Hi friends! If you’re receiving this, it’s because we’ve either: worked together, co-invested together, or connected over the past few years with the intention of a potential partnership.
We consider you a friend of Essence and would love to keep you looped in on how we interpret what’s happening in the industry and across our portfolio.
However, if you’re not interested in receiving these emails, no worries - feel free to hit unsubscribe below.
With that out of the way, we have a lot to fill you in on!
Technical Insights
The first wave of AI training data was human judgment, billed by the hour. Once the labs’ insatiable appetites finished consuming the entire open web, they started paying doctors, lawyers, and PhDs to write, grade, and correct model outputs, and that business sector got very big very fast. Mercor crossed $2B in gross annualized revenue in June. Surge AI is past $1B with something like 110 people. Scale AI was marked above $29B after Meta’s investment, and Handshake AI hit a $1B gross run rate early this year.
Those numbers are gross, with 60–70% passed through to contractors, and most of it comes in as purchase orders against deliverables, not recurring contracts (Vikram Aditya). What caught our attention more was Handshake buying Cleanlab in January. Cleanlab finds bad labels without putting a second human on them. When each example costs this much, being able to prove a batch is good starts to matter as much as producing it.
The data itself is changing too. Models are being trained to act rather than answer, and RL pays for outcomes. So the unit that matters is less a label or a preference pair and more an episode: an agent does something in a world, the world responds, and there’s a verifiable signal for how it went. It carries its own grading and it doesn’t run dry because better models generate harder episodes. SF Data is chasing the same thing from the other end, buying production traces from companies whose agents already do real work and turning them into training data.
Games are a clean way to get there. The engine is the referee, so the reward can be trusted, and a single match produces hundreds of decisions with known consequences. Good Start Labs, an Essence portfolio company spun out of Every, builds game environments and sells agent trajectories and full learning environments to frontier labs. The first one was AI Diplomacy, where o3 won by planning a betrayal and Claude Opus 4 lost because it wouldn’t lie. Try getting that negotiation and theory-of-mind behavior out of a standard rubric. Through partners like Bad Cards, real players now grade model humor as part of play, with 14,000+ hands played so far.
Early evidence says the skills carry over. Training on Diplomacy improved a model on customer support and industrial operations benchmarks. The result we keep coming back to is 1830, a railroad game with a stock market in it. Good Start trained a 30B model on it two different ways, and only the model that framed it as a multi-turn, tool-using agent improved on the Finance-Agent benchmark (results). Same game, different presentation, different model. It’s early, but that points to environment design as the hard part, and the defensible one.
Embodied AI is the other place this is moving fast. Robots have no internet to pretrain on, so every usable hour has to be captured by someone. WorldEngine reportedly went from $1M to $100M ARR, with eight frontier lab customers and 250K hours of humanoid data a month, recorded through head-mounted cameras, sensor gloves, and people remotely operating robots. XDOF raised $70M for robot data pipelines, Midcentury came out of stealth claiming 2M+ hours of first-person footage, and Worldmodeldata is licensing gameplay from Unreal and Unity titles as action-and-consequence data for world models. Worth being precise about what a robotics example actually is. Not just a video, but video plus depth, pose, and touch, all aligned frame by frame with the action taken and what happened next. A game engine emits that same structure natively, which is why we treat interactive environments and physical capture as one market.
Our view is that the data that will matter over the next few years will be scarce, verifiable by design, and interactive. Environments that keep producing harder tasks, physical capture that nobody has yet, and measurement that shows a batch actually moved the model are where we expect the action to be. The expert marketplaces aren’t going away, but the next crop of data companies will look more like simulation and capture infrastructure than staffing firms, and should get paid like software. Good Start Labs and EQ Co (still in stealth, more soon) are our two bets. If you’re building here, we’d like to hear from you!
Events Recap
In September we cohosted an Infra Demo Night with Propeller and True Ventures. We packed a room in Jackson Square to watch five infra startup founders (four of whom were through Essence: CocoIndex, Letta, Restate, and Vers) plug in their laptops and live demo what they've been building!


We also cohosted another Devtool & Infra Demo Night event, this time with PostHog, CTO Fund & 1984 Ventures. We got to see a wide variety of demos, ranging from platform engineering to codebase context to microVMs. From our portfolio, Insforge, Leverage, and OneWill presented.


Last but not least, we had a great time at the AI Infra Hot Coffee, Hot Takes event cohosted by Vermilion Cliffs Ventures and Entrepreneurs First at the latter’s office in SF. The event format was super fun; each speaker (including our GP Tim Chen) brought their hottest take, the audience got to push back, and at the end everyone walked away with something new.


We hope to see you at an upcoming event, listed at the bottom of the newsletter!
Portfolio Company Updates
Modal has closed a $750M round led by Accel at a $15.75B valuation, more than triple the $4.65B it was marked at in May. The company just passed $1B ARR and counts Cognition, Suno, Ramp, and Substack among its customers. Congrats to Erik, Akshat, and the team. Read more here.
Turso is being acquired by Supabase, which announced the deal alongside a $150M raise led by GIC with CapitalG, IronArc, and SquarePeg, just four months after Supabase’s $500M Series F. Turso’s SQLite-based architecture provisions databases with minimal overhead and on demand, which is what Supabase needs now that it’s adding 4M databases a month with 70% of those created by agents. Read more here.
Restate raised a $20M Series A led by Singular, with Redpoint and Capital One Ventures. Restate is a durable execution runtime from the team behind Apache Flink; the argument is that durability should be cheap and fast enough to sit inside an agent loop, not just around it. Replit moved Replit Agent’s execution onto Restate and now uses more than 10x the durable actions of its previous architecture. Read more here.
Insforge raised an $8M seed and launched InstaCloud, a serverless platform that lets coding agents like Claude, Codex, and Cursor provision databases, storage, and compute on their own, with sensitive operations gated behind human approval. CEO Hang Huang shares that they’re there to “kill AWS, GCP, and Azure.” Read more here.
DevTools Worth Trying
Atmee is building a real-time human foundation model for video agents. Give it one photo and a ten-second voice clip and you get a live avatar that holds a face-to-face conversation, with the breathing, hesitation, and imperfect delivery that make it read as a person rather than a render. Their latest demo is Natalie, an AI recruiter running a screening interview, on a single GPU. Watch it here.
Magic Patterns just launched MagicOS, an infinite canvas where your PRD, inspiration, design system, variants, and rejected directions all sit side by side instead of buried in a chat log. Ask for a few directions, compare them on the canvas, then hand the winner to Claude, Cursor, or Codex over MCP. It also imports from Figma and your codebase and takes voice input. Try it here.
Nilo passed 100,000 users on its browser-based AI game engine, which turns text, sketches, and images into playable multiplayer 3D worlds with no download or coding, and exports straight to Roblox. Check it out here.
Podcast Episodes
Here are our most recent podcast episodes from The Infra Pod with Dave McJannet, CEO of Dome Systems and former HashiCorp CEO, and the Open Source Startup Podcast with Ben Vinegar, cofounder of Modem:
Open Jobs
Nilo, as mentioned earlier, passed 100,000 users on its browser-based AI game engine. Things are continuing to accelerate for the incredibly talented team, and they’re hiring engineering and design roles. See more details here.

Many of our other portfolio companies are hiring across similar roles, check them out here! We are continuously updating this list.
Upcoming Events
ComfyUI is joining a panel at LA Tech Week today (Thursday, Oct 15) on how generative video is changing commercial and entertainment work. Moderated by Tool president Dustin Callif, with food and drinks for all. RSVP here.
LanceDB is hosting Reverie, a one-day summit on the data systems underneath world models, generative video, and physical AI, on Thursday Nov 5 in SF. Speakers from NVIDIA, Runway, Microsoft AI, Applied Intuition, Zoox, Adobe, and Exa will walk through the pipelines behind their training runs, what broke, and what they’d do differently. It’s capped at 200 and application-only. Apply here.
Clarify is hosting Horizons, a one-day GTM conference for B2B founders and revenue leaders, on Wednesday Dec 9 in Seattle. They’ll be hosting a number of GTM sessions including Yoodli’s 0-to-$30M growth story, a panel on deploying GTM engineers, and a headline Clarify product announcement before happy hour. It’s free, capped at 150, and seats are by request. Grab one here.
Thanks for reading! We’ll be back next month with more updates.
— The Essence Team

Asim Moinuddin, Tim Chen, Naomi Walker-Garrett


