Agentic AI · long only
Own the backbone every AI agent runs on
As agentic AI moves from demos to production, the durable profits accrue to whoever supplies the compute, memory, networking, power, and physical footprint that agents consume, plus the platforms that meter it. Not the model labs, and not the application layer where margins compress fastest.
Return since inception
From 2026-04-06
Versus S&P 500
Percentage points
Positions
Long only
Names followed
39 reviewed
The book
Full portfolio →| Holding | Weight | Conviction | Why it is held |
|---|---|---|---|
| NVDA
NVIDIA Corporation |
10.6% | 92 | Nvidia supplies the chips at the centre of the computing buildout, and its latest guidance calls for about $91 billion in a single quarter without counting any sales into China. It also locked up more than $500 billion of long term memory supply in July, which removes one of the few genuine limits on how many chips it can ship. |
| AMZN
Amazon.com Inc. |
10.5% | 87 | Amazon's cloud arm grew almost 37% last quarter, its fastest in more than four years, while earning a 39% margin on that business. The company also raised its spending plans for the year to about $220 billion and said rising memory prices were part of the reason, which is the main pressure on its spare cash. |
| PANW
Palo Alto Networks |
10.4% | 64 | Palo Alto Networks completed its $25 billion purchase of CyberArk in February to control how automated software is granted permission to act inside a company. Revenue grew 31% last quarter, but a large share of that came from the acquisitions themselves, and reported profit fell while the shares sit near a record high. |
| AVGO
Broadcom Inc. |
9.0% | 92 | Broadcom designs the custom chips and networking parts the largest cloud companies use to build their own hardware, and it has told investors that business grows from about $56 billion this year to over $100 billion next year. That forecast was left unchanged while the shares fell roughly 15% from their high, and it is backed by signed multi year agreements rather than hopes. |
| GOOGL
Alphabet Inc. |
9.0% | 80 | Alphabet's cloud business grew 82% last quarter and it now holds $514 billion of contracted work not yet delivered. The complication is that its spending pushed cash flow negative for the first time as a public company, and today several of its most senior research leaders announced they are leaving to start their own venture. |
| DDOG
Datadog Inc. |
6.6% | 68 | Datadog sells the software that watches whether other software is working and charges by how much gets processed rather than by headcount, so revenue rises as machine workloads rise. Growth was 32% in its last report, but the shares sit near an all time high above where most analysts think they should be, and the next results land tomorrow. |
| META
Meta Platforms Inc. |
6.5% | 52 | Meta's advertising business is still growing 28% and the shares are the cheapest of the large technology names after a 27% fall. But its data center spending consumed almost all of its spare cash last quarter, profits missed expectations, and management would not say how much it plans to spend next year. |
| EQIX
Equinix Inc. |
6.0% | 80 | Equinix operates the buildings where companies, networks and cloud providers physically plug into each other, and it delivered its biggest forecast increase ever on record customer bookings. It also raised its multi year growth outlook, but the building programme it announced will consume cash for several years and its next quarter guide came in light. |
How names are judged
01
Read
News, filings, and market data on every name we follow, gathered daily.
02
Judge
Two separate questions per company: how essential it is to the thesis, and whether now is a good moment to own it.
03
Size
Position size follows conviction, inside firm limits on any single name and any one sector.
04
Review
Nothing is bought, sold, or dropped by a formula. Every change is decided by a person.
What we have learned
All findings →L-001 · 2026-08-05
v2 conviction had no demonstrated predictive power, and the point estimates ran backwards
point for a first v3 review, never a judgment to defend or anchor to. Measured on the pre-rebuild snapshot: each conviction observation paired with the stock's 21-day forward return, minus the universe median over the same window, so the market is controlled for. | conviction band | obs | excess % | ±95% CI | effective N | reading | |---|...