The Afterlife of a Data Center: Resold GPUs, Mounting E-Waste, and Who Pays for the Cleanup

The equipment inside a data center has a shorter working life than the building around it, and the companies that own it do not agree on how much shorter. Meta extended the assumed life of most of its servers and network equipment to 5.5 years in January 2025, up from a previous range of four to five years, a change its annual report says cut 2025 depreciation expense by $2.92 billion and raised net income by $2.59 billion. Amazon moved part of its fleet the other way: effective January 2025, a subset of its servers and networking equipment went from a six-year life to five, which its annual report attributes to the pace of technology development in artificial intelligence and which added $1.4 billion in depreciation expense. Both figures are accounting estimates, not measurements, and they decide how much of the cost of AI hardware shows up in earnings each year.

The disagreement is sharpest for GPUs. Critics, including investor Michael Burry, argue that AI chips should be depreciated over less than three years, and that their true economic life may be one to three, a view Yardeni Research summarizes alongside the point that Nvidia has moved to a one-year product cycle. Defenders, including Yardeni, say older GPUs keep earning as they shift from training the largest models to less demanding inference and other work, so their useful lives run longer than the critics assume. The answer determines when hardware leaves a data center, and where it goes next.

Part of that answer is a resale market. According to price ranges published in May 2026 by Caladan Semiconductor, a used-GPU dealer, used Nvidia H100 SXM5 cards were trading at roughly $18,000 to $24,000 against $25,000 to $32,000 new, or about 72 to 75 percent of the new price. Used 80GB A100s fetched about 55 to 65 percent of new, down from more than 80 percent in 2022, according to the same source. Caladan reported that most used H100s were coming from failed startups and distressed sales rather than enterprise upgrades, while A100 supply was growing as companies that bought clusters in 2021 and 2022 moved to newer chips. The figures come from a dealer, and the article does not say whether they are asking or completed-sale prices, so they are a rough guide and not audited market data. Companies such as Reconext, a remarketer, now sell processes that wipe drives and test GPUs in place so racks can be resold without being pulled apart.

What does not get resold becomes electronic waste, a category already growing faster than the world's ability to handle it. The UN's Global E-waste Monitor 2024 counted a record 62 million tonnes generated in 2022, up 82 percent since 2010 and on track for about 82 million tonnes by 2030. Only 22.3 percent was documented as properly collected and recycled, and the report found generation rising five times faster than documented recycling. Researchers writing in Nature Computational Science in 2024 estimated that servers built for generative AI alone could add up to 5 million tons of e-waste cumulatively through 2030, and as much as 2.5 million tons a year by 2030 if nothing is done to reduce it, about 3 percent of the UN's projected 2030 global total. They also estimated that reusing and remanufacturing components, extending equipment life, and better recycling could cut that output by up to 86 percent.

The building itself is a separate problem. Ashwin Murthy of Columbia Law School's Sabin Center for Climate Change Law wrote in August that older abandoned data centers often sat vacant for long stretches, kept up by a skeleton crew while owners looked for buyers, and that modern hyperscale facilities are harder to repurpose because they can span millions of square feet and are specialized by design. Idle sites, he warned, risk becoming abandoned warehouses that waste land and leave hazardous material behind. There is also a cost that can fall on others: because data centers require heavy grid upgrades, he wrote, ratepayers may end up paying for infrastructure built for a company that fails and leaves a facility idle.

Some local governments are writing cleanup rules in advance. Murthy points to Susquehanna County, Pennsylvania, which requires decommissioning within twelve months after a data center's useful life ends, including removal of all hazardous materials and contents, backed by bonding, and notes that Smithfield Township has a draft ordinance. He argues that states could go further by requiring detailed decommissioning plans before permitting that bind any future owner, and bonds or similar financial assurances that cover the full cost, modeled on oil and gas bonding rules. Whether a locality can impose such rules varies by state, and he notes that rules aimed only at data centers could face equal-protection challenges, though he expects those to fail.

For now, how long a data center lasts depends on assumptions in annual reports, a resale market priced by dealers, and a patchwork of local ordinances. Who ends up paying when facilities are retired, whether shareholders, customers, or the town that hosted them, will depend on which of those holds up.