Venture Capital Backing GPUs is Fast Becoming an Outdated Concept

As the demand for AI compute accelerates, companies are raising billions of dollars to beef up data centres and procure GPUs. However, the mechanisms used to finance those assets still resemble the venture capital model that funded software startups.

Most investors today gain exposure to AI through equity investments in companies building models, applications, or infrastructure.

However, GPUs are capital-intensive, with NVIDIA H100 chips costing between $25,000–$40,000, according to some estimates. They also face rapid technological obsolescence, requiring financial structures to account for high depreciation rates. 

Uniquely, US-based Compute Labs is building a system that allows investors to buy and earn income from GPU hardware directly, rather than investing in AI companies.

“There are a lot of investors who are trying to get into this space. They want to get into the growth of AI and the market,” Co-founder and Chief Business Officer Nick Filichkin says in an interaction with AIM. 

However, the equity route often requires betting on valuations and company performance rather than the underlying demand for compute.

“People can invest in AI companies directly—which is fine—but we don’t know if those companies are overvalued,” he remarks, adding that most AI companies are private and purchasing assets in the secondary market might not be straightforward.

In a whitepaper co-authored by Filichkin published last year, he argues that GPUs should be treated as infrastructure assets rather than venture-backed startup expenditures. He contends that when GPUs are deployed in clusters inside data centres and leased to AI developers, they generate predictable revenue streams based on compute usage, similar to other income-producing infrastructure such as energy projects or equipment leasing. 

Because GPUs are tangible hardware with resale value and multi-year useful lives, they can serve as collateral in structured financing arrangements. In this framing, venture capital—designed for high-risk equity bets seeking outsized returns—is poorly suited to fund GPU infrastructure. 

Instead, he suggests that asset-backed financing models, where investors fund hardware directly and earn returns from leasing compute capacity, are a more appropriate way to finance the growing demand for AI compute.

GPUs as Collateral

Compute Labs proposes a model where investors own the hardware itself. The startup purchases GPUs using investor capital and deploys them into data centres operated by cloud providers that need additional compute capacity. Revenue generated from the use of that hardware flows back to the investors who funded it.

“We are giving access to investors who want to have infrastructure as opposed to investing in just another AI company,” he says.

In this structure, Compute Labs acts as the operator and asset manager. Investors provide the capital but do not manage procurement, installation, or hardware deployment. It then deploys it with neoclouds.

Neoclouds, such as CoreWeave, Lambda Labs, Nebius, Together AI, and Crusoe Energy Systems, are smaller cloud providers focused on AI workloads. These firms rent GPU capacity to developers and enterprises building machine learning systems, but often lack the capital to purchase large hardware clusters outright.

Compute Labs deploys investor-owned GPUs into these environments and leases them to those operators.

“We allow these companies to use that equipment so that they can then service their workloads, their customers, and then scale and grow their businesses without having to incur any of the heavy capital expenditure,” he adds.

The model assumes that GPU infrastructure behaves more like an income-producing asset than a venture investment. “What we do is actually raise capital from the private market and allow those people to access the cash flows from it.”

Factoring High Depreciation

However, GPU hardware depreciates quickly because each new generation of accelerators delivers major improvements in performance and efficiency. When new chips arrive, older hardware can lose value rapidly, reducing both resale prices and the value of assets used as collateral in financing structures.

For companies whose business depends on renting GPU compute, that necessitates a delicate balance. Revenues must remain strong enough to service the financing used to acquire the hardware while the hardware itself steadily loses value over time. Depreciation is therefore a central concern for any investment model based on GPU ownership. 

There has also long been a debate about GPU depreciation schedules, particularly whether hyperscalers are justified in spreading the cost of multi-billion-dollar GPU fleets over five or more years when new architectures from NVIDIA—such as Hopper, Blackwell, and its successors—arrive roughly every 12–24 months.

Filichkin says depreciation is built directly into Compute Labs’ financial modelling.

“What we’re projecting is about a 15–20% depreciation annually,” he explains, adding that the hardware is treated as fully depreciated after five years, even if it may still have operational or resale value.

The company distinguishes between accounting depreciation and the hardware's market value. Even after the asset is written off on financial statements, the founders believe that used GPUs retain economic value because they continue to run many AI workloads.

“On the residual value, when you actually sell the asset, we are considering that you will be able to retain about 30–45% [value],” he adds.

Compute Labs argues that demand for older GPUs will remain strong because many workloads, such as inference and training smaller models, do not require the latest chips. 

Albert Zhang, the company’s other co-founder, says this demand is already visible among infrastructure partners. “There is always going to be a demand for the older version of GPUs. In fact, we’re talking to many partners who prefer to do their business only on used GPUs because the ROI there will be higher.”

The company is also working on a system designed to track the operational history of GPUs deployed through its platform. Zhang describes the concept as a data record for hardware utilisation.

The system would track metrics such as utilisation levels and workload history, enabling buyers in secondary markets to evaluate used hardware more precisely. Zhang says this type of data will improve liquidity in secondary GPU markets and make hardware-backed investments easier to price.

The founders also point to regional differences in compute supply as a factor that could extend the life of older GPUs beyond typical accounting assumptions. In markets where access to high-end compute is limited, earlier generations of hardware may remain economically viable for longer periods.

Treating Compute as Commodity

Recently, Don Wilson, Founder and CEO of trading firm DRW, said on Bloomberg’s Odd Lots podcast that the rapid growth of AI workloads is pushing GPUs toward becoming a new kind of commodity market, similar to energy or raw materials. He argued that if global spending on AI compute continues to rise, the scale of demand could eventually rival some of the world’s largest commodity markets. 

“Within the next 10 years, the world will spend more per year on GPUs than it does on crude oil,” Wilson proclaimed.

He also described how financial markets could rally around GPU infrastructure, allowing operators and buyers of compute to hedge price risk through derivatives tied to GPU pricing indices. “A neocloud could buy the GPUs, sell a strip of futures contracts, and I envision that these will be traded kind of like electricity futures,” he said.

Meanwhile, companies like Silicon Data are creating benchmarks for compute pricing. It publishes daily pricing indices tracking the rental rates of GPUs such as NVIDIA’s A100 and H100 across cloud providers, data centres, and private compute markets. This effectively creates a reference layer for compute pricing. 

“Think about S&P 500, think about crude oil futures. We provide GPU future indices,” Carmen Li, founder and CEO of Silicon Data, told AIM in an earlier interaction. 

As GPU-backed lending and infrastructure financing expand, Li said financial institutions increasingly need visibility into how GPU assets are priced and depreciate over time. “Banks need to understand GPU depreciation and repricing, and that's where indexes come in.”

The AI boom may ultimately force a shift in financial architecture, from speculative equity financing toward markets built around pricing, leasing, and hedging compute capacity.