AI compute research and GPU market publications

Procurement

The Basis Gap in Compute Acquisition Structures

GPU procurement is not a search for the cheapest hour. Buying, leasing, and reserving capacity each package three exposures differently: replacement-price risk, unused-capacity risk, and residual value. The essay’s thesis is that the hidden variable is basis—the gap between what a contract promises and what a workload later needs. A cheap commitment without reserved machines can be right on rate and wrong on access; a reservation without a matching spec can strand capacity while the buyer still pays for a second, better-fit block. At thousands of GPUs those mismatches become project-finance problems, because collateral values and contracted cash flows move on different clocks. The market question is which risks a fleet owner should keep, and which to push to a lessor or lender, when hardware, utilization, or demand diverge from the term sheet.

Compute markets

Compute in Space. Why, How, and When?

Ground datacenters cannot expand as fast as AI demand because power interconnects, water for cooling, and suitable land all take years. The essay argues that orbit is becoming a real alternative, not a slogan: launch costs have fallen sharply, solar collection is far stronger above the atmosphere, and radiative cooling can reject heat without consuming rivers. Early orbital H100 hardware and reusable-rocket economics are the proof points. Radiation, servicing, and extra latency still decide whether space is a hedge or a science project, especially for delay-sensitive inference. The practical question for buyers and financiers is when orbital capacity should enter a terrestrial fleet’s cost and obsolescence model—and how to hedge the chance that ground hardware is stranded if compute migrates off-planet.

GPU economics

On GPU Depreciation

Investor alarm that five-year GPU lives overstate earnings treats an accounting schedule as if it measured physical wear. The essay’s thesis is that this conflates two different things. GAAP depreciation allocates capex, moves reported earnings, and creates a tax shield. Economic loss of GPU value is driven mainly by new architectures and training regimes, not by parts failure. Older accelerators can remain cheap for batched, throughput-heavy work long after they look obsolete on a latency benchmark, and hyperscaler retirement histories imply useful lives closer to seven to nine years. The market question for operators and lenders is how to report and finance useful life without using a depreciation table as a residual-value forecast—and whether economic obsolescence should be hedged rather than modeled as linear wear.

Memory

Memory: How It Works and Why It Sets the Cost of Compute

Frontier models wait on memory more than they wait on arithmetic. Prefill and decode keep weights and a growing key-value cache in high-bandwidth memory, so HBM capacity and bandwidth set how much context, occupancy, and therefore usable compute a GPU can deliver. The essay’s thesis is that memory, not transistors, is now the price-setting input. HBM cannot expand smoothly: stacking yield, CoWoS packaging, and a three-vendor allocation market reprice at contract resets rather than on a continuous spot curve. Because memory already dominates the bill of materials on leading accelerators, a discrete HBM step change reprices whole clusters at once. The practical question is whether buyers and lenders can treat HBM as a scarce, hedgeable commodity instead of a background component cost.

Compute markets

The AI Chip Wars: From Supplier Dominance to Contested Equilibrium

Headlines about OpenAI training on Amazon Trainium, or NVIDIA licensing Groq inference designs, invite a winner-take-all reading of the chip race. The essay’s thesis is narrower: large buyers are building credible outside options so they can contest pricing power even if NVIDIA remains the performance leader. Hyperscaler silicon does not need to win every workload; it needs to be viable enough to change contract terms, optionality, and where margins sit in the stack. Competition will appear first in cloud pricing and flexibility, not in merchant-GPU share. The market question is how to budget and hedge GPU-hours when the price of compute fragments across architectures, regions, and bargaining positions rather than remaining a single supplier quote.