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

Perplexity forced Google's AI mode redesign

Srinivas claims Perplexity changed Google.com more than any internal PM, evidenced by AI mode copying Perplexity's interface, inline citations, and follow-up suggestions.

Offense-only mentality drives survival

Coming from a lower-middle-class background in India, Srinivas maintains an 'attack, attack, attack' philosophy, believing defense and comfort lead to rapid obsolescence in AI.

Answer engine as strategic lead generation

Perplexity's search product served as a gateway to build frontier agent products, as resting on that success would have prevented development of deep research and computer use capabilities.

Token value per watt is the key metric

Power is the only unsubsidizable constraint in AI, making 'token value per watt per user' the critical economic metric for sustainable value creation.

Orchestration harness beats raw models

Following the view that 'the model is not the product,' Srinivas emphasizes that agent harnesses—which orchestrate models, tools, and connectors—capture value that raw model tokens cannot.

Multi-model orchestration creates moats

Perplexity differentiates by orchestrating across competing models (GPT, Claude) within its harness, unlike OpenAI or Anthropic who restrict their interfaces to proprietary models.

Chat advertising corrupts user trust

Srinivas is bearish on advertising in conversational AI because it undermines trust in objective answers and fails to serve the discovery-based behavior seen in travel and shopping.

Power users drive the token economy

The highest-value customers are not mass consumers but power users spending up to $10,000 monthly on sophisticated agent workflows that run continuous cron jobs and monitoring tasks.

Enterprise developer budgets shift to AI

Enterprises like Salesforce already spend 3.8% of developer salaries on AI tools, indicating a massive shift from headcount budgets to token-based infrastructure costs.

OpenAI's dominance is temporary

Srinivas believes OpenAI is neither financially ready for an IPO nor secure in its leadership, as even dominant consumer products can be disrupted within 6-12 months without continuous innovation.

Frontier has shifted to autonomous work

The value frontier has moved from question-answering to agents performing autonomous work, requiring companies to constantly reinvent products rather than defending existing features.