StanfordMS&E435EconomicsoftheAISupercycle｜Infrastructure,CapstoneCase-Class notes

Course Information

Institution: Stanford

Original Course: Stanford MS&E435 Economics of the AI Supercycle | Spring 2026 | Infrastructure, Capstone Case

Instructor Bio: This session is led by **Apoorv Agrawal**, Adjunct Lecturer in Management Science and Engineering at Stanford University and Partner at Altimeter Capital. Apoorv Agrawal holds a Bachelor’s degree in Computer Science from the National University of Singapore and an MBA from Stanford University. At Altimeter Capital, he leads the firm’s investments in artificial intelligence and enterprise software, with deep experience evaluating infrastructure investments across the AI value chain.

Course Description: This capstone case session applies the economic frameworks from prior lectures to a real-world, large-scale AI infrastructure project. It walks through end-to-end cost-benefit analysis, return on investment calculations, payback period modeling, and risk assessment for a flagship compute cluster deployment. The session demonstrates how to evaluate infrastructure economics under different AI growth scenarios and market conditions.

compute capacity. It also includes a very candid discussion about Intel’s turnaround and, later in the talk, a breakdown of how OpenAI thinks about compute supply, utilization, and the economics of inference.

What makes this lecture especially useful is that it does not stay at the level of “AI needs more GPUs.” Instead, it connects product demand, model capability, datacenter infrastructure, supply constraints, and semiconductor manufacturing into one investment and operations story. As a listener, I found some parts persuasive, some clearly speculative, and a few places where I wanted a concrete number or example.

Opening context: from networking to agents
The session opens with Professor Kati being introduced not only as OpenAI’s current compute lead, but as someone who has worked across networking, semiconductor strategy, and AI. He was previously Intel’s CTO and head of AI before joining OpenAI in November (1:16–1:24). That career arc matters because the lecture is not a narrow hardware talk; it keeps returning to how each layer—from fabrication to systems to models—depends on the others.

The first substantive topic is Intel. The speaker frames its recent narrative shift as more than a marketing relabeling. He argues that two forces are creating tailwinds for the company (1:35–1:51):

The world is heavily supply constrained, so firms that can actually manufacture—not only design—benefit.

CPUs are becoming more important again as AI systems begin to rely on agents.

This is an interesting reframing. Much of the public AI story focuses on Nvidia and accelerators. Here, the claim is that inference infrastructure and agentic workflows may make general-purpose CPU capacity valuable again, particularly for orchestration, control, and non-GPU workloads. It is a bold statement, and one I will return to later.

The speaker also emphasizes that Intel remains “pretty much the only leading edge American company left” with serious manufacturing capability (1:56–2:02). That may be slightly overstated depending on how one classifies other US firms and their partners, but the underlying point is clear: in a supply-constrained environment, domestic manufacturing capacity can become a strategic asset even if a company has struggled in product execution.

OpenAI’s compute chart: the central claim
A large portion of the lecture revolves around a chart released by OpenAI at the beginning of the year. On one side it shows compute capacity over the prior three years; on the other, the organization’s target capacity by the end of the decade (2:49–3:05). The chart is striking because OpenAI’s compute growth appears closely tied to revenue growth.

The lecturer’s job context helps explain why this chart is more than a marketing slide. At OpenAI, he runs “industrial compute,” meaning his team is responsible for delivering the compute needed for both training and inference (3:26–3:40). In his words, making the numbers go up is essentially the job. That is a useful reminder: when a frontier lab talks about compute scaling, it is not merely describing a research preference. It is describing an operational mandate.

Revenue as a lagging indicator of deployed compute
The most important analytical move in this section is the claim that, for frontier labs, revenue behaves as a lagging indicator. The speaker says it is basically a function of two variables: how much compute the organization has and how well that compute is utilized (3:55–4:10).

That is a deliberately simplified model, but it has real explanatory power. If more capacity comes online, users can be served, new products can be launched, and token consumption can rise. If utilization also rises—meaning the installed hardware is not sitting idle—then revenue should follow with some delay. The speaker supports this with OpenAI’s recent track record: compute has tripled year over year, and revenue has done the same (4:13–4:19).

This logic also explains the aggressive target shown on the chart. If the historical relationship holds, then adding capacity is not just an engineering ambition; it is the mechanism expected to support future revenue. But the lecture is careful not to present this as a permanent law. The speaker says there is no end in sight to the correlation “yet” (4:23), which leaves room for saturation, competition, pricing pressure, or a shift in demand.

One thing I found slightly hard to evaluate: the speaker treats utilization as if it can continue rising smoothly. In practice, utilization likely depends on product-market fit, model efficiency, request batching, demand peaks, and whether new capacity arrives in usable clusters. A single aggregate “utilization” number may hide a lot of operational complexity.

GPT-5.5 and the demand side
The discussion becomes more concrete when the speaker turns to GPT-5.5. He reports that Codex usage, and presumably revenue, has seen meaningful double-digit growth in the two weeks following its release (4:29–4:35). More importantly, he says people are no longer using it only for coding. It is increasingly being used for general knowledge work.

This matters because it broadens the potential demand base. Coding tools can be highly valuable, but the market for general-purpose knowledge work is much larger. If true, this could support the claim that token usage and task complexity will keep rising. The speaker describes a feedback loop: more capable models attract new uses, those uses generate more tokens, and more tokens require more compute.

That said, the lecture gives no specific figures for current token volume, revenue, or marginal cost per token. So while the direction of the trend seems credible, a listener cannot easily determine whether demand is growing faster than supply, or whether unit economics are improving. This is a case where a single example—such as one enterprise workflow or a comparison of tokens per task—would have made the argument easier to quantify.

Why supply constraints shape the whole industry
A recurring theme is scarcity. The speaker believes any company with serious manufacturing operations, especially fabrication rather than design alone, will benefit from today’s environment (1:42–1:53). This argument links Intel’s fortunes to OpenAI’s capacity needs: both are shaped by the same bottleneck, even though one supplies hardware and the other consumes it at extreme scale.

For OpenAI, the constraint is not simply “chips exist.” It is whether usable compute can be delivered fast enough across the full stack: facilities, power, networking, accelerators, memory, software, and deployment. The speaker’s role covers compute for both training and inference, which means the organization must plan for two very different workload profiles. Training tends to be concentrated and bandwidth-intensive, while inference depends more on latency, availability, and cost per request.

The lecture implies—but does not fully elaborate—that inference may become the dominant capacity concern. As products move from limited research previews to agentic systems that perform longer tasks, the amount of compute required per user interaction can increase. Agents may reason, call tools, browse, run code, and retry actions over a longer horizon. So a modest-looking product release can still translate into a large infrastructure footprint.

This is where the CPU comment becomes easier to understand. Accelerators may do the heavy neural-network computation, but a complex agent system also needs scheduling, context management, IO, safety checks, and coordination. The speaker’s claim that CPUs are “making a comeback” should probably be read as a systems-level observation rather than a prediction that CPUs will replace GPUs (2:04–2:12).

Strengths and weaknesses of the “compute equals revenue” argument
The lecture’s core argument is compelling because it starts from observed behavior rather than abstract theory. OpenAI tripled compute and tripled revenue; its leadership believes the next wave of products will continue driving usage; therefore, it plans for a sharp capacity increase by decade’s end. The mechanism is intuitive: more compute enables more users and more complex tasks, which produces more revenue.

It also explains why frontier AI firms often appear obsessed with infrastructure. If compute is both a production input and a constraint on growth, then delaying a datacenter or cluster can directly delay product expansion. From this perspective, “industrial compute” is not back-office operations. It is core strategy.

However, there are clear limits. First, the revenue relationship could weaken if market demand becomes saturated or if competitors narrow OpenAI’s product advantage. Second, revenue is not the same as profit. Even if the compute–revenue link persists, the marginal economics depend on model efficiency, energy costs, pricing, and utilization. Third, “compute capacity” itself is not a perfectly comparable metric over time. A newer cluster may be more capable per dollar or per watt than an older one, so raw capacity can mean different things in different years.

The speaker mostly acknowledges these uncertainties through careful wording. He says the correlation has no visible end “yet,” and he repeatedly frames the future as a research-driven environment (5:05–5:11). As a listener, I appreciated that restraint. Still, the chart itself is more confident than the caveats, and anyone using it for investment or strategic purposes should separate forecast from historical fact.

Intel: a turnaround built on execution
The Intel segment is short but conceptually important. The speaker identifies two bullish drivers: supply scarcity favoring real manufacturing, and CPU relevance returning through agentic AI (1:42–2:14). He also praises Intel’s CEO and says he loved working with the team (2:27–2:29).

This is a persuasive narrative, but it is not a complete investment thesis. The speaker acknowledges that “a lot of execution still [has] to be done” and notes that markets often move ahead of results (2:15–2:24). That is an important qualification. A favorable structural position does not automatically translate into timely product wins, foundry customers, or margin recovery.

I also found the joke about the speaker’s departure and Intel’s stock chart revealing (2:32–2:37). It lightly dismisses the idea that his move caused the recent stock move, while also reminding us that leadership changes are not sufficient explanations for share-price movements. The market is reacting to a broader set of expectations.

One unresolved question is how much of Intel’s opportunity depends on OpenAI-style demand specifically. If AI workloads increasingly concentrate on accelerators, Intel could benefit from CPU and systems demand without necessarily regaining leadership in every AI segment. The lecture suggests this possibility but does not resolve it.

What the lecturer seems to believe about agents
Although the transcript is truncated, the agent discussion is clearly central. The speaker describes OpenAI as “still a research lab” and appears to be setting up a distinction between pure research and deployed products (5:05–5:14). That distinction matters for infrastructure: research requires flexible, cutting-edge clusters, while a widely deployed product needs reliable, cost-effective serving infrastructure.

Agents change the workload because they can turn a single user request into many internal actions. This increases context length, tool calls, retries, and wall-clock time. In turn, it can increase demand for coordination compute—the kind of work the speaker associates with CPUs. It may also increase total accelerator demand if each agent step involves substantial inference.

The interesting tension is whether agents make systems more or less efficient. On one hand, better models can accomplish tasks with fewer steps. On the other hand, agents may unlock more ambitious tasks that were previously uneconomic to automate. The lecture leans toward the second effect: more capability leads to more total usage, not merely the same usage with fewer tokens.

That is plausible, but it is not guaranteed. A stronger version of the argument would distinguish between token volume, useful task completion, and revenue. It is possible for token use to grow while average revenue per user falls. So the key question is not just “Are people using more compute?” but “Are they using enough compute to justify the price they pay?”

Key takeaways for someone who missed the lecture
If I had to compress the talk for a classmate, I would emphasize five points:

Compute is the operating constraint.​ The lecturer’s team exists to supply OpenAI’s training and inference needs, so the capacity chart reflects daily operational priorities rather than a vague growth wish.

Revenue tracks deployed compute—for now.​ The simple model is compute multiplied by utilization, with revenue following. The historical evidence cited is three years of tripling compute and tripling revenue.

Demand is broadening.​ The GPT-5.5 discussion suggests usage is expanding beyond coding into general knowledge work, which could support higher token consumption and more complex agent tasks.

Supply scarcity favors real manufacturing.​ Intel’s potential recovery is tied to its fabrication capabilities and the return of CPU-relevant workloads, but execution risk remains substantial.

The central assumption is continued utilization growth.​ If users, tokens, and task complexity keep rising, the forecast makes sense. If demand saturates or efficiency gains outpace value capture, the relationship can break.

The strongest part of the lecture is its systems view. It avoids treating AI progress as simply a matter of model cleverness or chip performance. Instead, it shows why researchers, infrastructure teams, semiconductor manufacturers, and product designers are all part of the same economic chain.

The weakest part is the limited disclosure of operational detail. We hear that compute and revenue tripled, and that a recent product release drove double-digit growth, but we do not receive enough data to test the model independently. That is understandable in a public setting, but it means the chart should be treated as a strategic narrative rather than a fully specified financial forecast.

Final reflection
This talk is most valuable because it makes infrastructure feel like the central character in the AI story. The lecturer is not merely describing a hardware shortage; he is explaining why a frontier lab behaves the way it does. Compute determines what products can launch, how many users can be served, and how aggressively the organization can pursue agents.

At the same time, the argument rests on a continuing relationship between capacity, usage, and revenue. That relationship is well supported by the recent past, but it is not a law of nature. Energy limits, supply delivery, model efficiency, competition, and user willingness to pay could all change the slope of the curve.

For a classmate catching up, I would recommend focusing less on the exact target by 2030 and more on the mechanism: AI supercycle infrastructure is valuable when compute can be deployed, utilized, and turned into useful work. That is the thread connecting OpenAI’s datacenters, Intel’s factories, CPU demand, and agentic products.

Content Disclaimer:

This article is for general reference only and does not constitute professional R&D guidance, production process advice or quality certification. All material performance data has specific test premises; readers should verify parameters against actual equipment and working conditions.