AI Stops Waiting to Be Asked
Hey there, investors and innovators!
136 startups have now launched in F26. We read the last week of investor launch emails and Bookface digests, and the filter is the one we always use: founder quality first, real traction where it exists, and a space we understand better than most. This week was a product story. The strongest companies stopped building another box to type into. They are building AI that already has the context, does the work, and hands back an outcome.
What the Pattern Says
- AI is leaving the app and moving into the device. Vexo is a woven bracelet with a microphone and no screen. Cyclon rebuilds the phone so that AI is the interface rather than an app on it. Two teams with hardware backgrounds at Tesla, AMD and Google are betting that the next assistant is something you wear or hold, not something you open.
- Small and specific is beating big and general on cost. Kailash Labs says its 2 billion parameter open model, Marlin, matches Gemini 2.5 Flash on dense video captioning and temporal grounding at a fraction of the cost. When a task runs at industrial volume, a model trained for that one job is cheaper than renting a frontier model for everything.
- Outcomes, not software. Ageospatial deploys inside insurers with its own engineers, so the carrier gets priced risk rather than another tool to learn. Vexo books the table and sends the follow up rather than reminding you to. The buyer is paying for the job done.
- Privacy is becoming a feature you can name. Vexo never keeps your audio, and only you see your memories. Kailash Labs builds models that run on premises, where sensitive video already lives. Products that listen and watch now have to say what they keep.
- The uncomfortable one: consumer hardware is the hardest seed bet there is. Two of this week's five are new devices, and the last wave of AI hardware is still recent and expensive history. Two of the five are also solo founders. Conviction here is a bet on people, because there is little yet to measure.
Eight Capital tracks every YC batch, cohort to cohort. Reach out if you'd like a deeper cut on any theme or company above.
Fund Manager Picks
Vexo
An AI bracelet that remembers your day and acts before you ask
Why it's interesting
Vexo is a woven bracelet with a microphone and health sensors, and no screen. It turns what it hears into private memories, then acts on them through the apps you already use: booking the restaurant a friend mentioned, sending the intro you promised, ordering the groceries. It never keeps the audio. Haris Malik worked at Tesla and K-Scale Labs and studied computer science at Toronto. Farqaleet Kiani worked at Tesla, AMD, GM, Geotab and Molex, with mechatronics from McMaster. Vikram Narra came from Amazon and Shopify. Three founders who have shipped hardware and consumer software. Team of 3, founded 2026.
Kailash Labs
A factory for small and specialized video reasoning models
Why it's interesting
The world records more video than machines can read, and frontier models are too costly at that volume and cannot run on premises. Kailash Labs takes a customer's video and goal, distils a frontier model, and trains a small open model that is best at that one task. Its first model, Marlin, has 2 billion parameters, and the company says it matches Gemini 2.5 Flash on dense captioning and temporal grounding at a fraction of the cost. Shubham Sharma was a visiting researcher at KAIST and published a NeurIPS spotlight paper on offline reinforcement learning from work at Tsinghua. Aerospace engineering at IIT Bombay. Based in Bengaluru, team of 2.
Ageospatial
Physical intelligence for every decision with Geospatial AI
Why it's interesting
Insurers price buildings they have never seen. Ageospatial reads satellite and street imagery, maps, public records and the insurer's own data to say what a property really is, what hazards surround it, what changed, and whether it is worth inspecting. It says 97% of commercial properties are never inspected and puts US premium leakage at $36 billion a year. Maaz Sheikh has spent seven years building geospatial software for insurance, energy and governments, and was hired at 19 to support wartime and disaster decisions from satellite imagery. Geomatics engineering. Solo founder, and the engineers deploy inside each carrier.
Melty
Mashup any game
Why it's interesting
Melty lets players combine games into new ones. Its YC page says little beyond that line, so the pick rests on the founder. David Miller is the solo founder and says he previously coached millions of gamers, which means he knows what players want and how to reach them. Gaming is where generative AI meets an audience that already makes and shares its own content, and a founder who already has that audience starts ahead.
Cyclon
A phone where AI is the interface
Why it's interesting
Cyclon is rebuilding the phone around AI. You say what you want and the phone does it, and with one touch you can start agents that control the phone or run tasks in the background. Adam Zinebi worked at Oracle, studied at Berkeley, EPFL and CentraleSupelec, won YC's Call My Agent hackathon against more than 90 teams, and built and sold a music generation business to artists in Morocco and Switzerland. Thomas Kiefer worked on Android at Google and studied at Stanford and ETH Zurich, which is the right background for rebuilding a phone's operating layer. Team of 2, founded 2026.
