The missing piece · Michelle Chen

The frame

Understood backwards

She opens with Kierkegaard, and the quote sets up the whole talk: Cloudflare is a 16-year-old company that made a run of decisions which only add up in hindsight. The talk follows the quote's shape, backwards through the history, forwards through a demo, then onward.

From the notes: the Kierkegaard line, noted as she said it.

Evidence: transcript 00:00:56

The history

A stack with a hole in it

Each layer was built on the one before. A CDN has to be everywhere, so Cloudflare put data centers as close to users as possible, unlike a hyperscaler that centralizes compute in a few regions. With computers everywhere, it added compute: Workers in 2017, serverless functions that start in milliseconds, then Durable Objects, which are stateful, and containers in 2025. Then storage, for full applications.

In 2023 came Workers AI: GPUs spread across the same network. The bet was that models would shrink and network latency would matter more. The opposite happened. Models grew from seven billion parameters to trillions, and Cloudflare pivoted to serving giant models such as Kimi, GLM and DeepSeek in March 2026.

By 2026 it had every layer of an agent cloud: the network, compute that agents run on (a Durable Object lets an agent work on a task, then sleep), storage, and large models.

“but it never felt quite right.”00:06:55

The giant models fit the developer platform, not the infrastructure. The distributed edge had nothing that only it could serve well.

From the notes: “workers ai: gpu in the cloud”, and two open questions: what Durable Objects are from Cloudflare's side, and how Cloudflare really works with agents. Her answer to the first: the stateful compute primitive an agent runs on, spinning up for a task and sleeping between. The deep dive is still on the backlog.

Evidence: transcript 00:01:45, 00:03:47, 00:04:24, 00:06:16, 00:06:55

The turn

System One made Cloudflare a provider

The missing piece arrived from outside. In September 2026, TypeSafe launched Jev and the idea of System One models. An LLM generates token after token; a System One model takes one forward pass over the input and returns an output you can make a decision with. Not everything has to go to an LLM: the small, fast models take work off the big ones, and the two work together.

That brought Cloudflare back to its original bet.

“And so Cloudflare was not wrong here. We actually were just early with these small and fast models.”00:07:50

Small, fast models are exactly what a distributed network serves well. That's why she calls becoming a lab the missing bit: Cloudflare had the infrastructure, the compute, the data and the large models, but to use its own infrastructure it had to ship System One models itself.

From the notes: “system 1 model was the thing that make cloudflare a provider!!!” This is the hinge of the talk. Serving other labs' giant models made Cloudflare a host. Training the small models that fit its own network made it a provider.

Evidence: transcript 00:07:50, 00:13:18

The first model

Clef: their own Jev, in a weekend

Jev is closed source, so Cloudflare couldn't host it on its own infrastructure. They started experimenting three days after its launch, deliberated for a week over whether to become a lab at all, then decided on a Friday evening.

“And we ended up training the model over 48 hours over the weekend, and then we launched it the following week.”00:09:53

The result was Clef, a decision model. Agents can decide on their own, without an LLM reasoning for minutes and without waiting on a person's sleep schedule. As an inference provider, it takes pressure off the LLMs, which are kept for meaningful work.

“Clef gave power back to the agents to be able to make decisions on their own.”00:09:53

From the notes: “clef is their version of jev”. Jev returns in three talks today: Diogo Almeida's, Glean's, and Scott Roberts's at Workday, where it picks the action card that has the data while you type, and runs inside a SQL statement in DuckDB.

Evidence: transcript 00:09:53

The second model

Clairic: unstructured in, typed JSON out

Decisions weren't the only work too expensive or too slow for an LLM. Clairic, launched three weeks later, is an extraction model. Emails, tweets and Slack messages don't arrive as forms; Clairic reads unstructured input (text or images), extracts, computes, estimates and forecasts values, and returns a typed object an agent can act on.

“and it gives you back a perfectly typed JSON object.”00:11:35

She calls it System One in the sense that matters: just as fast and cheap, one forward pass.

From the notes: you weren't sure extraction to JSON was “really” System One. Her answer is that it's System One by how it runs (one pass, fast, cheap; her slide adds “no output tokens”), not by what it does: deciding and extracting are both work you stop sending to an LLM.

Evidence: transcript 00:11:35

Forwards

The CI triage demo

Her demo is an agent triaging a failed CI build with forty tools, the agent running in a Durable Object. The LLM-only version goes back and forth with the model, resending the same inputs. The System One version lets Clef prune the tool list to the ten it needs, uses Clairic to pull the PR number from the input and the exact error from hundreds of log lines, and calls the LLM once, to act. Same answer; in her earlier run, almost twice as fast, five LLM calls down to one, and about 94 percent fewer tokens.

“It was 72 percent less expensive.”00:14:06

From the notes: “pi durable comes back again how does that really work”. Her slide answers the first half: it's headed “CI triage agent · a Durable Object running Pi”, so the transcript's “PyDurable” is the Pi agent running inside a Durable Object. How that works is still the agents deep dive; the talk doesn't explain it.

Evidence: transcript 00:14:06

Onward

She closes by asking for reactiveness. When the hypothesis was wrong, Cloudflare pivoted to serving giant models; when the east wind came, it pivoted again and became a lab.

The reading: this is own the stack from the infrastructure side. Cloudflare owned every layer except the model, and that one gap kept the platform from fitting together until it trained the small models itself. For a client, the lesson is the System One split: not every request deserves an LLM. Decisions and extraction go to small, fast models you can own; the LLM is kept for the work that needs it.

Evidence: transcript 00:13:18, 00:18:56