Aravind Srinivas | 20VC Summary

1. Key Themes

The Orchestration Layer Is the Ultimate Prize in AI

Aravind argues that the real value in AI is not at the model layer but at the orchestration layer — the system that routes intelligence across models, tools, connectors, devices, and chips to maximize output value per watt per user. He explicitly positions Perplexity Computer as this conductor.

"If you can solve this problem, you will capture the most economic value in AI long-term. Short-term, it might look like, oh, like this other lab's revenue is growing exponentially... but long-term, this is the one objective that truly matters." 00:28:34

Token Value Per Watt Per User Is the North Star Metric

Aravind proposes a single defining metric for the AI era that reframes how to think about model selection, infrastructure investment, and product design. Whoever produces the most valuable output tokens with the least power wins.

"The most important metric in AI is token value per watt per user." 00:15:21

Power, Not Compute, Is the True Bottleneck

While the industry debates GPU supply, Aravind identifies power — specifically the ability to build and permit data centers — as the binding constraint on AI progress. He estimates 40% of planned data centers are not being developed due to public resistance.

"I think right now, 40 out of a hundred are not being developed because of public resistance." 00:42:43 "I think the biggest problem is actually in power." 00:30:41

The Model Is No Longer the Product — The Interface Is

Aravind and Greg Brockman's shared view: model builders who are pure token resellers have no durable business. Value accrues to whoever owns the interface where valuable tokens are generated.

"I would even argue that building models is a way to stay at the frontier, but you have to own an interface in which valuable AI output tokens are generated... This is the single most important thing to unlearn for most founders." 00:15:42

Export Controls May Have Accidentally Supercharged China

Rather than hobbling China, export controls forced Chinese companies like DeepSeek to develop vertically integrated, hardware-efficient architectures that could prove superior long-term. They've innovated on KV cache, attention layers, and training algorithms specifically because they couldn't use the Nvidia stack.

"By forcing them to go out there and build all this, you're converting them into a far more potent competitor... If AI is not just digital, that's also physical AI. You've got to build fabs, robots, chips, and harness the energy really well, package it into local devices. I think they have a lot more advantages than America." 00:46:01

The Agentic Economy Rewards Power Users, Not Mass Users

The AI token economy will be driven by a small number of extremely high-spending power users running continuous agent loops — not hundreds of millions of casual users. This fundamentally rewrites the growth playbook from consumer social to something closer to enterprise software, but more extreme.

"There are users in Perplexity Computer. There's one user I think who spends upwards of like $10,000 a month... Their business runs using agent loops that are running inside these harnesses." 00:16:36 "These products are not going to be used by a hundred million people, but they will generate revenue that's going to be higher than the advertising revenue of Google or Meta." 00:17:59

Advertising Will Not Become the Business Model for AI Chat

Aravind is structurally bearish on advertising working inside chat interfaces. The subjective, exploration-based nature of high-value ad categories (travel, fashion, shopping) doesn't fit the chat paradigm. Objective tasks will be agent-based; subjective ones remain ad-based.

"I am bearish on advertising to really take off in the chat interface... The interface is less about conversations and more about exploration. So when the decision-making is more subjective and vibes-based, you don't need an objective answer engine." 00:12:35

Perplexity's Business Has Tripled in 2025

A specific, underreported financial data point: Perplexity's revenue has more than tripled since the beginning of 2025, while also cutting burn by more than 50%, driven largely by model progress from Anthropic and cost competition from OpenAI.

"Our revenue has more than tripled since the beginning of the year. Tripled since the beginning of the year. And a lot of thanks to model progress made by Anthropic. And we also brought our burndown thanks to OpenAI competing with them and bringing down the cost for the same capability." 00:29:02

The Dario/Anthropic "Jobs Are Gone" Messaging Is Actively Harmful

Aravind argues that Anthropic's doom-and-gloom labor messaging directly contradicts their need for public support to build data centers and creates conditions for the public resistance that is already throttling 40% of planned data center builds.

"You can't win by saying that and also complaining about not being able to build data centers fast." 00:48:28

Physical Infrastructure Is the New Industrial Revolution Bet

Aravind says if given unlimited capital, he would build data centers — not train more models. He frames this as the return of the industrial age.

"I would build data centers... I think physical infrastructure build-outs is like the return of the industrial age again. Like the forefathers who built the industrial revolution, oil pipelines, steel bridges, factories." 01:09:25

2. Contrarian Perspectives

Micron Will Overtake Meta in Market Cap

Most investors view Micron as a cyclical memory supplier, not a structurally dominant AI company. Aravind argues the opposite: Micron supplies HBM (high bandwidth memory), which is the current chokepoint in AI infrastructure. Whatever is the bottleneck commands the price.

"It might not be inconceivable that Micron, the supplier of HBMs, might be more valuable than Meta in the next six to 12 months. It's already at like a trillion, and Meta is like 1.3 to 1.4 trillion." 00:33:13

Export Controls Helped China More Than They Hurt It

The consensus view is that US export controls on advanced chips have meaningfully set China back. Aravind's view is more nuanced and ultimately darker: being forced off the Nvidia stack is turning China into a vertically integrated AI superpower with advantages in the physical layer that America doesn't have.

"The jury's still out. Short term, it's helping... But there is a chance that because of that, they now get really good at the physical layer... Power is not a problem. Permits are not a problem. People are not a problem. Labor is not a problem. Expertise is not a problem. And so by forcing them to go out there and build all this, you're converting them into a far more potent competitor." 00:45:26

OpenAI's Consumer Dominance Is in a Commoditized Category

At a time when the consensus still treats ChatGPT's user numbers as a powerful moat, Aravind argues that consumer AI chat has already been commoditized — and OpenAI knows it, which is why they're pivoting hard to Codex and agentic products.

"Do you perceive them as a dominant leader right now? Yes. In what? Consumer search. Well, except there's no money there, right? It's been commoditized." 00:09:52

The Trillion-Dollar AI Revenue Opportunity Is Not in Advertising

While investors broadly assume that AI interfaces will eventually monetize like social media or search through advertising, Aravind argues the entire basis for ad-supported AI is structurally flawed — the chat interface doesn't capture the exploration-based user behavior that drives advertising, and ads would corrode the trust that makes AI products valuable.

"I am bearish on advertising to really take off in the chat interface. I'm happy to be proven wrong there, but I'm bearish on that." 00:12:35

CPUs Are Quietly Becoming a Major AI Bottleneck — and AMD and Intel Are the Beneficiaries

Everyone is focused on GPU scarcity. Aravind points out that agent harnesses run almost entirely on CPUs — and agents are using CPUs more than humans ever did — making Intel and AMD surprise beneficiaries.

"Agents are using CPUs more than humans. And so suddenly there's a rise in enterprise CPUs. And the beneficiaries of these are like Intel and AMD. So then they get to be the bottleneck." 00:34:14

3. Companies Identified

Perplexity AI answer engine and agent orchestration platform. Mentioned as the subject company — 400 people, $20B valuation, 45M users, 1B+ searches/month, revenue tripled in 2025, burn cut by more than 50%.

"A 20 billion dollar company. 45 million users. Over a billion searches a month. Built in three years by 400 people." 00:00:00

Anthropic Frontier AI lab, maker of Claude models. Mentioned as a key model provider whose model improvements directly drove Perplexity's revenue growth. Also cited for buying a wet lab, for Claude Code, and for lobbying export controls.

"Anthropic bought a wet lab. Could be for the talent, could be for the infrastructure to run like wet lab experiments. But imagine taking all those tokens and putting it in the mid-training instead of just tokens from GitHub." 00:22:16

OpenAI AI lab and maker of ChatGPT and Codex. Mentioned for its IPO readiness debate, consumer dominance, Codex product, and competition that drove down model costs.

"Why are they going all in on Codex? Because that's where the money is." 00:09:52

DeepSeek Chinese AI lab. Cited as a prime example of vertically integrated AI development forced by export controls — innovating on KV cache, attention layers, and non-Nvidia hardware.

"The deep seek is not building with the Nvidia stack. They're building with the Huawei stack... Their whole stack is getting vertically integrated to their hardware and their chips and their fabs." 00:44:12

Micron Semiconductor company, primary supplier of HBM (high bandwidth memory). Cited as potentially more valuable than Meta within 6–12 months due to being the current AI infrastructure bottleneck.

"It might not be inconceivable that Micron, the supplier of HBMs, might be more valuable than Meta in the next six to 12 months." 00:33:13

Nvidia Chip designer. Cited as the dominant GPU supplier and also noted as owning 5% of Intel.

"Nvidia and SoftBank own 5% each [of Intel]." 00:46:33

TSMC Semiconductor fab. Cited for its $150B investment in US fabs, including one in Arizona — framed as critical to US AI competitiveness.

"TSMC is investing like $150 billion into building American fabs. They've already invested $40 billion or something like that, $60 billion last time I checked." 00:46:33

SpaceX Space infrastructure company. Aravind's top pick among SpaceX, Anthropic, and OpenAI for a 10-year hold. Cited for Starlink connectivity, point-to-point travel, and space-based compute.

"SpaceX is the only company building space infrastructure for connectivity... That's just one aspect of the business." 01:10:30

CoreWeave AI cloud infrastructure provider. Cited as an example of a company that could be sustainable by focusing operationally on data center build-outs.

"Why is CoreWeave more successful at building data centers than OpenAI? It's operationally intensive. You got to focus." 00:35:34

Nebius AI cloud/NeoCloud. Discussed as a company that must build software on top of compute capacity to achieve sustainable margins, not just rent GPU capacity.

"The challenge that he has, which is there's a huge amount of money that wants just capacity and compute with the awareness that he needs to build a full-step product if he wants to have a long-term sustainable business." 00:37:09

Cursor AI coding tool. Mentioned alongside Perplexity and OpenAI as having been voted "most likely to fail" at a San Francisco meetup — and all three are now thriving.

"We were voted the most likely to fail. Cursor was voted the second most likely to fail. OpenAI was voted the third or something... Cursor, I think it's getting sold." 01:07:26 (Note: Aravind says "called SpaceX" but likely means Anysphere — the Cursor acquisition)

AMD Semiconductor company. Cited as a surprise beneficiary of the agent era because agent harnesses run heavily on CPUs.

"AMD is doing really well because CPUs became a bottleneck again." 00:33:46

Intel Semiconductor company. Cited as the US domestic fab investment play, with the US government owning 10% and Nvidia/SoftBank each owning 5%.

"That's why American government owns 10% of Intel. Nvidia and SoftBank own 5% each." 00:46:33

SK Hynix / Samsung Korean memory and semiconductor companies. Cited to illustrate that any company can become a trillion-dollar company — Samsung started as a grocery selling dried fish.

"SK Hynix and Samsung are worth a trillion last couple of weeks. Did you know Samsung started off as a grocery store? They started selling dried fish." 00:59:50

Meta Social media and AI company. Cited as a comparison point for market cap and discussed for its advertising-dependent model and need to launch subscription/cloud products to re-rate.

"Maybe once they do that, the narrative might change. But to go back to my point, it might not be inconceivable that Micron, the supplier of HBMs, might be more valuable than Meta in the next six to 12 months." 00:33:13

Google Search and AI company. Cited extensively as having been forced to redesign its homepage because of Perplexity, but also as behind on frontier coding models.

"I or the company Perplexity changed google.com more than any product manager at Google has ever done." 00:07:12

Cloudflare Internet infrastructure company. Referenced for the striking data point that agent traffic has now overtaken human traffic on their network.

"I saw the Cloudflare announcement that now agent traffic has overtaken human traffic for them." 00:55:28

Open Router Model routing and API aggregation platform. Discussed as a reliable token supply business whose real value is rate limit management and model fallbacks, not true model-level routing.

"Open router would go and they would pay for capacity for like one year ahead with the funding they have and secure the rate limits and multiple endpoints across multiple different providers." 00:39:36

Salesforce Enterprise software company. Cited for spending $300M on Anthropic, equating to ~3.8% of developer salaries. Also cited as a company that extended its lifespan by continuously acquiring new capabilities.

"Mark Benioff said they spent 300 million on Anthropic, which works out to be about three [percent of developer salaries]." 00:18:12

IBM Technology company. Cited as an example of a legacy enterprise staying relevant through acquisitions (Red Hat, HashiCorp, Confluent) even as its brand fades.

"IBM is still around because they went and bought Red Hat and HashiCorp and now they're buying Confluent." 01:04:15

Huawei Chinese tech/chip company. Cited as the hardware stack DeepSeek is building on instead of Nvidia.

"The deep seek is not building with the Nvidia stack. They're building with the Huawei stack." 00:44:12

Crusoe AI cloud/data center company. Named as one of several players competing in the AI infrastructure space.

"There's also other players like Crusoe and Firebird and there's a bunch of companies." 00:35:04

Fireworks AI / Base 10 AI inference providers. Cited as examples of the hosted inference business model that NeoCloud players could adopt.

"That's a business model of certain other companies like Fireworks and Base 10 and all that." 00:36:57

AWS (Amazon Web Services) Cloud platform. Cited as the model for how infrastructure companies must build software orchestration on top of raw capacity to achieve software margins.

"You have to actually build some software on top, kind of like how AWS did. It's called Amazon Web Services, not Amazon servers." 00:36:28

Booking.com / Expedia Online travel companies. Cited as top Google advertisers ($16B+ spend) to argue that high-value ad categories depend on exploration-based interfaces that chat cannot replicate.

"Number one advertiser on Google? Amazon. Number two, Booking.com... How much do you think Booking.com spends on Google? $16 billion, something like that." 00:10:51

WeChat Chinese super-app. Cited as the only example of ads in a messaging interface working — but attributed to a uniquely Chinese gamified economy, not applicable to US behavior.

"It works out in China in WeChat because there's no other way for them to fund the whole thing." 00:12:09

Stargate OpenAI's data center initiative. Cited as an example of a model lab attempting to do operational infrastructure work it isn't well-suited for.

"You could argue that OpenAI can do all the work that CoreWeave was doing. And that's kind of what they wanted to do with Stargate." 00:35:34

4. People Identified

Jensen Huang (Nvidia) CEO of Nvidia. Praised intensely for his truth-seeking intensity and operating with existential urgency despite running a $5T company.

"Jensen is so truth seeking. It's insane... He wakes up every day and tells himself that he sucks... He tells everybody around him that they're 30 days away from going out of business. Think about it. $5 trillion guaranteed to make $500 billion in revenue in the next two years, has the most advanced chips in the world. And he operates with that mentality." 01:15:12

Elon Musk (SpaceX / Tesla / xAI) Mentioned for his ability to focus on the single bottleneck problem and ignore everything else, as well as his structuring of SpaceX compensation around Mars colonization rather than personal wealth.

"His style is to just always look at the limiting problem and just ignore everything else... You have to be really good at ignoring even important things, which are distractions to your core objective right now." 01:14:32

Greg Brockman (OpenAI) President of OpenAI. Cited for his notable statement that "the model is no longer the product" — remarkable given his incentive to say the opposite.

"Greg Brockman recently tweeted, the model is no longer the product. And it's funny because you know that as a leader of a frontier lab, he has all incentive to say the model is the product." 00:13:01

Dario Amodei (Anthropic) CEO of Anthropic. Criticized for contradictory messaging on AI and jobs, and specifically for the framing that AI will eliminate jobs — which Aravind argues actively generates the public resistance blocking data center construction.

"I think so. I mean, I think they have contradictory messages in their own like different social engagements... There needs to be a consistent communication around this." 00:48:50

Mark Benioff (Salesforce) CEO of Salesforce. Referenced for disclosing $300M in Anthropic spend as a benchmark for enterprise AI token budgets.

"Mark Benioff said they spent 300 million on Anthropic, which works out to be about three [percent]." 00:18:12

Satya Nadella (Microsoft) CEO of Microsoft. Referenced for making a statement about how water-efficient data centers actually are, in contrast to public perception.

"Satya even made the statement that it's like a can of water or something in terms of how efficient these companies are." 00:41:29

David Deutsch Physicist and author. Cited for the idea that humans are uniquely capable of asking new questions about already-familiar things — used to argue there's no asymptote to human curiosity or economic value creation.

"David Deutsch is famous for saying this, right? Like we are the only species capable of being curious about what is already familiar." 00:23:59

Anne Bordetsky Referenced as someone who vouched for Perplexity's AI-pilled internal culture before the interview. (Excel partner, per Harry's reference to Samir at Excel.)

Roman (Nebius co-founder) Co-founder of Nebius. Referenced for his clear-eyed articulation that consolidation among frontier model providers is Nebius's biggest existential risk.

"Roman at Nebius said, he said if consolidation happens and there's Anthropic and OpenAI or two or three dominant providers, that is the biggest threat to Nebius." 00:38:19

Jeff Dean (Google) Legendary Google engineer. Used as an analogy for why companies will always pay a premium for frontier capability — would you hire five average engineers or one Jeff Dean?

"Would you rather hire that person and not hire people who are medium engineers but not Jeff Dean level with the same amount of budget you have?" 00:19:49

5. Operating Insights

The Cron Job Test for True Agent Adoption

Aravind identifies a precise behavioral signal that distinguishes superficial AI usage from transformative adoption: whether teams run repetitive, event-triggered workflows versus one-off tasks. This is an immediately actionable diagnostic for any company assessing its own AI maturity.

"Single biggest differentiation between those who use agents a lot and those who don't is whether they run repetitive cron jobs, whether you use AIs as one-off tasks... versus the AI is continuously monitoring something for you, continuously triggering based on certain events and going and doing certain things." 00:17:03

Build the Orchestrator, Not More Features

Aravind's internal architecture philosophy for Perplexity Computer is to act as the conductor, not a musician. The strategic frame is: stop optimizing individual features and instead build the system that routes intelligence optimally across all available resources. Applied to any product company, this means the leverage is in the harness design, not the underlying model selection.

"Computer is the orchestra conductor... What it orchestrates keeps evolving... It changes from models to files, to tools, to chips, to devices, but it doesn't even matter. Like you don't care as long as it orchestrates things correctly and maximizes the token value for what? For the user." 00:28:07

Train Your Own Models on Top of Open Source to Drive Down COGS

Perplexity's explicit path to margin expansion: post-train on open source models to handle what's already in production, and reserve expensive frontier model spend only for discovering new capabilities. This is a concrete cost architecture applicable to any AI product company.

"We're training our own models, post-training it on top of amazing open source models. And that will bring down the cost that we currently spend on frontier model tokens. We expect to continue to use frontier models for designing new experiences... But whatever exists today in our products right now, we expected to completely rely on models we own and serve ourselves." 01:05:42

Use Compute Credits as a Demand Generation Tool

Perplexity is giving $1M in compute credits to any team with a credible path to building a billion-dollar company, explicitly trying to fund 1,000 such companies. This is both ecosystem building and distribution — the companies built on your credits become your users.

"We're funding this thing called the billion dollar build, where we're giving a million dollars of computer credits to any group of people who have a credible path to building a billion dollar company. And I want like 1,000 such companies to be built." 00:49:59

Focus Only on the Limiting Constraint

Aravind names this as the most important leadership skill he observed in Elon Musk: the ability to identify the single bottleneck and ignore everything else, including things that are genuinely important but are not the current constraint.

"His style is to just always look at the limiting problem and just ignore everything else. That's very hard to do because you actually have to be really good at concentration. You have to be really good at ignoring even important things, which are distractions to your core objective right now." 01:14:32

6. Overlooked Insights

The 24/7 Always-On AI Cannot Be Realized on the Server — It Requires a Hybrid Local/Server Architecture, and Whoever Builds It Owns the Most Valuable Real Estate in AI

This point was made briefly and not picked up on by Harry, but it is enormous. Aravind is describing a specific architectural vision: the 24/7 AI agent that most people assume will run on the cloud is actually economically impossible at cloud token prices. The company that solves local-server orchestration — building a continuously learning local model that handles routine context and offloads to server-side frontier only when necessary — will own the most intimate, always-on relationship with users. This is essentially the OS-level position of the next computing era.

"Nobody's going to be able to afford a cron job at the fidelity of few seconds that runs all the time. The bottleneck there is actually orchestration and local compute... One needs to build a continuously learning local model that can save you on context windows, and try to preserve as much compute locally, and rely on the server side frontier only when necessary... That model is not just a model, it's a model plus the harness plus the local chip and the compute and the ecosystem of devices it controls. That system is going to be your own intelligence, essentially the data center moved to your local device." 00:26:10

Perplexity's "Billion Dollar Build" Is an Undisclosed Distribution Strategy, Not Just Philanthropy

Aravind mentions the "billion dollar build" — giving $1M in compute credits to teams building billion-dollar companies — in a single throwaway sentence, and Harry moves on. But this is structurally identical to what Amazon did with AWS credits for startups, which created a generation of companies that were deeply dependent on, and loyal to, AWS. Perplexity is doing the same thing: seeding 1,000 companies on Perplexity infrastructure creates a massive captive demand base, a distribution network, and a moat against competing orchestrators. The comparison to Amazon's startup credit programs is made explicitly, and the fact that Perplexity's own founding was enabled by ~$1M in cumulative cloud credits makes this a deliberate, experienced playbook move.

"When we started Perplexity, we had like around $200,000 worth of Amazon credits and GCP credits and Azure credits that almost like together, cumulatively, this was worth like a million dollars in compute credits. Now, in today's world, it's going to be like a million dollars of computer credits. And we're doing that. Like we're funding this thing called the billion dollar build." 00:49:30