This past week, a framework for understanding GPU infrastructure moved quickly through investor circles. Alex Song, who spent more than a decade trading asset-backed securities and later structured more than $2 billion of debt financings, argued that a GPU neocloud runs on airline economics. Chamath Palihapitiya amplified the comparison, noting that Ken Griffin and Sridhar Ramaswamy have made the same point.
The analogy is sharp, and it points to a problem that financing alone cannot solve. Airlines did not improve only by financing aircraft better. They also learned to match the aircraft to the route. GPU infrastructure needs the same discipline, and today's fixed-server architecture makes that very hard.
The Challenge in Airline Terms
Song's core observations, laid out in his longer post, are straightforward:
- Capacity is perishable. An empty seat is gone once the door closes. An idle hour on an energized GPU rack is gone the same way, and the debt service accrues whether the rack is busy or not.
- Fixed costs dominate. Servers, ten-year leases on liquid-cooled rack space and 10% to 14% debt service account for well over 85% of the cost base. The marginal cost of running a GPU is mostly electricity.
- Utilization is everything. Airlines break even at roughly an 80% load factor. A debt-financed GPU cluster needs roughly 65% to 75% utilization. In Song's words, the returns live in the last 25 percentage points.
- The operator is squeezed. The dominant GPU supplier plays the role of Boeing and Airbus combined, while powered shells and grid interconnects are the new airport slots.
- The GPU is the engine, not the airframe. The data center lasts 15 to 30 years. The GPU “comes off the wing every few years,” and there is no shop visit that resets it.
Every point comes back to one question: how much installed GPU capacity is doing useful, paid work for the life of the asset?
The Hidden Constraint: A Fixed-Cabin Airplane
Imagine an airline that could fly only one aircraft type, with the cabin welded in place at the factory. Every route, whether it carried 40 passengers or 400, would get the same plane. Load factors would suffer because the equipment could not match the demand — or the airline would fly only the routes that happened to fit the aircraft.
That is how most GPU infrastructure is built today. GPUs are bolted inside servers in fixed groups, usually eight per box. The ratio of GPUs to CPUs, memory, networking and storage is decided when the hardware is purchased, not when the workload arrives. A job that needs two GPUs strands six. A job that needs 32 must be spread across four servers and pay for the network hops between them. A cluster can be fully allocated and still poorly used.
The data reflects this. In VentureBeat Intelligence's September 2026 VB Pulse survey, 55 of the 68 respondents running AI in production who gave a utilization figure (81%) reported average GPU utilization of 50% or less. That is a self-reported sample, not an industry-wide measurement, but it is consistent with production research. A six-month Alibaba study of 155,410 GPUs found a median GPU SM utilization of only 6% for online-inference jobs, a sign of how far claimed capacity and productive capacity can diverge.
If returns live between 75% and 100% utilization, a fixed machine shape is a structural tax on those returns.
The Software-Defined Machine
Corespan builds what we call the Software-Defined Machine (i.e. server). The idea is simple: stop fixing the shape of the computer at the factory, and let software define it when the workload arrives. The architecture has three layers:
- Hosts with FIC 2500 cards. Standard servers supply CPU, memory and networking. FIC 2500 cards connect them to the fabric.
- An optical PCIe interconnect. Native PCIe runs over optical links, so GPUs and storage no longer need to sit inside the server that uses them.
- PRU 2500 resource pools. Dense chassis hold GPUs and NVMe storage as shared pools rather than as captive parts of any one server.
Corespan Composer is the control plane above all three. It decides which GPUs and storage attach to which host and carves machines sized to the job: 1, 2, 4, 8, 16 or 32 GPUs per server today, with 64 on our roadmap. When the workload changes, Composer changes the machine. A simple way to put it: orchestration software decides where a workload runs, the inference engine decides where a token goes, and Composer decides what machine exists beneath them.
Mercury Rack packages this at rack scale: a 48U rack built from up to nine PRU 2500 chassis, with 72 double-wide PCIe Gen5 GPU slots, up to 111 kW in a standard air-cooled rack with in-rack direct liquid cooling available, up to 1.2 PB of storage per PRU 2500, and support for NVIDIA, AMD and Intel GPUs. Customers buy a rack. They operate it as a pool.
Filling All Three Cabins
Song describes neoclouds selling capacity in three cabins: first class (multi-year take-or-pay contracts for large training clusters), economy (8 to 64 GPUs rented by the hour or month for fine-tuning and inference) and standby (spot and batch capacity sold at a deep discount).
Fixed servers lock capacity into one cabin, and economy demand rarely arrives in neat multiples of eight.
Composable infrastructure turns the cabin layout into a software setting. The same pool can serve a 32-GPU fine-tuning job in the afternoon, be re-carved into sixteen 2-GPU inference servers in the evening, and run batch work overnight. The operator shapes the box to the customer instead of waiting for a customer who fits the box. That is the equivalent of changing aircraft gauge route by route, and it goes straight at the stranded capacity that keeps clusters below breakeven.
Our investor framing models raising useful utilization on the same assets from the 35% to 45% range toward 70% to 80%. That is a planning target, not a measured result, until customer deployments prove it. It is also exactly the band where Song says the returns are earned.
Changing the Engine Without Replacing the Airplane
Song's most important insight may be the last one: GPUs behave like aircraft engines, not airframes. They wear out economically every few years, and most GPU debt is sized as if the collateral were a slow-depreciating whole aircraft.
In a fixed server, the engine is welded to the wing. A new GPU generation usually means replacing the whole server, including CPUs, memory and networking that may still have years of useful life.
Disaggregation separates the engine from the airframe. Because Corespan pools GPUs apart from hosts, an operator can add a new GPU generation to a PRU 2500 pool, keep existing hosts in service, and move older GPUs into workloads where they still earn, such as inference, batch and development. Heterogeneous pools also free the operator from a single vendor's refresh cycle, with NVIDIA, AMD and Intel accelerators in the same rack.
For lenders, this changes the underwriting conversation. Collateral that can be redeployed across workload classes, and partly refreshed rather than replaced, is a more defensible asset than a fixed box whose value depends on one generation and one customer.
More Revenue from Every Powered Megawatt
If powered shells are the new airport slots, the scarce input is not GPUs. It is megawatts. The metric that matters is revenue per megawatt, which depends on how many productive GPU-hours each megawatt produces.
Supporting up to 32 GPUs per host instead of eight means fewer hosts, fewer host CPUs, fewer network ports and less power spent on overhead per GPU. Higher utilization means more of each megawatt is turned into paid tokens rather than idle baseline draw. Both levers point the same way: more output from the same constrained slot.
Where the Analogy Breaks and Why It Helps
Not everyone accepted the comparison. Critics on the thread pointed out that AI can create its own demand in a way air travel cannot, and that compute may remain in shortage.
We agree, and that strengthens the case for flexibility. AI demand is not only growing. It is changing shape: from training to inference, from dense models to mixture-of-experts, from short prompts to long-context agents. Each shift changes the ideal ratio of GPUs to CPUs, memory and storage. A fleet with a fixed machine shape is a bet that today's ratio will hold for the life of the debt. A software-defined fleet can follow demand wherever it goes.
What This Means
Wall Street has brought the aircraft financing playbook to GPUs. The industry now needs the aircraft operating playbook too: right-size the equipment to the demand, keep every seat sold, and refresh the engines without scrapping the airframe.
That is what the Software-Defined Machine is built to do. Pool every GPU. Compose any server. Earn the last 25 points.
Corespan will be at SC26, Booth 5003, November 17 to 19, with Mercury Rack on display and a preview of Composer 4.0. Come see how the machine changes shape.