THE LABS WON'T WIN THE APP LAYER @mignano (Mike Mignano), General Partner, Union Square Ventures, interviewed by @HarryStebbings (@20VC) Summary: Mignano's argument is that the AI infrastructure buildout is largely finished, and value now shifts to the application layer, the way broadband once gave way to internet apps. He thinks the frontier labs cannot capture that layer, because markets rarely crown a single winner and specialized startups keep beating incumbents at the hard, regulated, context-rich problems. The takeaway for builders: move first, stay mission-driven, and spend tokens like the advantage they are. 1. The App Layer's Turn. The infrastructure is built, and now the applications get built on top of it. Mignano compares this moment to the early internet, when fiber and broadband were laid down and then an application layer arrived to use them. Trillions in value came from the labs' buildout, but the next wave is software, and there will be so much of it that you cannot place a bet unless you know exactly what you are looking for. That is the whole case for a thesis-driven fund over a consensus-driven one. 2. Obliterate, Don't Automate. USV backs companies that reinvent how something works, not ones that make an existing process incrementally faster. The example is Doctronic, which USV seeded on the idea of putting an AI doctor in everyone's pocket rather than helping practices process insurance claims. Automating a workflow usually means selling to a middleman and making incumbents a bit faster. Reinventing the model is where the enormous outcomes live. 3. Token Maxxing. If Mignano ran a startup today, he would still pound the table to maximize token spend on the things that matter, especially coding. A great engineer will pick the startup that says spend whatever you need on frontier models over an incumbent that hands them a constrained budget. Big companies like Salesforce, Microsoft, Meta, and Uber have to rein in spend because they carry tens of thousands of employees; a startup does not. Token spend is an advantage, and a small team should use every dollar of it against a giant. 4. The 3.8% Question. The entire bull case for Anthropic comes down to what share of developer salaries gets spent on tokens. Marc Benioff spent $300 million with Anthropic on his dev team, which works out to roughly 3.8% of those salaries. If that figure climbs toward 20% or 100%, Anthropic is wildly undervalued and its exponential revenue holds; if it stalls or spend migrates to open models, the story changes completely. One ratio decides whether the most valuable private company in the world is cheap or expensive. 5. Frontier Only For Code. Roughly 80% of non-coding enterprise tasks can run on models that are nowhere near the frontier. Summarization, drafting docs, and routine operations do not need the best model; coding does. That split creates room for a routing layer that sends each job to the model with the best price-to-capability fit. Open-source models are catching up fast enough that the frontier is only worth paying for when the work demands it. 6. The Rebel Alliance. Mignano is planting USV's flag in open-weight models, open harnesses, distributed compute, and human-aligned agents. Teams go where the incentives are, and as open options become genuinely competitive, smart teams drift toward them. China's open-source ecosystem is evolving at a startling rate, which pulls even more talent into the open camp. Publishing a thesis like this is a bat signal that tells the right founders who to call. 7. Who Is Your Agent Working For. As people hand agents their credit cards, their messages, and their agency, they will start asking whose incentives the agent actually serves. A lab's model is built to make the lab's model smarter, and a user may want a harness aligned with their own goals instead. Not everyone has to care about this for it to matter; enough people caring keeps a few good actors honest and holds the rest in check. Alignment with the user turns into a product feature and a real reason to pick one harness over another. 8. The 30% Rule. Markets almost never hand one company the whole thing; the winner usually takes about 30% and leaves 70% up for grabs. Coding assistants prove it, with Cursor, Lovable at $500 million in revenue, and Cognition all thriving at once. Anthropic put a whole team on design to go at Figma, and Figma still does billions with a trusted brand intact. Mignano changed his mind on this in the past year: even the biggest labs cannot do everything, just as Google and Apple never did. 9. The Context Moat. The durable advantage in AI products is the context they build up once they are inside an organization. Granola wins by doing one thing, meeting notes, and doing it best, which gets its foot in the enterprise door without asking anyone to rip out Gmail or Docs. Once a company's history of notes lives in the product, nobody wants to give that context up. Being first and staying focused is how a startup builds a moat that even Microsoft's bundling struggles to pry loose. 10. The Energy Floor. No matter which model wins, intelligence runs on power, so USV has been betting on energy since 2021. The portfolio includes Radiant's factory-line small nuclear reactors, Fuse, and Rune's micro data centers that sit next to wind farms to solve energy portability. These bets are capital-intensive at scale but cheap in the earliest days, when a team is running science experiments before anyone else is paying attention. The edge of energy innovation is exactly where a venture investor should place early bets. 11. Founder Over Market Over Product. Mignano used to rank product first; now he ranks founder, then market, then product. Early startups almost always pivot, so what matters most is whether the founder is resilient, can execute, and can adapt. The trait he underweighted is communication, which touches recruiting, fundraising, product vision, and storytelling to the market. A founder who cannot communicate cannot align a team or raise the capital to build. 12. Price As A Litmus Test. Fred Wilson's rule is never pass on price, and Mignano now uses price as a test of his own conviction. For the best founders, you would pay double and still feel good about it in hindsight. His hardest lesson as a former operator was to stop projecting his own plan onto founders, because even when your plan is right, it is their company and betting on your version is how you misjudge the team. The discipline is to trust the founder's judgment, and to let price tell you how much you actually believe.

I've been using Kimi K3 for ~16 hours now. The model is clearly good at a lot of different things (especially frontend), but non obvious reason why people are enjoying it so much is that it clearly does not follow the same rules in terms of safeguards and copyright. Kimi will happily clone MacOSX. If you ask it to help you improve another AI model, it will do it with a smile on its virtual face. Ask Fable to do the same thing? It literally starts to perceive you as a criminal committing a war crime (like no bro, all I want to do is fine tune an open source model). After using all three recent releases, Fable, GPT 5.6, and now Kimi, it's clear that the full power of the models has been significantly held back by the safeguard restrictions caused by last months debacle with the USG -- leading to the top models being quite literally lobotomized in some areas, which leads to subpar results as the safeguards pollute its entire thinking and problem solving abilities. The funny part? Is that you could have predicted this outcome 2-3 years ago when you started to see the rise of Chinese EVs and smartphones compared to western alternatives. They quite literally tried to copy the Tesla Model S and iPhone as hard as possible and then eventually it started to diverge to the point where their EVs and phones are just genuinely better (which is why we have export controls banning their EVs, because they would literally drive all US manufacturers to ZERO) There is a very clear behavior difference in Chinese capitalism and American capitalism. American capitalism tries to protects copyright, patents, etc (oh no, you can't download a book through LibGen, that's ILLEGAL!). Versus Chinese capitalism actually just does not give a fuck. "Hey you want a video gen model (Seeddance 2.5) trained on every single anime ever? And you want the main character to look exactly like Messi? Sure, here you go!" You see what I mean? When one half of the competition is being held up by regulators and restrictions on people who don't understand the technology and the other half has a leader who quite literally today said they are going to set up AI centers around the world to help other countries onboard to their open-source AIs, this is the sort of results that you will start to get. These models were not smart enough to have this difference in philosophy matter -- but the newest class of models is where this difference makes a big deal. If these models are finally at the point where they are smarter than 99% of humans, why would you want to use the American one who tries to impose its world view onto you versus the Chinese one who will just do what you say without asking any questions? And this isn't a full on bullpost on Kimi, the model is clearly not as smart as Fable / GPT 5.6 on things like math and science, but it's lack of handcuffs means that it can show the world what the frontier labs are gatekeeping from you and that starts to build customer resentment and loyalty towards the East, which is probably not what the USG wants. Interesting times. Interesting times, indeed.

Today, we are introducing Inkling. Inkling reasons efficiently across text, image, and audio modalities. We are making the full weights available. Available today for fine-tuning on Tinker. Play with it in the Inkling Playground. 🧵

Today we share the worldview behind our mission. Human values don't average out. Local knowledge can't be centralized. The good future has many AIs, raised in different places, shaped by the people they serve, disagreeing with each other the way we do.