Theory Ventures' Tomasz Tunguz: inference is now the biggest software market, agent pricing at par with human labor

Tomasz Tunguz on inference markets, agent pricing, and the Bending Spoons question

Tomasz Tunguz, founder and general partner at Theory Ventures, makes a structural claim worth paying attention to: inference is now the largest market in software, having displaced databases. He draws the analogy directly — just as the database market fragmented into fast and slow, image, video, and specialized variants, inference is fracturing the same way.

Theory Ventures is investing into that fragmentation. The firm has backed Sail Research, which targets inference workloads where users will accept a five-to-ten-minute delay in exchange for dramatically lower compute costs — a wedge into the large class of batch-oriented enterprise use cases that don't need real-time responses. It has also backed Ollama, which has around 10 million users running models locally. A Stanford study released roughly three to four weeks before the interview found that 90% of white-collar AI use cases can be handled on a MacBook, which Tunguz cites as validation for on-device inference as a serious deployment path, not just a hobbyist one.

“The single biggest market in software today is inference. It used to be the database market. Now it's inference... We made a prediction at the end of twenty five that 2026 would be the first year where employers would pay agents at market level for a person or more. And that happened the first time we noticed that was actually in March or April of this year.”

Voice and edge

On voice AI, Tunguz maps the stack as four layers: phone, carrier (Twilio), speech-to-digital conversion, and the model itself. Latency is the critical optimization variable, and it's materially different from how large language models are typically tuned. He expects the infrastructure to start in centralized data centers and move progressively to the edge, consistent with the direction companies like Cloudflare are already pushing.

Agent pricing at par with human labor

Theory Ventures made a public prediction at the end of 2025 that 2026 would be the first year employers would pay agents at market rate for a human worker, or above it. Tunguz says that threshold was crossed in March or April 2026, when a portfolio company began charging at parity with a human employee — and is likely to move to a premium.

The economic logic: agents carry no management overhead, no healthcare costs, and no long-term employment commitments. Tunguz draws the analogy to expert freelancers, who command a day-rate premium over full-time hires precisely because the engagement is flexible and immediate. If agents price like premium freelancers, the multiples on AI-native vertical software companies start to make more sense — though he acknowledges the 100x to 150x current ARR at which the fastest-growing AI companies are currently trading is "enormous," and that some are subsidizing those multiples with compressed gross margins to drive adoption before improving unit economics.

The Bending Spoons opportunity

On the question of whether an American equivalent to Bending Spoons — a firm systematically acquiring legacy software businesses at distressed multiples — makes sense, Tunguz is direct: the arbitrage is real. Net dollar retention and gross dollar retention at many legacy software companies remain strong even as their market multiples have compressed. The gap between intrinsic value and trading price is wide enough that he's surprised more established acquisition firms haven't been more active. He points to Constellation Software and Danaher as existing public-market analogs that could step into the opportunity, and says that cleaning up older fund vintages is now among the top two or three topics in every LP conversation Theory has.

Personal agent monetization

On consumer agents like Meta's Muse and Instinct, Tunguz frames the current phase as data acquisition first, cost scaling second, monetization third. The agent trajectory data being generated — what users ask for, how tasks unfold — is, in his view, a new targeting dataset comparable in strategic value to what search queries were for Google or social graphs were for Facebook. He estimates roughly $10 billion is being spent this year on data to train models of this type.

The revenue potential he points to is straightforward: the average U.S. Google user generates around $120 in ARPU annually, and he believes agentic systems could double or triple that figure through new targeting mechanisms built on behavioral trajectories. Whether the affiliate model works for agents executing explicit user instructions — where the retailer isn't crediting the agent with driving demand — remains an open structural question Tunguz doesn't fully resolve, though he expects the monetization model to crystallize once distribution and data scale are established.