Mike Mignano on AI Applications, Open Models and Venture Strategy

Union Square Ventures general partner Mike Mignano joins Harry Stebbings to discuss AI applications, open models and venture investing. Mignano previously backed Suno and Granola at Lightspeed; USV has invested in companies he discusses, including Abridge and Doctronic.

Asked whether fear of failure or the desire to win motivates him, Mignano says both matter. He recalls doing some of his best work at Anchor when the company had only three months of cash left: the constraint clarified what the team needed to change.

The two revisit an old disagreement over podcast production. Mignano once favored a natural conversational style, but now believes high production value and strong ideas help content stand out when basic recording and editing have become easier.

Stebbings argues that a show should deliver value and a hook early. Both suggest content can succeed either through a very informal style or a highly produced one; this is their editorial judgment, not a measured rule for every creator.

Mignano says he moved from Lightspeed to USV after a long relationship with the New York firm, including investing as a limited partner. He admires its willingness to take thesis-driven positions, while acknowledging that a mistaken thesis can cause an investor to miss opportunities.

He believes AI has gone through a massive infrastructure buildout and that the application layer now has more to build on. He also says the infrastructure buildout is not finished: his prediction is that substantial value can accrue to both layers, not that investment in infrastructure has stopped.

On always-on AI, Mignano expects products to seek more context and become more continuously present. He compares the change with adoption of voice assistants and meeting-note tools such as Granola, an investment he previously backed, while saying the direction could have benefits and drawbacks.

Stebbings argues that mid-sized seed funds are squeezed by larger funds and smaller collaborators. Mignano emphasizes networks, publicly sharing ideas and taking early bets on people. A clear thesis, he says, can help founders know which investors may be interested.

He calls USV's interest in open-weight models, open-source harnesses, distributed compute and human-aligned agents the 'Rebel Alliance.' He argues investors should be willing to publish imperfect ideas and revise them, while recognizing the risk of being wrong.

Mignano lays out two possible AI futures. In one, a lab reaches a form of recursive self-improvement and pulls far ahead; in another, progress follows an S-curve and plateaus, allowing competitors to catch up. He presents both as possibilities, not established trajectories.

He describes recursive self-improvement as AI doing research that improves future AI, but notes possible limits in architecture, data or hardware. He does not know whether such a system will exist or which constraint, if any, would stop its progress.

If a breakthrough occurs, Mignano thinks existing labs' compute and infrastructure could give them an advantage, although he leaves room for a new architecture from another team to leapfrog them. If capabilities plateau, he expects more competition on cost, products and model choice.

In the latter scenario, organizations might route tasks among models according to cost and capability, including open-weight options. Mignano defines a harness broadly as an application closely coupled to a model, and cites Claude Code and other tools as examples.

Mignano argues that as agents receive access to personal information and act on users' behalf, their incentives matter. He wants agents aligned with the people they serve, while acknowledging that people may become comfortable delegating even when privacy concerns remain.

Stebbings predicts most people will readily hand over control, comparing agents with previous adoption of online payments. Mignano concedes history may support that view but sees a difference when a system can buy things or send personal messages rather than simply display information.

He does not claim everyone must choose a human-aligned agent. His hope is that enough users and competing products could give consumers alternatives and discipline the market; he says he does not know whether this will happen.

Stebbings questions Anthropic's messaging on job displacement. Mignano sees a mission-driven company and says such firms can succeed. Statements in the discussion about company valuation and future revenue reflect the speakers' views at recording, not audited forecasts.

The speakers consider how much companies may spend on tokens relative to developer salaries. Stebbings proposes different spending scenarios and their implications for model providers; those are hypothetical calculations, not actual future spending or valuations.

Mignano thinks very large incumbents may impose tighter token budgets because unlimited usage across thousands of employees could strain economics. For a small startup, he would encourage substantial spending where it gives a competitive advantage, especially coding, while using cheaper tools for simpler tasks.

He predicts engineering organizations may become smaller teams of highly capable engineers as agents handle more lower-level work. He does not know exactly how teams will evolve, and the employment effect remains uncertain.

Mignano estimates that non-frontier models can handle 80% of non-coding enterprise tasks, such as summaries and briefs, while he currently prefers frontier models for coding. This is his estimate and preference, not a measured benchmark for all enterprise workloads.

On China's open-model ecosystem, he expects talented teams to follow incentives as open models and tools become more viable. The conversation does not establish a general security conclusion about models by country of origin.

Mignano sees model routing as useful for selecting a suitable balance of capability and cost. Dedicated providers such as OpenRouter and companies elsewhere in the stack may both compete to offer routing, in his view.

A small markup on model usage may be hard to defend, he says. He mentions a speculative alternative in which a routing provider is rewarded for selecting an effective model, but says he has not necessarily seen that idea implemented.

Stebbings doubts routing alone can produce a $50 billion company. Mignano responds that a routing product deeply embedded in developer workflows could become hard to replace; neither claims such a valuation has been achieved.

Asked whether USV missed the model layer, Mignano says the firm pursued different opportunities, including energy. He corrects himself during the interview: USV's energy thesis began in 2021, not 2001. He expects AI demand to make energy supply and portability important, while the full investment outcome remains unknown.

Mignano acknowledges that energy can be capital intensive but says early scientific experiments can offer venture entry points. He cites USV investment Radiant, which is developing small nuclear reactors; discussion of testing and commercial prospects is forward-looking, not proof of regulatory approval or deployment.

He also describes USV portfolio company Rune as building small data centers near generators to reduce the distance between power and compute. This is his description of its proposed approach, not evidence of its performance at scale.

USV's phrase 'don't automate, obliterate' means the firm prefers businesses that reinvent a market or workflow over software that only makes an existing process faster. Mignano says USV makes relatively few enterprise investments for that reason, though he acknowledges automation can create valuable companies.

As an example, he describes USV-backed Doctronic's ambition to make AI medical assistance widely available. It is an investment thesis and product ambition, not a claim that an AI system can replace a clinician or provide safe care in every situation.

Stebbings asks whether model providers will capture application markets, including health. Mignano cites USV-backed Abridge, arguing that years spent developing relationships, partnerships and navigating a regulated industry can be an advantage that a new entrant cannot instantly reproduce.

He compares specialized startups facing large incumbents with Spotify's competition against bigger firms. His claim is that focused teams can build advantages; he does not say every startup will win or that competition from the labs is inconsequential.

Stebbings points to competition around design and coding as evidence that more than one company can succeed. Mignano agrees that a large market can support several meaningful businesses, offering a rough market-share illustration rather than a general empirical law.

Mignano expects USV's focused theses to attract relevant deal flow, particularly around energy and open AI. Smaller funds, he says, need to choose areas deliberately; he had also published ideas and pursued themes at Lightspeed.

Although he believes startups have room, Mignano still worries about model providers entering their markets. He notes that Granola, which he backed at Lightspeed, faces products from OpenAI and Notion, and that competition from well-funded companies is a persistent startup risk.

Stebbings questions whether Granola can expand into large enterprises. Mignano argues that a focused meeting-notes product can get a foot in the door before expanding, whereas a broad platform may ask a customer to replace too many tools at once. This is the investor's assessment of a portfolio company.

Mignano also acknowledges the threat of Microsoft bundling and cites Slack as an example of a company facing such competition. He thinks accumulated organizational context can make a product more valuable and harder to replace, and emphasizes speed and early adoption in the AI application market.

The two debate whether smaller outcomes still matter when a few companies reach extraordinary valuations. Mignano says fund objectives and size change the answer: a small fund can produce strong returns from outcomes that may not move a multibillion-dollar fund.

Stebbings argues Series A now requires much larger funds. Mignano says some rounds are priced around $80 million to $150 million post-money, making ownership harder to acquire, though the examples are his description of the market rather than a complete dataset.

USV values ownership at seed and Series A, Mignano says, because later rounds can become expensive and dilute early investors. At a later stage, if a company appears to be a potential market leader, he may focus more on the multiple of cash invested than on the percentage owned.

The speakers disclose a shared investment in FOMO, a trading application. Mignano says USV accepted a later-stage entry without focusing on ownership because it saw potential upside, while noting its fund size may limit participation in every future round.

He says USV seeks large markets and that later-stage investments need especially large potential outcomes. Its current core fund is $275 million, he says, so its return arithmetic differs from that of funds managing several billion dollars.

Stebbings expects large platforms to deliver venture-like returns as outcomes expand. Mignano agrees that firms able to fund capital-intensive infrastructure companies have done well, but predicts a proliferation of less capital-intensive application startups where smaller, focused funds can also participate.

How elastic should investors be on price? Mignano recalls USV partner Fred Wilson's advice to 'never pass on price,' but adds nuance: USV can stretch more for a later-stage company it believes could be a market leader, while its smaller early-stage funds still face limits.

He tells a story about Anchor accidentally posting from Wilson's account shortly after launch. Mignano says he apologized by email and Wilson responded graciously; Stebbings similarly recalls experienced investors giving him time early in his career.

Mignano says a key investing lesson is not to project his own product ideas onto founders. Even a plausible plan is a poor investment basis if the team intends to pursue a different one; he looks for founders whose own judgment and execution he trusts.

His priorities have shifted from product first to founder first, then market and product. Early-stage companies often change direction, he says, making resilience, adaptability and execution important. He identifies communication as a founder trait he has sometimes underestimated.

Stebbings says he passed on Suno because the small stake did not meet his ownership preference, and on Granola because he struggled to imagine the unformed product. Mignano says the investments came from different approaches: Granola was a founder-led bet, Suno a thesis about democratizing music creation.

Mignano had known Granola founder Chris for years and had seen him operate. For Suno, he says, he sought out teams building AI music tools and then became convinced by Mikey Shulman and his colleagues' technical understanding and passion for music.

Asked how to justify Suno at a $5 billion valuation, Mignano sees large potential in a platform that democratizes a medium. As an investor in the company, his view is a valuation thesis, not a guarantee of future returns or evidence that a particular outcome is inevitable.

He says he once assumed Suno needed both a creator and a consumer marketplace, but now sees users making music for the enjoyment of creation itself. He calls this 'creative entertainment' and compares it with making images, code or games for fun.

Asked about a missed investment, Mignano names Substack, which Lightspeed met but did not back. He believes self-publishing and creator control over media will continue to grow; his praise is a personal investment view.

He argues that traditional media has been displaced in many respects by independent publishing, and admits he underestimated how large that opportunity could become after leaving Spotify in 2022. He expects television to continue unbundling, while acknowledging this is his interpretation of an ongoing change.

In the quick-fire round, Mignano praises Brin Putnam, founder of tabletop gaming company Board, which USV backed. He values her clear vision and relevant experience building Mirror, acquired by Lululemon; this is his assessment of a portfolio founder.

He names Factorial and Haystack as seed funds he admires, describing Factorial's model of supplying additional capital to angel investors. He also credits former Lightspeed collaborators Nat Friedman and Daniel Gross for high-signal introductions, including the introduction to Suno.

Mignano praises his former firm's growth fund and says a lesson from Fred Wilson is that a trusted relationship with founders matters both for helping them and for an investor's reputation. He describes the CEO role as lonely and says reliable support can make a difference.

On parenting, he values the support his parents gave his interests without pressuring him into a particular path. He tries to offer his children the same combination of encouragement and freedom to choose.

Mignano says he has changed his mind about model providers doing everything: even the largest companies cannot pursue every opportunity, although some do expand successfully. He closes by prioritizing enjoyable collaboration, strong founder relationships and fund returns, in that order.