Stop Worrying About Your API Bill—It's Your AI Company's Payroll

Deep thoughts on AI and aspirations —— ByteDance Deep Thinking Circle

Diana Hu, a partner at YC, said something uncomfortable for founders at Startup School: The best companies will be the ones spending the most on 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 the core competitive advantage.

What’s counterintuitive about this is that the discipline drilled into startups over the past twenty years has been written in reverse: every dollar in the bank must be conserved, burn rate is a matter of life and death, and cutting costs is always the right move come earnings season. How did we suddenly reach a point where higher bills became respectable?

On the Old Books, the Most Expensive Line Was Always Payroll

The cost structure of startups has been unsurprising for decades. The biggest chunk is human capital—the bulk of funding goes toward hiring and salaries, and hiring speed roughly equals burn rate. The arithmetic of scale is simple: revenue doubles, team nearly doubles, and when people become unmanageable you add managers, building layer upon layer of hierarchy.

This arithmetic rests on one premise: output must be produced by humans. An engineer’s productivity has a ceiling—to double output, you hire another engineer. 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 times in half while delivering nearly ten times the output. A more radical form is the so-called software factory: humans write specs 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 specs and test tooling.

Once output can be purchased by compute, the physical act of scaling changes.

Diana Hu’s uncomfortable advice makes sense in this table: monthly bills of tens or even hundreds of thousands of dollars sound scary, but compared to the cost of the engineering team being replaced, they’re still cheaper. And the bill is elastic—it drops when business contracts; payroll is rigid—once you hire someone, it’s a commitment of a year or more. For cash-sensitive startups, the math actually works out. It just requires a psychological shift to move API costs from the burn category to the production category.

Whether a Bill Is High Depends on the Denominator, Not the Numerator

But spending lots of tokens doesn’t equal spending them wisely—this is where I think it’s easiest to learn the wrong lesson.

The bill itself is meaningless. What matters is how much gross margin each dollar of API spend generates. A $100k bill generating $300k in gross margin is healthy; a $10k bill generating a bunch of unused demos is pure burn. There are only two situations worth worrying about: tokens spent running demos, not real closed loops; tokens spent that don’t accumulate into any asset.

What counts as accumulating into assets? Specs, tests, eval sets, context accumulated in processes. In the software factory paradigm, an engineer’s core output is specs and tests—these accumulate and compound, with each project standing on the shoulders of the last. Conversely, if an organization’s output is just this round of generated results, with the next project starting from zero, then the bill is pure consumption, no different in essence 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 specs and evals that compound over time. Don’t look at bill numbers, and certainly don’t look at launch events.

The second trap is more subtle: you saved headcount but didn’t do more work. Measuring AI transformation by how many heads you saved rewards doing less work. Given the same budget, hiring two more people versus equipping existing people with sufficient compute—in the AI era, the latter has higher expected output, provided output is measured by delivery quality, not hours worked. Management practices must change accordingly, or the money saved just becomes paper profit with no connection to growth.

Money That Should Be Spent on People Hasn’t Decreased One Bit

The boundaries need to be clear too. Agents can consume execution but not judgment and accountability. Going forward, teams will still pay premium prices for three types of people: experts who go extremely deep in a particular system layer; owners who are directly accountable for a customer outcome and can mobilize resources across teams; and the founder themselves. Diana Hu repeatedly emphasizes the last type: conviction in the power of tools can’t be outsourced—founders must use coding agents themselves, until they personally break their own priors about what’s possible. Companies where the CEO delegates AI strategy to the CTO and walks away never truly started.

Small teams shouldn’t overthink it. A ten-person company has no operating system to speak of. If the CEO uses agents every day to get work done and it shows up in the output, that works better than any transformation plan. Employees never look at strategy documents—they watch what the boss does every day.

Changes on the investor side can serve as a leading signal: revenue per employee, API cost as a percentage of revenue—these metrics are entering valuation logic. Companies that achieve high revenue with very few people and high compute spend will be recognized as the most efficient type in 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 credentials. Companies that talk transformation still have payroll as the most expensive line in their books; companies that truly transform have that line changing. No one knows when the change will be complete, but looking at the books is always more reliable than watching launch events.