Stop Worrying About API Bills—They're AI Companies' Payroll

Deep thoughts on AI and aspirations —— ByteThink Circle

YC partner Diana Hu said something uncomfortable for founders at Startup School: The best companies will be the ones spending the most tokens. What she meant is that a company armed with AI should willingly shoulder an uncomfortably high API bill, because it replaces what would otherwise be a more expensive, more bloated headcount. Saving tokens without saving money is core competitiveness.

What makes this counterintuitive is that twenty years of startup discipline has been written in the opposite direction: every dollar on the books must be saved, burn rate is a matter of life and death, cutting costs is always the right posture during earnings season. How did high bills suddenly become respectable?

On the Old Ledger, the Most Expensive Line Was Always Salaries

The cost structure of startups has had no suspense for decades. The biggest chunk is labor. The bulk of funding goes to hiring and payroll. Hiring speed roughly equals burn rate. The arithmetic of scale is simple: double revenue, nearly double the team. When people become unmanageable, add managers. Management layers grow this way, one tier at a time.

This arithmetic rests on one premise: output must be produced by people. An engineer’s productivity has a ceiling. To double productivity, you hire another one. The software industry has continuously improved marginal efficiency through reuse and tooling, but has never shaken the starting point: work is done by people, and the productivity ceiling is determined by headcount.

Output Can Now Be Purchased by Compute

AI has moved that starting point. An engineer surrounded by agent systems can deliver what previously required an entire team. Diana Hu observed that teams adopting closed-loop intelligence layers cut engineering sprint time in half while delivering nearly ten times the output. A more radical form is the so-called software factory: humans write specifications and define success tests, AI generates implementations and iterates until tests pass. Some companies have pushed their codebases to the point where they contain no hand-written code—only specifications and testing tools.

Once output can be purchased by compute, the physical mechanics of scaling change.

Diana Hu’s uncomfortable advice makes sense in this table: monthly bills of tens or even hundreds of thousands of dollars sound frightening, but compared to the cost of the engineering team they replace, they’re still cheaper. And bills are elastic—they drop when business contracts. Payroll is rigid—once you hire someone, it’s a commitment of a year or more. For cash-flow-sensitive startups, this math actually works out. It just requires mentally moving API fees from the burn category to the production category.

Whether Bills Are High Depends on the Denominator, Not the Numerator

But spending more tokens doesn’t equal spending them wisely—this is what I think is easiest to mislearn.

The bill itself is meaningless. What matters is how much gross margin each dollar of API spend generates. A hundred-thousand-dollar bill that produces three hundred thousand in gross margin is healthy. A ten-thousand-dollar bill that produces a bunch of demos no one uses is pure burn. There are really only two situations worth worrying about: tokens spent running demos instead of real closed loops, and tokens spent that don’t accumulate into any asset.

What counts as accumulating into an asset? Specifications, tests, evaluation sets, context accumulated in processes. In the software factory paradigm, engineers’ core output is specs and tests. These two things accumulate and compound—the next project stands on the previous project’s shoulders. Conversely, if an organization’s output is only this iteration’s generated results, with the next project starting from zero, then bills are pure consumption, fundamentally no different from hiring a few contractors. To judge whether a company is truly treating tokens as investment, look at whether its repositories and knowledge bases contain specifications and evaluations that are rolling thicker. Don’t look at bill numbers, and certainly don’t look at launch events.

The second trap is more insidious: people saved, work not increased. Measuring AI transformation by headcount reduction is equivalent to rewarding doing less work. Given the same budget, hiring two more people versus giving existing people sufficient compute allowance—in the AI era, the latter has higher expected output, provided that output is measured by delivery quality, not hours worked. Management practices must change accordingly. Otherwise, the money saved only becomes paper profit, with no relationship to growth.

The Money That Should Be Spent on People Hasn’t Decreased by a Dollar

Boundaries must also be made clear. Agents can handle execution but not judgment and accountability. Going forward, three types of people will remain expensive: experts who have gone extremely deep in a particular system layer; accountable owners who are directly responsible for customer outcomes and can mobilize cross-team resources; and the founder themselves. Diana Hu repeatedly emphasizes the last type: conviction in the power of tools cannot be outsourced. Founders must use coding agents hands-on until they personally shatter their prior assumptions about what’s possible. Companies where the CEO delegates AI strategy to the CTO and walks away have never truly begun.

Small teams don’t need to overthink it. A ten-person company has no operating system to speak of. The CEO using agents themselves every day, demonstrating it in their output—that’s more effective than any transformation plan. Employees never look at strategy documents. They look at what the boss does every day.

Changes on the investor side can serve as a forward-looking signal: revenue per employee, API cost as a percentage of revenue—these metrics are entering valuation logic. Companies achieving high revenue with minimal headcount and high compute costs will be recognized as the most efficient category of this era. The era of showing off headcount is exiting. The era of showing off bills may not be a bad thing, provided the denominator can support it.

Moving API bills from the cost category to the capital equipment category is the fastest way I’ve seen to judge a company’s AI substance. Companies that talk transformation still have payroll as the most expensive line on their ledgers. Companies truly transforming are in the middle of that line item switching. No one knows when the switch will be complete, but looking at the ledger is always more accurate than looking at launch events.