There is a moment in every technology cycle where the symbol of progress quietly stops being the thing that produces progress. We are living through that moment in the AI infrastructure super-cycle right now.
For the past three years, the GPU has been a trophy. Press releases are written around GPU counts. Funding rounds are sized against GPU commitments. Data center announcements lead with megawatts and rack densities as if the hardware itself were the achievement. The accelerator has become the scoreboard.
That worked, briefly, because the industry was solving a single, well-bounded problem: how fast can we train a dense transformer on a dense matrix multiplication? In that world, the answer really was "more GPUs, packed more tightly, wired more directly." The trophy and the outcome were the same object. That world is ending — and the operators who recognize it first are going to spend an order of magnitude less to deliver an order of magnitude more useful AI.
01 The Trophy Era Was a Dense-Matrix Era
Step back and look at what the standard AI rack assumes. Two to eight GPUs per CPU. A fixed ratio of CPU to accelerator. A bill of materials frozen at procurement time. Workloads route to whichever box has spare cycles, regardless of whether that box's shape — its memory, storage, and interconnect topology — matches what the workload actually requires.
That architecture made sense when every workload looked the same. Dense pretraining is a beautifully regular problem: predictable memory access, contiguous tensor shapes, synchronous all-reduce traffic that maps cleanly onto legacy interconnect schemes. If your only job is to multiply enormous matrices, the standard eight-GPU box is a reasonable unit.
The workloads that actually generate revenue have already diverged from that profile. Inference dominates real spend. Retrieval-augmented generation pulls more data from storage than from HBM. Agentic systems traverse graphs of tools, memories, and intermediate results in patterns that look nothing like a flat tensor. Reasoning models hop through sparse, dynamic state the silicon was never designed to navigate efficiently.
The result is a quiet scandal hiding inside every AI data center: utilization numbers no operator wants to publish. We have seen GPU fleets reported running at 38% — or as low as 11% — effective utilization, and inference clusters where average occupancy sits below 50%. Hyperscalers do better, but not by as much as they would like you to believe — and they do it largely by routing dense pretraining jobs that happen to fit the building block.
The trophy era in one sentence: we built our infrastructure around the workload we were proudest of, not the workload that pays the bills.
02 Compute as a Utility Is a Different Mental Model
Electricity went through this exact transition in the early twentieth century. Factories used to own their generators. Owning a generator was, for a while, a sign of seriousness — you were a real industrial operation. Then the grid arrived, and within a generation the question stopped being "how big is your generator?" and started being "what are you doing with the power?" Generators became commodity inputs. Outcomes became the trophy.
Compute is at the same inflection point. The question is no longer "how many GPUs do you have?" It is "how many useful tokens, embeddings, traversals, and inferences can you deliver per dollar and per watt?" That is a utility question. It cannot be answered by stacking more accelerators inside the same fixed-shape box, because the bottleneck is not raw FLOPs. The bottleneck is the mismatch between rigid hardware geometry and fluid workload demand.
A utility model has three properties that the trophy model never had:
- Resources are pooled, not partitioned. In a utility, capacity sits in shared reservoirs and is drawn down on demand. In trophy infrastructure, capacity sits in fixed servers and is stranded the moment a workload doesn't match the box.
- Allocation is dynamic, not procurement-time. A utility responds to demand on the timescale of the demand itself. Trophy infrastructure responds on the timescale of a six-month bill-of-materials decision.
- Measurement is in outcomes, not inventory. Utilities are judged on delivered service. Trophies are judged on what is sitting in the cabinet.
Every one of these properties requires a rethink of the physical substrate. You cannot pool resources you cannot move. You cannot dynamically allocate across boundaries you cannot cross. You cannot measure outcomes if your infrastructure can only report inventory.
03 What the Substrate of a Utility Actually Looks Like
The architectural shift required to turn compute into a utility is not subtle, but it is precise. Three things must change at the same time.
The first is the interconnect. The standard PCIe bus inside a server is the silent villain of the trophy era. It assumes everything an accelerator needs is within a few inches of the accelerator. The moment you try to share a GPU across hosts, or extend its memory beyond the chassis, you collide with that assumption. The fix is to move PCIe itself onto an optical fabric — to take the SerDes that have always lived on a copper backplane and run them over photonics, so the bus becomes a data center fabric rather than a chassis-bound wire. This is not a network protocol grafted onto PCIe. It is PCIe, end to end, with optics as the physical layer. The accelerator does not know it has crossed a rack boundary. The workload does not know it has crossed a host.
The second is the resource unit. If the bus is fluid, the box no longer needs to be. Instead of eight GPUs welded to a CPU host, you can build dense resource pools — Photonic Resource Units, in our case — that hold ten or twelve accelerators and expose them to whichever host needs them. CPUs become thin front ends. GPUs, NVMe, NICs, and future accelerators sit in shared reservoirs and get composed into the shape a workload requires. A single host can pull twenty-four to thirty-two GPUs when it needs a large memory domain, and release them when the job is done. The same hardware serves training in the morning, inference at midday, and a retrieval-heavy agent workload at night, without a single cable being touched.
The third is the control plane. None of this matters if a human has to schedule it. The whole point of a utility is that the consumer does not negotiate with the substrate. A composer layer has to sit above the fabric and present resources as a single pool to Kubernetes, to ML platforms, to whichever orchestration system the operator already uses. Workloads request the shape they need. The fabric assembles it. When the job finishes, the resources return to the pool.
Put these three together and the economics invert. A consumer-grade GPU in a composable pool can outperform a premium GPU stranded in a fixed server, because utilization is the multiplier that matters. A cluster of mid-tier accelerators that runs at 87% utilization will deliver more useful work than a cluster of flagship parts that runs at 38%. The trophy era treated those numbers as inevitable. The utility era treats them as a design failure.
04 The Harder Problem the Industry Has Not Yet Solved
We want to be honest about what composable infrastructure does and does not do.
It solves the fleet-level waste. It eliminates stranded capacity. It collapses the procurement-to-deployment cycle. It lets operators buy hardware once and reshape it as workloads evolve. It is, in our view, the necessary substrate for the utility era.
It does not, by itself, solve the problem of GPUs sitting idle inside a single workload because the workload's memory access pattern is hostile to the silicon. Sparse graph traversals, hypersparse reasoning frontiers, and dynamic retrieval patterns starve GPU cores in ways that no amount of optical switching can fix. That problem lives in the software layer — in execution planes, in data representations, in the willingness to stop pretending every problem is a dense matrix multiplication.
We are seeing early signals from researchers and founders who are starting to attack that layer directly — people rebuilding execution planes from the ground up, compressing sparsity before it ever hits the bus, isolating computation to the active frontier of a reasoning graph rather than materializing the whole tensor. That work is essential, and it is going to land on a substrate. The question is whether that substrate is going to be another rigid eight-GPU box, or a fluid pool of resources that can be reshaped to whatever the new execution plane needs.
We are betting on the latter, because the alternative is to keep buying trophies in the hope that the next one will finally be the right shape.
05 What Changes for Operators
If you operate AI infrastructure, the transition from trophy to utility is not an abstract debate. It shows up in your numbers.
Capital efficiency improves because you stop buying duplicated components inside duplicated boxes. Depreciation slows because you can upgrade card by card, vendor by vendor, instead of replacing a whole server every cycle. Time to deployment collapses because new workload shapes do not require new procurement. Vendor lock-in dissolves because the fabric is agnostic to whose logo is on the silicon.
The operators who internalize this first are going to look very different from the operators who do not. They will run smaller fleets that produce more useful output. They will offer GPU-as-a-Service with margins their competitors cannot match. They will adapt to new model architectures in days rather than quarters. And they will stop having to explain to their boards why a multi-million-dollar cluster is running at half utilization.
The trophy era was a necessary phase. It built the muscle, proved the demand, and funded the silicon roadmap that the next decade will run on. But it was always going to end the same way every infrastructure era ends: with the realization that the inventory is not the point. The output is.
Compute is becoming a utility. The substrate that delivers it has to be composable, photonic, and dynamic. Everything else is just a trophy in a more expensive cabinet.