Stanford MS&E435 Economics of the AI Supercycle — Building AI Factories — Annotated Lecture Notes & Key Takeaways

Course Information

Institution: Stanford

Original Course: Stanford MS&E435 Economics of the AI Supercycle | Spring 2026 | Building AI Factories

Instructor Bio: This session is led by **Apoorv Agrawal**, Adjunct Lecturer in Management Science and Engineering at Stanford University and Partner at Altimeter Capital, with guest speaker **Chase Lochmiller** (CEO, Crusoe Energy). Chase Lochmiller is a pioneer in AI compute infrastructure and energy optimization, leading Crusoe Energy’s mission to deploy low-cost, sustainable AI data centers by tapping stranded energy assets. He has deep expertise in the capital and operational economics of large-scale AI compute facilities.

Course Description: This lecture explores the economics of building and operating AI compute factories — the massive data center clusters that power generative AI models. It covers capital expenditure requirements, ongoing operational costs, energy efficiency tradeoffs, and site selection logic for AI infrastructure. It also evaluates constraints on compute supply expansion, the impact of energy costs on AI growth, and emerging business model innovations in the compute infrastructure space.

Setting Up the Problem

 

This talk, given as part of Stanford’s MS&E435: Economics of the AI Supercycle, zeroes in on a deceptively simple question: what does it actually take to build the places where AI gets made? The title — Building AI Factories — frames a data center not as a neutral piece of real estate but as a production facility for intelligence. That metaphor runs through the entire conversation.
 

The session opens with a striking chart showing capital expenditures by the five major hyperscalers on AI, climbing steeply upward. The host puts this in historical perspective, comparing it to massive undertakings like the space program, the interstate highway system, and the Manhattan Project. The comparison is meant to signal scale, not equivalence. We are told this wave of investment is among the largest ever made — second only, by this framing, to the U.S. defense budget.
 

Into that context steps Chase Lochmiller, founder and CEO of Crusoe. Before the technical discussion, we get a personal detail: Lochmiller is an avid mountaineer who has climbed five of the seven continental summits, including Everest. The host uses this to suggest that Crusoe’s planning culture resembles expedition planning — multiple contingency plans, layered on top of one another. It is a memorable way to introduce a CEO, even if it risks overstating how literally mountaineering maps onto data center construction.
 

The AI Factory as Physical Infrastructure

 

Lochmiller starts by grounding the conversation in something tangible. A data center, he says, is the physical manifestation of the AI boom. It is the infrastructure required to power, operate, and cool GPUs so they can run the enormous compute workloads behind training, fine-tuning, and large-scale inference.
 

This is an important reframing. When we talk about AI, the conversation usually orbits models, benchmarks, or product features. Here, the focus shifts to the basement: the buildings, electrical systems, and cooling apparatus that make those models possible. Every time a student in the room admits to using a model like Gemini, ChatGPT, or Claude, the underlying implication is that a data center somewhere is serving that request in real time.
 

The central claim is that AI cannot scale without a corresponding scaling of physical capacity. Intelligence, in this account, is not just an algorithmic achievement. It is an industrial one. That is why the talk calls these facilities “AI factories” rather than merely server farms.
 

A Useful First Equation

 

To make the economics manageable, Lochmiller breaks AI production into four components:

1. Data​ — including acquisition and the growing industry of data labeling.
2. Algorithms​ — spanning backpropagation, neural networks, transformer architectures, and newer recursive learning techniques.
3. Compute​ — especially high-performance GPU infrastructure capable of parallelizing workloads and performing tensor math efficiently.
4. Energy​ — the power required to run those GPUs.
5. Data centers​ — the physical buildings that house and operate all of the above.

 

He describes this as a kind of “basic equation,” which is a helpful pedagogical device even if it leaves out plenty of operational complexity.
 

One point that was hard to follow here is whether these components should be understood as roughly equal contributors to cost. They are clearly not, but the lecture does not yet give us a clear hierarchy. That omission matters, because a listener trying to model the economics needs to know which variables dominate.
 

The $650 Billion Question

 

The central financial puzzle arrives in a single question: if hyperscalers are spending $650 billion building data centers, where does that money go, and how much of it goes to companies like Crusoe?
 

That number is deliberately provocative, and it anchors the economic argument of the talk. The speaker is not asking whether AI is valuable. The question is structural: how is this enormous pile of capital actually being transformed into physical assets, and who captures the value along the way?

What the Money Buys

The lecture walks through several cost categories without assigning them neat percentages:

• Compute hardware, especially GPUs, repeatedly emerges as the star of the show.
• Memory, interconnects, power delivery, and cooling​ are presented as equally necessary, even if they receive less attention.
• Labor and logistics​ enter the picture when the discussion turns to construction timelines and supply chains.
• Land, permitting, and the grid​ become central once the conversation moves from a single building to a fleet of facilities.

The phrase “AI factory” starts to make more sense here. A factory implies not just expensive machines, but coordination: supply chains, utilities, maintenance, throughput, and yield. If a GPU sits idle because cooling failed or a substation was delayed, the economic loss is not merely technical. It is capacity that was paid for but cannot be used.

An Unanswered Allocation Question

What the talk does not do is give a clean breakdown of the $650 billion. We hear that GPUs matter most, while power and construction are major constraints, but we never get a pie chart or even rough ranges.

This is understandable in a live conversation, and Crusoe’s own business interests may make full transparency unlikely. Still, from a listener’s perspective, it leaves a gap. If the goal is to understand the economics of the supercycle, the missing middle is precisely this: what fraction of capital expenditure flows into chips, infrastructure, energy, land, and operations?​ Without those figures, “$650 billion” remains more rhetorical than analytical.

Power and Cooling as the Real Bottlenecks

As the discussion moves from components to systems, the bottlenecks shift from compute availability to infrastructure delivery. GPUs may be the most visible constraint, but a data center cannot function without enough electricity, adequate cooling, and a building designed to support both.

Energy as a First-Class Input

Lochmiller treats energy almost as a raw material. The logic is straightforward: no amount of algorithmic progress can run a training workload if the local grid cannot deliver stable power. This gives rise to a key argument in the lecture — that AI infrastructure is increasingly constrained by the physical world, not just by semiconductor supply.

This is where the “factory” language feels most apt. A chip company can promise faster hardware, but a factory still needs a reliable utility connection. Delays in substations, transformers, or transmission can stall a project even after the GPUs have arrived.

Cooling and System Design

Cooling receives special attention because modern high-density racks generate far more heat than traditional data center equipment. The lecture presents cooling as an integrated design problem rather than an afterthought.

The argument is stronger than a generic “AI uses a lot of power” claim. It identifies a specific scaling issue: as performance density increases, managing heat becomes part of the core architecture. A data center that cannot reject heat efficiently cannot fully utilize its compute.

One part that would benefit from an example is the relationship between power availability and model training. A concrete scenario — such as a delayed substation pushing back a training run by months — would make the economic stakes clearer. Without that illustration, the listener is left to infer how costly these delays really are.

The AI Factory as an Integrated System

Beyond individual components, the lecture invites us to think of an AI factory as a complete production system. Chips matter, but so does everything required to keep them running near full utilization.

From Chips to Throughput

There is an important conceptual move here. The value of a GPU is not just its specification on paper. It depends on whether power, networking, storage, cooling, and software can feed it work consistently. In factory terms, this is a throughput problem: how much useful compute can the system deliver over time?

This framing also exposes a risk. If one layer of the stack is overbuilt while another lags, capital sits underutilized. A company can own cutting-edge accelerators but still fail to achieve strong economic output because the supporting infrastructure creates a bottleneck.

Vertical Integration and Contingency Planning

Crusoe’s approach, as described here, emphasizes integration across planning, construction, and operation. The mountaineering anecdote resurfaces as a cultural claim: the company prepares multiple fallback plans rather than relying on a single critical path.

Whether this is genuinely distinctive or simply good project management is hard to judge from the outside. Most large infrastructure firms would also claim to plan for contingencies. Still, the underlying point is serious. In a market where delays can erode the value of expensive, time-sensitive hardware, operational resilience may matter as much as design elegance.

Crusoe’s Role and a Healthy Skepticism

The conversation naturally turns to where a company like Crusoe fits into the hyperscalers’ massive spending plans. This section is interesting not only for what it says, but also for what it cannot say without drifting into corporate promotion.

Enabler, Builder, or Partner?

The lecture suggests that Crusoe’s value lies in helping turn capital expenditure into usable capacity. Rather than focusing only on owning data centers, the company positions itself around accelerating deployment.

That distinction is economically meaningful. If demand for AI compute is surging and the bottleneck is how quickly infrastructure can be delivered, then speed and reliability may be more valuable than simply owning more square footage. The company’s pitch is essentially: we help turn ambitious CapEx plans into operational reality.

A Note on Bias

It is worth stating plainly that this is a CEO speaking about his own company. The talk is not an independent audit of Crusoe’s market position, cost structure, or competitive advantages. Claims about best-in-class execution should be treated as self-characterization rather than settled fact.

That does not make the technical observations unhelpful. Even if one discounts the corporate narrative, the lecture still offers a useful map of the dependencies between compute, energy, construction, and operation.

Strengths and Weaknesses of the Argument

Stepping back, the lecture is strongest when it treats AI as an industrial system rather than a purely software phenomenon. Its core insight — that massive AI investment ultimately depends on physical infrastructure — is compelling and accessible.

The “basic equation” of data, algorithms, compute, energy, and facilities gives listeners a memorable scaffold. The emphasis on power, cooling, and project delivery explains why simply designing a better model is not enough. In this sense, the talk succeeds as an introduction to the economics of the supercycle.

Its weakness is primarily one of precision. The conversation gestures at huge figures and major constraints but offers limited quantitative grounding:

• We hear about $650 billion in spending without a detailed allocation.
• We are told energy and cooling are bottlenecks without clear comparisons to historical data center loads.
• We learn that Crusoe helps accelerate deployment without a transparent metric for how much faster that deployment actually is.

These gaps do not invalidate the lecture. They simply mean it should be treated as a conceptual orientation, not a complete financial model. The lecture is especially valuable for students who have previously thought of AI only in terms of models and datasets. It forces the question: once you have invented the algorithm, how do you turn terawatts and concrete into usable intelligence?

Key Takeaways

• AI data centers are best understood as factories, not merely buildings full of servers.
• The core inputs are data, algorithms, compute, energy, and physical facilities — but their cost shares are far from equal.
• Hyperscaler capital expenditure is enormous, yet the lecture leaves the exact destination of those funds partially unspecified.
• The most credible bottleneck may not be chips alone. Power delivery and cooling capacity​ are physical constraints that can delay or cap utilization.
• GPU performance matters less in isolation than the system’s ability to keep those GPUs productive.
• Listening to a builder like Crusoe adds operational texture, but the account should be balanced against independent data and competitive perspectives.