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

Robots have no internet to pretrain on, so every usable hour of data has to be captured by someone. Over the past year that has made robot data vendors some of the fastest revenue ramps in any category. Mecka closed a $60M Series B led by Sequoia in October, with NVIDIA, Microsoft’s M12, Qualcomm Ventures and Samsung coming in as new investors, at the roughly $500M valuation TechCrunch had reported earlier, and says it was running above a $100M annual run rate as of June. microagi says it reached a $100M annualized run rate six months after launching Shift, a company figure with no disclosed calculation behind it. XDOF was in late-stage Series B talks at about $1.2B in early September, three months out of stealth. On the model side, Physical Intelligence is reportedly in talks at close to $20B, nearly double its July mark; Skild raised $1.4B at over $14B in January and said in September it had passed a $100M run rate ten months after its first commercial deployment; Generalist extended its Series B to $600M at $3B in August.

The implicit bet behind all that capture is the LLM playbook: find the scaling law, pour in more hours, wait for generalization. The infrastructure underneath that bet, the part that decides which hours are worth capturing at all, strikes us as the most interesting and least built-out piece of the picture.

Start with the fact that a token isn’t an action. There’s essentially one way to read a paragraph. There are a hundred ways to pick up a water bottle: the grasp, the approach angle, the lighting, the surface, the weight, what’s next to it. Robotics data has enormous variance along dimensions text never had, so a million random hours doesn’t map cleanly onto the task your robot actually has to do.

The scaling evidence is real but early. Dyna-2, announced in August, is trained on over a million hours of data and claims the first human-to-robot scaling law, with performance improving predictably as human video is added. Dyna’s stronger claim is that scaling human pretraining data also improves zero-shot evaluation on robot data the model has never seen, and that this carries to on-robot evaluation after post-training across embodiments. Generalist has published in the same direction. These are the best public results anyone has, and they are also self-reported, unreplicated, and new enough that nobody outside the labs has tested them on their own robots. Two pressures make this matter now: capture supply is exploding, and compute is getting harder to come by. Both mean you have to be far more deliberate about what you train on.

The other complication is that the only eval is the robot. In LLM land you can run an eval or an RL environment ten thousand times in sim. In robotics, you put the policy on the robot, put the thing in front of it, and see if it works. That changes the shape of the data business. The labs aren’t buying hours. They’re buying fixes for last week’s failures, on a weekly training cadence. Robot foundation models are small next to frontier LLMs (single-digit billions of parameters), so teams retrain constantly and need the targeted data delivered before the next cutoff.

All of which argues for selection over volume. Eventual, the team behind the open-source Daft engine (and an Essence portfolio company), is built around this idea. Cofounders Sammy Sidhu and Jay Chia met at Lyft Level 5, building perception and ML data infrastructure for self-driving. Sammy was Chief Architect at DeepScale before that, through the 2019 Tesla acquisition. They’ve spent years on the problem robotics teams are now running into: petabytes of sensor logs, and finding the five minutes that actually matter. Eventual’s product does scenario mining and visual understanding over robot video and telemetry. It answers questions like “show me every failed grasp on a deformable object,” and then curates and versions datasets from the answer. Its open-source daft-physical-ai library and its writing on hindsight labeling are worth a read.

Sammy’s framing is “luxury data”: lower volume, much higher value. A video of someone making a hot dog is cheap and getting cheaper. What it feels like to squeeze one, the forces on each finger, is not. As egocentric video commoditizes (Build AI released roughly a million hours openly in April, after going from 10K hours in November 2025 to 100K in December), the value moves to sifting, targeting and the hard-to-capture modalities like tactile data.

Testing is the other half of the loop. Data tells you what to train on. Testing tells you whether the change broke anything, and for most hardware teams that still happens in a warehouse staged like a movie set, with fake rooms, actors and vehicles, every time a revision ships. It’s slow, it’s expensive, and it covers a sliver of what teams want to test. Snow and rain mostly don’t get tested at all. The pitch for simulation here isn’t a better physics engine. It’s CI/CD for hardware: bring the real sensors and environment into the cloud and test every change. Most mobile robots are, at bottom, moving perception systems, so sensor simulation built for cameras, lidar and radar carries straight over to embodied robots. The history is a warning, though. Sim testing companies rarely struggled to land a first big contract. Repeatability is what got them, with long sales cycles and scenarios that change city by city. The teams that got traction picked scenarios matching where a robot actually operates and built on existing sim tools instead of rebuilding the engine.

Meanwhile the tooling layer is consolidating fast. Nominal announced a $240M Series C at a $3B valuation on Nov 9, led by Lightspeed and co-led by Insight, eight months after it first crossed $1B. It started in test and evaluation for defense and aerospace, now counts more than 80 commercial and government organizations including five of the seven largest defense contractors, and has been buying its way outward, acquiring Fid Labs for its agent team and Pull Systems to push into automotive. Foxglove has moved in the opposite direction, from a visualization tool into an agentic data platform. Its August release added an agent sidebar and MCP server for working with robot data in natural language, plus Semantic Search built on NVIDIA’s Cosmos video-text embedding model for finding behaviors in unlabeled data, all on top of a $40M Series B from Bessemer last November. Add Rerun, Voxel51, Encord, and NVIDIA’s Cosmos and Isaac stack, and the old lines are blurring. Development-time tools are moving into production monitoring, and observability tools are moving upstream into curation. The more useful split may turn out to be environment rather than lifecycle stage. Robots in controlled settings like warehouses and factories can lean on statistical, telemetry-driven reliability. Robots in open-ended public settings will keep paying seven figures for full-blown scenario simulation. Foxglove is a good reminder that this market rewards patience: it found its footing as robot data volumes finally got big enough to need it.

Robotics is splitting into “robot brain” companies that train generalist models and “robot deployment” companies that ship robots into specific environments, and many deployers send their data to the brain companies instead of training their own. The data needs look like they’re getting more diverse, not less: new environments, new failure modes, new modalities. What we’re curious about is what happens as specialized and generalized deployments both reach real production. Does the industry move faster once robots are learning from actual work instead of staged demos? That seems to come down to the infrastructure that decides which data matters, measures whether it worked and closes the loop every week, which is the part of this we actually know something about. If you’re building there, or you’re a researcher who wants to swap notes, we’d love to hear from you.

Events Recap

We cohosted the Data Outpost Kickoff Party with MotherDuck, Theory, and Zetta on Nov 3rd in SF, bringing together founders, engineers and data leaders the night before the conference opened.

Data Outpost is the next evolution of Small Data SF, two days built around agentic data infrastructure, AI interfaces and workflows, memory and provenance as trust infrastructure, and building with taste.

We hope to see you at an upcoming event, listed at the bottom of the newsletter!

Portfolio Company Updates

Trajectory launched its post-training platform and raised a $60M Series A led by Sequoia at a $300M valuation, with participation from Nvidia, Bessemer Venture Partners, and Bain Capital Ventures. They automate post-training so lawyers, doctors, and financial analysts can keep improving an open-source model from their own feedback instead of hiring a research team. Harvey, Rogo, Decagon, and Mercor are already customers. Congrats to Ronak, Michael, Arjun, and the team. Read more here.

Phonic announced a $27.5M Series A led by First Round, with participation from Lux Capital, Felicis, and Definition, and launched Phonic-1 alongside it. The speech-to-speech model took the top spot on the Artificial Analysis Speech to Speech Index, with Phonic being the only model that can call tools while the caller is still talking. Phonic now powers calls on six continents for companies like Ashby, Maven AGI, and Rho. Congrats to Moin, Nikhil, and the team. Read more here.

Letta released FrontierMemory, a benchmark that scores agent memory by how much it helps on later tasks instead of how well it recalls past facts. Agents study trajectories of past work and write memory from them, and that memory is graded on whether it helps another agent solve coding and knowledge-work tasks. Read more here.

Fearn launched The Thicket, a weekly digest that reads the patent record for you. Follow companies, technologies or single applications, and every Monday you get what moved, from new filings to the prototypes that never left the lab. Free while in beta. Start a watchlist here.

Unify released Continual-ARC, a benchmark for whether an agent gets better with experience or just remembers. It serves the same ARC puzzles again with fresh inputs, charges for every example and wrong answer, then tests the frozen learner on puzzles it met and 100 it never met. Harnesses that keep notes (OpenClaw, Hermes Agent, Prime Agent) saved nothing measurable as shipped, and only training the model’s weights got it better at puzzles it had never seen. Read it here.

Products Worth Trying

InstaBox, from Insforge, is out of beta and open to everyone. Each agent gets its own Linux machine in the cloud, a 4 vCPU microVM with a persistent disk, Docker, a desktop and Chromium, started with one command. You can fork a running box in seconds, give a port a public HTTPS URL, and run Android 14 in a box with adb. The team reports more than a million boxes created during the first month of beta. Try it here.

Trajectory, mentioned above, has opened up its platform for post-training your own models. A few agent prompts get you a full suite to train, evaluate, and deploy a model, and a research preview called continual grading lets the model learn what good looks like from far more examples than you could grade by hand. Join the waitlist here, and see Ronak’s post here.

Phonic, also mentioned above, released their new speech-to-speech model Phonic-1. It is trained on 50+ languages, can switch languages mid-conversation while keeping the same voice, and in one demo carries a 44-minute hotel booking call with 123 tool calls. New signups get $250 in free credits through Nov 30. Try it here.

Atmee shipped v2 of its real-time avatar model, with API access by request. One image and ten seconds of audio produce a character that can run as a live video call, voice only, or generated video, with expressions and gestures that respond to the conversation. The team reports it runs up to 50x faster and 10 to 100x cheaper than video models like Seedance 2.5 and Wan 3.0. See the launch here.

Phota is now live on the App Store! It's an always-on AI photo assistant for iPhone that picks your best shots, fixes them in one tap (open eyes, relight, straighten, clean background), and builds albums after a day out without being asked. TechCrunch covered the launch and how co-founders Cecilia Zhang and Zach Xia built their own personalization model so edits keep people looking like themselves, here. Here’s an example of before and after of Essence’s resident puppy Maeve’s first birthday! Download the app here.

Podcast Episodes

Here are our most recent podcast episodes from The Infra Pod with Dome Systems CEO Dave McJannet (ex-CEO Hashicorp) and the Open Source Startup Podcast with Semgrep CEO Isaac Evans.

Open Jobs

Magic Patterns launched MagicOS last month, an infinite canvas where your PRD, design system, and design variants sit side by side. The team is now hiring across engineering, design, growth, and enterprise sales, with roles in San Francisco and remote across the US. 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

Flarion is taking its drop-in Rust engine for Spark on the road. In New York the team is hosting Flarion After Hours, a private dinner at COQODAQ on Wednesday Nov 18, then running a booth at Data + AI World Tour at the Javits Center on Thursday Nov 19. In Las Vegas they're opening The Bodega, an invite-only suite at the Palazzo with food, coffee, and a spot to watch the keynote, from Monday Nov 30 to Friday Dec 4 during AWS re:Invent. Find your city here.

Workbrew is joining the Los Angeles Mac Admins Meetup on Thursday Nov 19 at Highland Park Brewery. Come meet the local Mac admin community over a few beers and hear what's new in Workbrew 2.0. Grab your spot here.

Crash Override is cohosting Cloud & Cocktails during AWS re:Invent on Tuesday Dec 1 in Las Vegas, alongside Lucid Software, Cortex, incident.io, Miggo, and Tailscale. Expect drinks, food, and free games at PLAY Playground inside the Luxor from 6 to 8 PM, with people from platform engineering, DevOps, cloud architecture, and incident response. RSVP 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! If you’re looking for other great sources of technical insights, RadiantOps publishes Far Field, a weekly roundup of the sharpest practitioner writing on where software is heading. Recent issues ask who reviews the code now, whether your agent fixed the code or the test, and who maintains what an agent wrote in an hour. You can subscribe here.

We’ll be back next month with more updates.

— The Essence Team

Forwarded this email? Subscribe here

Asim Moinuddin, Tim Chen, Naomi Walker-Garrett