Library/How Export Controls Helped Not Hurt China & Power is the Bottleneck to AI | Perplexity CEO

20VC with Harry Stebbings

How Export Controls Helped Not Hurt China & Power is the Bottleneck to AI | Perplexity CEO

1h 35mJun 15, 2026
Key Themes
AI orchestrationAgentic workflowsPower bottlenecksInference economicsExport controlsAI-native startupsFounder mindset
Summary

Perplexity CEO Aravind Srinivas on why AI’s next fight is agents, orchestration, power, and physical infrastructure

Aravind Srinivas argues that AI search answers are becoming commoditized and that the next major value layer will be products that do work: agents, deep research, coding, computer use, and enterprise workflows. Across the conversation, he presents Perplexity as an orchestration layer that routes across models, tools, devices, chips, local compute, and context rather than betting solely on one model. The episode also digs into AI economics: token costs, power constraints, inference infrastructure, open-source models, export controls, China’s strategic response, and the possibility of small AI-native teams building unusually large companies. It closes with reflections on founder mindset, differentiated AI labs, SpaceX as a long-term infrastructure bet, and leadership lessons from Elon Musk and Jensen Huang.

1
AI value is moving from answers to action

Srinivas argues that basic question-answering and consumer search are becoming less differentiated. The frontier, in his view, is shifting toward tools that perform work for users: deep research, coding, computer use, monitoring, triage, and persistent agent workflows. That reframes AI products around outcomes rather than chat responses.

2
The orchestration layer may matter as much as the model

A recurring claim is that the model alone is no longer the whole product. Durable AI systems need harnesses, tools, files, connectors, routing, local compute, frontier models, and user context working together. Perplexity is presented as an example of a company trying to become that coordinating layer.

+1 more
3
Power is presented as AI’s defining bottleneck

The conversation repeatedly returns to physical infrastructure: power, land, cooling, permits, turbines, memory, CPUs, and data-center operations. Srinivas argues that buying GPUs is only part of the problem; the companies and regions that solve the physical bottlenecks may shape how fast AI can advance.

+1 more
4
AI economics will be shaped by token value, not just token volume

Srinivas introduces “token value per watt per user” to connect model output, user value, and power consumption. He argues a smaller number of power users running high-value agent workflows could generate major revenue, while enterprises will increasingly scrutinize token budgets and demand smarter routing or local inference.

+1 more
5
Export controls may buy time while also forcing China to build deeper capability

The episode presents a nuanced view: export controls may preserve a short-term gap between frontier models and open-source competitors, but they may also push China to vertically integrate across chips, memory, fabs, data centers, and model architecture. That could create a more self-sufficient and resilient competitor over time.

+1 more
6
AI could make small teams far more capable

Srinivas argues that AI-native companies may reach major scale with far fewer people and that one or two motivated founders can now attempt projects that once required larger organizations. He links this to a more optimistic jobs narrative: AI as a tool for new company creation, not only labor displacement.

+1 more
7
AI adoption may divide objective and subjective decisions

Srinivas suggests agents will be strongest where decisions are objective and measurable, such as certain transactions or operational tasks. More subjective categories—fashion, furniture, travel inspiration, and lifestyle choices—may remain tied to browsing, discovery, and advertising because preferences are harder to reduce to agent optimization.

+1 more
8
Founder durability is framed as speed, focus, and truth-seeking

Beyond AI strategy, Srinivas emphasizes an aggressive operating posture: move fast, stay curious, focus on the limiting bottleneck, and avoid becoming comfortable after early success. His praise for Elon Musk and Jensen Huang centers less on wealth and more on focus, paranoia, and relentless problem-solving.

+1 more
Select any chapter text to Deep Dive with AI
01Intro: Founder Mentality, AI Search, Agents, and Where the Money Is

The episode opens with Aravind Srinivas describing a combative founder mindset shaped by his background and a belief that he has little to lose. He argues Perplexity pressured Google to change search, with Google’s AI Mode now looking similar to Perplexity’s answer-engine interface. He also says basic question-answering is no longer the frontier: the paid value in AI is shifting toward agents, deep research, coding, and computer-use workflows.

Srinivas frames his motivation as winning and staying on offense rather than fear of failure.
Perplexity is introduced with reported scale: a $20B valuation, 45M users, over 1B searches per month, and roughly 400 employees.
He says Perplexity influenced Google’s search roadmap and sees Google AI Mode as validation of answer-engine design.
He argues consumer search answers are becoming commoditized, while higher-value AI products will do work for users.
He is skeptical that chat interfaces can become massive ad businesses because ads may undermine trust and because many ad categories rely on exploration.
02Model Is Not the Product: Agent Harnesses, Power Efficiency, and AI Revenue

Srinivas argues that raw model access will be commoditized and that durable value will come from harnesses that combine models with context, tools, connectors, interfaces, and workflows. He introduces “token value per watt per user” as a core metric and predicts that intense power users running persistent agents could create enormous revenue even without consumer-scale user counts. The chapter also broadens the frontier beyond coding into enterprise workflows, chips, drugs, robotics, climate, and scientific discovery.

The “frontier” is defined as the best achievable AI outcome, not necessarily one specific model.
Agent harnesses, context, tools, and connectors are described as the product layer that turns intelligence into useful work.
Perplexity’s differentiation is framed as orchestration across both models and tools.
Power efficiency is tied directly to AI economics through the idea of maximizing token value per watt per user.
Persistent agent workflows such as monitoring, email triage, debugging, and research are described as the current frontier.
Open-source models may lower costs for today’s tasks, but users may keep paying for whichever capabilities remain at the frontier.
0324/7 AI Agents, Orchestration, and the Power Bottleneck

The conversation explores what it would take to make always-on personal AI practical. Srinivas argues that continuous frontier-model use would be too expensive, so systems will need to route tasks among local models, device-side chips, private context, tools, and cloud frontier models. The discussion then shifts to infrastructure, where power, cooling, permitting, land, memory, CPUs, and operational execution become the real constraints behind AI progress.

The challenge with 24/7 agents is framed less as rogue behavior and more as the cost of keeping powerful models running continuously.
Local and device-side AI are presented as ways to reduce cost, preserve privacy, and make persistent agents viable.
Perplexity is described as an “orchestra conductor” across models, tools, chips, devices, files, and connectors.
Power is identified as a larger data-center bottleneck than simply acquiring GPUs.
Bottleneck suppliers in memory, CPUs, storage, and power may gain pricing power as agent workloads grow.
Neoclouds may be durable if they solve hard physical operations and add software layers, not merely rent GPU servers.
04Inference Economics, Power Bottlenecks, Export Controls, and AI Jobs

Srinivas weighs whether inference and routing companies can become very large standalone businesses. He argues inference can be meaningful if open-source models stay competitive and providers execute on hosting, capacity, and operations, while routing is more about reliable token supply than clever model choice. The chapter also covers public resistance to data centers, export controls that may help in the short term but strengthen China long term, and a more optimistic story about AI-enabled entrepreneurship.

Inference businesses could be very large if they reach major revenue scale with strong margins and operational excellence.
Model routing is framed as reliability, fallbacks, rate limits, and endpoint access more than high-margin intelligence.
Power, permitting, politics, and physical supply chains are presented as durable AI bottlenecks.
Public resistance to data centers is linked to environmental concerns, grid-price fears, inequality, and negative AI sentiment.
Export controls are described as helpful short term but potentially counterproductive if they push China to vertically integrate across chips, memory, fabs, and architectures.
Srinivas criticizes doom-heavy jobs messaging and argues for highlighting AI-enabled company formation and compute-credit programs.
05AI-Native Companies, Token Budgets, and the Rise of Agentic Work

The discussion turns to how AI may change company-building and work. Srinivas argues that small teams can now build much larger companies with fewer employees, while workers should begin using AI through curiosity rather than obligation. Enterprise adoption may run into token-budget constraints, making hybrid/local inference and orchestration layers more important. The chapter also covers agent traffic on the internet, objective versus subjective transactions, and the strategic question of what leaders would do with thousands of agents.

AI-native companies may reach large valuations with far fewer employees than historical companies required.
AI may lower the barrier to entrepreneurship for small teams and individuals.
Workers are advised to get started with AI through curiosity and experimentation.
Token budgets and cost scrutiny could slow enterprise AI adoption unless hybrid inference and orchestration reduce spend.
Google is described as advantaged in low-cost token production but not necessarily at the coding-model frontier.
Agent traffic could reshape the web, especially for objective transactions, while subjective categories may remain more ad-driven.
A strategic leadership question is how much faster goals could be pursued with thousands of agents and sufficient compute.
06AI, Wealth Mobility, Model Costs, and Perplexity’s AGI Ambition

Srinivas connects AI to opportunity and wealth mobility, arguing that people with curiosity and agency can use AI to create income and businesses. He then discusses Perplexity’s growth, top-line priorities, margin strategy, and plan to reduce frontier-model token costs by training and serving more of its own models. The chapter closes with his view that speed matters more than early moats, Perplexity’s ambition to operate with semi-autonomous agents, and data centers as the industrial-scale infrastructure buildout he would pursue with unlimited capital.

AI is framed as widening opportunity for people who are willing to experiment and take repeated shots.
Perplexity is described as growing quickly, above a $500M revenue run rate, and focused on top-line growth while keeping a path to profitability.
The company plans to lower model costs by post-training open-source models and serving more capabilities internally.
Frontier model providers are said to stay relevant only if they keep delivering new capabilities.
Srinivas says Perplexity responded to skepticism by tripling revenue and cutting burn by more than half.
He argues startups should prioritize speed and market feedback over obsessing about a premature moat.
Perplexity’s internal ambition includes semi-autonomous agents running parts of the company.
With unlimited capital, Srinivas says he would build data centers because AI requires a new physical infrastructure wave.
07SpaceX as the 10-Year Bet, Differentiated AI Labs, and Founder Mindsets

The episode ends with a rapid-fire discussion of long-term bets, AI careers, differentiated labs, and founder psychology. Srinivas chooses SpaceX over OpenAI and Anthropic as a 10-year hold because he views it as a unique space-infrastructure and connectivity company. He advises graduates to stay curious, argues that AI labs need real foundational differentiation, and frames Perplexity’s long-term path around accuracy plus orchestration. He closes by praising Elon Musk’s focus on bottlenecks and Jensen Huang’s truth-seeking, paranoia, and willingness to keep building despite success.

Srinivas picks SpaceX as a preferred 10-year bet because of its unique space and connectivity infrastructure position.
Starlink is cited as a tangible product that can reset expectations for in-flight internet.
Future tech jobs may look like reincarnations of existing roles, with forward-deployed and quality-control-oriented work becoming more important.
Graduates are advised to stay curious and avoid short-term AI FOMO.
He is skeptical of undifferentiated AI labs and favors foundational bets such as alternatives to transformers, GPU dependence, or robotics models.
Perplexity’s trillion-dollar path is framed around accuracy and orchestration across devices, chips, models, tools, files, and connectors.
Elon Musk is admired for focusing on the limiting bottleneck; Jensen Huang is admired for truth-seeking and operating with urgency despite NVIDIA’s scale.