Scott Wu on Cognition's fundraise, AI solving Navier-Stokes, and why cybersecurity is 10% of Devin's sessions and growing fast

Scott Wu on Cognition's fundraise, cybersecurity demand, and AI solving Navier-Stokes

Cognition AI, the company behind the Devin coding agent, has closed a new funding round priced at a $46 valuation — the number a deliberate nod to the perfect score in Math Counts, the competition that shaped Wu's early identity as a mathematician.

Where the business actually is

Wu frames the moment in three distinct stages. At Devin's launch roughly two years ago, the product was a prototype and agents were widely dismissed as a buzzword. A year ago, the product worked for specific use cases but most enterprises hadn't crossed into genuinely delegating background tasks. Now, he says, the capabilities are good enough that the question has shifted from can it work to which use cases do we measure.

That last point is where Wu thinks enterprise conversations have landed. The early phase of counting tokens and tracking how many engineers had access lasted about four months. Customers are now asking about concrete outcomes tied to their top three or four priorities, and Wu argues the companies that don't build measurement frameworks around specific use cases will never know what's actually working.

A 50,000-person software org, he notes, will not figure out coding agents at the same speed as a three-person YC company. Security reviews, onboarding, and internal guardrails are the real friction, not capability.

“Agents are just getting really, really good... We just solved Navier Stokes with AI. Like that is insane. It's absurd... Cybersecurity is a small but meaningful part of our business — around 10% of the Devin sessions and ACUs that get spent today are on security, but our security product's only about two months old.”

Cybersecurity

Security is roughly 10% of Devin sessions today and growing quickly. Wu flags that the product itself is only two months old, which makes the share notable. His concern is asymmetric: large enterprises take time to adapt, but small attacker teams aren't waiting, and model capabilities relevant to offensive security keep improving.

Model routing and the intelligence ceiling

Wu argues that raw model intelligence is no longer the primary bottleneck for most tasks. ChatGPT releasing a smarter model doesn't move retention metrics much, he says, because the models already get most things right. What matters increasingly is context, style, cost, speed, and task-specific training.

That shifts the value of orchestration. Wu distinguishes between two things: bringing in real-world tooling — logs, test runners, browser interactions — so an agent can verify and correct its own work, and routing tasks to the right model on the cost-performance curve. He expects enterprises to become more aggressive about using multiple models rather than defaulting to the largest one, and sees that routing problem as durable and growing in importance.

Ramp building a router and Stripe acquiring OpenRouter fit Wu's read that payments and spend management infrastructure will need to be rebuilt around agents and token consumption.

AI and mathematics

Wu calls AI solving Navier-Stokes in 88 hours "insane" and "absurd" — a fundamental fluid dynamics problem that has absorbed enormous human effort, resolved inside a week by a model working through an orchestrated system. He holds a bet that the Riemann hypothesis will not fall by end of 2026, though he puts the probability at below 50% for this year and thinks 2027 is likely if not sooner.

His practical read on the math controversy is that disputes over credit and priority are standard in academia — Newton versus Leibniz over calculus is still argued today — and the AI accomplishment itself is what will be remembered.

For Cognition's own benchmarks, Wu is explicit that solving open mathematical problems is not the priority. The benchmarks that matter to customers are things like finding security vulnerabilities in real-world codebases, not IMO problems.