Stanford CS153: Frontier Systems · Lecture recording

AI Scaling, Bottlenecks, and Why Compute Isn’t a Commodity Yet

Speaker: Anjney Midha — Partner @ AMP PBC (Co-instructor) Date: Spring 2026 · Duration: ~55m

TL;DR

Anjney Midha sets the stage for CS153 by framing the “great transition” happening across the entire AI infrastructure stack — from capital and chips to models and policy. The lecture centers on two big ideas: context (the environments and feedback loops that determine which AI systems keep improving) and compute (why GPU prices are rising, not falling, and what historical infrastructure cycles tell us about what comes next). The core thesis: we are in a pre-standardization era for compute, analogous to early electricity or steel, and the students in this room will help shape the standards and institutions that resolve it.

Key Concepts

Anj’s Life Scaling Laws

Opening personal advice from Anjney — take life seriously but don’t forget to have fun. The most important people in this class are the students around you, not the speakers. Invest in relationships. “Maybe go to the real Coachella while you still can!”

The AI Infrastructure Stack

The industry is organized in layers, and every layer is being disrupted simultaneously — something Anjney calls “the great transition”:

  1. Capital — flexible, goes everywhere
  2. Land, Power, Shell — energy production, data center construction
  3. Chips — hardware (NVIDIA, AMD, custom silicon)
  4. Cloud — software that makes chips usable
  5. Models / Agents — training and post-training
  6. Applications / Solutions — products built on top
  7. Governance — safety, security, trust frameworks

For the first time in Anjney’s career, assumptions at every layer are being revisited simultaneously. This creates extraordinary opportunity for people who understand the full stack.

The Pre-Training → Post-Training Shift

The production pipeline has industrialized:

  • Base model training: ~2x/year on 100K+ B300-equivalent GPUs
  • Mid-training (adding capabilities): 2-4x/year, ~10% of training compute
  • Continuous post-training (SFT + RL): ongoing
  • Critical shift: RL post-training now consumes almost as much compute as the entire rest of the pipeline combined

Bottlenecks on AI Capabilities Progress

Four key bottlenecks:

  1. Context — data and deployment strategy
  2. Compute — infrastructure and efficiency
  3. Capital — of various kinds to secure all the above
  4. Culture — perhaps the most important of all

Reinforcement Learning — Why It’s Working Now

  • RL itself is old (70+ years). What’s new: initializing RL with an LLM that has strong enough priors about the real world.
  • Result: systems learn much faster than before, and capabilities keep scaling with more compute + more context.
  • Historical RL (chess, Go) would surpass humans but then plateau. LLMs broke through that ceiling because the priors are general enough to keep learning.

The Intelligence Manufacturing Recipe

A simple, repeatable loop that has been empirically validated over 4 years:

  1. Raise money → buy compute
  2. Augment with data → pre-train a model
  3. Ship a product (state-of-the-art enough that people want to use it)
  4. Deploy → inference generates revenue + context feedback
  5. Pipe context back through RL → improve capabilities
  6. Repeat — the two flywheels (revenue + context) reinforce each other

Key data point: Anthropic went from $9B to $20B in revenue. Each time they brought up new compute, capabilities jumped 60-90 days later, followed by a revenue jump. The correlation is strong and predictable.

Context as the New Moat — “Context Feedback Loop Wars”

The most important concept of the lecture. Anjney argues that the question “who wins in AI?” comes down to who controls unique, defensible context:

  • Context = the environment in which an agent learns. Like a park where you train a dog — every factor (the grass, the rain, the kids running around) influences learning.
  • Verifiable contexts are where RL progress is fastest. Code is verifiable (unit tests pass or not). Material science is verifiable (a superconductor either works or doesn’t). Aesthetics, beauty, love — not verifiable, hence models are terrible at long-form creative writing.
  • Winners = teams with unique, defensible access to context
  • Losers = teams locked out of essential contexts

Real-world example: When OpenAI tried to acquire Windsurf (IDE), Anthropic immediately cut model access to Windsurf users. Context leakage — if your competitor can observe how your model helps customers, they can distill from that. This was the first visible “context war” shot.

The sovereign context angle: Mistral was founded on the insight that governments and mission-critical workloads (defense, national records) cannot run on US cloud infrastructure due to the Cloud Act (US gov can access data on US company servers globally). This is driving “sovereign AI” — local compute, local models, open-source weights. It’s the first time in 15 years that the relentless consolidation of cloud infrastructure is reversing.

The Big Question: What Are the Limits of RL?

  • Philosophical: Agents should be able to learn anything
  • Empirical: Life is messy. Progress is fastest in easily verifiable domains.
  • Open question: Does RL generalize across domains? Anjney’s view leans empirical — progress is fastest in easily verifiable narrow domains. It’s not clear that a coding agent can bootstrap itself into material science.

Recursive Self-Improvement (The RL Flywheel)

Anjney thinks about recursive self-improvement at the systems level (a team that keeps getting better), not just the model level. SOTA Mission + Research Compute → Leading Technology & Product → Frontier Flywheel.

Compute Is Not a Commodity

Key data point: Anthropic’s revenue growth is “beautifully correlated” with compute buildout — $30M → $6.2B as power capacity scaled from near-zero to 400+ MW.

The hoarding cycle: Big tech is spending more on infra in 2024-2026 than in the preceding 30 years combined. Combined hyperscaler CapEx: $23B (2015) → $354B (2025) → $605B (2026E).

The infrastructure trade: $1 of compute (hard assets trading at 3-4x revenue) → $1 of software revenue (trading at 30-40x revenue). A predictable ~10x value transformation. This is what’s driving the spending frenzy.

H100 prices are rising, not falling. Average was $1.73/hr 2 years ago. Today: significantly higher and climbing. On-demand is sold out.

Historical Infrastructure Cycles

“What happens next? Let’s look at history of infrastructure for clues.”

Anjney draws parallels to previous general-purpose technology cycles:

ResourceBoom PeriodPattern
Steel1867–1895Hoarding → Panic of 1873 → crash → standardization
Fiber opticsLate 1990sOverbuild → dot-com bust → eventual stabilization
DRAMMultiple cyclesViolent semiconductor cycles, repeated booms and busts
Shipping (Baltic Dry Index)2003–201294% crash in six months after peak
Uranium1970sNuclear boom → Three Mile Island → 65% collapse

Average cycle duration: ~2.8 years (digital), ~6.3 years (physical). AI is both digital and physical, which is confusing markets.

Is Compute Just Another Commodity?

This is the most contrarian and data-heavy section of the lecture:

  • Compute is not fungible: H100 ≠ GB200 ≠ B300. Even chips from the same manufacturer are not interchangeable. Current GPU market pricing (March 2026) shows B300 at $3.90-$4.40/hr vs AMD MI325 at $1.55-$1.70/hr.
  • Forecasting is broken: Training is spiky (small experiments → hero runs). Inference is cyclical (heavy during day, dead at night). No stable consumption pattern exists.

The punchline: We are in the pre-standardization era of compute. What’s needed:

  1. Standards (like AC/DC for electricity, TCP/IP for networking) — common units, delivery interfaces, interconnection, metering
  2. Institutions to enforce those standards — reallocating power from hoarders to public benefit

Industry Insights

  • Anthropic’s origin story (from Anjney’s perspective): Dario and Tom called him while running research at OpenAI. They pitched “leave and start a new lab.” Anjney made 22 investor introductions on Sand Hill Road. 21 said no. “Sounds theoretically cool, but do you have empirical proof?” Four years later: $20B revenue.
  • Amp’s internal rule: After Anjney sent a blog post to a founder friend who immediately replied “Did you use Claude for this?” — Amp now has a rule: no AI-generated documents sent internally. “We sit. We write. And we share even if it’s raw.”
  • GPU as drug dealing: A founder who raised ~$700M-1B messaged that morning saying “We’re in a compute crunch. Need H100s ASAP. Price not a problem.” Anjney’s comment: “It’s a good time to be a drug dealer.”
  • Recursive self-improvement: Anjney thinks about this at the systems level (a team that keeps getting better), not just the model level. “You can have a company hitting takeoff because they’ve figured out how to keep recursively improving themselves.”
  • Claude Code usage: Commits by the Claude Code agent on GitHub are “beautifully correlated with the compute buildout” — real usage growth, not just revenue pumping.

Memorable Moments

  • “AI Coachella”: A viral tweet warned students to “be wary of taking classes that sound like AI Coachella.” Anjney leaned into it: “I think AI Coachella is pretty fun.”
  • Getting emotional: While telling students to invest in their relationships, Anjney teared up unexpectedly. “I don’t know why I’m getting so emotional. I need some water.” He met his wife Viv as a sophomore at Stanford; they’ve been together 13 years. Both his companies were co-founded with former Stanford roommates.
  • Life scaling laws: “I’ve found a pretty simple heuristic for navigating life: just have fun. With people you enjoy hanging out with. That’s pretty much it.”
  • 3Blue1Brown connection: Grant Sanderson (3Blue1Brown creator) was Anjney’s undergrad drawmate at Stanford. They spent the night before this lecture talking about what it would take to create world-class educational content for any domain using AI.
  • RTX 5090 as course prize: Last year Jensen Huang signed five RTX 5090s as prizes for the best project. Anjney: “I’m not going to ask Jensen this time… they’re a little bit more valuable this time around.”

Notable Quotes

“Take life seriously but not so seriously that you forget what’s important. Don’t forget how to have fun and remember what makes life worth living.”

“Where is there context that can be reliably measured and verified when you’re working with an agent? That’s the question I would be asking if I was you.”

“Dollar of compute in — hard assets trading at 3-4x revenue — being turned into a dollar of software revenue, which trades at 30-40x revenue. We have developed a predictable way to transform one input into another that humanity considers 10 times more valuable.”

“Anybody who told you chips are a commodity should probably get a phone call from you and ask them what they think about this. Because chip prices are not going down. They’re going up.”

“Think of yourselves not just as students, but as active participants. You are extraordinarily lucky to be alive at this moment in time.”

Reading Materials

Official Readings (from Discord #general)

  1. Chinchilla: Training Compute-Optimal Large Language Models (Hoffmann et al., 2022)

    • arxiv.org/abs/2203.15556
    • Key paper by Arthur Mensch (Mistral co-founder, upcoming guest speaker) on optimal scaling of training compute vs. model size
    • Directly referenced in lecture — Anjney assigns this before Arthur’s talk

  2. CLOUD Act — Clarifying Lawful Overseas Use of Data Act

    • AWS explainer
    • US policy allowing government access to data on US company servers globally
    • Central to the “sovereign AI” argument in this lecture — why Mistral and sovereign compute exist

Unofficial / Supplementary

  1. AI Compute Demand — Stanford MSE 435

  2. Economics of the AI Supercycle — Stanford MSE 435 (full course)

    • mse435.stanford.edu
    • Sister course covering the economics angle of the infrastructure themes in CS153

Connections

  • Part of CS153: Frontier Systems — Stanford, Spring 2026
  • References Lecture 0 (not recorded?) where scaling laws and the Anthropic revenue flywheel were first discussed
  • Upcoming speakers mentioned: Jensen Huang (NVIDIA), Lisa Su (AMD), Sam Altman (OpenAI), Satya Nadella (Microsoft), Liam Fedus (co-creator of ChatGPT, Periodic Labs), Arthur Mensch (Mistral — read the Chinchilla paper before his talk), Andreas Blattmann (Black Forest Labs / Stable Diffusion)
  • Assigned reading: Chinchilla scaling laws paper (before Arthur Mensch’s lecture)
  • The Bitter Lesson (Rich Sutton) — referenced from Lecture 0
  • Course project: “The One-Person Frontier Lab” — build something that creates real-world value using frontier AI tools over 10 weeks