Now is the Time for the App Layer | OpenAI & Anthropic Won't Win the App Layer | Mike Mignano, USV

Why does it matter? Because the app layer window is open right now, and most founders are still a cycle behind.

The most expensive belief a founder can carry today is that OpenAI or Anthropic will inevitably eat the application layer. Mike Mignano — USV's newest general partner and the founder who sold Anchor to Spotify — has watched this exact fear cycle play out before, with Google, with Apple. He's done being afraid of it. The AI infrastructure phase is over. The application phase has started.

• The AI infrastructure buildout mirrors the internet's fiber/broadband era — the next massive wave of value gets built at the app layer, not beneath it. • Agents will be your second self, buying things and sending personal messages — "who is your agent working for?" is the defining question of the next decade. • Startups can out-spend constrained incumbents on frontier tokens right now — that asymmetry is both a product weapon and a recruiting signal. • Labs won't win the application layer; specialists have always found a way through, and the historical base rate hasn't changed.

The AI infrastructure phase is over — the app layer window is open right now, not in three years

Now is the moment, not the warm-up. "We are coming out of a time in the market where we have undergone a massive infrastructure buildout in AI." OpenAI, Anthropic, xAI — vast capital, magical technology, and that phase is largely complete. What comes next maps directly onto a pattern Mignano has watched before: "It reminds me a lot of the early days of the internet when we built out fiber and we built out broadband... now it's time for the applications to be built."

The broadband era didn't reward the DSL providers. It rewarded Google, Amazon, and Facebook. The category-defining returns sat one layer above the infrastructure. When Brandon at McCormick argues the next 24 months still favor infrastructure, Mignano's counter isn't a denial — it's a timeline separation. Both can be true simultaneously. But for founders choosing where to build, the infrastructure window is narrowing. The application window is widening. The companies that move now own the compounding advantage.

Agents will be your second self — and the 'who is your agent working for?' question will define a new class of startup

Every technology before agents had a ceiling on intimacy. Chrome learned your browsing habits, "but it wasn't out buying stuff for us... it wasn't out sending very personal messages to family members and loved ones. It wasn't a second self." Agents cross that line — autonomously, in real time, with your credit cards and calendar and relationships.

Mignano published a piece titled: "Who is your agent working for?" For most agents being built today, the answer is the lab. Lab incentives are to make its models smarter, better, faster. Those aren't your incentives. "I want that thing to work for me. I want that thing to be completely aligned with my incentives."

His bet at USV is the "Rebel Alliance" — open-weight models, open-source harnesses, human-aligned agents. He doesn't think every agent needs to be explicitly aligned; he just thinks enough players have to care that market forces keep the rest in check. The startup opportunity lives precisely in the gap between what labs are building and what users will eventually demand.

Startups should be maxing frontier token spend — incumbents are being forced to cut back, and that gap is a structural weapon

Salesforce, Microsoft, Meta — they're already constraining developer token budgets. Too many employees; the math doesn't work at scale. "If I were the CEO of a startup right now I would actually still be pounding the table to maximize token spend." For coding especially, go to the frontier. "As a startup you need every advantage you can get right now."

Mark Benioff reportedly spent $300 million on Anthropic for his dev team — 3.8% of developer salaries on tokens. The question of what happens when that moves toward 20%, or 100%, is where Anthropic's revenue trajectory lives. But that's a future problem. The present one is the gap.

The talent angle sharpens it further: "If you're the best dev, are you going to go to a big incumbent who are going to give you a budget and really constrain your abilities in terms of model usage or are you going to go to a startup where fundamentally they say, 'Hey, it's a free-for-all. Be your best self.'" Unconstrained frontier access is the recruiting pitch incumbents cannot currently match.

OpenAI and Anthropic won't eat the app layer — we said the same thing about Google fifteen years ago, and Google couldn't do it either

Fear is causing founders to build defensively or not build at all. Mignano used to share that fear. Then he remembered: "10 years ago, 15 years ago, we thought the same thing about Google. We thought the same thing about Apple. We thought these companies are going to just take everything. And the reality is they can't and they don't."

The argument isn't optimism — it's pattern recognition. Regulated industries build moats through years of relationships and compliance that trillion-dollar balance sheets can't shortcut. Deep vertical context accumulates in ways horizontal labs can't replicate without spending the same time. Spotify went up against a behemoth and could have been killed fifty different ways — it won by specializing early.

The market winner in any given category typically captures around 30% of the market. The remaining 70% is live. If the market is big enough, that remainder produces generational companies. Building in fear of lab encroachment means surrendering that 70% to someone less afraid.

USV's filter rejects most enterprise AI being built right now: don't make the old model faster — obliterate it

Most enterprise AI makes existing workflows faster. Useful. Not fundable, in Mignano's framework. "We like to bet on businesses that literally obliterate markets and existing business models. We want to invest in businesses that literally reinvent the way something is done."

The test is blunt: does your product make the incumbent faster, or does it make the incumbent irrelevant? USV backed Doctoronic — not AI to make medical practices more efficient, but a doctor in everyone's pocket, bypassing the medical practice entirely. "We can put doctors in everyone's pocket and totally reinvent the model with AI."

Every vertical has a professional gatekeeping access to a service. The obliterate thesis asks whether AI can bypass the gatekeeper entirely. That's a categorically different company than one that helps the gatekeeper work at twice the speed — and only the former deserves the urgency of a venture-backed startup.

If AI models plateau and commoditize, context becomes the only durable moat — and whoever accumulates it first wins

The day models stop meaningfully improving is the day context becomes the only weapon left standing. Models follow S-curves. "If it plateaus then we're at a point where all the other labs can catch up and they can all have the same technology. And if they all have the same technology, then we're going to have a lot of competition" — on price, on product, on everything.

What survives that commoditization is context. If Granola gets inside an organization and everyone uses it, the accumulated institutional memory becomes irreplaceable. "That's not something you as an enterprise want to give up." The switching cost isn't the feature set — it's the organizational history built up over time.

This reframes the entire product strategy question. You're not just building a product. You're building a context accumulation engine. Being first and moving fast aren't just growth tactics — they're how you construct the only moat that survives model commoditization entirely.

Traditional media is structurally dead — and the man who built podcast infrastructure still underestimated how far independent media would go

The founder who democratized audio left Spotify in 2022 thinking the independent media opportunity was "baked." Done. Finished. Four years later: "I probably underestimated just how big independent media could become... It's gotten so much bigger since then. And that was only four years ago."

Every major TV personality and media institution has made the leap to self-publishing. The unbundling of television hasn't peaked — it's still in its early innings. The content strategy that cuts through is binary: Logan Paul/MrBeast production scale, or a teenager pointing a phone sideways and hitting record. The mid-tier firm producing from its conference room gets crushed. The production value floor has risen so high that "decent" is invisible; excellence is just the baseline.

Your best work happens three months from running out of money — constraints aren't a problem to solve, they're the engine

Asked when he did his best work: "I believe I did my best work when we were three months out of cash." Near-failure clarifies in ways that abundance cannot replicate. "It's so clarifying to know that you're going to fail unless you turn the thing around."

The framework is a complete motivational system: failure is the ultimate constraint — it forces ruthless prioritization, eliminates distraction, and generates the specific urgency that produces breakthroughs. Pair it with an impossibly ambitious mission pulling you forward, and you have both forces working simultaneously. "I strongly believe that great great things, great companies, great products, great people come from constraints."

Premature removal of constraints — too much runway, too much comfort — is how good companies stay good instead of becoming great. Don't solve for comfort. Solve for the company.

The next decade belongs to whoever earns the right to be someone's second self

Every thread in this conversation — agents as second selves, context as the real moat, obliterate over automate — points at a shift Mignano doesn't quite name directly: value in technology is migrating from raw capability to trust. Which model is smartest matters less and less as intelligence commoditizes. Which product the user trusts with their agency, their institutional memory, their actual life — that's where the next generation of generational companies lives. The labs can commoditize intelligence. They cannot commoditize the trust built through years of accumulated context. Build for trust.

---

Topics: venture capital, AI, application layer, agents, independent media, token spend, USV, startup strategy, open source models, creative AI, Suno, Granola, media disruption, founder advice