When the Machines Improve Themselves, Who's Holding the Hardware?
Recursive AI improvement changes the math on GPU depreciation, and makes the case for compute derivatives urgent
Dean W. Ball published an excellent piece this week on the automation of AI research, arguing that frontier labs will scale their effective workforces from thousands to hundreds of thousands within a year or two, not by hiring, but by deploying armies of automated AI researchers whose sole objective is to make themselves smarter. He calls this the most important development in AI over the coming year, and he’s probably right.
But Ball’s piece is focused on capabilities and policy. I want to talk about what happens to the physical infrastructure underneath all of this, specifically, the GPU when the software running on them starts improving itself at an accelerating rate.
The short version: recursive self-improvement, if it unfolds anything like what Ball describes, fundamentally changes the economics of GPU ownership. It compresses the useful economic life of chips, destabilizes residual value assumptions baked into billions of dollars of financing, and makes the absence of a functioning compute derivatives market look less like a market gap and more like a structural vulnerability.
The Depreciation Problem
A GPU’s physical life and its economic life are two very different things. Physically, a data center GPU can run for five to seven years before failure rates become problematic. That’s the number hyperscalers and neoclouds use for accounting purposes. It’s the depreciation schedule that shows up on balance sheets, the assumption embedded in asset-backed securities, and the basis for debt covenants.
But economic life, which is the period during which a GPU generates cash flows that justify its operating costs, is a function of what the software demands from the hardware. And that’s where Ball’s thesis becomes a balance sheet problem.
Today, Nvidia releases new GPU architectures on roughly a 12-to-18-month cadence. When Blackwell ships, H100s move down the stack, from frontier training to fine-tuning, from fine-tuning to inference, from premium inference to cost-sensitive batch jobs. This is the “value cascade” that bulls cite when defending five-year depreciation schedules. The chip doesn’t become worthless; it just migrates to lower-margin work.
This cascade has held up reasonably well so far because algorithmic progress, while fast, has moved at a pace roughly synchronized with hardware generations. Dario Amodei has cited approximately 400% annual algorithmic efficiency gains, meaning the same compute delivers four times the capability year over year, even without new chips. That’s rapid, but it’s been predictably rapid. Hardware vendors, cloud operators, and lenders have been able to plan around it.
Now imagine Ball’s world comes to pass. Labs deploy tens of thousands of automated researchers, each grinding through the low-hanging fruit of algorithmic efficiency: architecture tweaks, training optimizations, data pipeline improvements, inference tricks. Ball himself sketches a range: maybe the automated workforce doubles the efficiency gains to 800% annually. Maybe it’s 4,000%. Even the conservative end of this range breaks the synchronization between hardware cadence and software capability that makes today’s depreciation schedules defensible.
Why This Matters for Financing
This isn’t an abstract accounting question. It’s a structural risk embedded in hundreds of billions of dollars of committed capital.
Consider the neocloud model. Companies like CoreWeave have financed massive GPU fleets through asset-backed lending structures, borrowing against the future cash flows of their hardware. These structures assume useful lives of five to six years and residual values that hold up in secondary markets. The debt covenants, interest rates, and recovery assumptions all flow from that baseline.
If recursive improvement compresses a GPU’s economic frontier relevance from, say, 18 months to 9 months, the cascade still happens, but it happens faster and steeper. A chip that was doing premium inference work in month 18 under the old regime is now doing discount batch processing. The revenue per GPU-hour degrades more rapidly. Secondary market prices fall more sharply when new generations arrive, because the performance delta between generations widens.
The result is a duration mismatch: long-dated liabilities backed by assets whose economic life is shrinking in unpredictable ways. This is the classic setup for financial stress. (Duration mismatch killed WeWork, for example.) This is not necessarily a crisis, as the underlying demand for compute remains enormous, but it is a repricing of risk that the current financing ecosystem isn’t equipped to handle gracefully.
Hyperscalers like Microsoft and Google are partially insulated because they can repurpose hardware internally and subsidize transitions with other revenue streams. But pure-play infrastructure companies, GPU lessors, and anyone holding concentrated hardware positions face a different risk profile entirely. And the investors behind them—the pension funds, sovereign wealth funds, and credit funds buying into GPU-backed ABS—are underwriting a depreciation curve they may not fully understand.
The Circular Financing Wrinkle
There’s an additional layer that makes this worse: the vendor financing loop. Nvidia invests in companies (directly or through associated funds) that then turn around and buy Nvidia GPUs. Those customers finance the purchases with debt backed by the GPUs themselves. The revenue Nvidia books from these sales supports its valuation, which supports its ability to invest in the next round of customers.
In a world where algorithmic progress accelerates sharply, this loop becomes more fragile. Faster obsolescence means the collateral backing the debt degrades faster. If enough GPU lessees face writedowns or covenant issues simultaneously, the knock-on effects ripple back through the chain. We’ve seen this movie before. It was called the telecom equipment bubble of the early 2000s, where vendor financing created a similar circular dependency between equipment makers and their overleveraged customers.
Recursive self-improvement doesn’t cause this problem, but it accelerates the timeline on which it could become acute.
Enter Derivatives
Here’s the thing: every risk I’ve just described is, in principle, manageable. We know how to handle uncertainty about future asset values, volatile spot prices, and duration mismatches. We do it in oil, natural gas, electricity, interest rates, and dozens of other commodity and financial markets. The tool is derivatives—forwards, futures, options, and swaps—and they don’t exist yet for compute.
The absence of a compute derivatives market means that every participant in the GPU economy is forced to hold unhedged exposure to exactly the risks that recursive self-improvement amplifies.
A neocloud financing a $2 billion GPU cluster can’t lock in a forward price for the compute revenue that cluster will generate in 18 months. A lender underwriting GPU-backed debt can’t buy a put option on residual values. An enterprise committing to a three-year cloud contract can’t swap floating compute costs for fixed ones. Everyone is simply hoping that the depreciation curves they underwrote don’t shift beneath them.
Here’s what a functioning market could look like in practice:
Forward contracts on compute pricing. A GPU lessor could sell forward contracts on $/GPU-hour for delivery 6, 12, or 18 months out. This locks in revenue and gives the lessor certainty to service debt, while the buyer—an AI lab or enterprise—locks in costs. The spread between spot and forward prices would itself become an information signal about the market’s view on obsolescence risk and demand trajectory.
Options on residual values. A lender financing a GPU fleet could purchase put options that pay out if secondary market prices for, say, H100s fall below a strike price at a specified date. This is essentially insurance against the scenario where algorithmic breakthroughs render a generation of chips less valuable faster than expected. The premium on these options would be a direct market-priced measure of obsolescence risk—something that doesn’t exist today.
Fixed-for-floating compute swaps. An enterprise locked into a long-term cloud contract at variable pricing could enter a swap to pay a fixed compute rate, transferring the volatility risk to a counterparty willing to bear it (perhaps a market maker or a fund with a view on pricing trends). Conversely, a neocloud with fixed-cost hardware could swap into floating exposure to capture upside if demand spikes.
Basis swaps across GPU generations. A holder of older-generation hardware could manage the spread risk between, say, H100 and Blackwell compute pricing. If recursive improvement makes newer chips disproportionately more valuable, this spread widens, and a basis swap lets the holder hedge against that specific risk without liquidating the physical position.
None of these instruments require inventing new financial theory. They’re standard structures adapted to a new underlying asset. What’s needed is price discovery: transparent, real-time pricing data for compute across chip generations, geographies, and contract tenors. Companies like Ornn and Silicon Data are beginning to provide some of this with spot pricing APIs, but the infrastructure for a full derivatives market—clearinghouses, standardized contracts, regulatory frameworks—remains nascent.
The Urgency
Dean Ball’s piece ends by noting that 2026 is likely to be “a more rapid year of AI progress than all years that have come before” and calls this the “conservative forecast.” I agree. But notice what that implies: the risks I’ve outlined aren’t hypothetical tail scenarios. They’re baseline expectations.
If algorithmic efficiency gains accelerate, even modestly, from 400% to 800% annually, the depreciation curves underwriting tens of billions in GPU financing are wrong. Not slightly wrong. Structurally wrong. And every month that passes without price discovery mechanisms and hedging instruments is a month where risk accumulates unpriced.
The irony is that the AI industry is building some of the most sophisticated predictive systems in human history, yet the financial infrastructure surrounding it remains pre-modern. We’re financing a rapidly depreciating, algorithmically volatile asset class with static five-year schedules and a prayer. The telecom bubble had the excuse of being a first-mover into infrastructure-heavy tech financing. We don’t have that excuse. We’ve seen this pattern before, and we have the financial tools to manage it.
What we lack is the market.
The case for compute derivatives was already strong on supply-and-demand grounds alone. Recursive self-improvement makes it existential. When the machines start improving themselves, the value of the hardware they run on becomes a moving target, and the only responsible way to operate in that world is to give market participants the instruments to price and hedge that movement.
Someone’s going to build this market. The question is whether it happens before or after the first big writedown.
If you enjoy this newsletter, consider sharing it with a colleague.
I’m always happy to receive comments, questions, and pushback. If you want to connect with me directly, you can:
Brilliant piece on the duration mismatch problem. The telecom bubble analogy is spot on but what really sticks is that we're literally building pricing infrastructure for eveyrthing except the actual compute underneath. Worked in structured credit and the idea of ABS backed by assets with accelerating obsolesence curves makes me deeply uncomfortable, espeically when the triggr is baked into the system itself.