Hi friends! Welcome back to another installment of our newsletter, we have a lot to fill you in on.

For most of the last cycle, model providers treated infrastructure as part of the moat. The API was the product. Orchestration, routing, fine-tuning, the dev surface, etc. was supposed to live downstream of the model, owned end-to-end by the lab that trained it. That has broken down faster than almost anyone expected. The ecosystems pulling ahead right now are the ones that gave up trying to own the full stack and let it go open.

ComfyUI is the cleanest example. It started as a node-based UI for image generation and turned into something much bigger: a programmable layer for stitching open models, proprietary APIs, custom inference, and increasingly agentic steps into workflows you can actually reuse.

The point was never that it made Stable Diffusion easier to run. The point is that as model capabilities commoditize (faster than we can keep up with), the leverage moves up to the control plane sitting on top of them. ComfyUI is that layer for a meaningful slice of the generative ecosystem. They have 60,000+ community nodes and a workflow surface that has quietly become the default place open-model experimentation actually happens. That kind of adoption is hard to manufacture and harder to displace.

The tell is that the model providers themselves are leaning into this rather than fighting it. Weights are getting released. Tooling is getting more interoperable. APIs are being designed to plug into orchestration systems rather than wall them out. The reason is simple: distribution has moved into the ecosystem, and an endpoint that does not compose well with the rest of the stack is a worse endpoint, regardless of how good the underlying model is.

OpenRouter is the same pattern on the language side. Instead of betting on a single winning model, OpenRouter abstracts model access entirely and lets developers route across hundreds of models and providers based on cost, latency, and capability. Open-weight models are taking a meaningful and growing share of production traffic on the platform. Developers are not picking a team, they are optimizing.

We think this is the bigger industry shift, and the one most people are still underweighting. The center of gravity in AI is moving from vertically integrated model companies to the open coordination layer sitting above them. Every previous platform cycle has gone the same way; the winning infra came from the open ecosystem, because open ecosystems iterate faster and accumulate developer surface faster than any single closed system can. Open infrastructure is now the connective tissue for AI-native teams and companies, and increasingly, it is where the real platform value compounds.

Warp is now open source with OpenAI as the agent sponsor! The bottleneck in shipping Warp isn’t writing code, it's now speccing and verifying it. Agents can do the implementation now (workflows in Warp now run GPT models), so the team is opening up to let the community focus on the higher-leverage work and deliver more business value. Read more here.

ComfyUI raised $30M at a $500M valuation, and the interesting thing is what the round implies about where generative AI is splitting. Closed prompt-based tools (Midjourney, DALL-E, Sora) won consumer mindshare, but the professional layer (VFX houses, ad agencies, game studios) has consolidated almost entirely on ComfyUI's node-based workflows because working creatives need reproducibility and granular control. 4M users, 60K+ community nodes, 150K+ daily downloads, and the first primarily AI-generated Super Bowl ad (SVEDKA, via Silverside) all ran through it. Read more here.

InsForge is an open-source backend platform built specifically for AI coding agents, exposing databases, auth, storage, edge functions, and a model gateway through a semantic layer (and MCP) so agents can provision and operate the entire backend without a human clicking through a dashboard. Read more here, or try it out!

CocoIndex is a Rust-based incremental compute engine that turns codebases, PDFs, Slack, inboxes, and meeting notes into continuously fresh context for AI agents — re-running only the delta when sources change. Their CocoIndex Code MCP server gives Claude Code, Cursor, and Codex AST-aware semantic search that cuts ~70% of tokens per turn. Check it out here, or try it out!

Here’s our most recent podcast episodes from The Infra Pod with Benny Chen from Fireworks AI and the Open Source Startups Podcast with Batuhan Taskaya from Fal:

Cline has continued to see adoption grow after rocket ship growth last year:

To support their continuing growth, Cline is currently hiring across engineering, product, marketing, and more. See more details here.

Many of our other portfolio companies are hiring across similar roles, explore more here!

Upside is hosting AI for GTM tonight (Tuesday, May 12th) from 6-9pm PT. There are no panels and no pitches, just operators opening their laptops and walking through the AI workflows they actually use day-to-day. Watch the livestream or join in person in SF, with operators from Cursor, Vanta, Zapier, Exa, and Wonderschool. Expect demos on lead-leakage agents, an AE companion agent, a GTM state machine, and much more. Register here.

Essence VC is co-hosting Infra Demo Night next Tuesday (5/19) in Menlo Park alongside Propeller, Mozilla VC, and 1984 Ventures. We're bringing together the Bay Area's infrastructure software startup community for food, drinks, and live demos. Founders will demo in front of technical leaders including Gagan Kanjila (ex-Capital One, SVB, OnDeck) and Yazan Risheq (Google) for candid feedback. We'd love to see you there!

Obvious is running Frontier Build, a 5-day intensive (2 days remote, 3 days in person) powered by Autobuild. Two of your engineers work alongside the team that built it to ship roughly a quarter of your roadmap in a week. Autobuild decomposes a spec into a dependency DAG and runs parallel agents that build, test, and merge PRs autonomously, with humans reviewing at the release-branch level. See upcoming sessions at obvious.ai/autobuild.

Thanks for reading! We'll be back next month with more updates.

— The Essence Team