Key Takeaways

  • Mike Mignano, a new General Partner at USV, maps out two contrasting futures for the AI model landscape: one where frontier labs achieve 'recursive self-improvement' and become uncatchable, and another where AI intelligence hits an S-curve plateau.
  • The 'recursive self-improvement' path implies exponential, unending growth in AI capabilities, allowing leading labs to continually enhance their own research and models without external help.
  • The S-curve future, which Mignano suggests aligns more with historical tech adoption, leads to model commoditization, intense price competition, and a focus on cost optimization.
  • In the commoditized S-curve world, builders will prioritize open-weight and open-source models to manage 'token spend' and will seek to differentiate at the application layer, using "harnesses" for human alignment.
  • For USV, this S-curve outcome underpins their investment thesis, pushing them to back application-layer innovations that can thrive on a diverse, cost-efficient base of underlying AI models.

The Two AI Futures: Exponential or S-Curve?

If you're building in AI today, you're betting on a future that's still hazy. USV's new General Partner, Mike Mignano, laid out two very different possibilities on 20VC with Harry Stebbings. The first scenario is a rapid, exponential sprint where today's frontier labs achieve something close to superintelligence. Imagine an AI that doesn't just learn but actively improves its own AI research. Mignano described this potential, saying, “Once they reach it, if they reach it, once they reach it, we will have some form of recursive self-improvement such that we hit this exponential growth of intelligence.” This isn't just faster iteration; it's a system that, once deployed, “can just continually improve itself on the exponential forever until it plateaus, until it self-limits in some way,” Mignano explained. In this future, a handful of leading labs become effectively uncatchable, setting the pace for all innovation.

But Mignano also sketched a second, more historically consistent path: the S-curve. Every new technology adoption, he notes, has followed this pattern: a slow start, then a fast ramp-up, and finally, a plateau. “That's potentially the second future for AI,” Mignano mused. If AI intelligence hits such a plateau, the game changes entirely. The exponential growth phase ends, and suddenly, the playing field levels. “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,” Mignano said. This isn't about one dominant superintelligence; it's about competitive parity.

Building for Commoditization

For ambitious founders, the S-curve future implies a very different strategy. Once core AI technology plateaus and becomes widely accessible, competition heats up. Mignano predicts we'll see intense rivalry “in terms of price, in terms of product experience, and in terms of the various components that make up the new intelligence stack.” What does that mean for your startup? It means the underlying models become commoditized. The differentiating factor won't be raw intelligence, but how you apply it, how efficiently you run it, and how well you align it with human needs.

Founders and enterprises in this world will obsess over cost. “I think they're probably optimizing for cost,” Mignano affirmed. They'll scrutinize “the trade-offs between intelligence and spend and how they can optimize their token spend.” This points directly to the rise of open-weight and open-source models, which offer greater control and cost efficiency compared to proprietary frontier models. USV's investment strategy reflects this belief, focusing on the application layer and "obliterating" existing markets by building atop accessible, cost-effective AI. They're betting on the "harnesses" – the tools and interfaces that make AI intelligent and useful, aligned, and safe for specific human problems.

What to Do With This

Don't pick a future and bet everything. Instead, prepare for both. This week, audit your current and planned AI dependencies. Identify core workflows that rely on specific frontier models. Then, sketch out a contingency plan: how would you migrate those workflows to open-weight or open-source alternatives if model pricing becomes prohibitive or a proprietary leader suddenly shifts terms? Focus on building unique data moats and workflow IP that transcend any single model provider, ensuring your value isn't locked into one of Mignano's potential futures.