Perplexity CEO: Micron Will Be More Valuable Than Meta & How Export Controls Helped Not Hurt China

Summary

• Micron more valuable than Meta within 6-12 months is Aravind’s headline call: Micron is already near $1T against Meta’s $1.3-1.4T, and “whatever is the bottleneck will command the price.” Memory has 5x’d in price and still isn’t fully priced because it’s still the bottleneck — same logic explains AMD’s run, since “agents are using CPUs more than humans” for the download-munge-host work around GPU-generated tokens.
• Consumer search is commoditized — “there’s no money there.” OpenAI is still the dominant leader in consumer search: Aravind says yes when Harry asks about dominance, and Harry specifies consumer search. Aravind then dismisses its monetization. That’s exactly why everyone (Codex, Claude Code, Perplexity Computer) is racing to agents that do work: “the frontier is where the money is.” He told Harry pre-show OpenAI isn’t ready for an IPO.
• The model is not the product; the single most important metric in AI is “token value per watt, per user.” Pure token resellers have no business and even model builders get commoditized; value sits in the harness-plus-orchestration layer. Perplexity’s edge: it orchestrates across models — “you wouldn’t find GPT-5-5 inside the Claude Code harness… you would find both these models inside Perplexity Computer.”
• Bearish on chat advertising — a direct challenge to the OpenAI ads thesis (“yet to be proven”). Google’s top advertisers are Amazon, Booking (~$16B/yr), Expedia; travel and fashion are subjective, exploratory decisions where “the interface is less about conversations and more about exploration,” and ads inside an answer engine “fundamentally corrupts the trust.” His split: objective transactions go agent-based, subjective ones stay ad-based.
• Power is and will remain the bottleneck: he estimates ~40 of 100 data centers aren’t being developed because of public resistance, and expects resistance to worsen as anti-AI sentiment channels through inequality, climate, grid prices, and RAM prices. Given unlimited money, the one thing he’d do is build data centers.
• Export controls: “jury’s still out” — short-term they’re the only reason a frontier/open-source gap exists (Anthropic lobbied hard for them), but they’re forging China into “a far more potent competitor” that’s vertically integrating around the Huawei stack. He puts 20-30% odds on another DeepSeek moment — a radically more efficient architecture that strands US capacity.
• Dario has done AI “a disservice” with jobs-doom messaging that’s also internally contradictory (“there is no evidence that AI is taking over jobs”): “You can’t win by saying that and also complaining about not being able to build data centers.” His counter-program: $1M compute credits to anyone with a credible path to a $1B company.
• Perplexity by the numbers: 400 people, ~$20B valuation, revenue “more than tripled” this year, ARR “far, far more” than $500M, burn down 50%+, IPO hopefully before 2028. The coming IPO wave could get partly funded by reallocation — he sketches Vanguard/BlackRock moving $30-40B out of Microsoft/Salesforce into Anthropic. His 10-year buy-and-hold among SpaceX/Anthropic/OpenAI: SpaceX, the only company building space infrastructure for connectivity.

Deep dive

1. “Attack, attack, attack” — and the claim that Perplexity redesigned Google

• The origin story frames everything: lower-middle-class in India (“not even like lower middle class in UK or the US”), a family for whom a Google engineering job was the ceiling of ambition. “I have nothing to lose. I came from nothing… Be on the offense all the time. Attack, attack, attack. That’s my motto.” Playing defense, he reminds himself, is “the stupidest thing to do.”
• His boldest claim, made without flinching: “Perplexity changed google.com more than any product manager at Google has ever done.” Nobody inside Google wanted to tinker with the interface making them $250B a year — and now AI Mode “looks exactly like Perplexity. There’s not even any difference — the font, the citations, the specific bolding of inline text, inline hyperlinks, suggested follow-ups… Except it’s still not as good.”
• The answer engine “was always a lead gen for the frontier products” — and the discomfort is universal: “If Anthropic thinks Claude Code is already a win, in 6 or 12 months from now, they won’t even be around. No one can relax.”

2. Consumer search is commoditized — the money is at the frontier

• Harry surfaces Aravind’s pre-show remark that OpenAI isn’t ready for an IPO, then asks whether OpenAI is still a dominant leader. Aravind says yes; Harry specifies “consumer search,” and Aravind then guts the concession: “except there’s no money there, right? Because it’s been commoditized.” The proof is revealed preference: why else is OpenAI all-in on Codex, Anthropic on Claude Code, Perplexity on Computer, Meta launching “Hatch” (as heard) at $200/month?
• Frontier doesn’t mean frontier model. He endorses Greg Brockman’s tweet that “the model’s no longer the product” — noting the tell that a frontier-lab leader has every incentive to say the opposite. In non-advertising revenue, “the money is in whatever is the frontier. And today the frontier is about going out there and doing things for you.”

3. Bearish on chat advertising — the category-by-category takedown

• Asked whether OpenAI builds a $100-200B ad business: “Yet to be proven.” Then the walk-through: Google’s #1 advertiser is Amazon, #2 Booking.com (~$16B/year), #3-4 Expedia. Where does Harry book flights? Google — because “I’d like to see the options.” Aravind: “Exactly. The interface is less about conversations and more about exploration. When the decision-making is subjective and vibes-based, you don’t need an objective answer engine.” Fashion and DTC budgets go to Meta because you’re doom-scrolling, not asking.
• The deeper objection is trust: recommending protein shakes with sponsored ones appended “fundamentally corrupts the trust that people have” in an accuracy product. Ads-in-messaging has never worked outside WeChat, where the whole economy was gamified around it. “I’m bearish on advertising to really take off in the chat interface. I’m happy to be proven wrong there.”
• His durable split, prompted by Cloudflare reporting agent traffic overtaking human traffic (faster than Harry expected): “Anything where the transaction is based on objective judgment gets disrupted by agents. Subjective things will still be ad-based.” You’ll buy the mic on objective specs; the table depends on the aesthetics of the room. The advertising internet doesn’t die — it splits.

4. The model is not the product — maximize token value per watt, per user

• His definition of the layer that matters: a harness is “rules for how the agent loop should run” — skills, sub-agents, connectors, tools. Without it “you don’t capture and convert the intrinsic intelligence in the model into valuable output tokens.” Pure token resellers have no business; even model builders get commoditized; infra earns something; the value is in grounding a model in context and orchestrating it as one unified system.
• The differentiation claim: Perplexity orchestrates across competing models — “you wouldn’t find GPT-5-5 inside the Claude Code harness. You wouldn’t find Claude Opus 4.7 or 4.8 inside the Codex harness… you would find both these models inside Perplexity Computer.”
• The thesis in one line, repeated deliberately: “The single most important metric in AI is token value per watt, per user.” Price is fundamentally power-denominated, and nobody but a government can subsidize watts — so whoever produces the most valuable output tokens with the least power expended has the most pricing power. “Short-term it might look like this lab’s revenue is growing exponentially… long-term, this is the one objective that truly matters.”

5. A few power users propel the token economy — and they’ll pay for the frontier

• Unlearn the billion-user mentality: an engineer got Amazon spending “half a billion dollars a month” via a runaway Claude Code agent loop; real engineers at Meta spend “$10 million a year per engineer” on coding tools; one Perplexity Computer user spends upwards of $10,000/month — not waste, “their business runs using agent loops.” The single biggest divider between heavy and light users: “whether they run repetitive cron jobs” — continuous monitoring, inbox triage, latency root-cause — versus one-off delegation. “These products are not going to be used by 100 million people. But they will generate revenue higher than the advertising revenue of Google or Meta. It’s going to happen.”
• Harry’s sharpest number: Benioff’s $300M Anthropic spend is 3.8% of Salesforce’s developer salaries. Stay at 3.8% and the labs are never $5T; go to 100% (as Brandon McQuaid predicts within a year) and they’re $10T. Aravind: “they can certainly be $10 trillion companies” — but the bigger market is non-developers (finance, corp dev, sales reps, research analysts), which Perplexity Computer targets: “Think of it as Claude Code multiplied by 10.”
• On whether token costs fall (Harry: “we thought agent costs would go down — they’ve gone up”): “Yeah, for now.” You pay for the frontier the way you’d pay $1M for one Jeff Dean over five $200k engineers. Thought experiment: an open-source model as good as Opus 4.8 at 10x cheaper in 12 months kills spend on today’s tasks — but the frontier moves to autonomous software engineers, AIs designing chips and drugs, curing cancer: few users, enormous reach. “It feels like a contradiction. It’s not.” Anthropic bought a vet lab; he speculates about using its tokens for vet-lab experiments.

6. 24/7 AI is coming — and the real constraint is cost, not safety

• Four competing objectives — intelligence, accuracy, privacy, cost. Everyone worries an always-on agent “does something crazy”; “the real concern actually is the cost. Nobody’s going to be able to afford a cron job at the fidelity of a few seconds running all the time” on server-side frontier compute. The answer is a continuously learning local model plus harness plus the local chip — “essentially the data center moved to your local device” — with server-side frontier used only when necessary.
• The positioning metaphor, fully built out: Computer is the orchestra conductor; sub-agents are the musicians; models, tools, and connectors are the instruments; the symphony is the work. What gets orchestrated keeps changing — models, files, chips, devices — “you don’t care as long as it orchestrates things correctly.”
• Who’s best positioned? “I believe it’s us” — because the orchestrator is positive-sum at every layer: “If Jensen produces a better chip, it’s great for us. If Dario produces a better model, it’s great for us.” Evidence: revenue “more than tripled since the beginning of the year” — partly thanks to Anthropic’s model progress, with burn down because OpenAI competed with Anthropic and brought down the cost of the same capability.
• On Google as the low-cost “token king”: all the advantages exist, “but they underestimated the importance of coding models, and so they’re far behind the frontier right now. Totally capable, totally competent team. But today they’re not quite at the frontier.”

7. Power is the bottleneck — Micron over Meta

• A data center isn’t chips from Dell — it’s land, turbines, grid deals, cooling, permits, all far slower than silicon. Today’s deployed models were trained on Hopper; the first Blackwell-generation model (“Mito,” as heard) is “already scary,” and Vera Rubin data centers will be used next year. That physical build-out time is why markets give infra companies higher P/E ratios than Meta, which builds infra but gets valued as software.
• Aravind’s call, highlighted at the show’s open: “It might not be inconceivable that Micron, the supplier of HBMs, might be more valuable than Meta in the next 6 to 12 months. It’s already at like a trillion. And Meta is like 1.3 to 1.4 trillion.” Why isn’t Micron fully priced after a 5x in memory costs? “Because it’s still the bottleneck. Whatever is the bottleneck will command the price.” Same for AMD: “Agents are using CPUs more than humans” — the tokens come off GPUs, but the downloading, transforming, and hosting all runs on enterprise CPUs.
• Three years out, power still rules: he estimates “40 out of 100 are not being developed because of public resistance” — anti-AI sentiment channeling through wealth-inequality anger, climate fears, grid prices, even RAM prices — despite the water claims being “untrue” (Satya’s “can of water”). Build-out migrates to resource-rich, regulation-friendly countries; “Elon’s going to space to do that.” Asked what he’d do with unlimited money: “I would build data centers. Physical infrastructure build-out is the return of the industrial age.”

8. The neocloud value stack — where margins live and die

• Aravind says Nebius/CoreWeave can be sustainable, but doesn’t know which will win — and only above bare-metal: “If you’re just a GPU server rack renter… there’s not a lot of value. It’s called Amazon Web Services, not Amazon servers.” Why does CoreWeave out-build OpenAI at data centers (the Stargate ambition)? Focus and operational intensity — permits, power, supply chain, TCO.
• $100B companies spanning inference, server capacity, and data-center build-outs are possible — $10B revenue at 30-40% gross margins — but only if open source stays within ~12 months of the frontier. Stretch the gap to 15-18 months and “I don’t think these companies really have a business model.” He endorses Emad Mostaque’s warning that lab consolidation is their biggest threat: “you don’t control your own destiny if you’re those companies.”
• OpenRouter as a $100B model-routing business? “Probably not.” The real product isn’t picking the cheaper model — it’s reliable token supply: pre-purchased rate limits across Bedrock/Azure/OpenAI endpoints, fallbacks when APIs error, and keeping your tokens out of China on open-source models. The margin is the spread between bulk discounts and list price — real value, “not a high gross margins business.”

9. Export controls helped — and forged a more potent China

• His DeepSeek-moment probability: “20%, 30% chance” that a vastly more efficient, differently-architected model strands US capacity. The mechanism is the controls themselves: DeepSeek is building on the Huawei stack, denied HBM and 3D NAND — so they shrank the KV cache enough to host on SSDs, innovated on the attention layer and interconnect-light training, and are vertically integrating down to fabs. “That’s a very different bet from what America’s making.”
• On whether controls helped or hurt: “Jury’s still out. Short-term it’s helping — he believes export controls are the only reason there is even a development gap between open source and frontier. Anthropic lobbied very hard for it.” But long-term: “by forcing them to go out there and build all this, you’re converting them into a far more potent competitor” — in China, “power is not a problem. Permits are not a problem. Labor is not a problem.”
• Do we still underestimate China? “I think so” — because AI is physical too: fabs, robots, chips, energy harnessing, local devices — “a lot more advantages than America.” The US response he tallies: TSMC’s $150B investment in American fabs ($40-60B already in), the government’s 10% of Intel with Nvidia and SoftBank at 5% each, Elon building a terrafab. His policy prescription: fund physical infrastructure, be fact-driven about data centers, “not fear-monger.”

10. Dario’s disservice, the entrepreneurship gospel, and the IPO wave

• Has Dario’s all-jobs-are-going messaging been a disservice? “Yeah, I think so” — and it’s inconsistent: “the most recent one I heard was there is no evidence that AI is taking over jobs.” The kill shot: “You can’t win by saying that and also complaining about not being able to build data centers fast.” His alternative gospel comes with receipts: Perplexity’s “billion dollar build” gives $1M in compute credits to any group with a credible path to a $1B company (echoing the ~$1M in AWS/GCP/Azure credits Perplexity itself started on); on Altman’s $2M YC token grants — “we should do more of that.” The proof-of-agency story he insists is real: an SF Uber driver who built new apps with AI off his YouTube interview and now makes more passive income than driving for Uber. Harry’s pushback stands un-smoothed: “I don’t think that many people have agency.”
• The head-count math: 400 people built a ~$20B company, so “with 40 people I could probably build a billion or two billion dollar company… we could be worth two trillion dollars with 10,000 people.” He’d rather the 100,000 people a typical $2T company employs split into a thousand few-billion-dollar companies. Internally he wants to “turn this company almost into an AGI” — semi-autonomous divisions with human scaffolding.
• The IPO wave (SpaceX, Anthropic, OpenAI) could get partly funded by reallocation: picture Vanguard/BlackRock moving “$30-40 billion” of their Microsoft/Salesforce holdings into Anthropic as an enterprise-AI hedge. Public SaaS must “weather the storm” by buying the next thing — IBM survives on Red Hat, HashiCorp, now “Confluence” (as said; likely Confluent). And the stakes explain the paranoia: Anthropic at “1 to 1.5 trillion” — Meta’s valuation, built in 6 years versus Meta’s 20 — means “anyone who’s winning today can lose tomorrow, including the model providers.”
• Perplexity’s own scoreboard, delivered with relish after being voted “most likely to fail” at an SF meetup (Cursor second, OpenAI third): revenue tripled since that judgment, burn cut by more than 50%, ARR “far, far more” than $500M, IPO hoped sooner than 2028. Quick-fire: his 10-year hold is SpaceX — “Anthropic and OpenAI can claim they do whatever each other does,” but SpaceX is the only company building space infrastructure for connectivity. And the operator he models: Jensen, despite the $5T company and most advanced chips, “operates with that mentality that he could be 30 days away from going out of business.”