I got into looking at these lectures by looking at the lineup of speakers they had. It was completely stacked with big names from industry (Sam Altman, Ben Horowitz, Jensen, etc.). But the comment section says its just ‘AI Coachella’, meaning nothing of substance is being said, and its mainly for maintaining the interviewee’s image.

I decided to take a look at it myself, is there really of anything of substance said?

The first lecture is:

Lets takea field-trip into the 2026 Frontier AI factory

In 2026 the way we make infrastructure for AI is changing.

The AI scaling loop is get compute -> get data -> model results, put the money back into compute. Bottlenecks include how much compute you have, the public’s perception, and how much capital you have Most companies (like in YC) follow these three steps:

1. incubation (focus on what part of the system you want to attack)
2. SOTA release
3. expand model to more capabilities / diversify offerings

Its fun to take apart new companies and see how many are actually following this roadmap.

THE ultimate scaling law:

• More compute turns into better models, which turns into better revenue From a systems perspective, you can convert between infra and compute: $1 of infrastructure turns into $1 of revenue. GUARANTEED. WOW what a claim.

Think about it. Literally every big company is exponentially investing in infrastructure. For a reason right? Example: The H100 GPU is a 2 year old chip, but the cost to use them is peaking NOW.

• Computing chips is not just a commodity that depreciates over time

examples of products that followed this law:

Steel: Hoarded until industrialization caused prices to crash. Fiber Optics: Over-expansion led to internet transit prices dropping 90% in 5 years. DRAM: Panic/world events trigger infrastructure sell-offs before the market stabilizes. Uranium: Value peaked, crashed after major incidents, and was ultimately stabilized by government intervention. Stabilization Timeline: Historically takes 3 years for digital commodities and 6 years for physical infrastructure.

Compute is Not Yet a Commodity

Unlike traditional commodities, compute is currently non-fungible and incredibly difficult to forecast. This means chips are not interchangeable; an H100 GPU cannot be swapped for a H50.

Path to Standardization of Compute

To turn compute into a true, tradeable commodity, the industry requires trusted standards and institutions to establish fungibility.

4 Requirements for Compute Fungibility:

1. Common Unit: A universal metric of measurement.
2. Standard Delivery Interface: Uniform access points.
3. Interconnection / Pooling: The ability to combine resources.
4. Commercial Infrastructure: Reliable metering, control, and settlement.

Current State: We are currently in the pre-standardization era of compute—akin to the internet before the adoption of the HTTPS standard.

extra thoughts

Anjey seems seriously happy go lucky and optimistic about how AI will 100% without a doubt scale. Though his points are not nearly as thorough as Ronald Coases’ theories, it is likely just a result of the times we live in. Attention spans are going down so seriously anything with a moderate amount of value gets popular.

Never mind -> its not just attention spans are going down. What is throwing me off is that instead of providing examples, he should explain in more depth what it means that infra can directly convert to revenue.

on the positive side -> he makes the concrete claim that AI is in the prestandardization era. which means AI will get standardized too, and turned into something fungible. and that AI infra directly translates to revenue.