Early Riders published an essay called The Singularity Stack, and it is the most coherent framing of the AI buildout I have read this year. The argument is that the whole thing is one conversion chain. Electricity becomes compute, compute becomes models, models become harnesses, harnesses become agents, and the work those agents do becomes payments that need somewhere final to land.
Nine layers in total: energy, compute, models, harnesses, agents, identity, coordination, value transfer, settlement. Each one is the physical input to the layer above it. A constraint anywhere reprices everything downstream. An improvement anywhere compounds through the rest.
The layer that actually holds value
The part I keep coming back to is harnesses, which is a word almost nobody uses outside agent engineering. The essay argues that the switching cost in AI sits a layer above the model, not inside it. The code is easy to copy. Open source implementations set the floor price near zero. What accumulates is the operating context built up around real work: memory, tools, credentials, evaluations, approvals, budgets, routing, and audit trails.
That is exactly what I have found building agent setups. The model is the part you replace every few months. Everything you built around it is the part that hurts to lose. Models get swapped underneath on cost and latency, and the harness is what makes those swaps routine instead of a rebuild.
Where I get off the train
The conclusion is where the essay stops being a useful map and starts being a position. It argues Bitcoin is the purest expression of the stack, the position with the fewest ways to lose. The reasoning is monetary rather than technical.
Technological progress produces deflation, the essay says, because efficiency means doing more with less, and money is where the difference gets saved. Under expandable money that deflation has to be offset. Loans are written in fixed nominal dollars while the borrower's revenue follows whatever prices the market will bear, so falling prices shrink income while the principal stays put. The essay traces the spiral to Irving Fisher in 1933 and cites roughly 9,000 US bank failures and a money stock that contracted by a third between 1930 and 1933.
If you accept that frame, holding the settlement asset is the cleanest claim on the entire chain. No earnings to miss. No shares to dilute. No competitor able to ship a better version of a fixed supply.
If you do not accept it, the essay is a well-built argument resting on a contested premise. Plenty of people look at the same deflation and conclude the opposite, that central banks can manage it or that the debt spiral is overstated. The layer analysis stands on its own. The Bitcoin conclusion depends on the monetary argument, and that is a separate fight.
The numbers that carry the weight
The acceleration case rests on a few figures. The essay cites Epoch AI estimating frontier training compute growing 4 to 5 times a year, a doubling roughly every six months. Moore's law doubled every two years. It also puts ChatGPT reaching 100 million users in two months.
Those are the load-bearing claims. If compute growth decelerates, the acceleration argument weakens with it, and the rest of the essay becomes a description of a normal technology cycle rather than a singularity.
The honest thing about the ending is that it admits what it cannot prove. Every conversion in the stack ends in money, so the settlement asset claims a share without naming which layer or which company books the margin. That is a real property. It is also untestable on any useful timeframe, which makes it easy to believe and hard to check.
Read it anyway. The layer framing will change how you read AI news, even if you reject where it lands.