Tim Davis's recent essay is one of the sharpest things written about AI infrastructure this year. Read it before you read this. His argument — that GPU debt is being priced on software portability, not silicon; that today's investment-grade GPU financings require a hyperscaler co-signer because the asset cannot yet stand on its own; and that a vendor-neutral software layer is what turns compute into financeable collateral — is correct, and it lines up with everything we see in the field.
We want to build on it, because there is a second half of the same argument that we think is under-discussed. If software portability is what lets a workload move across silicon, hardware composability is what lets the silicon move across workloads. A lender's recovery scenario needs both. Every GPU loan is really a software loan — and every GPU loan is also a topology loan.
The Chassis is Collateral Too
Davis's causal chain — workload portability → operational transferability → asset liquidity — is the right frame, but "asset" in that chain is not a GPU. It is a rack and in 2026, most of the AI racks being financed at scale are HGX-class servers: eight accelerators welded to a single baseboard, one NVLink topology, one power envelope, one cooling design, one generation.
Think about what that means for the repossession scenario Davis walks through. A lender takes the keys to a hall of Hopper HGX boxes. Even in a world where the software stack is fully portable across vendors, the new operator inherits:
- A fixed 8-way topology whether their workload wants 8, 4, or 72 GPUs in a domain
- A cooling and power design tuned to one specific SKU that is no longer the newest one
- Zero ability to insert a next-generation accelerator without a forklift upgrade of the entire chassis
- No way to reconfigure the fabric to serve prefill on one node and decode on another as inference disaggregates
- A depreciation curve set entirely by the vendor's roadmap, exactly as Davis describes
Software portability solves "can someone else program these chips." It does not solve "can someone else redeploy this rack against a different workload mix, a different generation of silicon, or a different service tier." That is a physical problem, and it is where a large share of the residual-value risk actually resides.
Put another way: Davis is right that CUDA is a moat for one vendor rather than a standard for an asset class. HGX is the physical version of the same moat. It ties the collateral value of a rack to a single vendor's next product decision. When Blackwell ships, Hopper HGX doesn't lose value because CUDA broke. It loses value because the box can't be reconfigured to host what comes next.
What Lenders Actually Need to See
Davis lays out the four questions a lender asks in a recovery: (i) what will someone pay for the hardware, (ii) how likely is it that a new operator can use it, (iii) how quickly can it be put back to work, and (iv) what will that transition cost. Software portability compresses questions two and four. Hardware composability compresses questions three and four, and it is the only thing that touches question one directly, because it changes what "the hardware" even means in a resale.
A composable rack is not a fixed 8-GPU appliance. It is a pool of accelerators, memory, and fabric that a new operator can carve up on the fly — a 72-GPU domain for a large-model serving tenant on Monday, sixteen 4-GPU inference partitions for a batch tenant on Tuesday, a mixed prefill-and-decode topology on Wednesday. The same physical asset serves a much wider set of workloads and a much wider set of buyers. That is what "many qualified operators" looks like on the hardware side of Davis's aircraft analogy.
It is also the only way the "long tail" argument Davis makes for older GPUs actually pays out. Yes, an A100 can still earn revenue at year six — but only if the rack it lives in can be re-provisioned to serve the batch, fine-tuning, and mid-tier inference work that keeps it useful. An A100 stranded in an HGX chassis that can't share memory pools with newer silicon, or can't be repartitioned to smaller inference shards, doesn't get to run that tail. The chip is capable. The chassis is not.
Composability Over Concentration
This is the thesis we have been building Corespan against. The Corespan PRU 2500 is an air-cooled chassis with some direct-liquid-cooled options for a GPU/SSD utility chassis. The FIC 2500 fabric card provides scale-across PCIe over photonics. Corespan Composer is the orchestration layer that lets an operator carve, recombine, and re-target that pool of GPUs and memory without touching the physical build. The design goal is boring and specific: make the rack outlive the generation of silicon inside it, and make the topology outlive the workload it was originally sold to serve. That has three consequences a debt investor should care about.
Generation-neutral collateral.
A composable rack can accept multiple generations of accelerators. When Blackwell displaces Hopper, or when the next Instinct or the next custom accelerator ships, the chassis, cooling, power, and fabric survive the transition. Depreciation of the silicon does not drag the whole capital stack down with it. The lender is no longer underwriting a bet on one vendor's roadmap.
Vendor-neutral operator pool.
Because the fabric is PCIe-native and the orchestration layer is vendor-neutral, the pool of qualified operators for a repossessed rack is not defined by "who has this exact HGX SKU deployed today." It expands to anyone running a heterogeneous fleet — which is where the market is going, not away from.
Utilization as a credit property.
Davis flags median GPU utilization figures in the single-digit percentages at Alibaba, low double-digits at xAI, and 50-to-60 percent at typical inference shops. Composability is the mechanical answer to that. A pool that can be re-partitioned in software delivers steadier utilization, which delivers steadier cash flows, which — exactly as Davis argues on the software side — are what actually get financed at investment-grade spreads.
None of this replaces the software argument. It sits underneath it. A portable software stack running on a rigid, generation-locked chassis still leaves the lender exposed to the physical roadmap risk that today shows up as a 45-to-69 percent year-three price band on Hopper. A rigid rack running a portable stack is still a rigid rack.
The Container Analogy, Finished
Davis closes with the shipping container: it did not make shipping valuable, but it made shipping bankable. That is exactly right, and we would extend it one step. The container worked because two things standardized at the same time — the box, and the way it moved through ports, ships, cranes, and trucks. Standardize the box without standardizing the handling equipment and you get a warehouse full of steel that no crane can lift. Standardize the handling without standardizing the box and every port needs bespoke gear.
Compute needs both. A software layer that lets workloads move across silicon is the standard container. A composable rack architecture that lets any generation of silicon share fabric, memory, cooling, and power is the standard port. You need the pair for compute to actually become bankable on its own credit rather than a hyperscaler's balance sheet.
We think Davis is right about the destination. We would just add that the road there runs through the rack, not only the runtime. Software makes compute programmable. Composability makes compute redeployable. Together, they make compute financeable.
That is the layer we are building. If you are a lender, an operator, or a hyperscaler thinking about how to underwrite the next hundred billion dollars of GPU capacity without a co-signer, we would like to talk.