20VC: Micron Will Be More Valuable Than Meta | How Export Controls Helped Not Hurt China | Power is the Bottleneck to AI | Why Dario Has Done a Disservice to AI with his Labour Replacement Messaging with Aravind Srinivas, Founder @ Perplexity

What It Covers

Perplexity CEO Aravind Srinivas argues that the AI race is fundamentally an orchestration problem, not a model race. He covers power as the primary bottleneck to AI infrastructure, why Micron could surpass Meta in valuation, how export controls inadvertently strengthened China, and why Dario Hassabis's labor replacement messaging damages the broader AI ecosystem.

Key Questions Answered

• •Orchestration over models: The most valuable AI metric is token value per watt per user. Companies that orchestrate across multiple models, tools, and devices — rather than building a single model — capture more economic value. Perplexity routes across both Anthropic and OpenAI models simultaneously, something neither lab can do for the other, tripling revenue since early 2025 as a direct result of this multi-model architecture.
• •Power as the real bottleneck: Roughly 40% of planned U.S. data centers are not being built due to public resistance rooted in misinformation about water and energy consumption. This resistance — not chip supply or capital — is the primary constraint on AI infrastructure scaling. Srinivas argues factual public education, not fear-mongering about job losses, is the lever to unlock faster build-out.
• •Micron valuation thesis: Memory bandwidth is the current hardware bottleneck in AI infrastructure, with HBM prices rising 5x in cost of goods. Srinivas predicts Micron could surpass Meta's $1.3–1.4 trillion market cap within 6–12 months because whoever controls the bottleneck commands pricing power — the same dynamic currently benefiting AMD as agent loops drive renewed CPU demand in enterprise environments.
• •Export controls paradox: U.S. export controls on NVIDIA GPUs and HBM chips created a roughly 12-month gap between open-source and frontier models, which Srinivas credits as a short-term win. However, the controls forced DeepSeek to build a vertically integrated stack on Huawei hardware with innovations in KV cache efficiency and attention layers — potentially producing a more memory-efficient architecture that could disrupt NVIDIA-dependent U.S. infrastructure at scale.
• •24/7 agent economics: Continuous always-on agents are cost-prohibitive if run entirely on server-side frontier models. The viable path requires a hybrid architecture: a continuously learning local model handles repetitive, low-complexity tasks while routing to server-side frontier models only when necessary. Srinivas frames this as the orchestration problem — maximizing intelligence and accuracy while preserving privacy and controlling cost across local and cloud compute simultaneously.
• •Messaging damage from doom framing: Dario Amodei's public statements linking AI directly to mass labor displacement contradict Anthropic's own data showing no current evidence of AI-driven job losses, creating contradictory public messaging. Srinivas argues this framing actively suppresses data center permitting, increases regulatory friction, and discourages entrepreneurship — while the more accurate story is that agentic AI enables small teams of 20–40 people to build billion-dollar companies that previously required hundreds of employees.