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

ELI5 / TLDR

Aravind Srinivas runs Perplexity, the “answer engine” that spooked Google into copying its homepage. His big argument: the money in AI is not in owning the smartest model — models get cheap and swappable — it’s in being the layer that wires everything together (the orchestrator). Along the way he makes two spicy claims: that Micron, which makes the memory chips AI needs, could be worth more than Meta within a year, because whatever is the bottleneck commands the price; and that the real ceiling on AI isn’t chips or money, it’s electricity — and the public is increasingly fighting the data centres that supply it.

The Full Story

The pitch: stop trying to own the model

The central idea Srinivas keeps circling back to is that everyone is fighting the wrong war. The instinct is to build the smartest model and win. He thinks that’s a trap.

The output tokens if you’re literally just a reseller of model tokens you have no business because the model will get commoditized.

His mental model goes like this. A raw model is just intelligence sitting in a box. To turn that into something a customer pays for, you need a harness — think of it as the rules and plumbing around the model: which tools it can reach, which files, which sub-tasks it breaks the job into, when it hands off to a different model. The harness is what converts raw smarts into useful work. Perplexity’s bet is to be the conductor that orchestrates across all of it — multiple models, your files, your tools, even the chip in your laptop.

That last point is his differentiator. OpenAI’s coding tool won’t quietly run Anthropic’s model inside it; Anthropic’s won’t run OpenAI’s. They can’t, because they’re at war. Perplexity, owning no flagship model of its own, can route to whichever is best for the task — and pocket the gains every time any layer of the stack improves.

If Jensen produces a better chip, it’s great for us. If Dario produces a better model, it’s great for us. If Apple produces a better device, it’s great for us.

The metric he says matters most is a mouthful — “token value per watt per user.” Strip the jargon: how much useful work can you squeeze out per unit of electricity per customer. Because in the end, he argues, the cost of AI bottoms out at the cost of power, and power is the one thing nobody can fake.

Why advertising won’t save the chatbots

Stebbings pushes on whether OpenAI builds a giant ad business like Google’s. Srinivas is openly bearish, and his reasoning is clean. A chat box is for objective questions — what’s the best protein powder, factually. But most ad money rides on subjective decisions: which hotel feels nice, which jacket looks good. You don’t want a single confident answer there; you want to browse options. That’s why hotel bookings still happen on Google and fashion ads live on Instagram.

When the decision making is more subjective and vibes based, you don’t need an objective answer engine.

Worse, slipping ads into a tool people trust for accurate answers poisons the well. His split: objective transactions get eaten by agents; subjective ones stay ad-based. So the chat interface, he thinks, was never a natural home for advertising in the first place.

Where the money actually is

If not ads, then what? Power users running agents continuously. He describes engineers spending millions a year on coding tools, and one Perplexity customer burning $10,000 a month — not wastefully, but running their whole business through automated agent loops.

The dividing line he draws is between one-off tasks and cron jobs — recurring automated jobs that just keep running. (Imagine the difference between asking an assistant a question once, versus having one that watches your inbox all day and acts every time something lands.) Every inbound email triaged automatically, every latency spike traced to the engineer who caused it. That’s the frontier — and Perplexity is aiming it at non-developers: finance teams, sales reps, research analysts. He sizes that as “claude code multiplied by 10.”

The Micron-over-Meta claim

Here’s the headline bet. He frames it as a simple rule: whatever is the bottleneck commands the price.

Micron, the supplier of HBMs, might be more valuable than Meta in the next 6 to 12 months.

HBM is high-bandwidth memory — the fast memory stacked next to AI chips, and currently in desperate short supply (he notes memory prices have spiked sharply). The same logic, he says, explains AMD’s run: agent loops run on ordinary CPUs, and agents now use CPUs far more than humans do, so CPUs became a bottleneck again. Memory, storage, CPU compute — the picks-and-shovels of the AI build-out — get priced above the companies merely renting out data centres, because the suppliers are the choke point.

It’s a coherent story. Whether Micron specifically overtakes Meta in twelve months is a market call, not a logical certainty — and that gap matters (more in Claude’s Take).

Power is the real wall

Ask him the bottleneck three years out and the answer doesn’t change: power.

I actually believe that there will be a lot of resistance to building data centers.

His read is that the public backlash isn’t really about water or grid strain (he calls those claims overblown) — it’s a lightning rod for everything people fear about AI: job losses, wealth inequality, rising electricity and RAM prices. He cites a figure that roughly 40 out of 100 planned data centres aren’t getting built because of local resistance. The escape hatch: build them in countries with friendly regulation, abundant sun and natural resources — or, in Elon’s case, in space.

China, and the DeepSeek risk

The most genuinely technical stretch is on China. US export controls block China from buying Nvidia chips and the high-bandwidth memory — so DeepSeek is forced to build on Huawei’s stack and engineer around the memory it can’t get. They’ve shrunk the “KV cache” (the running memory an AI keeps mid-conversation) small enough to live on cheap storage instead of expensive HBM, and reworked the model’s attention and training to sip less interconnect bandwidth.

The risk for America: by forcing China to vertically integrate the whole physical stack — chips, fabs, power — the controls might forge a more dangerous competitor, one that can build data centres without fighting over permits or power at all. He puts a “DeepSeek moment” — a vastly more efficient architecture that runs on local devices and strands all the over-built capacity — at 20-30%.

If AI is not just digital, that’s also physical AI… I think they have a lot more advantages than America.

Everyone is two years older and nobody is safe

A recurring drumbeat: no one relaxes. He’s blunt that Anthropic’s coding tool, OpenAI’s lead, even Perplexity itself, are all temporary. He invokes Jensen Huang waking up daily convinced Nvidia is 30 days from going under, despite a $5 trillion valuation. Perplexity, he notes, was voted “most likely to fail” at a San Francisco meetup; it has since tripled revenue and cut its burn more than half. His tell on the doubters:

Most of those people who sit on these meetups and vote don’t actually build anything useful.

The optimist’s coda

On the social anxiety — won’t AI gut jobs and widen inequality? — he flips it. Instead of ten-thousand-person companies, expect thousands of tiny ones: a few friends, a million dollars of compute credits, a real shot at a billion-dollar outcome. He cites an Uber driver who built an app from one of his videos and now earns more passively than from driving. His critique of Dario Amodei’s “all the jobs are going” messaging is sharp: you can’t fear-monger about job losses and complain that you can’t build data centres fast enough. Pick a story.

Key Takeaways

The orchestrator beats the model. A model is commoditised intelligence; value lives in the harness that wires models, tools, files, and chips into finished work. Perplexity’s edge is being model-agnostic — it routes across OpenAI and Anthropic, which they can’t do to each other.
“Token value per watt per user” is his single most important metric: useful output per unit of electricity per customer. Cost of AI ultimately floors at cost of power.
Whatever is the bottleneck commands the price. Hence Micron (HBM memory) possibly over Meta within 6-12 months, and AMD/Intel benefiting as agent CPU usage explodes.
Ads won’t migrate to chat. Objective queries suit answer engines; subjective, vibes-based purchases (hotels, fashion) stay on Google and Instagram. Ads in a trusted answer tool also corrode trust.
Power, not chips or capital, is the binding constraint — and public resistance is blocking ~40% of planned US data centres. Build-out shifts abroad or to space.
Export controls are a double-edged sword: they buy America a ~12-month model lead but may forge a vertically integrated Chinese rival that can build physical infrastructure unimpeded. ~20-30% odds of a disruptive “DeepSeek moment.”
Costs go down, but you always pay for the frontier. Open-source models will match today’s best; spend just shifts to the next frontier task (autonomous software engineers, AI designing chips/drugs).
Cron jobs separate power users from dabblers — continuous, event-triggered agents, not one-off prompts.
CoreWeave/Nebius-style “neoclouds” can be durable only if they add software margin on top (the AWS, not “Amazon servers,” lesson); pure GPU-rack rental has thin value. Their existential risk is model-provider consolidation.
Perplexity: ~400 people, ~$20B valuation, revenue tripled and burn halved this year; IPO floated for 2028, possibly sooner.

Claude’s Take

This is a genuinely sharp founder talking his book, and the trick is separating the two. The orchestration thesis is the strongest part — and conveniently, it’s the layer a company with no flagship model must occupy to matter. That doesn’t make it wrong; “we win when any layer improves” is a real structural advantage and the AWS analogy is apt. But it’s also the most self-serving framing available to him, so weight it accordingly.

The “Micron > Meta” line is doing exactly what the title wants: inviting scrutiny. The logic — bottlenecks command the price — is sound and well-illustrated. The prediction is a leveraged bet on memory staying scarce. Memory is cyclical; the entire history of semiconductors is supply catching up and prices collapsing. He half-admits the symmetric risk himself with the DeepSeek scenario: a more memory-efficient architecture would knock the legs out from under exactly the HBM scarcity his Micron call depends on. So the bottleneck framework is durable; the specific twelve-month price target is a coin-flip dressed as a thesis. Treat it as a vivid teaching example, not a forecast.

Best non-obvious insight: the subjective-vs-objective split for what agents disrupt. It’s a clean, portable mental model that explains more than the ad question — it’s a decent lens for guessing which businesses survive an agent-saturated internet.

Where to keep the salt handy: the relentless optimism on entrepreneurship (“anyone can build a billion-dollar company in 12 months”). Stebbings, to his credit, pushes back that most people don’t have that agency, and Srinivas doesn’t really answer it — he just insists you have to help them. The Uber-driver anecdote is one data point standing in for a structural claim. And the China section, while the most informative, is also where you should verify rather than absorb — the technical details (KV cache on SSDs, attention innovations) are plausible and roughly track public reporting, but they’re delivered fast and confidently in a domain where confidence is cheap.

Score: 7. Dense, opinionated, low on filler, and several of the frameworks (orchestration, bottleneck pricing, objective/subjective) are worth keeping. Docked for the headline being a stretch he can’t fully defend and for the long stretches of motivational founder gospel near the end that add energy but little signal.