Tailscale's Remy Guercio on the shift from token maxing to ROI maxing

At the AI Engineer World's Fair 2026, Michael Grinich caught up with Remy Guercio of Tailscale at the WorkOS booth. It's become something of a recurring conversation — the last one was around six to eight months earlier, near re:Invent, when Tailscale was still rolling out Aperture. This time the topic was money: who's spending it on AI, why the bills got so big, and what to actually do about it.

Remy works in strategic projects at Tailscale on Aperture, the company's AI gateway. The full conversation is worth watching, but here are the ideas that stuck with us.

From token maxing to ROI maxing

Remy's framing for the moment we're in: the last 12 months were about token maxing, and the next 12 will be about ROI maxing.

Remy uses "token maxing" approvingly. It's the work of figuring out what you can do once you have a million-token context window — stuffing everything into a model's context, with tools like Claude Code and other coding agents finding the right parts of files to load. That exploration has been genuinely productive. It has also been expensive.

The pushback, Remy says, is coming from people seeing the bills — an OpenAI bill, an Anthropic bill, a Gemini bill, all landing at once, and finance teams asking questions. Few of those people want granular per-PR accounting, he adds. Most just want to understand what's generating the bills in the first place.

The Concorde problem

Cheaper per token doesn't mean cheaper per task. Remy pointed to the Artificial Analysis benchmark tracking cost and intelligence per task, where the trend is heading down as open source models improve. But the picture isn't uniform: when Sonnet 5 came out, it was cheaper per token yet more expensive on a per-task basis.

Michael reached for the Concorde as the analogy. It burned enormous amounts of fuel, but because it flew at Mach 2 and crossed the Atlantic in about three and a half hours, it was efficient per mile. A model can burn tokens fast and still be the efficient choice if it gets to the answer quickly. The only way to know is to measure the task, not the token.

Seeing what's actually happening

Aperture is Tailscale's answer to that measurement problem. It's an AI gateway that acts as both an LLM gateway and an MCP gateway, sitting between your agents and your LLM or MCP endpoints. It does model routing, but Remy says a big part of it is governance: seeing what agents are doing inside the org, keeping logs and sessions, and adding guardrails like stripping PII on the way out to providers.

On cost, it can show the average cost per million tokens on a per-user, per-model, or per-harness basis over any period — enough to spot, say, a harness that isn't caching the way it should. It also handles the access side: who and which teams get which models and budgets, so an organization can understand its AI spend rather than just absorb it.

Don't consolidate on one provider

The tempting response to a pile of bills is to consolidate everything onto one provider. Remy thinks that's the wrong move — you trade many bills for one very large bill, and you lose the ability to compare.

His preferred approach is to let a thousand flowers bloom: give engineers and teams room to experiment across models, then verify the results. His example was an engineer trying GLM 5.2 and being able to confirm in Aperture that it genuinely worked better or cheaper, instead of going on vibes. He also flagged the routing tradeoffs teams run into — vLLM's semantic router, and a project called headroom for prompt caching and optimization — noting that switching models mid-flight blows your cache.

What this looks like inside Tailscale

Tailscale has lived the curve it's now building for. Over the last 12 months, AI usage went from tab autocomplete to engineers running agents constantly, measured in billions of tokens rather than tens or hundreds of millions.

The part Remy finds most interesting is cultural. A subset of engineers always had more ideas than time; AI let them try many paths, review them all, and keep the useful ones. His hope is that this produces better software — teams exploring several approaches and shipping a good one — rather than simply more software.

Michael pushed on the ROI math. Building ten things is less efficient than building one, in a narrow sense. But if the goal is finding product-market fit, the calculation changes.

Capital allocation for tokens

That leads somewhere Tailscale didn't originally set out to go. Michael pointed out that a company that started by building software for engineering teams is now describing something closer to capital allocation across the org — token allocation. Remy agreed, with a caveat that they aren't targeting finance teams directly.

To support it, Tailscale is adding a feature that lets you send arbitrary labels through Aperture, so you can do ROI analysis by PR, by bug report, or by whatever unit matters — and see what a given experiment or fix actually cost.

The economics explain the interest. AI is a good bit more expensive than the fully specced $4,000 MacBook Pro that used to be everything an engineer needed, so finance, IT, and security are all paying attention to a much bigger line item. Remy's summary of Tailscale's posture: an adult, measured, thoughtful approach to AI usage, where cost is a huge component but not the only one.

Watch the full conversation for the rest — including where Remy thinks model routing and per-task efficiency go from here.

This interview was recorded at the AI Engineer World's Fair 2026 in San Francisco.