Europe cannot GPU its way to an AI advantage

Europe's AI infrastructure prospects cannot be measured in GPUs alone, according to Ken Claffey, CEO of high-performance data platform provider Vdura, who said at the Xcelerated Compute Show that the region has "built nowhere near the same level of infrastructure" as the U.S. but can leverage its high-performance computing (HPC) workforce, available power, and industrial sites — including a former automotive manufacturing site in Germany with gigawatts of power being converted into a data center. FarmGPU CEO Jonmichael Hands said networking is the hardest part of building AI clusters, with 800-gigabit fabrics accounting for between 20% and 30% of a cluster's cost, and warned that a three-month delay can put an operator out of business. The shift from training to inference is driving competing architectures and new storage demands for key-value caches used by long-context and agentic applications.

LONDON – Europe may lack the hyperscalers and 100,000-GPU clusters found in the U.S., but its AI infrastructure prospects cannot be measured in accelerators alone. Its expertise in high-performance computing HPC could provide an advantage as AI systems become more complex. Turning that expertise into commercial infrastructure, however, will require Europe to combine processors, storage, networking, cooling, and software into reliable systems that can adapt as workloads change. “We have built nowhere near the same level of infrastructure,” Ken Claffey, CEO of high-performance data platform provider Vdura, said during the Xcelerated Compute Show. The U.S. has hyperscalers and a growing neocloud sector, while Europe is pursuing a different mixture of neoclouds, AI factories, and public-private gigafactories. Claffey nevertheless argued that the region has no choice but to compete. Europe’s advantages include an experienced HPC workforce, available power in some markets, and industrial infrastructure that could be redeployed. Claffey cited a former automotive manufacturing site in Germany with gigawatts of power that is being converted into a data center. The difficulty is doing it quickly enough. “Moving fast in Europe, for lots of obvious reasons, is really difficult,” Claffey said. “It is one of the things that I’ve been blown away by in the U.S., how fast things are happening.” A GPU does not a cluster make Europe’s HPC heritage matters because building AI capacity involves far more than acquiring GPUs. “The hardest part about building AI clusters is networking,” said Jonmichael Hands, CEO of neocloud provider FarmGPU. High-end AI clusters increasingly use 800-gigabit network fabrics, while networking can account for between 20% and 30% of a cluster’s cost, according to Hands. The physical infrastructure is also less forgiving than conventional data center operators may expect. “One speck of dust could mean a couple of decibels of loss on these links and optics,” he said. “It’s just completely different how you design, architect, and build these high-performance AI networks.” If networking problems delay a cluster by several months, its expensive GPUs remain idle while the operator continues carrying the financing costs. “You can’t be off by three months,” Hands said. “You’ll be out of business.” These requirements should be familiar to HPC specialists accustomed to balancing processor, storage, and network bandwidth across large systems. Claffey said he routinely encounters people he has known from HPC for decades now building infrastructure for neoclouds. But commercial AI changes the consequences of failure. Academic HPC teams might receive an angry email when a system goes down. Operators running infrastructure costing billions of dollars face demanding service-level agreements and potentially millions of dollars in customer credits. “In HPC, you were tasked with building something that went fast but may not be super reliable,” Claffey said. “We’d all build the Formula One race car. It would go down. No one gets fired.” In other words, commercial AI must provide Formula One performance without making the same reliability trade-off. Do not optimize for one moment Infrastructure design is becoming more difficult as attention shifts from training toward inference. Hands said this has prompted competing architectural approaches, including disaggregated inference using specialized engines and tightly integrated Nvidia systems designed to scale up. Storage requirements are changing too. Training and checkpointing already require the high-speed parallel file systems familiar from HPC, while inference creates new demand for storing key-value caches used by long-context and agentic applications. The danger is assuming that today’s dominant workload will remain dominant throughout the infrastructure’s economic life. “A year or so ago, it was training, training, training. Now it’s inference, inference, inference. What will it be next year? Who knows?” Claffey said. AI infrastructure is commonly financed and depreciated over five or six years. Claffey advised operators against committing all their investment to infrastructure specialized for inference when workloads, models, and hardware requirements are changing so quickly. “I’m not a gambler,” he said. “I would rather have an infrastructure that can pivot and be flexible and adaptable.” Enterprise demand is also likely to remain hybrid. Claffey pointed to oil and gas companies that continue to run traditional simulation and modeling across large CPU estates while adding GPUs for AI-native and AI-augmented applications. New hardware does not necessarily make previous generations worthless either. GPUs can move from training and fine-tuning into inference as they age, extending their economically useful life. Government should finance, not design Europe’s higher cost of capital presents another hurdle. Building AI infrastructure requires coordination between an operator, an anchor customer or off-taker committed to buying capacity, a GPU supplier, and lenders willing to finance equipment whose future value remains uncertain. Claffey said European public-private partnerships could help by providing guaranteed demand or financial backstops, reducing lender risk and potentially lowering borrowing costs. The involvement should stop short of government specifying the architecture or choosing suppliers, he warned. Traditional public HPC procurement can involve research proposals followed by requests for information and proposals, creating a cycle lasting 18 months to two years. That timetable is poorly suited to an AI infrastructure market changing from month to month. “Government should not be involved in what’s built, how it gets built, or what vendors get selected,” Claffey said. “Where they should be is building the ecosystem and some of the financial mechanisms that enable European and U.K. industry to move fast.” Europe should also choose which parts of the stack it can realistically control. Claffey questioned the value of trying to eliminate every supply-chain dependency by developing European silicon. Instead, he advocated open, multivendor systems supported by a diverse range of providers, with Europe concentrating on the software and data layers. That includes adapting and fine-tuning frontier and open-weight models for industries in which European organizations already possess specialist knowledge. Europe’s AI advantage, if it develops one, will not come from collecting the largest possible pile of GPUs. It will come from applying decades of advanced-computing expertise to systems that remain reliable, commercially viable, and useful after the infrastructure fashion has changed again.