving toward this 1→0 direction.
P.S. This essay is a bit different than my other ones. I think this is a relatively mild take all things considered, I wrote a much more "general audience-friendly" version for a class, but thought the data + thesis is good enough for a quick X version release. Hopefully it was an amusing read :)
My data is YC only since they had great data I could scrape. I'm using it as a proxy for early-stage investors. ↩︎
YC's public directory via the yc-oss API, snapshot Sep 6, 2026: 4,500 companies, 20 batches, Winter 2019 to Summer 2026, using YC's own industry labels. Code and data here. ↩︎