THE MODEL IS NOT THE PRODUCT
@AravSrinivas, Founder & CEO of @perplexity_ai , interviewed by @HarryStebbings (@20vcFund)
Summary: Aravind Srinivas argues the money in AI sits in orchestration, the system that turns every layer's progress into the most valuable output tokens per watt, while the binding constraint on the whole industry stays physical: power. He runs Perplexity on offense (400 people, a $20B valuation, revenue tripled past $500M this year) and bets that whoever conducts models, tools, chips, and devices wins more than any single model builder. If he is right, the winners pay for the frontier, fewer people build bigger companies, and Micron matters more than Meta.
1. Attack, Attack, Attack. Srinivas plays offense because he has nothing to lose. He grew up lower-middle-class in India, where his family counted a job at Google as the win, so everything past that already feels like a bonus. Whenever he catches himself playing defense to avoid failure, he treats that as the dumbest move available and goes all in. The line he repeats: be on offense all the time.
2. The Model Is Not the Product. Srinivas builds on Greg Brockman's line that the model is no longer the product. A pure reseller of model tokens has no business, because the model itself gets commoditized within months. The value sits in the orchestration around it: the agent harness, the connectors, the tools, the grounding in real context that turns raw intelligence into output people pay for. Even a frontier lab has to own an interface, or its application-layer business disappears.
3. Token Value Per Watt. Srinivas names one metric above all others: token value per watt per user. Dollars trace back to power, and power is the one input no company can subsidize except a government. Whoever produces the most valuable output tokens for the least power has the most pricing power. Everything else in the business is downstream of that ratio.
4. Orchestrating Across Models. Perplexity routes across competing models, something OpenAI and Anthropic structurally cannot copy. You will not find GPT-5 inside Claude Code or Opus inside Codex, but you will find both inside Perplexity. Because of that, every improvement anywhere helps the product: a better chip from Jensen Huang, a better model from Dario Amodei, cheaper open-source inference, a faster device from Apple. Perplexity tripled revenue this year partly on progress its competitors paid to make.
5. The Frontier Premium. Given a fixed budget, Srinivas would hire one Jeff Dean over five average engineers, and he applies the same logic to tokens. People pay up for frontier capability. The catch is that what counts as frontier keeps moving, and last year's frontier gets roughly 10 times cheaper once open-source models and a good harness catch up. So spend on today's tasks falls while new frontier work appears: autonomous software engineers, AI designing chips, AI designing drugs.
6. Bottleneck Economics. Srinivas states it as a pricing law: the bottleneck commands the price. Memory is the bottleneck right now, which is why he thinks Micron could be worth more than Meta in 6 to 12 months. CPUs became a constraint again because agent loops run on them, lifting AMD and Intel. His takeaway for investors: price follows the bottleneck, so track the constraint.
7. The Power Bottleneck. A data center is land, permits, turbines, grid deals, and cooling, not a rack of chips bought from Dell. Srinivas estimates 40 of every 100 planned data centers are stalled by public resistance, much of it driven by false claims about water and power use. This physical buildout time caps how fast frontier capability can grow, and he does not see it easing in 3 years. The buildout may simply move to countries with friendlier permits and spare energy.
8. The Local Compute Bet. The real risk of always-on AI is cost, not a rogue model doing something crazy. Nobody can afford continuous frontier inference running on a server at the fidelity of a few seconds. Srinivas wants a continuously learning local model, paired with its own chip and harness, that handles most work and calls the server only when it has to. The data center moves onto your device, and your most sensitive tokens stay with you.
9. The Export-Control Backfire. Short term, Srinivas credits export controls with the roughly 12-month gap between open-source and the frontier. Long term, forcing China onto the Huawei stack pushes a vertically integrated, memory-efficient architecture, the kind of KV-cache and SSD-inference work DeepSeek already shipped. He puts maybe a 20 to 30% chance on another DeepSeek moment that leaves overbuilt American capacity stranded. China can also build the physical layer, power and fabs and robots, faster than the US.
10. Fewer People, Bigger Companies. 400 people built Perplexity's $20B valuation, and Srinivas wants 1,000 startups of about 40 people each worth billions. The number he watches: token spend as a share of developer payroll. Salesforce reportedly spends about 3.8% today, which keeps the labs at $5T, and near 100% would make them $10T. Perplexity aims at the non-developer work, finance, corp dev, sales, research analysts, which he calls Claude Code times 10.
11. The Cost Of The Doom Narrative. Srinivas thinks Dario Amodei's "all the jobs are going" message is wrong and self-defeating, especially while the same labs ask the public to let them build data centers. He offers a counter-example: a San Francisco Uber driver watched one of his interviews, built a web app with AI, added billing, and now earns more in passive income than from driving. You cannot win by telling people they are about to be screwed and then complaining that you cannot build fast enough. Honest optimism about agency does more for the buildout than fear.
12. Velocity As Humility. In the first year or two of a company, Srinivas calls hunting for a moat a trap, since the only real edge is velocity. His line: moving fast is a way of expressing humility, because you keep making contact with the world and questioning your own assumptions. He sets the bar with Jensen Huang, who runs a $5T company and still wakes up telling himself he is 30 days from going out of business. For the one buy-and-hold pick among the coming IPOs, Srinivas takes SpaceX, the only one of the three he calls an n-of-1.
