In a pointed critique of Nvidia, Tory Green, CEO of GPU-power aggregator io.net, told Capacity that Nvidia’s strategic reliance on high-end GPU models risks making it the IBM of this generation.
“Its $30,000 GPUs are not the future of AI – they are a luxury solution for a small slice of workloads,” Green said. The issue isn’t just GPU scarcity; it’s a mismatch between workload and compute.”
According to Green, decentralisation means distributing computing workloads across a broader spectrum of GPU resources rather than consolidating them into hyperscale data centres dominated by costly, high-end GPUs.
By aggregating GPU capacity from smaller data centres, enterprises, edge providers, and even individual operators, workloads can be intelligently routed to precisely match their performance requirements, significantly reducing costs and inefficiencies.
He argues that the future of scalable AI computing hinges on intelligently pairing workloads with the right GPU resources. If Nvidia overlooks this shift, Green suggests, it risks ceding market leadership to competitors who embrace decentralised models.
“Over the long term, if decentralisation continues to gain momentum, Nvidia risks becoming the IBM of this cycle - dominant in the early stages but outpaced by more flexible architecture. The upside lies with those who can unlock and route IO computability and not just sell it.”
Green’s reference to Nvidia becoming the “IBM of this cycle” serves as a cautionary analogy. IBM once held unrivalled dominance through mainframes and centralised computing systems but subsequently lost its edge as the computing industry evolved towards more flexible and decentralised architectures.
Green warns that Nvidia, despite its current strength, could similarly lose ground if it fails to adapt to a decentralised AI computing landscape, which increasingly favours distributed, efficient, and affordable hardware solutions.
There has been a clear acknowledgement by Nvidia of a need for change, as the firm has recently joined the growing photonics movement through new networking switches and plans to double down on quantum computing.
CEO Jensen Huang even reversed course on quantum, unveiling plans to launch a dedicated research centre in Boston and setting up a special ‘quantum day’ at GTC, despite previously suggesting earlier this year that the technology’s usefulness was still decades away.
Despite this apparent willingness to adapt, albeit following a significant stock shock triggered by DeepSeek, Green remains unconvinced.
“Ultimately, decentralised computing fixes this by exposing a large range of hardware and matching jobs to the right performance tier,” he added. “This is going to unlock dramatic cost savings, especially for inference-heavy workloads.”
One area of Green's concern that Nvidia likely won't change is its pricing. Analysts previously told Capacity that Blackwell chips are more expensive to produce than previous Nvidia hardware, putting pressure on the company's profit margins.
Add to that the need to get them out the door, as Nvidia had to delay initial shipments due to a now-resolved design flaw.
“The future of AI is not just raw power – it is intelligent, permissionless access to global computing,” Green said. “The bottom line is this: if we want scalable, affordable AI, we can’t run it on these $30,000 GPUs. We have to find cheaper and more efficient alternatives.”
RELATED STORIES
Nvidia GTC 2025: All the updates including Blackwell Ultra & AI factories
Nvidia, xAI join $100bn AI Data Centre Fund Backed by BlackRock, Microsoft
Nvidia unveils AI-powered telco tools to drive 6G & RAN automation
