How Do You Price the GPU Boom?

High demand, tight supply, and abundant capital have made GPUs one of the most sought-after resources in the AI economy, making them the “new oil” after data and rare earth magnets.

But if rare earth magnets and their constituent elements can have a price index, why not compute?

Price indices help consumers, businesses, and investors compare performance and measure shifts in purchasing power, which is especially necessary for GPUs given their volatile pricing.

While some AI benchmarks measure how fast GPUs run or how efficiently they train models, Silicon Data is laser-focused on the financials—what the market actually pays for compute.

The US company publishes daily GPU pricing indices for different GPU variants based on rental and transaction data from cloud providers, colocation facilities, brokered cluster sales, and private rental platforms.

Building a Price Index for Compute

The index currently maps the hourly rental price for a GPU across a broad range of timelines. 

“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.

Prices are normalised for hardware configuration, rental terms, performance characteristics, and geography.

The index is available on platforms including Bloomberg and Refinitiv, placing GPU pricing alongside other financial reference data and making it accessible to buyers, operators, and financial institutions.

Li pointed to oil markets as an example of how a globally traded resource moved away from direct supply deals toward financial benchmarks and futures to manage price volatility. Instead of negotiating long-term contracts with producers, buyers hedge exposure through standardised futures that track market prices.

For now, Silicon Data’s GPU Price Index serves as a reference layer, allowing banks and trading firms to build swaps or forwards on GPU pricing. “Banks probably know less about [NVIDIA] B200 than you [AI companies/developers],” explained Li. 

They often rely on their debtors (AI firms) to understand how much money these GPUs will help them earn. 

“They now have this data from Bloomberg to tell them the B200 trading rate. They can put that in their database to figure out the risk profile for financing your product,” he added.

Silicon Data began with indices for widely deployed processors—NVIDIA’s A100 and H100—which form the backbone of much of today’s AI infrastructure.

This week, the company launched what it says is the world’s first rental index for NVIDIA’s B200, the next-generation platform expected to anchor frontier-scale AI clusters.

Alongside the B200 launch, Silicon Data updated its A100 and H100 on-demand rental indices, expanding provider coverage across hyperscalers, neoclouds, regional data centres, and specialised platforms. 

“As A100 and H100 markets mature, we’re seeing stabilisation curves and residual-value patterns that differ from short-term pricing narratives,” Li noted. “These updates ensure our indices continue to reflect how the market actually behaves.”

This focus on price sets Silicon Data apart from most benchmarks in the AI ecosystem.“In parallel, we incorporate secondary and refurbished market data—including resale prices, rental yields, and liquidity signals—to reflect how GPUs are actually priced and traded in the market,” he added. 

Pricing Meets Credit Risk

That distinction ties directly to the AI bubble debate. 

As GPU-backed lending rises, lenders now need to understand how GPUs behave in the real world—from depreciation and lifecycles to how new generations reprice the fleet. 

Neocloud operators such as CoreWeave and Lambda—built almost entirely around renting high-performance GPUs—are increasingly relying on multi-billion-dollar loans backed directly by GPUs, often raised through special-purpose vehicles.

However, lenders also need to consider GPU depreciation, which, according to brokerage Bernstein, lose more value in the first year, often due to high-performance workloads and burn-in losses.

CoreWeave depreciates GPUs over six years, Nebius over four, while engineers and project-finance specialists peg the actual economic life at three to four years. Analysts at Cerno Capital estimate that if Microsoft, Alphabet, Meta, and Amazon all extended the lives of data centre assets to six years, reported depreciation would drop by 54%, from $51 billion to $28 billion. 

To empirically determine the decline in asset value, Silicon Data earlier this year launched SiliconMark, which measures the actual performance, stability, and degradation of GPUs over time rather than relying solely on age or spec sheets. “This allows lenders to distinguish between nominal depreciation and functional depreciation,” said Li. 

What determines whether GPUs stay in service is how quickly the economics shift.

Jordan Nanos, an analyst at the research firm SemiAnalysis, told AIM that hardware would stay in service as long as it makes economic sense to rent. In hyperscale environments, once power, cooling, and maintenance are accounted for, the marginal cost of running a modern GPU often bottoms out around $0.30 to $0.40 per hour. As long as a chip can earn more than that, operators keep it in the fleet.

“If new GPUs are so much more performant than the old ones that it no longer makes economic sense for buyers to rent the old ones, they get replaced,” said Nanos. “But the converse is also true: if performance improvements are not realised for new GPUs, the market will demand the old ones for longer.”

In practice, hardware’s existence is tied to its value, and not mere physical functionality. It is decommissioned when contracts end, and no one is willing to pay enough to keep it operational.

“If that contract expires, and there is no one interested in renting the GPUs at the current price, a cloud provider will decommission the GPU systems and replace them with something else,” he said.

That leaves lifecycle risk outside operators’ direct control. Each new generation can either strand fleets early or extend their lives, depending on how sharply it resets performance per dollar and where market pricing settles.

The post How Do You Price the GPU Boom? appeared first on Analytics India Magazine.

Generated by RSStT. The copyright belongs to the original author.

Source