Fink Flags Compute Shortage That Decentralized Networks Are Racing to Fill

Larry Fink, CEO of BlackRock, seemed caught up in the first telling on 15 Jul, talking to CNBC's Squawk on the Street following BlackRock's quarterly earnings. "The demand for compute is not slowing down. It's growing faster," he said. "The problem we have as a country, we're not investing fast enough." His claim comes after the four biggest hyperscalers announced plans to spend $725bn combined in 2026 on building their data centre capacity, up 77% on the $410bn spent in 2025.

Meanwhile, research from IBM found average server utilization is around 12%-18% of capacity. A report published by CastAI in April that studied data from AWS, GCP and Azure found that although demand for compute is rising, average utilization of existing GPUs has dropped to 5% as AI teams request more computing power than they need, to allow for fluctuations. CIO Insights estimated the underutilization could amount to $30bn in unproductive capital.

While Fink and the hyperscalers race to build out capacity and outpace demand, the Web3 ecosystem is aggregating computing power from dormant GPUs to solve the problem on their own.

Serving small teams

What that scarcity does to price seemed to trouble Fink more than the build out of infrastructure. He noted that the price of computing power, while already expensive, could be set to increase.

Heightened costs could limit access for small to medium businesses, and with it, their ability to compete. The question he said he asks hyperscalers, he said, is, "how quickly can we bring down the cost of compute?"

Those most affected are startups and students, who have fewer resources to pay for rising prices. Greg Osuri, founder of Akash network, a marketplace for decentralized computing, saw the effect on demographics when cloud providers ran short of GPUs after ChatGPT arrived in 2023. "If you don't work for a big AI lab, you wouldn't get access to compute," he told Sandmark.

The inflation has since spread through the hardware stack, with Osuri citing memory prices up four times and hard disks doubling. "Anything that AI touches [is] just going up because of the demand."

Research from the Ethereum Foundation published in February 2026 found that centralization inevitably reaches bottlenecks in computing power, due to its reliance on a single entity. To Osuri, the "only way really" to solve the mismatch of supply and demand is through a distributed network linking idle computers and chips that aren't being used to their full capacity.

"There's a lot of dormant compute out there," said Osuri.

Idle power

Distributed compute networks such as Akash capitalize on the computing power that exists in idle computers and GPUs that sit in people's homes and institutions, which researchers found could collectively offer "compute capabilities comparable to the cloud."

The Ethereum Foundation's February research paper found blockchain to be well suited to aggregating computing power from disparate sources, both incentivizing sharing through rewards and making aggregation more secure by recording workloads and usage.

Akash, which published its white paper in 2017, runs an open marketplace matching buyers with owners of idle hardware sitting anywhere from office buildings to spare bedrooms. Users in need of computing power post jobs onchain, and the amount of capacity they need. Providers then use competitive pricing in an open auction to win the job, with the resulting lease settled onchain. Akash workloads typically run 60% to 85% below equivalent hyperscaler rates, according to the company website.

As both the orders and leases live onchain, the system is also permissionless, which Osuri says draws "a wide variety of participants" that traditional clouds cannot reach. "It's a free market for compute, and that's a core thesis of decentralized compute," he said.

Blockchain coordination 

Distributed networks suit some AI workloads better than others, however. Training has historically demanded enormous amounts of identical chips packed into a single location. The initial stages of AI model creation require vast amounts of data to be swapped between chips to compare results through each step of training. If one chip lags behind, the process is delayed meaning it's more efficient to have a block of the same type of chips in the same place.

Inference, the processes that AI uses to run daily operations, can work comfortably on smaller, scattered hardware. As AI agent usage grows, the permissionless nature of blockchain-based compute networks could become an important advantage over centralized solutions, allowing agents to bid on compute to power their inference autonomously.

As the agentic economy grows, a permissionless decentralized compute network could become even more useful. Adam Reeds, co-founder and CEO of crypto lender Ledn, told Sandmark he expects agents to utilize the networks to search and bid for the best compute prices. "I think we'll start routing AI compute programmatically, and that's an interesting use case for crypto, to settle that really fast," he said.

A UCL Institute of Finance and Technology paper published 30 Jun found that blockchain's programmable payments and auditable records "can reduce coordination frictions when agents transact with unknown or semi-trusted counterparties," letting software buy services "without opening accounts, negotiating contracts or routing every transaction through a human checkout process."

Scaling problems

While the potential for decentralized compute is theoretically there, researchers warn that networks could still run into scaling problems. They found that blockchains underneath these networks strain as transaction volumes grow, and pseudonymous payouts could create money laundering and tax exposure that sits awkwardly beside the enterprise customers the sector is courting.

Osuri acknowledged that the crypto branding itself still repels enterprise buyers. He saw the biggest risk to decentralized computers not fulfilling their potential at an institutional scale due to a "lack of enterprise bridge," where the decentralized supply could be packaged into large enough contracts to meet the needs of Fortune 500s.

He also claimed that the moves by large players, including BlackRock and OpenAI, to consolidate small data centres, could limit the amount of computing power available for decentralized networks.

The two worlds may yet meet in the middle. Fink expects compute to trade like any other commodity, telling CNBC, "We're going to have a futures market. This is going to be the next revolution in finance." Crypto's compute marketplaces have been running a version of that market for years, and for Osuri, the task now is proving they can carry serious volume. "It's crypto's growing up time now," he said. "You got to show fundamentals."