At eleven at night, Qiu Ziwei pulled up Sam Altman's conversation from the closing session of Y Combinator's “2026 Startup School,” dimming his laptop screen halfway so he wouldn’t wake the other two co-founders asleep in the next room. The three of them made up the entire team behind a customer-service automation tool for small and mid-sized e-commerce brands — days went to the product, so quiet hours like this were the only time left to catch up on reading. In the video, Altman was talking with YC’s current president, Garry Tan, about why there’s never been a better time to start a company.
The line itself didn’t do much for Qiu Ziwei — some version of ‘now is the best time’ gets said every year in startup circles. What actually made him turn the brightness back up and sit upright was the comparison Altman tossed out right after: a piece of development work that took him three months back in 2005 would now take an AI agent about seven minutes.
What Stopped Him Wasn’t the Pep Talk — It Was the Number That Came After It
Seven minutes versus three months struck him at first as an exaggerated bit of rhetoric — until he remembered something that had happened to his own team just the week before: a feature refactor they’d budgeted two weeks for got wrapped up in three days, with an AI agent helping write tests and catch duplicated logic. He’d been feeling that same time compression almost every month; nobody had just put a specific multiplier on it out loud before.
What came next in the video was Altman’s read on where inference demand for the whole industry is heading: he expects global demand for AI inference to keep growing roughly tenfold a year for years to come, with demand for high-quality AI essentially uncapped. He went further and offered a longer-range forecast — that by 2032, each person will be consuming 500 billion tokens a month.
Qiu Ziwei rewound that line to make sure he’d heard the unit right. He wasn’t doubting the number’s accuracy so much as registering what it was — Altman’s own long-range judgment call, not a statistic published by any official body. But the sheer scale of it was enough to surface a more immediate question: if even the direction of that call turned out to be right, did anyone on his three-person team actually know how many tokens they were burning through in a given month?
He Opened a Spreadsheet and Found the Team’s Bill Had Always Been a ‘Feels Fine’ Black Box
The answer was no. Three months earlier, when the team was just getting started, they’d put a few AI platform subscriptions on a credit card and never gone back to check the line items since — the monthly charge always looked roughly in range, felt like a good deal, and got filed away as a solved problem, automated like a utility bill.
What Qiu Ziwei now realized was that ‘feels like a good deal’ rested on an assumption nobody had stated out loud: that usage growth would happen to stay inside the subscription plan’s flexible range. If Altman’s direction was right — inference demand compounding at something like tenfold a year — that assumption would eventually break, just not on any predictable date. Up until that point the bill looks completely normal; only after it breaks does the question of ‘why did this month jump so much’ surface for the first time, and usually only after the fact. Right now, his team had no record at all that could answer something as basic as how many tokens the three of them had actually burned through the previous week, or on which feature.
Everyone Thinking You’re Crazy Turns Out to Be the Clearest Signal
Midway through the conversation, Altman mentioned that when OpenAI was founded, only around fifty people in the field, by his own account, genuinely believed AGI (artificial general intelligence) was achievable — and that period of being dismissed by the mainstream actually handed them two unexpected advantages: no major competitors racing to catch up, and enough time to just do the research properly.
Hearing that, Qiu Ziwei thought back to when his team first decided to build customer-service automation, and a few former colleagues had quietly told him the space was already saturated and a big company would eventually swoop in. He hadn’t given that comment much weight before, but this time it landed differently: if even Altman treats ‘most people think you’re doing something dumb’ as a signal rather than a warning, it might be worth going back through his own team’s calls to see which of them were actually underrated opportunities rather than reasons for self-doubt.
Small Teams’ Speed Is Wearing Down Big Companies’ Scale Advantage
Altman also described a pattern he’s observed: the faster the pace of technological change, the more easily large companies’ scale advantage gets neutralized — because that advantage was always built on spending more time and stacking more resources, and once execution speed itself becomes the deciding variable, small teams get a shot at things that used to be out of reach for them.
That line eased some of the anxiety Qiu Ziwei carried about being a team of three — but he was also clear-eyed about an unstated condition behind it: for a small team to actually move faster than a big company, they can’t be spending that time after the fact untangling why this month’s AI spend spiked, a problem they could have seen coming. A speed advantage is meant to be spent moving forward, not cleaning up usage that already got out of control.
Altman Says Mistakes Are Fine — He Decided Usage Wasn’t Going to Be One of Them
Near the end of the conversation, Altman gave a piece of very direct advice: startups are going to make a lot of mistakes, and the industry is actually quite forgiving of that, so founders shouldn’t hold back out of fear of getting things wrong. Qiu Ziwei agreed with that attitude and planned to genuinely apply it to product calls and market experiments — keep the vision clear, and accept that the first few steps shouldn’t be obvious.
But he also drew one small line in that moment: product direction could afford to be tried and wrong, but whether token usage was visible was not something he wanted left unmanaged any longer. That night, without waiting for a team meeting, he opened the plainest possible shared spreadsheet and asked the other two to log one line each time they ran something resource-heavy — which feature, roughly how long it ran. He knew the sheet was far from precise, but from that night on, ‘why did this month get more expensive’ stopped being a question they could only guess at after the fact.
FAQ
Q1: What is Y Combinator’s Startup School?
Startup School is an annual startup education event hosted by Y Combinator, a well-known startup accelerator, bringing in founders and investors to share their experience for people who are starting or planning to start a company. The conversation this article draws on was the closing highlight of the 2026 edition, between Sam Altman and YC’s current president, Garry Tan.
Q2: Is the ‘500 billion tokens per person per month by 2032’ figure an official statistic?
No. It’s a personal long-range forecast Sam Altman offered during the conversation to illustrate his own view on how fast AI inference demand is growing — it is not a statistic published by any research institution or official body. This article is simply relaying the speaker’s statement as given; it does not represent an endorsement by this site or AITK of the figure’s accuracy, and readers should still base their own usage planning on their actual usage data.
Q3: Can the claim that ‘only around 50 people believed in AGI early on’ be verified?
This is likewise Altman’s own account in the conversation, a personal recollection of the atmosphere around OpenAI’s founding. This site has not independently verified how that number was arrived at; it is presented here purely as part of the speaker’s narrative, not as a historical fact this site has confirmed.
Q4: Isn’t the claim that ‘three months of work now takes an AI agent seven minutes’ a bit of an exaggeration?
This is Altman drawing a comparison to his own personal experience from 2005 — essentially a rhetorically effective personal anecdote meant to convey the felt gap in development speed, not the result of a rigorous benchmark. Actual efficiency gains vary widely by task type, tooling, and team familiarity, and shouldn’t be applied directly to any specific project.
Q5: If you agree small teams should seize the moment, what’s the actual first step?
The point of this article isn’t that a team needs to make some major strategic pivot right away. It’s that a team should first make sure it can actually see what’s changing — especially whether AI tool usage and spend are being tracked as task volume grows. Seeing it is what makes it possible to act on real data when a real adjustment is needed, instead of going on a feeling.
Source Note
This article is based on a conversation published by Y Combinator’s official YouTube channel on July 28, 2026, titled “Sam Altman: "Never a Better Time to Do a Startup"”, recorded at the closing session of 2026 Startup School (July 25–26, 2026, Chase Center, San Francisco) between Sam Altman and YC’s current president Garry Tan, reorganized and told as a narrative rather than translated verbatim. ‘Qiu Ziwei’ in this article is an illustrative character used for explanation, not a real client; his team’s background and the scenario described are an illustratively reconstructed situation. The four claims discussed in this article — that people will consume 500 billion tokens per month by 2032, that global AI inference demand is growing roughly tenfold a year, that only around 50 people believed AGI was achievable in OpenAI’s early days, and that three months of 2005-era development work now takes an AI agent seven minutes — are all statements, recollections, or forecasts made personally by Sam Altman during the conversation, not official statistics or academic data independently verified by this site or any third party. They are presented solely to help readers understand the speaker’s viewpoint and do not constitute an endorsement by this site of their accuracy.
Take Action
If even part of what Altman describes turns out to be right — AI inference demand continuing to compound — then how many tokens your team is actually spending each month, and on what, will eventually stop being a question you can leave alone and start being one that shapes your unit economics. Rather than waiting for the bill to tell you first, lay it out now. Try AI Token King for free, and let us lay out exactly how many tokens every agent and every feature on your team is actually using, so you can set the right usage boundaries while usage is still manageable.
