By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.
Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.
At a glance
WHAT IT’S REALLY ABOUT
Why AI Compute Is Scarce and How Founders Navigate It
- Guido Appenzeller explains that exploding AI demand outstrips chip supply because of foundry capacity limits and bottlenecks in manufacturing chips and building cards, and new fabs take years and billions to build.
- Access to scarce compute often comes via pre-reserved, long-term commitments with large clouds or investment deals, while Appenzeller advises founders to shop around, including specialized clouds like CoreWeave or Lambda.
- Hardware choice depends on workload: memory needs, number of cards per server, networking, and whether the goal is training or inference, and Appenzeller says most founders are better off renting than owning until spending reaches very large scale.
- On moats and open source, Appenzeller says differentiated data can matter in niche domains, though fine-tuning on private data atop broadly trained models is likely, and open models are advancing but still lack a large open LLM at the time of recording.
- Appenzeller expects smaller models to increasingly run on devices, cutting inference costs for companies, and describes AI as a new type of compute that requires a rebuilt stack full of startup opportunities.
IDEAS WORTH REMEMBERING
5 ideasSupply cannot simply be scaled up quickly
Appenzeller says chips are made in capacity-constrained foundries like TSMC, capacity must be reserved long in advance, and building a fab takes a couple of years and billions of dollars.
Large-scale access requires commitments
To get newer chips at scale, companies often pre-reserve capacity through negotiations with big clouds, sometimes committing exclusively for two years, or through investment deals that bring capacity.
Consider consuming a service instead of hardware
Appenzeller suggests founders first ask whether they need hardware directly or just something running on top of it, such as a hosted model service like Replicate for Stable Diffusion.
Workload determines the right hardware and reservation style
Memory needs, cards per server, and networking fabric depend on model size and training versus inference; training and always-on inference tend toward long-term reservations, while batch jobs can use variable capacity.
Most companies should rent rather than own
Owning requires hiring staff and CapEx, so Appenzeller thinks even most mid and late stage founders are better off renting, with exceptions for specialized needs, sensitive data, or spending around $100 million a year.
WORDS WORTH SAVING
5 quotesThe biggest thing that is triggering that is just the, the, the crazy exponential growth of AI at the moment. I mean, AI has been booming since mid last year. I think nobody expected how quickly it would move, and that has just, uh, created a demand which at the moment the market can't fulfill.
— Guido Appenzeller
So you, you first go to elementary school to learn reading and writing and, you know, then you go to your, your, uh, you know, vocational training for the, for the specializ- specialized job that, uh, you know, you, you're going to do in the future.
— Guido Appenzeller
You can look at AI just as a new application, but honestly, I think it's, it's probably a better way to look at it as, as a different type of compute, right?
— Guido Appenzeller
And the big advantage is we don't actually need to know how to solve a problem as long as the network can figure it out, right? The neural network can figure it out, we're fine.
— Guido Appenzeller
As with any, any scarce, uh, scarce resource, right, there's a lot of, uh, deal-making going on
— Guido Appenzeller
High quality AI-generated summary created from speaker-labeled transcript.
