Value-First AI Spending

The core critique

"Token quotas and token budgets are no different than hiring quotas. Until they get efficient, they'll be inefficient and I think a lot of companies will just burn money," argues Adam Foroughi.¹ The analogy is precise: a fixed hiring quota, ten engineers a quarter, optimizes for headcount rather than value, since some of those hires prove productive and some do not, but the quota itself creates pressure to fill seats rather than to identify genuine value. A token budget applies the identical pressure to spend tokens rather than to create value with them. Leaderboards ranking teams by token usage produce the same failure mode in a different shape: "if you're creating leaderboards of token usage, you're incentivizing people to create slop to climb the leaderboard," optimizing for a leading indicator in a way that destroys the underlying value the indicator was meant to predict.

The alternative: link spend to KPIs

The proposed alternative is not to avoid budgeting but to link consumption to measurable outcomes: what KPIs is a team actually optimizing for, and how does token consumption connect to those KPIs. Where model improvement can be traced directly to a business outcome, at AppLovin, better ad-targeting models lead directly to ad revenue, the causal chain is visible enough to justify investing more without needing an arbitrary ceiling. "If you optimize to value creation, you'll want to invest in tokens because there's revenue on the other side." The underlying test is simple: can the line be drawn from tokens consumed in a given workflow to business value created. If yes, spend aggressively; if no, fix the measurement infrastructure before spending more, rather than either restricting spend by default or spending freely without any accountability at all.

Why the code-generation percentage is the wrong metric

At AppLovin, 80 to 90 percent of code is AI-generated, but Foroughi is explicit that optimizing directly for that percentage is a mistake: "you could just produce slop to hit that number." The question that actually matters is whether engineers are using the tools well enough to accelerate the things that create value for the company, since the share of AI-generated code is only a leading indicator, and the real indicator is whether model improvement translated into revenue or profit growth. The same discipline applies to engineering output and to token budgets alike: do not optimize for the proxy metric, optimize for the thing the proxy was supposed to measure in the first place.

Why it matters

The framework sits between two more extreme postures. One view holds that most companies simply underspend on AI and should be far less cautious about the cost; the opposite view holds that spending needs a hard ceiling to avoid waste. Value-first spending rejects both: it argues against stinginess for its own sake and equally against spending without measurement, proposing instead that a company build the infrastructure to trace token consumption to outcomes and then let spending scale with demonstrated return rather than with an arbitrary quota in either direction.