An AI buildout gets budgeted as compute, power, and people, with the hardware entered as a capital line that depreciates on a fixed schedule. Accelerators and the servers, power, and cooling equipment around them are being replaced well before that schedule runs out. The cost of retiring them now arrives in the same planning window as the purchase. Through 2050, the hardware coming off data center floors runs to well over a trillion pounds at the top of the range, far past what earlier forecasts allowed for. Equipment pulled that early still carries resale value and recoverable material, and both are now part of the purchase decision.
John Shegerian is Chairman and CEO of ERI, the electronics recovery and IT asset disposition company he co-founded in 2002. It has capacity above a billion pounds of material a year, and was the first company in its industry to reach carbon neutrality at all of its facilities. Shegerian has spent more than two decades on the back end of hardware cycles, through consumer electronics, then solar, and now the equipment coming out of data centers.
"The lifespan for this equipment is being compressed now, on the low side from two and a half up to five years maximum. The technology improves so fast that it actually outraces the obsolescence opportunity," says Shegerian. He has watched this same pattern run through consumer electronics and then through solar, where panels built to last three decades are superseded long before they fail. AI infrastructure is turning over faster than either of those did, across a far larger installed base. Retirement at that pace and scale becomes a line in the capital plan.
Obsolete on arrival
A short refresh cycle changes what a piece of equipment is worth on the day it is bought. Operators buying the newest silicon are the ones who replace it soonest, so they are trading peak performance now for a narrower window to recover value later. Justifying that trade is harder once the subscription and compute bills sit in the same budget as the hardware. Resale and material recovery decide how much of the purchase price comes back.
ERI buys and retires its own equipment on that same cycle. Shegerian needs shredders and testing lines that can handle whatever hardware arrives next, so he replaces them long before they wear out. "By not hanging on to equipment for five to seven years, we're turning it two to three years," he says. He can turn equipment that fast because what he retires still has a buyer waiting for it.
The AI labs and the robotics companies driving the cycle are Shegerian's clients. Their hardware reaches him sooner than their own plans called for. "Even the smartest and the best people who understand how amazing and how fast technology is changing, they can't predict well enough for all of us to be on the same page on the plan," he adds. "It's really moving faster than they ever thought." The answer his customers have landed on is to write the retirement plan at the same time as the purchase order.
Lifecycle in the capital plan
Shegerian's customers now settle how equipment will be retired before they order it through what's known as a lifecycle plan. Hyperscalers and OEMs bring him in while the purchase is still being decided, because resale value, reuse options, secure data destruction, and recovery economics all bear on whether it makes sense. Those numbers get weighed alongside the rest of the AI spending decisions in a buildout.
A lifecycle plan is only as good as the prices behind it, and secondhand values move constantly. Shegerian's larger accounts have stopped asking what their equipment is worth once a quarter, because the compute commitments those numbers feed into run for years. "They want real-time market intelligence now, not quarterly, not QBR type stuff, not annually," he says. "It's really become a real-time industry."
Resale is what holds the calculation together. When equipment an operator has finished with still has a buyer somewhere else in the world, a short service life turns into a recoverable cost, and Shegerian's customers make that case back to him. "Even if it becomes semi obsolete, it's not totally obsolete," he adds. "We mitigate our downside risks for future investments because we want to make future investments with shorter life cycles." Working equipment is one source of that value. The metal inside it is another.
Geology on an AI schedule
The next round of build-out needs more than chips and the power to run them. Rare earth magnets and critical minerals are built into the equipment itself, and most of the world's capacity to process them sits in China. Manufacturers and governments elsewhere depend on that supply.
Opening a new mine is the obvious answer, but also the slowest one. "The soonest you ever get any valuable material out of any new mines being dug is 15 years," Shegerian says. "When you talk to the top geologists in the world, sometimes the numbers go to 20 or 25." Recovery from equipment already in service is the only new supply that can be stood up in months, so the hardware a company retires this year helps supply the hardware it buys next year.
Shegerian handles the waste end of this industry and still talks about the buildout as a net good. He has seen the sequence once already, when the buildout behind the late 1990s internet produced the first e-waste crisis, and this time an industry exists to absorb it. "I believe AI is going to do great things for this planet," he concludes. "And the downsides can be mitigated, the downsides we can work on."
