AI, Crypto Mining Expose Global Compute Infrastructure Constraints

Europe is "entirely supply constrained" on digital compute infrastructure, General Tensor chief investment officer Victor Teixeira said at the 12th European Blockchain Convention in Barcelona, citing semiconductor fab lead times of five to seven years and foundries allocating capacity two or three years in advance. Teixeira said AI workloads and Bitcoin mining are driving a hardware split, with GPUs handling both AI and mining while specialized ASICs — including new AI inference and training chips emerging from China — take over high-volume operations. He attributed Europe's vulnerability to de-industrialization and the shutdown of domestic coal-fired power stations, compounded by U.S. industrial policy reorienting supply chains and China addressing its own manufacturing bottlenecks.

BARCELONA, Spain — At the 12th European Blockchain Convention https://eblockchainconvention.com/european-blockchain-convention-12/ , institutional investors and technology executives discussed the growing infrastructure needs of digital assets. Against that backdrop, AI workloads, specialized mining hardware, semiconductor constraints, and energy availability are putting pressure on global compute infrastructure. In an interview with EE Times, Victor Teixeira, chief investment officer at General Tensor https://www.generaltensor.io/ , explained how semiconductor supply and energy grid limitations affect digital compute, with Europe at the center of the squeeze. Geopolitical pressures and European supply limits Teixeira, whose background includes mergers and acquisitions in the energy sector across Europe and Australia, pointed to historical policy decisions as a root cause of Europe’s current vulnerability. He said that, as de-industrialization took hold, European policymakers shut down domestic coal-fired power stations on the assumption that industrial capacity could be transferred to manufacturing centers in China, Vietnam, and other developing countries. View All https://www.eetimes.com/category/sponsored-content/ He argued that this approach has broken down because of two geopolitical forces. On the one hand, U.S. industrial policy has reoriented supply chains toward domestic priorities. On the other, he said, China has deliberately addressed its domestic manufacturing bottlenecks to retain control of its key supply networks. “As far as Europe is concerned, we are now in a situation in which we are entirely supply constrained,” Teixeira stated. “We lack the manufacturing facilities, we don’t have the natural resources necessary for manufacturing, and our supply chains are not tough enough to protect themselves against such a situation.” The economy’s recovery after the pandemic revealed the fragility of the system, he said. Semiconductor manufacturing fabs can take five to seven years to go from construction to production, while leading foundries allocate capacity two or three years in advance. As a result, Europe faces persistent constraints on hardware availability. At the same time, its power infrastructure limits efforts to expand digital capacity. ASICs versus GPUs The computing landscape is becoming more specialized as processing architectures evolve. In the past, GPUs—originally designed for video games—provided essential hardware for early Bitcoin mining and AI models because of their flexible, general-purpose computing capabilities. “Video games are to tech what Formula 1 is to cars—it is always the hardest, weirdest solutions that come out of video games,” Teixeira noted. As mining difficulty increased, Bitcoin mining shifted from GPUs to ASICs, devices designed to execute a specific cryptographic algorithm efficiently by removing unnecessary graphics pipelines and operating system overhead. Teixeira compared this hardware optimization to traditional software engineering, noting how developers on the PlayStation 1 reduced system memory allocations to free up performance for more demanding programs. A similar architectural split is happening in AI. Although Nvidia and AMD provide general-purpose GPUs to large-scale data centers, specialized AI ASICs are emerging from China. These purpose-built chips have eliminated older GPU features to focus entirely on AI inference and training. As Teixeira explained, “Bitcoin ASICs are not suitable for carrying out AI workloads, but GPUs can handle both. There comes a stage when you end up splitting the work, with some tasks using one type of equipment and others using another.” GPUs will remain useful for a range of workloads, but Teixeira expects specialized hardware will gradually take over most high-volume operations. Foundry bottlenecks and supply chain The struggle over access to advanced semiconductor nodes has now become a key strategic arena. Cutting-edge processor production depends heavily on TSMC, intensifying competition among AI developers, cloud providers, and hardware designers. Companies with less ability to absorb high chip costs are being pushed away from leading-edge production. For instance, automotive manufacturers mainly rely on mature nodes—28-nm and above. Yet achieving very high energy efficiency can require advanced process geometries, such as 3-nm process technology. This is evident in recent ASIC developments. Japanese company Triple-1 taped out its third-generation Bitcoin-mining chip https://www.eetimes.com/crypto-mining-asic-goes-deep-sub-threshold-on-3-nm/ , the Kamikaze III, on TSMC’s 3-nm process. The chip aims to achieve an energy efficiency of 10.45 joules per terahash by operating at sub-threshold voltages of 0.26 volts. The company argued that its Japanese base gives it an advantage over Chinese developers that are restricted from accessing advanced nodes. Energy infrastructure and coin per watt Beyond silicon fabrication, energy infrastructure is a major operational constraint. In the U.S., data centers usually draw power from regional grids or independent generation in unregulated markets such as Texas. By contrast, in Europe, much of the power grid infrastructure was established in the mid-20th century and lacks the interconnections needed to accommodate rapidly growing data center loads. In the U.K., operators frequently curtail large amounts of wind power capacity. “Many wind farms are dumping 40% of the electricity they produce since they are unable to sell it on to the grid,” Teixeira said. Because of these grid limitations, crypto mining and AI data centers have to differ in how they assess location and power costs. As Teixeira pointed out, crypto mining operations tend to move toward the lowest-cost power sources, such as stranded hydroelectric power in the Nordic countries or excess renewable energy output. Worldwide, crypto mining represents 183 TWh of power capacity Dec. 2024 https://www.jbs.cam.ac.uk/faculty-research/centres/alternative-finance/publications/cambridge-digital-mining-industry-report/ , roughly 30% to 40% the entire data center industry’s power footprint https://www.iea.org/energy-system/buildings/data-centres-and-data-transmission-networks , at the time. Teixeira stressed that the basic economic incentives that govern crypto mining and AI are quite different. “As for Bitcoin, I mine it, then sell the Bitcoin—you have to achieve profitability right away, because if you don’t, you’re in effect paying just to mine and ending up with nothing,” Teixeira said. “With regard to AI, currently a great deal of it is operating at a loss when selling the product, on the basis that future revenue will be so large as to offset the loss incurred today.” With energy costs rising, infrastructure providers are exploring modular power options, such as small nuclear reactors and biomass generation, located next to their data centers. See also: Crypto Mining ASIC Goes Deep Sub-Threshold On 3 nm https://www.eetimes.com/crypto-mining-asic-goes-deep-sub-threshold-on-3-nm/ Why RISC-V + Blockchain Is the Conversation I’ve Been Waiting to Have https://www.eetimes.com/why-risc-v-blockchain-is-the-conversation-ive-been-waiting-to-have/ Understanding the Quickly Evolving AI Data Center Architecture https://www.eetimes.com/podcasts/understanding-the-quickly-evolving-ai-data-center-architecture/