The instructor's framing lecture for the quarter: AI has triggered a full-stack rewrite of how software is built, and the people who win will control verifiable context and stable compute.
Every layer of the tech stack, from capital and power down to chips, models, apps and governance, is being rebuilt at once, and that upheaval is the opportunity. The scaling recipe (raise money, buy compute, add data, train, deploy, feed the results back through reinforcement learning) now works so reliably that the real fights have moved to two inputs: who owns a context that can be reliably verified, and who can secure compute that is not fungible and keeps getting more expensive. Progress runs fastest in domains you can verify, like code and materials, and slowest in ones you cannot, like beauty and love.
The full stack
Midha lays out the stack every guest speaker maps onto: capital, then land/power/shell (energy and data centers), then chips, then cloud software, then models and agents, then applications, then governance and safety. For 15 years this stack was stable. AI has forced everyone at every layer to revisit basic assumptions about where they sit in the value chain, which he calls the great transition.
The scaling recipe
The path from research to product is now industrial, not bespoke. You raise money, buy compute, add data, pre-train a state-of-the-art model, deploy it for inference, and get two flywheels: inference revenue funds the next batch of compute, and usage generates context feedback that you pipe back through reinforcement learning. Base models train roughly twice a year on the order of 100,000 top-end GPUs; the reinforcement learning step now eats nearly as much compute as everything else combined.
Context is the moat
Since the recipe is repeatable, value accrues to whoever has unique, defensible, and verifiable context, meaning the environment an agent learns in. Ask where a task can be reliably measured: code passes unit tests, materials can be physically checked (Periodic Labs uses RL from physical verification to hunt superconductors). Midha's example of the fight: when OpenAI moved to buy the coding tool Windsurf, Anthropic cut off its model access within days, because a rival distilling your customer interactions is context leakage.
Sovereign context
Not all context is equal. Developer code is fine to send to a cloud server abroad; government records, defense, and national data are not. That is why Mistral, founded by the co-creators of Llama and the Chinchilla scaling paper, ships open-weight models that run locally under a customer's control. The U.S. Cloud Act, which lets the government reach data on U.S.-run servers anywhere, is why sovereign AI and infrastructure independence are now real markets breaking the cloud oligopoly.
Compute is not a commodity
The dogma that chips are a commodity is wrong right now. Anthropic's revenue tracks its compute buildout closely: about 60 to 90 days after new compute comes online, capability jumps, then revenue jumps. A dollar of hardware (land, power, chips, trading at 3 to 4x revenue) converts into software revenue worth 30 to 40x, roughly a 10x value transformation. Prices of the 2-year-old H100 are rising, not falling, and chips are not fungible even within one vendor's lineup.
History rhymes, so standardize
Steel, fiber optics, DRAM, shipping, uranium: every general-purpose technology runs a cycle of buildout, hoarding, panic, crash, then stabilization. AI is unusual because it marshals massive physical atoms (land, power, chips) to make a purely digital thing (intelligence). To move past the boom-bust, compute needs to become fungible, which takes agreed standards (like TCP/IP or AC/DC) and institutions to enforce them for public benefit. We are in the pre-standardization era, and the students' assignment is to help design that peaceful transition.
- The AI stack has clear layers: capital, land/power, chips, cloud, models, apps, governance, and every layer is being renegotiated at the same time.
- The scaling recipe is now repeatable, so competitive advantage shifts from the recipe to owning context that can be reliably verified.
- Reinforcement learning started working at scale about two years ago because LLM priors are general enough to keep learning instead of plateauing.
- RL progress is fastest in verifiable domains (code, materials) and weak where verification is hard (creative writing, aesthetics, love).
- Compute is not fungible today: H100 prices are rising, chips differ across and within vendors, and demand is spiky and hard to forecast.
- Anthropic's revenue lags its compute buildout by about 60 to 90 days, showing capability scales predictably with compute.
- Turning compute into a real commodity needs standards plus institutions to enforce them and reallocate hoarded capacity toward public benefit.
In their words
“We are now in the full-stack rewrite, and I need you guys to start thinking up and down the stack.”
“Anybody who told you chips are a commodity should probably get a phone call from you.”
“Progress is fastest in easily verifiable domains.”
Terms to know
- Context
- The environment and feedback an agent learns from; whoever owns verifiable context can keep improving a model in that domain.
- Context feedback loop
- Piping real usage results (like whether a coding agent solved a task) back through RL to keep raising capability.
- Reinforcement learning (RL)
- Training by rewarding a system for completing a task rather than telling it how, now the compute-heaviest stage of model training.
- Sovereign AI
- AI infrastructure a government or org runs on its own hardware so sensitive data never leaves its control.
- Fungible
- Interchangeable unit for unit; megawatts are fungible, but GPUs today are not, which is why compute behaves like a scarce commodity.
- Recursive self-improvement
- When the compute and context flywheels get good enough to propel themselves; Midha frames it at the system level, not one super-model.
Anjney Midha at Stanford CS 153: Frontier Systems
New to this? Come build with us.
Reading is good. Building with people is better. Our drop-ins are free and open to total beginners.