Agent = LLM + Harness
A model is not an agent. What turns capability into completed work is the scaffolding around it — and that is where the value accrues.
← All insights
Framework · Q4 2023 Letter
Sovereign AI
Compute became strategic infrastructure. Nations began buying it the way they once bought refineries and fabs.
Written January 3, 2024, and shared here as originally written. Views reflect that date.
Earlier in 2023, I wrote about the ‘Dawn of AI.’ The largest buyers of AI computing are the cloud hyperscalers, such as Amazon’s AWS, Microsoft’s Azure, Google’s GCP, and Oracle’s OCI. But another large and influential user base on the leading edge of supercomputing is government-funded scientific research centers.
Sovereign-funded AI supercomputer data centers are a powerful new driver that has emerged as most countries want to infuse their local culture, data, and startup ecosystems to advance their societies. What is incredible is the size of the installed base expansion occurring in just two years: ~300 exaFLOPs of new AI computing performance to come online between 2023 and 2025, seven times the roughly 50 exaFLOPs that existed before.
Announcing next-gen AI supercomputing: 350 exaFLOPS to drive scientific innovation
Cumulative AI performance of scientific supercomputers · exaFLOPS
Show the data
| System | Online | Cumulative exaFLOPS |
|---|---|---|
| Piz Daint | 2016 Q4 | ~1 |
| Tsubame3 | 2017 Q4 | ~1.5 |
| Summit | 2018 Q4 | ~4 |
| Pangea | 2019 Q2 | ~7 |
| Selene | 2020 Q2 | ~16 |
| JUWELS | 2020 Q4 | ~13 |
| Perlmutter | 2020 Q4 | ~18 |
| Polaris | 2021 Q4 | ~24 |
| Leonardo | 2022 Q4 | ~35 |
| MareNostrum-5 | 2023 Q1 | ~48 |
| Azure ND H100 v5 | 2023 Q2 | ~95 |
| Eos | 2023 Q4 | ~175 |
| Grace Hopper systems (LANL, CSCS, Jülich, Bristol UKAI, TACC, JCAHPC, NCSA, Cyfronet) | 2024–2025, announced | 350+ |
An exaFLOP measures a computer’s performance and is a unit of computing speed equal to one quintillion (1,000,000,000,000,000,000) floating-point operations per second, or a billion billion calculations per second. This measurement is used to quantify the performance of supercomputers. It indicates their ability to handle extremely large and complex calculations, often utilized in scientific research, weather modeling, simulations, and artificial intelligence. These new systems are spread out across the world at some of the most prestigious innovation centers, such as the Los Alamos National Laboratory (LANL) in Santa Fe, New Mexico, the National Composites Centre in Bristol, United Kingdom, the Swiss National Supercomputing Centre (CSCS) in Lugano, Switzerland, and the Joint Center for Advanced High-Performance Computing (JCAHPC) in Tokyo, Japan.
Technical experts due diligence
In the course of my investment due diligence research to identify winners and losers across the technology stack, I have spoken with dozens of technical experts across the computing landscape to better understand the changes of AI workloads. I have spoken with directors of cloud computing hyperscalers from Amazon, Facebook, Google, Microsoft, and Oracle, in addition to field experts putting AI to use, such as a principal investigator of AI development & implementation in the radiology/pathology department of a national hospital, a global solution strategy director of digital transformation, and a global leader of AI/ML, quantum computing and high-performance computing. These calls help me brainstorm new ideas and garner real-time feedback from customers, suppliers, and competitors about operations on the ground. Both the bull and bear cases are analyzed continuously to try to arrive at a ground truth based upon a range of expected future outcomes.
More frameworks
The Five-Layer Cake
Five layers stand between electricity and intelligence — energy, chips, infrastructure, models, applications. Each one gates the layer above it.
Disclosures
This page excerpts the framework discussion from Katam Hill LLC’s Q4 2023 Letter to clients, dated January 3, 2024. It omits portfolio performance and reflects views as of that date, not as of any later date. Certain economic and market information contained herein has been obtained from published sources prepared by other parties. While such sources are believed to be reliable for the purposes used herein, Katam Hill LLC does not assume any responsibility for the accuracy or completeness of such information. Unless otherwise indicated, market analysis and conclusions are based upon opinions or assumptions that Katam Hill LLC considers to be reasonable.
Securities or companies highlighted or discussed on this page have been selected to illustrate Katam Hill LLC’s investment approach and/or market outlook and are not intended to represent the adviser’s performance or be an indicator for how client accounts have performed or may perform in the future. Each has been selected solely for this purpose and not on the basis of performance or any performance-related criteria. Clients’ accounts are actively managed, and securities discussed may or may not be held in such portfolios at any given time. The securities referenced do not represent all of the securities purchased, sold, or recommended to clients. A complete list of all recommendations made by Katam Hill LLC within the past twelve months is available upon request.
Nothing on this page shall constitute advice or a recommendation or offer to sell or a solicitation to deal in any security or financial product. It is provided for information purposes only. Readers should not rely on this material in making any future investment decision. Past performance is not indicative nor a guarantee of future results. Actual returns for a particular investor will vary.