Nvidia (NASDAQ: NVDA) is shipping artificial intelligence hardware faster than customers can absorb it, yet the chips already delivered remain a quiet and growing competitive threat.
Nvidia’s data center business generated $47.5 billion in fiscal 2024, then surged 142% to $115.2 billion in fiscal 2025, and climbed another 68% to $193.7 billion in fiscal 2026.
Those figures combine to roughly $309 billion in shipments across just two fiscal years, and nearly all of that hardware is still operational inside data centers worldwide.
How long those chips remain in service is not a trivial question, as the answer shapes when Nvidia’s biggest customers will need to buy again.
Microsoft depreciates servers and network equipment over two to six years, while Alphabet generally uses a six-year assumption for the same category of assets.
Meta Platforms (NASDAQ: META) raised its estimated useful life for most servers to 5.5 years at the start of 2025, a change that cut that year’s depreciation expense by roughly $2.9 billion and added $1.00 to earnings per share.
Amazon (NASDAQ: AMZN) moved in the opposite direction, raising its server estimate from five years to six in early 2024, then reversing course a year later for a subset of servers and networking equipment.
Amazon attributed the shorter lives to “the increased pace of technology development, particularly in the area of artificial intelligence and machine learning,” and the reversal added $1.4 billion to its 2025 depreciation and amortization expense.
A machine purchased in 2024 on a five-year depreciation schedule does not retire until 2029 at the earliest, meaning almost everything from the 2024 and 2025 spending waves should still be working in 2028.
Chief financial officer Colette Kress said on the company’s late-August earnings call that Nvidia expects revenue to grow about 70% in fiscal 2028, describing that as a supply-constrained outlook rather than a demand-constrained one.
Kress also noted on a previous earnings call last November that the A100 chips Nvidia shipped six years earlier were “still running at full utilization today,” crediting the durability of its CUDA software ecosystem.
That longevity cuts both ways, as a chip that remains productive represents capacity Nvidia has already been paid for once and that now competes with whatever the company wants to sell next.
Once a server finishes its depreciation schedule, its owner can rent it out at any price that covers electricity and physical space, making older chips cheap competition for inference workloads that do not necessarily require the newest hardware.
A marketplace for used Nvidia chips opened this summer, though Kress confirmed in August that Nvidia’s computing capacity remains fully utilized across every cloud it serves, keeping supply as a bottleneck at least through fiscal 2028.
CEO Jensen Huang noted on the August call that each new generation carries more revenue per gigawatt of data center capacity, with the Hopper platform at roughly $18 billion and the new Vera Rubin platform at roughly $40 billion, and said “customers want to race to the next generation as fast as they can.”
The direction of future useful-life estimate changes from cloud companies will serve as an important signal, with extensions suggesting longer-lasting fleets and shorter estimates pulling replacement demand forward into nearer years.
