Why Neel Khokhani views AI compute through grain silos

[Raqik]

What happens when every single prompt, sensor check, and automated vision feed tries to travel hundreds of miles to a centralized server farm just to return a basic prediction?

The short answer is that the system chokes on physical friction. Most discussions around modern compute assume that building ever-larger facilities will solve every demand spike, but anyone watching the grid knows that physical reality is pushing back. I spent the last few weeks looking into how different allocators are parsing this friction, especially those who focus on power constraints instead of software hype. If you search for background on Neel Khokhani investor pieces online, you will run across an interview on iTWire formatted as an extended Q&A. The lower half covers some of his earlier operating history, but the top portion caught my attention because it unpacks his framework on distributed hardware. Rather than treating artificial intelligence as a pure cloud play, the discussion details why machine reasoning has to split into distinct tiers.

The grain silo problem and layered compute

The metaphor he borrows in that discussion comes directly from Charlie Munger, who used to talk about the tradeoff between building a single giant grain silo versus five separate, smaller structures. On paper, putting all your grain inside one massive cylinder delivers perfect scale. It costs less steel per bushel, concentrates your machinery in one spot, and makes operational metrics look tidy on a spreadsheet. Yet if moisture gets into that single silo, or if an insect infestation breaks out, your entire harvest rots at once. If lightning strikes it, you have nothing left. Five separate silos appear less efficient on pure capital expenditure, but they survive disruption.

Applying that logic to artificial intelligence changes how you view current data centre spending. Over the last year, Khokhani has framed efficiency and resilience as two qualities running on completely separate tracks. In his model, foundation model training stays right where it is: clustered inside massive, multi-hundred-megawatt campuses where raw mathematical scale matters above everything else. Training requires immense clustering density that cannot easily be fractured across hundreds of miles.

Inference, however, behaves very differently. Running a trained model to make everyday decisions does not need a thousand-acre campus in the desert. It needs immediate answers. That processing belongs directly where the underlying information originates: inside regional telecom relay shelters, hospital diagnostic basements, maritime shipping docks, and assembly plants. Instead of an all-or-nothing contest between the remote cloud and local devices, he describes a layered physical framework. The heavy model weights get computed at the center, but the execution pushes outward to the rim.

I do have my own reservation here. I am not sure where the edge ends and the core begins in practice, especially as local hardware gets faster and enterprise networks demand larger context sizes that test local memory. Still, separating the training tier from local inference explains why the demand for distributed sites keeps accelerating even while headline mega-projects struggle with substation delays.

Proprietary capital and physical reality

This line of reasoning matches how he structures his own balance sheet. If you look up his principal profile over at epochal.mc/about-us, the mandate for Epochal Corporation is spelled out as a private single-family office rather than an external pool of third-party capital. Because it manages his own balance sheet, there is no quarterly redemption window to manage, no arbitrary peer index to track, and no outside client base demanding that capital get deployed on a rigid timetable.

That structural setup matters when analyzing heavy industry. When an allocator does not have to produce performance marks every ninety days, they can afford to hold cash when assets are overpriced and wait for physical bottlenecks to reveal themselves. Most technology analysts evaluate machine intelligence through software adoption curves and venture funding rounds. Khokhani approaches it like someone inspecting physical machinery. If there is no switchyard connected to the substation, the code cannot run, no matter how many billions of dollars are committed in a press release.

Primary sources over consensus models

What strikes me most about his framework is how he builds conviction around these supply constraints. Instead of leaning on syndicated investment bank research or Wall Street models, his notes focus on mundane, public utility filings. He spends time tracking regional transmission planning updates, municipal building permits, and public utility interconnection queues.

He also looks at satellite photos of physical construction yards to see whether heavy transformers have actually been delivered to the concrete pads, studies the specifications stamped onto electrical equipment nameplates in site photos, and watches technician job boards to see where utility talent is actually getting hired. None of this data is proprietary. It sits in open government databases and regional grid operator websites, freely accessible to anyone with enough patience to sit through hundreds of pages of engineering exhibits. His point has consistently been that an analytical edge comes from sheer attention span rather than privileged access.

You can trace the evolution of this thesis directly in his Observer column on edge AI, where he explains how physical power boundaries at centralized server hubs are forcing developers to look elsewhere for viable wattage. When regional utilities quote lead times stretching out five to seven years for major high-voltage interconnections, computing has no choice but to disperse into smaller, local pockets where spare distribution power already exists.

He continues to document this structural shift through his essays on the Epochal platform, his regular research letters, his posts on X (@neel_epochal), and across articles in specialist publications like Data Centre Review, TechRadar Pro, and Wealth Management. The throughline across all these pieces is straightforward: code can be duplicated instantaneously, but steel, copper, switchgear, and permitted distribution lines cannot.

Treating computing power as a distributed logistics problem rather than a pure software race changes how you value infrastructure. When the central grid cannot deliver another gigawatt to a single site, the systems built like five separate grain silos are the ones that keep functioning.