Sierra's co founder says enterprise AI token spend will jump from 3.
Clay Bavor wants enterprise AI buyers to think about money spent, not cost saved. On a recent 20VC interview summarized by Podcast Alpha, the Sierra co-founder argued that AI token spend inside large engineering organizations is on track to jump from roughly 3.8% of engineering budgets to about 20% over the next several years. The reasoning is mechanical: reasoning models consume far more tokens per query than chat models do, and the move from chat to agent-shaped work multiplies that burn at the workflow level, not the prompt level.
When the public narrative is "models are getting cheaper," Bavor is saying the per-query cost is going up, and the buyer who already has a platform embedded inside their systems absorbs that rising line item before any competitor can. Token spend, not headline accuracy, is the variable that compounds.
Sierra's so-called "Brain" is a roughly 20-to-30-page grounding document built around each customer's operations, terminology, and policies. Every deployed agent runs against it, and the document's value rises with each customer interaction because the corrections and edge cases accumulate inside it. Replicating that anchor at a competitor is not a model swap; it is a rebuild. Sierra describes the approach and lists flagship deployments on its customers page, where the roster reads as a who's-who of regulated consumer finance and healthcare rather than a SaaS-style breadth play.
The clearest live example is Rocket Mortgage. The relationship started as a customer-support deployment and has since expanded into what Sierra describes on its Rocket Mortgage customer page as full-company AI infrastructure. The trajectory matters because it shows the mechanism running in production, not in a deck. A support bot is a single workflow; full-company infrastructure means the grounding doc, the agent layer, and the audit trail now sit underneath functions that have nothing to do with chat. Each new function pulls more queries through the same anchor.
The capital base backs the strategy. Sierra raised $950M in a round covered by TechCrunch in May 2026, the kind of check that funds an embed-first go-to-market rather than a sales-led land grab. Bavor's market read on the same interview is that the AI cloud market will consolidate around whoever is deepest inside the enterprise, not split the way rideshare did.
The reference case Bavor invokes is Palantir, a company whose commercial strategy has long been to install engineers inside customer environments until the customer's data plumbing runs through Palantir's ontology. Bavor's analogy is his own framing, not an independent comparison, but the operational similarity is concrete: both companies sell embedded deployment depth over a self-serve product, both price against the customer's existing IT line items, and both depend on operational lock-in rather than feature velocity. Palantir's enterprise footprint is documented across its own press releases; whether Sierra reaches comparable depth across the Fortune 50 is the open question.
The 3.8%-to-20% projection is a Bavor forward-looking call, not measured spend data, so it should be read as a Sierra thesis, not consensus. The "roughly 40% of Fortune 50" footprint figure is also a Sierra self-claim, not an independently verified census, and the public customers page lists a meaningful but smaller roster. The Podcast Alpha piece is a paid summary of the 20VC interview, so the original audio or transcript is the better citation for exact wording; the underlying shape of the argument, however, is consistent across the summary and the customer-facing pages.
The watch item is whether the Rocket Mortgage-style expansion pattern repeats. If Sierra can name a second customer that moved from a single-agent deployment to full-company infrastructure in the next two quarters, the embed-first thesis is doing work. If the next batch of named customers stays at the support-bot layer, the 20% projection is a pitch, not a forecast.