What Clayton Christensen’s Steam Shovels Teach Us About Photonic Switching and the Next Wave of AI Compute
In Chapter 3 of The Innovator’s Dilemma, Clayton Christensen tells a story that has become one of the most durable parables in business strategy. On pages 64 through 73, he walks through the history of the mechanical excavator industry, a world of cable-actuated steam shovels that for nearly a century defined how humans moved earth. It is a story about a dominant industry that did almost everything right, listened carefully to its best customers, improved its products on the metrics those customers demanded, and was nevertheless annihilated by a technology that, at its introduction, looked like a toy.
The parallel to today’s AI infrastructure build-out is so exact it is almost uncomfortable, and it is the reason I believe that photonic switching — the architectural vision on which we are building Corespan Systems — will not first win inside the gigawatt-scale frontier training campuses that dominate the headlines. It will win, as hydraulics did, somewhere else entirely.
The Cable Shovel Industry Had Three Decades of Record Profits Before It Died
Christensen’s data is worth recalling carefully. From the 1830s through the 1940s, the steam-shovel industry — Bucyrus-Erie, Northwest Engineering, Marion, and about two dozen others — executed a textbook run of sustaining innovation. They migrated from steam to gasoline, then to diesel-electric power, then to arched-boom designs. Twenty-three of the twenty-five leading steam-shovel makers successfully jumped from steam to gasoline engines. Bucket sizes grew roughly four percent a year, every year, for decades. Customers — sewer contractors, general excavation firms, and the mining majors — wanted bigger buckets and longer reach, and the industry delivered exactly that. As Compound Learning recounts the history, the steam-shovel manufacturers recorded all-time record profits as late as 1966.
When J.C. Bamford shipped the first hydraulic excavator in 1947, it could lift a quarter-cubic-yard bucket, reach six feet, and rotate only 180 degrees. A comparable cable shovel could lift ten to twenty times as much and rotate a full circle. No serious customer in the established value network wanted one. Bucyrus-Erie tried to split the difference with the “Hydrohoe,” a cable-hydraulic hybrid aimed at its existing customers, and as Shortform summarizes, the machine languished for ten years before the company gave up and went back to cables.
The hydraulic manufacturers did something different. They did not try to serve the miners. They found residential contractors who already owned tractors and who needed to dig narrow utility trenches for the postwar suburban building boom — a job no cable shovel could perform economically or even physically. Those customers did not care about bucket size. They cared about maneuverability, speed, and cost. Over twenty-five years, bucket capacity scaled from three-eighths of a cubic yard in 1955 to ten cubic yards by 1974, at which point hydraulic machines were simply better on every axis that once favored cables. Of thirty cable-shovel manufacturers, only four survived. Caterpillar and Deere, which had entered through the residential trench market, inherited the earth.
Christensen’s essential point is not that the incumbents were stupid. They were extraordinarily competent, and they failed precisely because they were competent. Their customers did not want hydraulics. Their margins did not justify hydraulics. Their resource-allocation processes, tuned to their most profitable buyers, could not rationally invest in hydraulics, and by the time hydraulics had matured enough to matter to those customers, the game was already over.
The AI Infrastructure Industry Is Deep Inside Its Cable-Shovel Era
Now look at what has happened over the last five years in AI infrastructure. Since 2020, the industry has executed a magnificent run of sustaining innovation along a single dimension: scale. Bigger clusters. Bigger models. Bigger power draw. Bigger capital outlays. Every year, the bucket gets larger. Nvidia’s Hopper generation begat Blackwell — the B200 delivering roughly five times the AI compute of an H100, with 192 GB of HBM3e at 8 TB/s, priced at thirty to forty thousand dollars a chip. A GB200 NVL72 rack now delivers 1.4 exaflops from seventy-two GPUs wired together as a single coherent accelerator. Nvidia’s data-center business crossed $215.9 billion in fiscal 2026 revenue. Individual training campuses are planned at the gigawatt scale — xAI’s Memphis sites alone are targeting one million GPUs and a full gigawatt of grid draw.
This is the steam-shovel trajectory in its purest form. Every year the industry’s best customers — the three or four hyperscalers and frontier labs who can write ten-figure checks — demand more of exactly the same thing, and the industry delivers. Rack power densities have moved from the traditional 10–15 kilowatts to 50–150 kilowatts per rack, with some AI training racks pushing past 100 kW. Individual chips that drew 150–200 watts a decade ago now draw 700–1,200 watts. U.S. data-center electricity consumption reached 183 TWh in 2024 — more than four percent of national electricity use — and is projected to reach 426 TWh by 2030. The Financial Times reports that grid-connected power coming online over the next three years will be able to supply only a fraction of the requested AI load.
There is an extraordinary amount of money being made right now selling bigger buckets. There will probably be record profits in this business through 2027, 2028 and maybe 2029. Yet the dimensions along which the industry is competing, absolute FLOPS, rack density, training-cluster size, are the classic markers of performance overshoot. Forty percent of existing AI data centers are projected to face operational constraints by 2027 due to power shortfalls. New 100 MW project requests are being quoted five-year grid-connection timelines. Training clusters are now generating power-load fluctuations of up to 70% within milliseconds, making them a first-order threat to regional grid stability.
The customer the industry is optimized around — the hyperscaler commissioning the next gigawatt campus — wants bigger, faster, denser. That customer does not want a fundamentally different architecture, and that is exactly the condition under which Christensen’s framework predicts a disruptive entrant takes hold.
The Analog-to-Digital Boundary Is Where the Hydraulics Live
If the sustaining trajectory is “more watts, more copper, more electrical packet switching,” then the disruptive trajectory has to live somewhere the incumbents cannot rationally invest. At Corespan Systems, our conviction is that it lives at the analog-to-digital boundary inside the data center — specifically, in replacing electrical packet switching and copper-based PCIe fan-out with photonic circuit switching and PCIe over optics, glued together by a software-defined composable photonic interconnect.
A brief primer on why this matters. Every bit that moves between a GPU, a CPU, a memory pool, or another accelerator in a modern cluster crosses the analog-to-digital boundary many times. Copper traces are analog transmission lines; they attenuate, radiate, and consume power to drive signals. Packet switches are digital machines that optically receive a signal, convert it to electrical form, buffer it, make a routing decision, convert it back to photons, and re-transmit it — the so-called optical-electrical-optical (OEO) penalty. As Lumentum has documented, their R300 MEMS-based optical circuit switch demonstrates a 98% reduction in switching latency versus an equivalent OEO packet switch, and because the fabric is bit-rate transparent, the same hardware carries 400G, 800G, or 1.6T links without replacement.
Corespan’s architecture composes this into a system architecture. Our DynamicXcelerator platform uses PCIe-over-optics, co-packaged optics in our 2500 Series, and vendor-neutral support for optical circuit switches to deliver disaggregated pools of GPUs, CPUs, FPGAs, memory, and storage that any workload can compose on demand across racks, rooms, or whole buildings. Instead of buying a statically configured eight-GPU server and accepting the stranded memory, stranded accelerators, and stranded bandwidth that always result, you compose the exact machine your workload needs, then dissolve it when the workload completes.
This system architecture, in hydraulic-backhoe terms, is a small, maneuverable machine. It lifts a smaller bucket than the GB200 NVL72. It does not out-FLOPS a Blackwell rack. On the metrics the hyperscale frontier-training customer cares about, it is not interesting — yet.
The Early Adopters Will Not Be the Frontier Labs
This is the part of the Christensen analogy that is most often missed. The hydraulic backhoe did not win by convincing Anaconda Copper to stop using cable shovels. It won because a new class of customer, the residential contractor with a tractor and a half-acre lot, had a job that cable shovels literally could not do. The early hydraulic market was not a worse version of the old market. It was a different market.
The early adopters of photonic-native composable infrastructure will follow the same pattern. They are not the three hyperscalers building gigawatt training campuses; those customers want bigger buckets, and they have the capital and the grid allocations to get them for another cycle or two. The early adopters will be:
- Sovereign AI deployments in countries that cannot secure, and do not want to depend on, hyperscale U.S. or Chinese capacity. They need 10–50 MW facilities, not gigawatts, and they need every watt to do useful work, which means eliminating stranded compute is existential, not aspirational.
- Enterprise and regional AI clouds building inference-first infrastructure where workloads are bursty, heterogeneous, and utilization-sensitive. These operators cannot afford the -twenty-fifty percent GPU utilization rates typical of statically provisioned clusters. Composability at the photonic layer turns utilization into a software problem.
- Specialized HPC, scientific computing, and digital-twin operators whose workloads need unusual mixes of memory, accelerators, and interconnect topology that no pre-built reference architecture serves well.
- Neocloud and GPU-as-a-Service providers whose margins depend entirely on squeezing more billable hours out of a fixed silicon pool — exactly the problem a composable fabric is built to solve.
- Edge, telco, and defense operators who need to stretch AI compute over kilometers of fiber rather than meters of copper, and for whom PCIe-over-optics is the only way to federate resources across sites without losing coherent accelerator semantics.
None of these will show up in the earnings call of the early AI winners as a line item. Each of them, individually, looks like a small market. That is exactly what the residential-contractor market looked like in 1950.
Why the Incumbents Cannot Follow
Bucyrus-Erie did not fail because it lacked engineering talent. It failed because its resource-allocation process, tuned to mining and heavy-construction customers, rationally rejected a technology those customers did not want. The same dynamic is now operating across the AI-infrastructure incumbents. The rational move for a company selling $50,000 GPUs by the hundred thousand is to sell more $50,000 GPUs or build a $75,000 GPU. The rational move for a packet-switch vendor is to ship faster packet switches. The rational move for a hyperscaler is to pour concrete for another gigawatt campus.
An architecture that eliminates OEO conversions, lets you pool and re-pool accelerators dynamically, and lets a 20 MW facility do the work a statically provisioned 40 MW facility does today is not a product any of those incumbents can rationally prioritize, because their best customers are not asking for it. Christensen’s law of resource dependence, as he laid it out in the book, predicts precisely this outcome.
That is the opening. It is also the reason the Corespan team is building what we are building, in the sequence we are building it — finding the customers whose jobs cannot be done with a cable shovel, serving them well, and letting the bucket size grow from there.
Hydraulics took twenty-five years to reach the miners. Photonic-native composable infrastructure will move faster, because software eats value-network boundaries in a way that mechanical equipment never could, but the shape of the curve is the same. The frontier labs will be the last customers to adopt it, not the first. And by the time they do, the architecture will already be the default everywhere else.
The cable shovel is still an impressive machine. It is just not the future.