Sequoia's Sonya Huang on AI application layer winners, humanoid robotics timeline, and enterprise AI battles

Summary

Sonya Huang, partner at Sequoia, argues that value in AI is accruing to the application layer — but with an important caveat: foundation models have earned the right to win a large portion of it themselves.

ChatGPT's daily-to-monthly active user ratio has climbed from roughly 14% a couple of years ago to levels now in line with Reddit and approaching Google. Huang says Sam Altman described three distinct usage patterns at Sequoia's AIScent conference: older users treat it as a Google replacement, users in their 20s and 30s use it as a life coach, and younger users treat it as an operating system. The memory and tool-use features OpenAI is building around that core reinforce the thesis that ChatGPT is becoming the front door to AI the way Google became the front door to the internet — a $2 trillion frame of reference Huang invokes deliberately.

Vertical vs. horizontal

Huang sees the enterprise AI market as firmly first-inning. Sequoia portfolio companies Glean, Harvey, and Sierra represent different bets on how deep verticalization needs to go. Glean plays horizontal — connecting across enterprise data — while Harvey and Sierra go deep into legal and customer support workflows respectively. Healthcare CIOs, Huang notes from recent reference calls, are choosing one transcription vendor and moving on, which suggests customer behavior is already creating lock-in faster than the commoditization narrative would predict. Gong is her cleaner example: transcription was supposed to commoditize, and it hasn't, because sales teams have standardized processes and data on top of it.

The margin compression risk is real but scoped. Companies with low switching costs and thin differentiation above what the underlying models provide will get competed down. Companies that are genuinely hard to build, or that are integrating deeply into customer workflows where buyers want to pick a champion, have a more defensible position.

MCP and the tooling layer

A poll at AIScent ranked MCP and tool-use as the single biggest driver of AI progress expected in the next 12 months. Huang treats MCP as a protocol the industry will standardize on, not a startup opportunity. Anthropic benefits from having steered it, but the open-standard nature limits how much any small company can monetize around it. Cloudflare, with infrastructure advantages already in place, is the one player she'd consider an exception.

Inference-time compute

On the pre-training wall question, Huang is an AGI maximalist and frames inference-time compute as the second scaling curve, not a replacement for the first. The AlphaGo analogy she uses: top humans sit at roughly 3,500 ELO, the best bots hit around 3,000 before inference-time compute, but letting the model think before placing a piece pushes ELO to 5,500 — far above human level. Translating that to LLMs, she credits Noam Brown and OpenAI's reasoning team with recognizing that the RL and LLM research tracks needed to converge, and Sam Altman with investing decisively in reasoning infrastructure when that wasn't yet obvious. Pre-training isn't dead; investment just follows the lowest marginal cost of the next unit of intelligence, and right now that's on the reasoning axis.

Humanoid robotics

Huang describes herself as a recent convert from bear to bull. Her timeline: humanoids in tens of thousands — possibly thousands — of households within two to three years. The data barrier that historically blocked robotics progress is breaking down through synthetic data pipelines and generative world models, which can produce tens of thousands of environment variations to train against. She draws the autonomous vehicles parallel: Waymo and Tesla required enormous economic engines to get there, and humanoid robotics is on the same capital-intensive path — not just GPUs, but hardware co-development and action-data collection across the board.

Valuation and the bubble question

Huang says there is already too much venture capital and that a meaningful cohort of companies is raising on what she calls "vibe revenue" — pilots counted as revenue, poor retention dressed up as growth. Underneath that layer, she sees a smaller group of companies growing high-quality revenue at rates she hasn't seen before. The TAM logic that justifies current valuations rests on outcomes-based pricing: if you're replacing services spend rather than software spend, the addressable market is in the trillions.

Energy as the binding constraint

Chase from Crusoe shared figures at a Sequoia event that reframed the infrastructure picture for Huang: the largest data center cluster in the US, in Northern Virginia, totals 4.5 gigawatts of capacity. Crusoe alone has 20 gigawatts in pipeline and more than 2 gigawatts built out. Chip availability, once the bottleneck, is easing. Power is the new constraint, which is why AI buildout is concentrating in West Texas around Abilene, where renewable overbuilding — particularly wind — has created stranded energy capacity. Huang's read is that AI infrastructure will continue to follow power availability as the primary site-selection variable.