The Longest Life in Compute

On Monday, Nvidia announced memorandums of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to mobilize more than $500 billion of third-party capital so that hyperscalers, frontier labs, and enterprises can borrow against AI hardware instead of paying cash for it. Jensen Huang told CNBC this is the first time technology chips have become an investable asset class and described the chips as productive, long-lived, fungible, and flexible. This is PROBABLY all correct, but I think it misses something cool.

In the same announcement, Nvidia told bond buyers that CUDA keeps extending the useful life of the hardware and improving its economics over time. EXTENDING THE USEFUL LIFE. Nvidia has just informed the largest capital allocators on earth that its software support policy is the collateral.

The pitch is a hundred and fifty years old

As always (it seems), everything old is new. Productive, long-lived, fungible, transferable across operators (aka STUFF) is the pitch for rolling stock — is decidedly pretty well understood. In the 19th century, equipment trust certificates put locomotives and freight cars into a trust that leased them back to the railroad, which meant that when the railroad went under, the equipment was not railroad property and did not go into the estate. (By the way, if you'd like to read an excellent book on the subject, let me recommend RailRoaded. In an era when American railroads failed constantly, equipment trust paper was among the safest debt you could hold. This is convenient! Because, many many many people think we're in a world where things are going to go bankrupt (soon-ish?), so the fact that the structure Nvidia is using having been stress-tested across a century and a half of bankruptcies is quite nice.

Now, one of the most important parts is that a freight car works whether or not the manufacturer of the freight car is in business. A freight car has residual value only if it can roll onto somebody else's track. AS AN ASIDE, for decades in America, a great deal of it was not interchangeable, because the southern roads ran a five-foot gauge while the north ran what became standard, so freight moving between them had to be transferred by hand at the break. Then on May 31 and June 1 of 1886, work gangs across the South moved one rail three inches inward on roughly 11,500 miles of track in about 36 hours. And while it was a really big financial feat, it was ALSO a big financial one. Overnight, a boxcar sitting in Atlanta became collateral worth something in Chicago.

Back to GPUs, this is not the case today! CUDA is HIGHLY hardware specific, and this is (or was anyway) a real blocker. Compute is fungible only to the degree that the software layer keeps accepting the hardware underneath it.

Nvidia pulls the lever

CUDA 13 removed offline compilation and library support for the Maxwell, Pascal, and Volta architectures (Volta is the V100, which shipped in 2017). That silicon still computes exactly as well as it did the day it was installed, though probably less power efficiently than even a very low end chip. But, because nvcc will no longer generate machine code for it, cuBLAS and cuDNN will no longer ship kernels for it, the architectures are marked feature-complete in the toolkit release notes, and PyTorch dropped them as build targets to match. This is bad, and (ultimately) sets a pretty severe deprecation schedule.

This is also one of the biggest questions in the market - exactly how long do cards still work? What ends the productive life of an accelerator? Until now, the answer was its the morning the framework stops compiling for it, and the company that wrote that framework had a significant incentive to NOT update. But no longer! NOW, Nvidia has every reason to keep the deprecation out as long as possible, because that makes the physical asset worth more, longer.

This is also a HUGE accounting impact. The hyperscalers moved server useful life from three or four years to six, which analysts estimate removed around $18 billion a year in depreciation expense from their income statements. Michael Burry's argument, which has moved from accounting newsletters into the mainstream over the past year, is that carrying GPUs on five and six year schedules while Nvidia ships a new architecture annually understates depreciation by roughly $176 billion across 2026 through 2028. Amazon already cut a subset of its servers and networking gear from six years to five, citing the increased pace of development in AI specifically, and absorbed roughly $700 million of lower operating income for the honesty. Meta went the other direction in the same window. So, with no settled convention to appeal to, everyone is tossing aronud vague ideas. Nvidia isn't stopping, of course, and Hopper landed in 2022, Blackwell in 2024, Vera Rubin hit full production this March, and Rubin Ultra is slated for the back half of 2027. The cadence that creates the problem is not slowing down.

What happens when things invert

For thirty years the vendor incentive ran in exactly one direction. Deprecate the old architecture, make the upgrade compulsory rather than attractive, and book the new generation. The installed base of five-year-old hardware was a support cost with no revenue attached, and every quarter you kept it alive was a quarter somebody didn't buy the replacement. The entire industry is built on that reflex.

However, with Nvidia's new financialization (and half a trillion dollars of paper written against that hardware), the sign changes. Whether or not Nvidia ends up signing an explicit residual guarantee (and the announcement conspicuously does not contain one), it now has a direct commercial stake in five-year-old racks remaining useful. And there are lots more examples. Meta handed its Hyperion joint venture a residual value guarantee covering the first sixteen years of operation. Broadcom agreed to cover 100% of any shortfall to the senior tranches on the $35 billion Apollo and Blackstone structure funding Anthropic's compute. Wall Street had already lent more than $11 billion against GPUs held by neoclouds, and the figure has only gone up since. Every one of those lenders is now exposed to a software decision made in Santa Clara; the only ones who caught it were the Motley Fool, though it framed it as a demand signal rather than a support obligation.

So, at the end of the day, we have something unprecedented: a hardware vendor's software support calendar is, in effect, a financial covenant. Now, the difference between a GPU supported for eight years and one supported for three is the difference between investment grade and junk on the same physical asset. Which is... pretty freaking huge from a software perspective.

Aviation figured this out the hard way. When Fokker collapsed in 1996, they left 1,130 aircraft already in service, with almost no support, and they solved it by building a standalone company created specifically to keep spares and product support flowing. The airframes were airworthy either way butwhat made them financeable was somebody agreeing, on paper, to keep supporting them. Orphan a fleet and the values go regardless of what the metal can still do.

What to actually ask for

If you are signing anything that involves accelerators over the next eighteen months, price per FLOP is no longer the interesting number, and neither is delivery date. Ask how long the toolchain will target this architecture, get a date rather than an adjective, and ask what specifically happens to a cluster you own outright on the day the compiler moves past it. Almost nobody currently has that number, in the same way that almost nobody running a real-time dashboard can tell you its actual latency. The number exists whether or not you know it, and it is your architecture restated in years.

There is a genuinely good outcome here! PARTICULARLY for open source and drivers, which have typically been underappreciated. If financing works and Nvidia does what the paper requires, that's architectures compiling for eight years instead of three. By 2030 that produces an enormous installed base of hardware that is uneconomic for frontier training and ENTIRELY adequate for inference, which is already the majority of AI compute and the part of the workload where the bill actually lands. Depreciated silicon that no lab wants to train on is exactly the silicon you want sitting next to a factory floor, a substation, a hospital basement, or anywhere else the work does not live in a data center. A financing structure designed to keep the buildout going may accidentally fund the distributed compute layer nobody could previously justify on a spreadsheet.

Nvidia spent a decade selling scarcity, resulting in a world where last year's rack becomes a bad bet. It now has to sell durability and a promise about software, made to people who will eventually ask for it in writing. The fact that they are SO incentivized to keep things running, particularly the hardware you already bought, really is a new world.

Want to learn how intelligent data pipelines can reduce your AI costs? Check out Expanso. Or don't. Who am I to tell you what to do.

NOTE: I'm currently writing a book based on what I have seen about the real-world challenges of data preparation for machine learning, focusing on operational, compliance, and cost. I'd love to hear your thoughts!