🪙 Get in the token path

Stop being a side dish on someone else's bill

A few weeks ago, Jamin Ball at Altimeter wrote a great piece called Get in the Token Path. The core idea: the biggest infrastructure winners of the cloud era monetized the core consumption primitive of the platform. In cloud, that primitive was compute. In AI, it’s tokens, the unit AI labs charge by, kind of like kilowatt-hours for electricity.

I loved this post, but I’d argue that it falls short in one area. The token is the entry point. Agent spend is the market.

From line item to leverage

I was looking at the spend breakdown of one of our customers at Browserbase recently and noticed that their browser bill was a much lower % of their token bill.

That stuck in my head because it forced a question I think a lot of AI infrastructure founders are dodging right now: am I actually in the path of the dollar, or am I just a small line item next to the real one?

At first, that sounds discouraging. If your product is a small part of the token bill, how big can it really get? To me, it clarified the opportunity: every line item around the agent can become a control point for the biggest line item: tokens.

If you’re building an agent today, your bill of materials looks something like this: tokens, browsers, sandboxes, search APIs, memory, evals, observability, payments, voice, and a dozen other things. Tokens are the loudest part of the bill, but they are not the whole bill.

So, use the token bill as the wedge into the rest of the stack. Tokens are the biggest lever, the most visible pain point, and the easiest place to prove you deserve more of the workflow.

Optimization is how you earn the spend

The mistake I see AI infra companies make is assuming customers will route spend through them just because they offer a better product. Great products are how you get in the door. Optimization is how you expand from a line item into a platform. Customers route spend through you when you actively make their existing spend more efficient.

This is the lesson I keep coming back to from watching Datadog. They figured out that if you’re sitting near someone’s bill, the best move is to help them spend less on the bigger line item above you. They built a whole cloud cost monitoring product so they could literally help customers reduce the AWS spend they were sitting on top of.

Imagine a company spending $100M on AWS and $10M on Datadog. Datadog can become ROI positive in two ways:

1. Prove that $10M on Datadog is cheaper than building and maintaining cloud observability in house.

2. Generate insights that create more than $10M of savings on the $100M AWS bill.

The second one can often be much easier than the first. Especially when dealing with CFOs.

The same playbook works for tokens, only better, because token spend is messier than compute spend. Companies are routing inference across multiple labs. They’re using the wrong model for the wrong task. They’re not caching. They’re not batching. They have no idea which agent flows are draining the budget. The optimization opportunity is enormous.

So if you’re an AI infra company, the wedge is: cut my token bill, and I’ll let you deeper into the agent stack. That’s how you go from a $5 line item to a $50 platform.

Three ways to get in the flow

There are three ways to earn this trust.

First: extend the token. Add capabilities the labs don’t offer. Memory layers, evals, prompt management, connectors, voice. Anything the lab can’t do but the agent needs. Become the value-added control plane around tokens.

Second: hide the token. Sell the outcome, not the token. My friend George runs Rely, which automates diligence for multifamily real estate. Rely’s customers aren’t buying tokens. They’re buying completed lease audits, vendor reviews, and utilities reconciliation. The token cost is real, but it disappears into the value of the outcome. The customer doesn’t care if the bill underneath is $10 of tokens or $50, because the alternative was a junior analyst spending three days on it. Cursor does this for code. Harvey does it for legal. Decagon does it for support. Selling the outcome is the cleanest way to own the entire agent spend, because the customer never sees the underlying tokens at all.

Third: optimize the token. This is the one I keep coming back to. If you can credibly walk into a customer and say “we’ll cut your token bill by 30%,” you’ve earned the right to also be their browser, their memory, their search, and their observability. The optimization is the wedge; the whole platform is sticky.

The path to $100M

There are so many ways to create value for your customers. If the core consumption unit is a token, the question to ask yourself is:

Can I save you tokens? Can I increase the margin on the tokens you’re already buying? Can I add capabilities to those tokens that the labs can’t?

Get customers to route inference through you, deliver more value than the token providers themselves, and you’ve put yourself in one of the most powerful positions in this entire wave. That’s what takes a company from tens of millions of dollars to hundreds.

Compute mattered for a decade. Agents matter for the next one. The infrastructure companies that win will walk in through a small line item, make the customer’s biggest line item smaller, and earn the right to route the rest of the stack.

The token is the wedge. Agent spend is the prize.