Stanford CS153 Frontier Systems: Anjney Midha on Frontier Systems

ELI5/TLDR

Anjney Midha co-founds AI labs for a living (Anthropic, Mistral, Periodic Labs were all in his orbit). In this Stanford lecture he argues two things. First: the AI recipe — buy compute, train a model, deploy it, feed what you learn back in — is now well understood, so the winners won’t be whoever has the recipe but whoever owns a slice of context nobody else can touch. Second: everyone says GPU chips are a commodity that gets cheaper over time, and right now they’re dead wrong — prices for two-year-old chips are rising, because compute behaves less like a settled commodity and more like steel or fiber-optic cable did in the messy early years before anyone agreed on standards.

The Full Story

The recipe stopped being the secret

A few years ago, turning AI research into a business was bespoke and rare — a new model showed up maybe once or twice a year. Now it’s an assembly line. Midha sketches the cadence: a base model trained twice a year on the equivalent of 100,000 top-end chips, then “post-training” (extra skills bolted on after) with about 10% of that compute, then continuous tuning that never really stops.

The recipe itself is almost embarrassingly simple. Raise money, buy compute, add data, train a model good enough that people want to pay for it, deploy it, and run the model — what’s called inference, the actual using of the model to answer questions. That gives you two loops spinning at once. One loop is money: the cash from people using the model buys your next batch of chips. The other is information: when the model does a coding task, you can see whether it got it right — the tests pass or they don’t — and you pipe that signal back in to make it better. Midha calls this second loop the context feedback loop.

When he pitched this to investors four years ago, he made 22 introductions and got 21 no’s. The objection was always the same: sounds nice in theory, where’s the proof? The proof arrived. Anthropic went from 9 to 20 billion dollars in revenue; the same pattern repeated at the other labs.

“The question I keep getting over and over again is, ‘Okay, well, who wins?’… If everybody’s going to be applying this scaling recipe around and it’s so easy and repeatable, then where does the value accrue?”

Why reinforcement learning suddenly works

Underneath all this sits a technique called reinforcement learning — RL. The idea is old and simple. Think of training a dog. You don’t explain how to fetch; you just reward the dog when it brings the ball back and withhold the treat when it doesn’t, over and over. RL does the same to a model: define the task, reward success, repeat.

For seventy years this hit a wall. A program could learn chess or Go, beat humans, and then plateau — it couldn’t keep getting better, and it couldn’t carry what it learned into anything new. What changed around two years ago: if you start the process with a large language model that already has decent common sense about the world baked in, it keeps learning instead of stalling. Smart starting assumptions — priors, in the jargon — turned out to be the missing ingredient.

But there’s a catch Midha is honest about. RL works beautifully where you can check the answer. Code is checkable — write a test, it passes or fails. So is material science, oddly: a new superconductor either superconducts or it doesn’t (he name-drops Periodic Labs running RL against physical lab experiments to hunt for new materials). The trouble is everything you can’t cleanly verify — beauty, taste, love, good writing.

“These models are not good at long form writing, at creative writing. They hallucinate, they make these clichéd hyphens… ‘It’s not just X, it’s Y.’”

He tells a small story: he sent a friend a blog post he’d outlined himself and asked Claude to flesh out. The friend replied in 30 seconds — “Did you use Claude for this?” — and didn’t believe the denial. His company now has a rule against sending each other AI-written documents. The lesson for students: the unverifiable domains are wide open precisely because the machines can’t grade themselves there, and your taste might be the only available judge.

Context is the real moat

If the recipe is commodity, the defensible thing is owning context — the environment an agent learns from. Train a dog to fetch in a park, and the park is the context: the grass, the kids running around, the rain. For an AI agent, the context is the codebase, the file history, the place where success or failure gets measured.

So Midha’s framework for where to spend a career: find a domain where (a) results can be reliably verified, and (b) you have unique, defensible access to the context. He illustrates with a concrete skirmish. When OpenAI moved to acquire the coding tool Windsurf, Anthropic cut off Windsurf’s access to its models within days — almost unheard of in the industry. The logic: if a competitor uses your model inside their product, they can watch how it helps customers and copy that. Context leakage. The comfortable old assumption — that an app company can always count on a model company to keep supplying intelligence — quietly stopped being true.

The same fight scales up to nations. This is the sovereign AI story, and the reason Mistral exists. Software engineering context, the argument goes, isn’t that sensitive — a developer doesn’t mind piping it to a cloud server somewhere. But government records, defense data, national systems? Under the US Cloud Act, data sitting on US-company servers anywhere in the world is reachable by the US government. For a lot of countries, that’s a dealbreaker — which is why a head of state and the world’s richest chipmaker will stand on a Paris stage next to a 33-year-old scientist and call open-weight local models the future of Europe.

The part where compute refuses to be a commodity

Then Midha switches to the half he loves most, and the more striking claim. For fifteen years, cloud computing got relentlessly cheaper — Amazon and Google piled up servers for their own needs, found they could rent the spare capacity out, and economies of scale did the rest. The whole industry assumes chips work the same way: a commodity, always depreciating.

He shows a chart of the H100 — a chip over two years old — and asks the room what the price is doing over the last 90 days. Going up. He recounts a founder who’d raised hundreds of millions messaging that morning in a “compute crunch,” needing chips immediately, price no object.

“It’s a good time to be a drug dealer.”

Why won’t compute behave? Because it breaks the two rules that make something a commodity. First, it’s not fungible — a megawatt is a megawatt anywhere, but an H100 is not a GB200 is not a B300, and a chip from AMD isn’t interchangeable with one from Nvidia. Second, it’s brutally hard to forecast. Training is spiky — you tinker small, then suddenly fire up a huge run. Inference is cyclical — everyone uses the chatbot during the day, nobody at night. Electricity we’ve forecast smoothly for 75 years; compute we can’t, so the big players just hoard land, power, and chips on the bet that some model breakthrough will pay off.

History rhymes, and it’s a warning

Midha’s instinct is to look backward. He runs through steel (1867–1895), fiber optics, DRAM memory, shipping, uranium — each a new general-purpose technology, each with the same arc. Prices climb, a few players hoard, then some small event triggers a panic, a self-fulfilling sell-off, companies collapse, and only afterward does society agree: this thing is genuinely valuable, let’s stabilize it. The stabilizing came from standards (AC/DC for electricity, TCP/IP for the internet) and institutions to enforce them — because at scale, humans don’t cooperate on their own.

What’s genuinely new about AI, he argues, is that it fuses two worlds that hate colliding. The output is pure bits — software revenue. The input is atoms — land, power, physical chips. Software revenue trades at 30–40 times earnings; the hard assets that produce it trade at 3–4 times. You’re turning a dollar of physical infrastructure into roughly ten dollars of value the market prizes. Coordinating spiky digital demand against slow physical supply is the puzzle nobody has solved.

His conclusion: we’re in the pre-standardization era of compute — the chaotic stretch before anyone agrees on a common unit, a delivery interface, metering, settlement. The students’ assignment for the quarter: what would it take to reach a peaceful transition on compute without the boom-and-bust, and what’s your part in it? His pitch is that they’re not just students but participants who can write, blog, and push the people building these institutions toward better standards.

Key Takeaways

The AI business recipe (compute → data → train → deploy → feed results back) is now well understood and repeatable; the edge has moved from having the recipe to owning context others can’t access.
Context feedback loop: deploying a model lets you observe success/failure on real tasks and pipe that signal back through RL. Whoever controls a verifiable, defensible context captures the value.
Reinforcement learning started working at scale ~2 years ago because LLMs supply strong enough priors to keep learning instead of plateauing — the old failure mode of chess/Go-era RL.
RL progress is fastest where answers are cheaply verifiable (code, material science) and weakest where they aren’t (writing, aesthetics, taste). Unverifiable domains are wide open for humans.
Open question Midha flags as unresolved: RL does not clearly generalize across domains — a coding model isn’t bootstrapping itself into biology. Billions are being bet on whether it will.
Anthropic cutting Windsurf’s model access after OpenAI’s acquisition attempt was a deliberate move to stop “context leakage” — a competitor distilling insight by watching the model help customers.
Sovereign AI: the US Cloud Act (US-company servers anywhere are reachable by the US government) is pushing governments toward locally-run open-weight models — the reason Mistral exists.
Contrary to industry consensus, GPU prices are rising, not falling — even for the two-year-old H100. Compute is behaving like scarce, hoarded infrastructure, not a depreciating commodity.
Compute isn’t a commodity because it fails two tests: it’s not fungible (H100 ≠ GB200 ≠ B300; AMD ≠ Nvidia) and it’s hard to forecast (spiky training, day/night-cyclical inference).
Anthropic’s revenue tracked compute with a ~60–90 day lag: new compute brought online, then a capability jump, then a revenue jump.
AI uniquely fuses atoms (land, power, chips) with bits (software revenue) — hard assets trade at 3–4x revenue, software at 30–40x, so the transformation is roughly 10x value creation.
Historical pattern across steel, fiber, DRAM, shipping, uranium: price run-up → hoarding → panic → sell-off → eventual stabilization via standards + enforcing institutions. We’re at the pre-standardization stage for compute.

Claude’s Take

This is a lecture, not an argument that has to survive peer review, and it’s worth reading as the worldview of someone with skin in every layer of the game — Midha says so himself, flashing a “disclosure list” of labs he’s co-founded or backed. The biases run in predictable directions. “Sovereign AI” and “compute scarcity” are both theses that happen to be good for a man whose portfolio includes Mistral and whose firm runs an internal GPU-price tracker. That doesn’t make him wrong, but the framing is downstream of the positions.

The strong part is the compute argument. The claim that chips aren’t a commodity — not fungible, not forecastable — is concrete, falsifiable, and goes against lazy consensus, which is the most useful kind of claim. The H100-prices-rising observation is the sort of thing that’s either true on the spot or not, and it reframes the whole “AI bubble” conversation usefully: maybe it’s a standardization problem, not a valuation problem.

The weaker part is the historical-analogy carpet-bombing — steel, fiber, DRAM, uranium, shipping all marched out to say “infrastructure is cyclical.” True, but it proves almost anything; every boom looks like a prior boom if you squint, and he even admits “this time may be so different.” The honest moment is when he concedes RL probably isn’t generalizing across domains — that single admission undercuts a lot of the more breathless AGI talk elsewhere and is the most valuable sentence in the hour.

A 7: genuinely good mental models (context as moat, compute non-fungibility, the bits-from-atoms tension) delivered in a loose, motivational, name-dropping package with a fair bit of filler — Coachella, dad jokes, dating anecdotes — that pads the runtime. Take the frameworks, discount the salesmanship.