Perplexity CEO: Micron Will Be More Valuable Than Meta & Power is the Bottleneck to AI

Why does it matter? Because the companies winning the AI race aren't building the best models — they're routing intelligence more efficiently than everyone else

Perplexity's CEO Aravind Srinivas has a theory that scrambles the standard AI investment thesis: the model is already a commodity, the orchestration layer is the only durable business, and a memory chip supplier will be worth more than Meta within a year. He says it flatly, then backs it with a specific metric, a specific stock call, and a direct indictment of every lab that thinks their weights are their moat.

• Whoever routes AI tokens most efficiently across models, tools, and local compute wins — not whoever builds the best model; the only metric that matters is token value per watt per user
• Power is the binding constraint on AI development, not compute or software — 40 out of every 100 planned US data centers are blocked by public resistance
• Chat interfaces structurally cannot support a $100B advertising business; the format destroys the user trust that advertising requires
• Export controls designed to slow China down may be producing the architectural innovations that make China a stronger long-term competitor, not a weaker one

The model is already a commodity — whoever wins the orchestration layer wins everything

Greg Brockman recently tweeted "the model is no longer the product" — which Srinivas finds remarkable precisely because Brockman runs a frontier lab and has every incentive to say the opposite. The fact that he said it anyway is its own signal.

The argument: if you're just reselling model tokens, you have no business, because models get commoditized. The real value is in the agent harness — the rules governing how the loop runs, which sub-agents and tools it accesses, how it converts raw intelligence into useful output. "Without the harness you don't necessarily capture and convert the intrinsic intelligence in the model into valuable output tokens."

Perplexity's specific differentiation is something OpenAI and Anthropic structurally cannot offer: orchestration across competing models. You won't find GPT-5 inside Claude Code's harness. You won't find Claude inside Codex. But Perplexity Computer runs both — which means every improvement anywhere in the stack, from Jensen's next chip to Dario's next model, makes their product better rather than threatening it.

The metric Srinivas wants to replace benchmark obsession with: token value per watt per user. How much value does each watt of compute produce for each user? That's the number that determines who captures AI's economic surplus. The orchestrator — not the model builder — is the net beneficiary of every advance at every layer of the stack.

40% of planned US data centers are dead on arrival — and power will stay the bottleneck for at least three years

Building a data center isn't a compute problem. It's a permitting problem, a power procurement problem, and increasingly a public opinion problem. "I think right now 40 out of 100 are not being developed because of public resistance."

People believe data centers consume enormous amounts of water and electricity. Srinivas says both claims are wrong — Satya Nadella reportedly compared the water consumption to a can of water. But factual accuracy doesn't matter if the narrative is already set. The resistance is channeling multiple anxieties at once: wealth inequality, climate fear, rising electricity bills, the sense that RAM prices have gone up because AI companies bought it all. "It's a lot of things" — and none of them are easily corrected with a press release.

The practical consequence: power will remain the bottleneck for at least three years. "If I gave you unlimited money, what would you do today that you're not doing? I would build data centers." The buildout is migrating toward countries with more permissive regulations, toward space (Elon Musk's long-term bet), toward wherever natural resources and political will coincide.

The companies and countries that solve permitting, power procurement, and the public narrative around data centers — not those that win the model race — will determine where frontier AI actually runs. That is a physical infrastructure and political problem, not a software or research problem.

Chat can't carry a $100B ad business — the format kills the trust that advertising depends on

Who's Google's number-one advertiser? Amazon. Number two? Booking.com, spending something like $16 billion a year. Where do you book hotels? Still Google. Why? Discovery. You want to see options, compare prices, browse. "When the decision making is more subjective and vibes based, you don't need an objective answer engine."

Fashion goes to Instagram because you're doom-scrolling, not querying. Travel goes to Google because you want a grid of results, not a recommendation. "The chat interface doesn't capture that user intent, that user behavior right now, which is why it was never a great fit for advertising."

And when a chat product tries to insert ads anyway, something else breaks: "You ask for the best protein shake, but by the way, here are some protein shakes you can check out — it kind of hurts the trust." Meta tried ads inside messaging apps and email. It never worked in America. WeChat only works because there's no alternative in China.

The bull case for OpenAI's ad revenue rests on replicating Google's model inside a format that structurally destroys what makes advertising work: browsing behavior, discovery intent, and user trust in the answers. Objective purchases get disrupted by agents. Subjective purchases stay on discovery platforms. There is no large advertising surface left for chat-native AI. "I'm bearish on advertising to really take off in the chat interface."

Micron could surpass Meta's market cap in 6–12 months — and the logic is airtight

Micron is already at roughly $1 trillion. Meta sits at $1.3–1.4 trillion. "It might not be inconceivable that Micron, the supplier of HPMs, might be more valuable than Meta in the next 6 to 12 months."

The reasoning is clean: whatever is the bottleneck commands the price. High bandwidth memory is the bottleneck for GPU inference. AMD is doing well right now because agent loops run on CPUs — "agents are using CPUs more than humans" — and enterprise CPUs became a bottleneck again. Whoever supplies the chokepoint sets the terms.

The counter is that Micron is already fully priced. Srinivas rejects this: fully priced assumes the bottleneck goes away. The Blackwell generation of models is just starting deployment. Rubin follows next year at full capacity. Each generation demands more memory bandwidth, not less.

The framing to walk away with: stop evaluating AI companies by application-layer multiples. Map the supply chain's current chokepoints — HBM memory, power, cooling, networking — and find where pricing power is actually concentrating. That's where value is being created, even when the narrative is pointed elsewhere.

The users running $10,000/month in agent loops are the real signal — not 100 million casual users

One Perplexity user spends upward of $10,000 a month. Their entire business runs on agent loops inside Perplexity Computer. This isn't an anomaly — it's the shape of the market Srinivas is building toward.

"The single biggest differentiation between those who use agents a lot and those who don't is whether they run repetitive cron jobs." Not one-off research tasks. Continuous workflows: triaging every inbound email, running root cause analysis every time a latency spike hits, monitoring systems and triggering actions on events around the clock.

Nobody can afford to run frontier AI 24/7 on servers — the cost is prohibitive at current pricing. The solution is hybrid: a continuously learning local model handling context compaction and routine tasks, calling server-side frontier compute only when necessary. Build the local/cloud orchestration right and you can realize the 24/7 agent without bankrupting your users.

"These products are not going to be used by 100 million people but they will generate revenue that's going to be higher than the advertising revenue of Google or Meta." The product design implication: most current AI tools optimize for single-session interactions. Transformative economic value accumulates in persistent, event-triggered loops — a different architecture, different pricing, and a fundamentally different customer profile.

Export controls are forcing China to build a vertically integrated AI stack from scratch — and it may end up more efficient than ours

DeepSeek isn't building with Nvidia. They're building with Huawei. Export controls on both GPUs and high bandwidth memory have forced their entire architecture to optimize around constraints that US labs don't face.

The results are architectural innovations that may outlast the controls themselves: KV cache innovations small enough to host on SSDs, no dependence on high bandwidth memory at inference time, different storage architectures because 3D NAND is also restricted. "These architectures that DeepSeek's building are far more memory efficient." The model architecture is different. The training algorithm is different. The whole stack is being vertically integrated to their specific hardware and fabs.

Meanwhile, China can build data centers faster. Power is not a problem. Permits are not a problem. Labor is not a problem. "By forcing them to go out there and build all this, you're converting them into a far more potent competitor."

Srinivas puts the probability of another DeepSeek-level architectural surprise at 20–30%. That's not a tail risk — that's a planning scenario worth modeling seriously. And if that surprise arrives, anyone who over-built US data center capacity for the current Nvidia stack has a real problem.

Frontier labs are on a 6-month capability treadmill — and their customers are already building exit ramps

"If Anthropic thinks Claude Code is already a win, in 6 or 12 months from now, they won't even be around." Srinivas says this about the company he currently depends on for model access. The point isn't hostility — it's structural reality.

"Frontier model providers will only remain relevant if they remain at the frontier. If for 6 months you're not seeing a new capability, it's bad for them." The assumed defensibility of current frontier lab valuations rests on a capability gap that open source is closing quarter by quarter. Enterprises are already planning the contingency: fine-tune an open model, cut the frontier dependency, keep the frontier API only for genuinely novel capabilities.

Perplexity's own plan is explicit: whatever features exist in their products today, they expect to handle entirely with models they train and serve themselves. Frontier APIs get reserved for designing new experiences that don't exist yet. The customers paying premium prices today are six months into the same transition.

The right lens for valuing frontier labs: R&D velocity — release cadence, capability novelty, time between genuine breakthroughs — not revenue multiples on current ARR. "No one's ever in a comfortable position. No one can relax."

AI just made a billion-dollar company achievable with 40 people — and that changes who gets to build

Perplexity: $20 billion valuation, 400 people. Srinivas runs the math live: with 40 people, you can probably build a billion- or two-billion-dollar company. With 4,000, maybe $200 billion. The compression factor is roughly 10x — and it's already happening.

"For the first time in history, you can get started on an idea with like one or two other friends and maybe have a real genuine shot at building a billion dollar company." Perplexity is running a program called Billion Dollar Build, distributing $1 million in compute credits to any group with a credible path to that outcome. The goal is a thousand such companies. "I would rather have those 100,000 people be split into groups of like thousand — and each of those thousand groups are worth a few billion dollars."

The unlock isn't redistribution. It's access to the infrastructure — compute credits, agent harnesses, orchestration layers — that previously required hiring 200 people to assemble. An Uber driver in San Francisco watched one of Srinivas's interviews, built a web app with AI assistance, added billing, and now makes more passive income from the app than from driving. That's one data point. But it's the shape of the bet.

The next AI competition will be won in the physical world, not the model lab

Perplexity tripled revenue since being voted most likely to fail in San Francisco. The model powering that growth is becoming a commodity. The power running it is becoming scarce. The memory chip storing it might soon be worth more than the world's largest social network. What this conversation signals about where things are heading: the decisive advantages in AI are no longer primarily about who writes the best training code. They belong to whoever secures electrons, permits, and memory bandwidth — then routes all of it most efficiently to users. Software ate the world. Now the world is eating back.

---

Topics: artificial intelligence, AI infrastructure, Perplexity, large language models, AI orchestration, semiconductor memory, data centers, advertising, China AI, startup strategy, agentic AI, token economics, GPU supply chain, OpenAI, Anthropic