Narrow Beats Broad
Hey there, investors and innovators!
121 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. What came through this week was not a sector, it was a shape. The companies with something to show are the ones that picked a narrow problem and went all the way down it, and the ones still describing a category are the ones without a demo.
What the Pattern Says
- The narrow products are the ones with customers. ByteAsk does one language family, C and C++. DeepMark does one layer of one medium, identity inside AI voice. Both have something concrete to point at. The companies in this group describing a broad platform have launch copy and little else. Narrow is not a smaller ambition, it is a shorter path to someone paying.
- Verification is becoming the product, not the feature. ByteAsk does not stop at a diff that looks right: it builds, debugs and tests every change against the real toolchain. DeepMark does not ask whether audio is fake, it answers who is authorised to speak. As generation gets cheap, the thing worth selling is proof.
- Research done before the batch is the hardest traction to fake. DeepMark's founders spent six years on audio watermarking at the University of Montenegro, published some of the earliest work in deep audio watermarking, and say it is cited by teams at Meta, Microsoft, Sony and Adobe. They signed their first platform licence with a voice AI provider before YC, and say it is in production. Nobody assembles that in eight weeks.
- Regulation is setting the clock in at least one category. DeepMark points at the EU AI Act marking rules and a 2 December deadline for systems already on the market. A deadline someone else set is the most reliable demand signal a seed company can have.
- The robotics money is moving to the layer between the model and the job. Deploy's pitch is that getting a generalist robot to do real work is the same problem in every industry, and that it builds everything between the model and the task. Nobody in this group is training a new robot foundation model. They are making the existing ones useful, which is where the revenue is.
- The uncomfortable one: credentials are doing a lot of work in the thinner pitches. Two of this week's companies lead with where the founders studied rather than what they have built, and one of those is running two products at once. Pedigree is not traction, and a batch that rewards it teaches the wrong lesson.
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
DeepMark
A live watermark inside AI voice that survives phone calls
Why it's interesting
DeepMark embeds a machine readable identity into the audio itself as an AI agent speaks, not into metadata, which is stripped the moment a file is touched. The mark survives compression, codecs and a bad phone line, and resolves to a trust record a bank or a call centre can check in real time. Slavko Kovacevic and Kosta Pavlovic worked together as researchers at the University of Montenegro for six years before founding it. They say their robustness benchmark landed in IEEE Access this year and their first platform licence with a voice AI provider is already in production. Team of 5, founded 2025.
ByteAsk
The AI coding agent for C and C++
Why it's interesting
ByteAsk writes C and C++, the code inside trading systems, cars, chips and operating systems, and it refuses to stop at a diff that looks correct. It builds, debugs and tests each change with the real toolchain until the change holds, runs from the terminal or inside VS Code, Neovim and Emacs, and is available fully on premises, which matters to the customers who own this code. Anirudha Kulkarni wrote low latency C++ at Optiver and Quantbox. He and Pratyush Saini previously founded LawSutraAI and sold it to Manupatra, India's largest legal tech firm, where they say it serves over 100,000 lawyers. Both studied at IIT Delhi.
Deploy
The generalist robot deployment company
Why it's interesting
Deploy's argument is that the hard part of robotics is no longer the model, it is everything between the model and the task, and that this problem is the same across industries. So it sells deployment rather than intelligence. Sahith Reddy Chada was a research engineer at Honda Research Institute building vision language action models and world models for robotics and autonomous vehicles. Irene Grace Karot Polson was an autonomy engineer at AeroVect deploying self driving vehicles at airports, after flight data work at Airbus. Both came through IIT Kanpur, with masters degrees from Minnesota and Penn. Team of 2, founded 2026.
Nxtcure Labs
The full-stack AI CRO for drug discovery
Why it's interesting
Nxtcure is building an AI native contract research organisation to carry therapeutics from discovery to market, and alongside it a consumer mobile product guiding cancer patients through treatment. The CRO market is large, slow and genuinely ready for this. Alessandro Hammond graduated Harvard in 2025, Phi Beta Kappa, is a Schwarzman Scholar at Tsinghua and is doing an MD/PhD at Yale. Keanu Clark is MIT computer science, 2025, previously a researcher at the Sharp Lab. The credentials are exceptional. The open question is scope: a full stack CRO and a consumer patient app are two companies, and the page gives no evidence of either one running yet.
Vyra
Autonomous video editing for non-editors
Why it's interesting
Vyra removes the timeline. You upload footage, describe what you want, and get a finished edit back in minutes. Caleb Pong was a freelance video editor and Sulan Zhang is a creator with more than 60,000 followers who studied computer science and visual arts at Brown, so the team knows exactly who it is selling to. That is also the risk: this is the most contested product category in the batch, and taste is the only moat on offer. Worth watching for whether the edits are good enough that someone pays, which will be visible fast. Team of 2, founded 2026.
