Earlier this month, Boris Cherny sat down with Sequoia's Lauren Reeder at AI Ascent 2026. Cherny is the engineer who built Claude Code at Anthropic, and the clip doing the rounds is the headline: he hasn't written a line of code by hand since October 2025. He ships dozens of pull requests a day from his phone, running several Claude Code sessions concurrently and a few thousand sub-agents overnight.

That's the part that gets shared. The more useful idea, for anyone allocating into early-stage software, sat one level below.

The framework
Reeder used a phrase I keep coming back to: product overhang. It describes the gap between what an AI model can already do and what shipped products actually let users do with it.

Anthropic's bet on Claude Code is a clean illustration. They started building it in late 2024 for a model that didn't yet exist. Cherny was candid that for the first six months it was barely usable, and he used it for around 10% of his own code. Then Opus 4 shipped in May 2025, and adoption inflected. Every subsequent model release moved the curve again.

The strategic insight is that the tooling layer is consistently underweight against model capability. There is always a window of six months or so where the model can already do something that no product has yet captured. Companies that build for the next model, or the next market behaviour, sit at the front of the curve when capability and adoption catch up. Companies that build for what's available today are sized for a market that's about to be obsolete.

Cherny's other big argument is the printing press analogy. Before the press, around 10% of Europeans were literate. Within fifty years, more had been published than in the previous thousand combined. He thinks software is on the same trajectory but compressed, and that the next decade will produce roughly ten times more disruptive startups, because tiny teams can now go head-to-head with incumbents.

Both ideas sound abstract until you start matching them against companies on the ground.

Bondio: building for a curve that hasn't arrived
Bondio is building the infrastructure layer for the eSIM revolution, effectively "Stripe for eSIM connectivity." Today, around 20% of devices globally are eSIM-enabled. The GSMA expects that to hit 50% inside two years. That curve is the overhang.

The incumbents servicing global roaming are 20-year-old Mobile Virtual Network Aggregators (MVNAs) built on physical-SIM batch ordering, slow APIs, six-to-eight-month integration cycles, and $20k minimum commitments. They were built for a market that is being replaced. Bondio's platform onboards a customer in minutes, runs at 0.2% eSIM issue rates against a 2% to 5% industry average, and 99.98% API uptime. They reached $3.26m gross ARR in 18 months from launch and were on a path to profitability without needing further capital as of November 2025.

Cherny would call the incumbents' position a negative asset. The processes and specialisations that made legacy MVNAs profitable are now the reason they cannot serve the wave of new eSIM retailers: travel apps, fintechs, airlines, neobanks. Bondio is built ground-up for that wave, with founders who spent years inside Vodafone, Plivo, AWS, and Opensignal. Deep domain, lean architecture, sized for where the market is going rather than where it is.

Querio: building for AI agents the rest of the stack hasn't caught up to
Querio is doing something structurally similar in a completely different category. They are building a code-first business intelligence platform where the company's analytics logic lives as versioned files in a Git repository, accessible to both the data team and to AI agents, on the same trusted definitions.

The bet is that within a few years, every mid-market company will want AI-powered self-serve analytics, and the existing BI stack cannot give it to them safely. Those tools were built for human analysts clicking through dashboards. Querio's architecture is built for an agent-driven workflow that is just now becoming production-ready.

The early signal is that customers feel the gap acutely. Five months after relaunching the product in November 2025, Querio is at $240k ARR with ~20% month-on-month growth, 93% pilot-to-paid conversion, and 250% net revenue retention. They are winning head-to-head against WrenAI, ThoughtSpot, and custom OpenAI agents. The longer a customer uses Querio, the more business logic lives in their Git repo, which makes replacement equivalent to replacing Git itself for analytics. That is exactly the kind of compounding moat you cannot retrofit onto a legacy BI vendor.

Same shape as Bondio. Deep domain founders (ex-Amazon, ex-Glovo), built for where the market is going, structurally faster than the incumbents because they aren't dragging legacy assumptions with them.

What we're looking for
Both companies fit the same pattern. A clear shift in how an enterprise function is going to be done. An incumbent layer that cannot adapt without dismantling the thing that made it profitable. Founders who know the industry from the inside. A product designed to be on the right side of the curve when the wave breaks.

That is what product overhang looks like in practice, and it is most of what we are looking at in current deal flow. If Cherny is even half right about the next decade, this is the layer that will compound.

The full interview with Boris Cherny is on YouTube ("Anthropic's Boris Cherny: Why Coding Is Solved, and What Comes Next"). Twenty-four minutes, worth the time.