Why Now is the Time for the App Layer | Why Startups Should be TokenMaxxing | Mike Mignano, USV

Summary

• The infrastructure buildout is advanced; the application layer is the trade. The core frame: the lab buildout (OpenAI, Anthropic, xAI) mirrors the fiber/broadband buildout that preceded the internet’s application layer — Harry says “now that infrastructure is built… it’s time for the applications to be built.” Pushed on Mercor’s Brandon claiming infra accrues more value over the next 24 months, Mignano hedges — “the buildout’s not done” — but sees massive value creation in the app layer as well, where so much software will ship that “you’re not really going to be able to make a bet unless you know what you’re looking for.”
• Two possible futures for models. Either a lab hits recursive self-improvement — “once you reach recursive self-improvement, you’re gone. No one can catch up,” and it’s probably an existing lab with the compute, though Thinking Machines or SSI could leapfrog with a new architecture — or AI follows every prior technology onto an S-curve plateau, models commoditize, and the game becomes cost optimization: open-weight models, routing layers, and human-aligned harnesses. He’s explicitly not hoping for the runaway scenario.
• Agents are a trust problem the labs can’t own. Harry argues, “We have never handed over so much of ourselves to a technology before than we’re about to do with agents” — Chrome watched us browse but “it wasn’t a second self.” Harry’s blunt counter: nobody but hardcore technologists cares about privacy, and handing over the keys will be “de facto in 5 years.” Mignano concedes history favors Harry but argues only enough people need to care “such that one player or a few players keep the others in check” — the thesis behind USV’s “Rebel Alliance” of open weights, open harnesses, and distributed compute.
• Startups should be TokenMaxxing while incumbents ration. Harry’s Mark Benoff math: his $300M Anthropic spend is 3.8% of developer salaries — at 20% Anthropic is “grossly undervalued,” at 100% it’s far bigger than anyone thinks, but migration to open models is “a very different game.” Mignano’s split: large companies must constrain token budgets, but “if I were the CEO of a startup right now, I would actually still be pounding the table to maximize token spend” — frontier for coding, Sonnet for ops — and the best engineers will follow the free-for-all budgets.
• Labs won’t eat the app layer — and he used to think they would. His changed mind of the year: “I thought the model providers could do everything… companies can’t do everything,” the same error made about Google and Apple. Moats: the healthcare example’s decade of regulatory clearance (USV, 2018), Granola’s accumulated org context enterprises won’t give up, Figma surviving Anthropic’s dedicated design team. Market structure: the winner takes ~30%, “the 70% remaining is totally up for grabs.”
• USV’s counter-position to missing the model game: energy since 2021. Whatever model wins, an energy layer underneath gets paid — and they’ve learned we need more energy and more portability than anyone thought: Radiant (factory-line small nuclear reactors), Rune (micro data centers beside wind farms). Plus “don’t automate, obliterate” — skip enterprise middleman software, back Doctronic-style full-model reinvention.
• Fund math: ownership early, cash-on-cash late, never pass on price — with nuance. Fred Wilson’s biggest lesson is “never pass on price”; Mignano’s caveat is small funds need a price limit at seed “or we can’t make the fund math work,” while post-Series-B you forget percentages and underwrite $25M to a 10x/100x/1000x (the FOMO deal). The barbell holds everywhere: large consensus platforms (Thrive’s growth numbers beat 90% of seed funds) or small opinionated funds like USV’s $275M core — Harry: you can’t do Series A today without a $400M fund, and the $50-100M middle “will get crushed.”
• Both men confess their worst misses on each other’s winners. Harry turned down Suno’s seed over 1% ownership on a $200-250k check and passed on Granola for “lack of imagination”; Mignano’s regret is Substack, never invested. His two Lightseed hits were opposite motions — Granola a pure founder bet on 15 years of knowing Chris, Suno a pure thesis bet on democratizing music — and Suno at $5B underwrites to “unlimited upside potential” on a new behavior he calls “creative entertainment”: making music purely for the joy of it, like Claude Code or Midjourney.

Deep dive

1. The infrastructure buildout continues — now applications get written

• The frame for the moment: the lab era — “OpenAI, Anthropic, xAI, SpaceX” — was a capital-intensive infrastructure buildout, directly analogous to fiber and broadband preceding the internet’s application layer. Harry says, “Now that infrastructure is built… it’s time for the applications to be built.”
• Harry’s test: Brandon at Mercor claims “much greater value accrue in the next 24 months at the infrastructure layer than at the application layer.” Mignano won’t fully disagree — “we’re not done. The buildout’s not done” — but sees massive value creation at the app layer as well.
• The kicker for fund strategy: there will be so many apps that “you’re not really going to be able to make a bet unless you know what you’re looking for” — which is why he joined thesis-driven USV over consensus-driven scale, even conceding “being consensus driven has been extremely valuable and extremely lucrative” recently.

2. Two futures: recursive runaway or S-curve commodity

• Future one is recursive self-improvement — self-improving AI doing its own research, compounding “on the exponential forever.” His categorical line: “Once you reach recursive self-improvement, you’re gone. No one can catch up.” If it happens, it’s probably an existing lab with the compute, frontier models, and chips — though researchers at Thinking Machines or Ilia and Daniel’s Safe Superintelligence “could leapfrog all of it” with a post-transformer architecture. “That’s not necessarily the future that I am hoping for.”
• Future two: every technology in history follows an S-curve — slow takeoff, exponential ramp, plateau. The self-limit is unknown: transformers may not scale past a point, data may run out, hardware buildout may not keep up.
• If it plateaus, “the technology might start to look somewhat more like a commodity” — labs converge, and competition shifts to price, product experience, and the intelligence stack: open-weight models, routing layers to optimize token spend, and harnesses.

3. Harnesses, and who your agent is actually working for

• His definition, since Harry admits he’s been afraid to ask: a harness is “the application that tightly couples with the model” — Claude Code and Claude Co-work as the flywheel case, plus Hermes (“takes over your Mac Mini”) and Pi from Arendelle.
• The incentive problem behind his “Who is your agent working for?” post: lab incentives are to make lab models smarter — but “if I’m outsourcing all of my agency and all of my personal information and… my credit cards to an agent, I want that thing to work for me.”
• Harry’s pushback, delivered as bluntly as promised: “I don’t think anyone gives a [expletive] about privacy other than hardcore technologists” — we got comfortable with credit cards online and finding spouses online, and handing agents the keys “will be de facto in 5 years.”
• Mignano’s two-part rebuttal, conceding “history says it won’t” go his way: first, Harry says “we have never handed over so much of ourselves to a technology before than we’re about to do with agents” — Chrome learned our interests but “it wasn’t a second self.” Second, not everyone must care: “enough people have to care… such that one player or a few players keep the others in check.”

4. TokenMaxxing: incumbents ration, startups floor it

• Harry’s Anthropic underwriting math: Mark Benoff says he spent $300M on Anthropic, equating to 3.8% of developer salaries spent on tokens. At 20%, “Anthropic’s grossly undervalued”; at 100%, “this is just so much larger than we ever thought”; if it stalls or migrates to open models, “it’s a very different game.” Mignano: “I think it’s a risk.”
• His two schools of token spend: a Salesforce or Microsoft can’t give 5,000-50,000 employees unlimited budgets — hence mentions involving Meta, Uber, and Microsoft. But “if I were the CEO of a startup right now, I would actually still be pounding the table to maximize token spend” — frontier models for coding, “Sonnet or something” for summarization and ops — “I want every advantage I can get against Salesforce.”
• Talent follows the budgets: the best dev picks the free-for-all startup, especially mission-driven ones — “there’s so much money and capital sloshing around… a lot of companies haven’t needed to be mission-driven. I think now they will.”
• On Harry’s scenario of newer models enabling the 100x engineer whose enormous token bill replaces ten mid engineers: engineering orgs will evolve toward smaller teams of higher-caliber engineers, with lower-level tasks delegated to agents — though “I don’t know exactly what it looks like yet.”

5. The open ecosystem and the routing layer’s business-model problem

• His sizing: “80% of non-coding tasks in the enterprise can be done with models that are not at the frontier” — summarization, docs, briefs — and open-source models “are catching up faster than they were previously.” On China owning the best open models: “startups and teams go where the incentives are,” and as the Rebel Alliance gets “a fighting chance,” smart teams shift open. Publishing that thesis is deliberate seed strategy — “you’re sending up a bat signal” to founders who haven’t surfaced yet.
• Routing matters now because enterprises want the right model per task on capability and cost — Open Router in New York is the pure play, and companies across the stack tell USV they plan to monetize their own routing layer.
• Harry’s skepticism — how do you avoid becoming a commodity pipe, and “it’s hard to see that $50 billion company built in routing alone”? Mignano’s most interesting answer is a bounty model: the router collects a fee when it picks the most efficient model — “haven’t necessarily seen it built out yet.” His hedge: infrastructure companies get “deeply embedded in developer workflows and you just can’t rip them out.”

6. USV missed the models — its answer is energy

• To Harry’s “did USV miss the model game?”: USV “was playing a different game.” The standing bet since 2021: no matter which model wins, an energy layer underneath gets paid — and the lesson since is “we only need even more energy than we thought and we need more portability of energy than we thought.”
• The portfolio expression: Radiant, building small nuclear reactors that come off a factory line, among the first to test in the dome in the US; Rune, micro data centers sitting directly beside generators and wind farms to solve portability. Harry adds Boom’s turbines-for-AI opportunity and Panthalassa’s data centers at sea; Mignano describes it as Adam Smith’s invisible hand — “the market solves itself.”
• His answer to capex-intensity: at the innovation edge, “in the earliest days they’re actually not that capital intensive” — the science-experiment stage is exactly where a venture fund should be.

7. Obliterate, don’t automate — and why the labs won’t eat the apps

• The USV filter: “we like to bet on businesses that literally obliterate markets and existing business models” — which is why they mostly skip enterprise automation (“selling to a middleman… making existing businesses faster”). The specimen: Doctronic, whose seed USV led when “AI putting a doctor in literally everyone’s pocket” sounded crazy.
• Against lab intrusion, the healthcare case: nearly ten years of healthcare relationships and regulatory clearance became the moat — “you can’t just walk in the door and say, hey, we’re going to do healthcare now.” USV invested in 2018; the inflection took five-to-seven years. And the romantic version: “what fun can we have as venture capitalists if we don’t believe that startups can take down the Goliath?”
• It’s not binary even when labs attack: Anthropic dedicated a team to design and Figma still does multiple billions in revenue. Harry’s own dev-tools mea culpa — he expected a runaway winner, yet Lovable is at $500M ARR alongside Cognition and the rest. Mignano’s market structure: “usually the market winner takes something like 30% of the market… the 70% remaining is totally up for grabs.”
• His biggest changed mind of the past year: “I thought the model providers could do everything… And I think fortunately what I’ve been reminded of is that companies can’t do everything” — the same thing once believed of Google and Apple, capped by Harry relaying a story from Roryo Driscoll about Microsoft once eyeing becoming a bank. “It works sometimes — AWS. But they can’t do everything.”

8. Granola and Suno: a pure founder bet and a pure thesis bet

• Granola was a pure founder bet: he’d known Chris 15+ years — Socratic’s office sat behind Anchor’s — “I know he can do it. I’ve seen him do it.” The competitive worry is real (OpenAI and Notion both launched directly competitive products), but the defenses are focus — “they’re just doing notes… we’re just going to be your second brain” to get a foot in the enterprise door — and context as moat: once an org accumulates that rich history, “that’s not something you as an enterprise want to give up.” His era-defining line: “this era of building AI products is in many ways about being first and about moving really really fast.”
• Suno was a pure thesis bet from the Anchor playbook — music had never been democratized before AI — so he “hunted down every team” building AI music before the founder lens kicked in with Mikey, a former musician, over an instant-connection dinner in Hoboken.
• Underwriting Suno at $5B: “unlimited upside potential,” benchmarked against the generational platforms that democratized a medium — YouTube, TikTok, Twitter. He initially thought Suno needed Spotify’s creator/consumer two-sidedness and still considers it a platform, but now isn’t sure; he sees it as “literally a new behavior” the team calls creative entertainment — “you’re making music to make music,” like Claude Code or Midjourney, “no greater aspiration for this content.”
• Harry’s confessional counterpart: he passed on Suno’s seed because a $200-250k check meant 1% ownership (“my LPs tell me high ownership”), and on Granola from “a lack of imagination” — both terrible misses, both Mignano wins.

9. Fund math: ownership early, cash-on-cash late, and the Fred Wilson rule

• The stage split: at seed and Series A ownership is important — later rounds get too expensive for a small fund to buy in — but past the Series B/C chasm “forget about percentages”: underwrite $25M as cash-on-cash — “is it a 10x, is it 100x, is it a 1000x.” The live example: the FOMO round (Paul’s next-generation trading app), done with Harry — “for this one, we’re not going to be ownership focused.”
• Series A economics have repriced: rounds are “$80 million to $100 million post… occasionally $150 million post,” and Harry told his LPs “you can’t do Series A unless you have a $400 million fund today.” His companion claim: the $50-100M seed fund is the worst seat — crushed by large funds, too big to be collaborative. Mignano’s structure agrees: be large and consensus, or small and opinionated — “I don’t think you can have a fund that’s in the middle.” Meanwhile “it will be the best time ever for the very large platform funds”: Thrive’s growth numbers beat 90% of seed funds.
• On price: Fred Wilson’s biggest lesson is “never pass on price.” Mignano’s nuance: later-stage market winners allow elasticity, but at the earliest stages “there will always be a limit for us on price just because otherwise we can’t make the fund math work.” Harry’s addition via Peter Fenton — “use price as a litmus test for your conviction”: for an Alan Chang or an Air Wallace Jack, “you could treble the price and I’d still pay it.” Mignano: on his best deals, in hindsight, “I’d probably pay double.”

10. Craft lessons, and traditional media’s obituary

• His formative constraint theory: “great companies, great products, great people come from constraints and failure is the ultimate constraint” — “I did my best work when we were three months out of cash” at Anchor.
• The biggest venture lesson, courtesy of a Jeremy Lou call-out: don’t project your operator ideas onto founders — even if you’re right, “it’s the founder’s company,” and a bet premised on the team executing your plan “is a bad bet.” He’s also fully flipped his ranking to founder > market > product, and traces his worst founder misreads to communication — it touches recruiting, fundraising, product vision, and story.
• The non-AI regret is Substack — self-publishing and “controlling their own destiny in terms of media.” His flat call: “I think traditional media in many respects is dead.” Leaving Spotify in 2022 he thought independent media was “baked” — “it’s gotten so much bigger since then… I don’t even think we’ve reached the peak yet. Television is still in the process of being massively unbundled.”
• Quickfire specifics worth keeping: best first founder meeting — Bren Putnham of Bor (ex-Mirror, sold to Lululemon), “force of nature founder”; favorite seed funds — Matt Hartman’s Factorial (“arms angel investors” like Hugging Face’s Clem with funds on top of their own capital) and Haystack; highest-signal deal-sender — Nat Friedman, which is how he met Suno’s Mikey; favorite growth fund — Lightseed Growth (Anthropic, xAI, SpaceX).