Benchmark's Everett Randle on house money, open-source AI risk, and why app companies building internal research teams is a mistake

Benchmark's Everett Randle on house money, open-source risk, and the app layer

Playing with house money

Everett Randle describes the current venture market as carrying a "feeling of inevitability" he hasn't seen since the summer and fall of 2021 — and the comparison worries him. The last 18 to 24 months have seen nearly every company marked up quickly, and firms are deploying aggressively into AI, space, defense, and nuclear on the assumption that things will keep rising.

The key structural difference from 2021 is Anthropic. Several firms have spread Anthropic exposure across as many as eight funds, meaning the coming IPOs from Anthropic and OpenAI will buoy entire fund portfolios simultaneously. That creates a "house money" dynamic: firms feel their job is already done, so they swing freely on incremental bets. In 2021, liquidity events like DoorDash, Airbnb, and Palantir existed but weren't distributed that way — Palantir, Randle notes, was concentrated, spent years stuck near $8 per share, and was broadly seen as a black sheep of venture.

The end of the $3 Uber era

Randle flagged in a January 2024 post that the subsidized-token era would end, and he thinks it largely has. The shift from ChatGPT-style interactions to long-running reasoning agents that spin up sub-agents has caused token costs to "absolutely balloon." He still sees some subsidization in the market — Claude Code offering roughly 20x the usage of API-via-Cursor through subscription pricing being the clearest example — but the gap is narrowing.

The forward-leaning company view, which he associates with Dylan Patel at SemiAnalysis, is that token costs reaching and exceeding human labor costs is actually the goal. Those companies want their competitors using open-source or "dumb" models while their own employees run on frontier intelligence. Randle doesn't dismiss that framing but acknowledges the picture is genuinely unclear.

“The last time I felt like investors had this much of a sense of inevitability... was the summer and fall of twenty twenty one. And we all kinda know how that ended up... If an app company like, app companies are either already at 80% plus open source usage for their own apps... or 90% of them are trying to get to that ratio.”

App layer: open source and the labs risk

Across Benchmark's portfolio and the broader market, Randle says 80% or more of app companies are already running their customer-facing workloads on open-source models, and the remaining companies are trying to get there. The logic is straightforward: well-understood, mature use cases don't need frontier intelligence, and fine-tuning open-source models on specific tasks outperforms the generic frontier model while costing far less. Internally, those same companies still use frontier models for coding, where incremental gains remain sharpest.

On labs risk, Randle is less bearish than the consensus. He uses legal AI as an example: Anthropic added more ARR in Q1 2025 than the entire medium- and long-term TAM of the AI legal market. Sending a SWAT team to grind out a seven-year, top-down enterprise sales cycle in legal, to capture some fraction of that ARR, makes no sense given that prioritization calculus. The same logic applied to Microsoft not crushing every niche SaaS market it could have targeted — it had Office and security doing tens of billions each, so it didn't bother.

The internal research team mistake

One pattern Randle expects the industry to look back on unfavorably is app companies standing up internal AI research teams, typically by hiring a handful of researchers from Meta or similar labs. The implicit pitch is "we have researchers, so we're not exposed to lab risk." Randle thinks this is largely a founder defense mechanism rather than a real competitive moat — there simply aren't that many opportunities for an in-house team to do groundbreaking work relative to what researchers with the most GPUs are producing inside frontier labs. He's careful to distinguish this from experimentation teams like Ramp Labs, which he sees as doing something genuinely different.

Western open source: a real gap, slowly closing

Randle has said publicly that the West has no good open-source models, a claim that drew pushback — but he points to the OpenRouter token rankings as his response: the top 10 are all Chinese models. He's not attributing Chinese open-source quality solely to distillation from American closed-source models, calling that view oversimplified and potentially tinged with xenophobia. The DeepSeek paper, he says, was "truly groundbreaking."

The more pointed near-term risk is a Beijing decision to restrict overseas access to Chinese AI models. Given that most Western app companies fine-tuning open-source models are building on Chinese foundations, a cutoff would remove or stall the primary training substrate for a large portion of the application layer.

On the Western supply side, he points to NVIDIA's Nemotron, Thinking Machines Lab's Inkling (which he says portfolio companies are bullish on despite it not yet claiming Chinese open-source parity), and Reflection AI's stated strategy of building the American open-source model. Three to four months ago, he couldn't point to much. Now there's at least something.

Where venture still works, and where it doesn't

Legal AI is his clearest example of a category that wasn't venture-investable before and now is — prior attempts at tech-enabled law firms like Atrium failed, and traditional legal SaaS faced constrained TAM and slow sales cycles. Harvey and Lagora now have "eye-watering numbers." On the other end, he sees entrepreneurs building highly niche data businesses — instrumenting medical clinics, capturing every conversation and video interaction to create proprietary datasets — that will probably top out at $50M to $100M in revenue over eight to ten years. Venture scale, no. Worth building, yes.