From silicon to agentic software, read through the bonds, the balance sheet and the physical constraints.
NSO Capital is my independent research and advisory practice.
The work is narrow by design: how the AI infrastructure buildout is financed, where the physical limits bind, and which parts of the capital structure are mispriced as a result.
I spent twenty-five years managing risk across fixed income, structured credit, equity-linked instruments, and multi-asset strategies, across the full economic cycle, navigating three bear markets.
My practice publishes its reasoning, advises a small number of counterparties, and holds no view it cannot defend on the numbers.
Coverage starts at the wafer and follows the chain up to the agent, not the other way round.
I hold seven pillars rather than a diversified index, because the reward is in being right about the few things that matter.
Equity, converts, senior unsecured, leases and vendor finance read as one document.
Every view is written up with the case against it stated first, and the level at which it is wrong.
Demand for intelligence is not the scarce thing. The capacity to build for it is.
Compute demand is being underwritten years in advance, while the inputs that satisfy it clear on much longer cycles: advanced packaging, high bandwidth memory, transformers, interconnection queues, grid capacity, skilled labour. That mismatch is where returns and losses are decided, and it shows up first in how the buildout is financed. Who issues, at what tenor, against what asset, with what residual assumption. Reading that record carefully explains more of the outcome than any forecast of end demand.
Substations, turbines and packaging lines have lead times that no amount of capital compresses. Those dates are knowable, and they are rarely in the model.
Tenor, security package and vendor terms reveal what management believes about residual value long before guidance does.
The binding constraint moves: memory, then power, then interconnect, then talent. Owning the constraint matters more than owning the theme.
The buildout is covered as a single system. Each pillar is tracked for capacity, pricing power and how it is funded.
Accelerators, foundry capacity and advanced packaging. The floor under everything above it.
High bandwidth memory supply agreements and conventional DRAM cycles, where pricing turns fastest.
Generation, transformers and interconnection queues. The constraint that binds longest.
Shells, cooling and land. Real assets financed on assumptions worth testing line by line.
Optics, switching and the interconnect that decides how much of a cluster is usable.
The hyperscalers and neoclouds funding the cycle, read through issuance and lease obligations.
Where inference turns into revenue, and where gross margin proves the rest of the stack.
One pillar in isolation says little. Held together and priced against the debt that funds them, the seven describe the cycle.
I follow these companies very, very closely. The information here is a result of my research. Contact me if you would like to talk about how I reached these numbers.
Each idea is dated and carries the price at publication.
noel@nso-capital.comCassandra Chained to Demand is a philosophy and lens that I developed in response to Michael Burry’s work in AI.
I believe the discerning feature is that the entire AI trade is ultimately tied to end demand, irrespective of how difficult that is to gauge or measure at any point in time.
I publish irregularly, where there is something to say.
I will be joining Substack soon.
To be notified when a new piece is published, write to the address in the contact section.
Engagements are limited to a small number of counterparties: allocators sizing exposure to the buildout, credit investors underwriting issuers inside it, and management teams testing how their own financing will read to the market. The output is written, specific and dated. Positioning support, diligence on a single issuer or structure, and standing coverage of the pillars that matter to the mandate.
The useful question is rarely whether the buildout continues. It is who is paid, in what order, if it slows.
Email is read and answered within two business days by either myself or my assistant. For anything time sensitive, say so in the subject line.