The power crunch is now the AI buildout's binding constraint — and it is reshaping the data center

The shortfall

The US power shortfall for data centers is projected to reach the mid-30 gigawatt range by 2028. Permitting delays in Texas and Ohio are causing a pause in data center growth rather than project cancellations. Grid optimization is unlikely to solve the shortfall, because utilities manage for peak demand rather than average utilization. The constraint is not a single failed project but a system that cannot add power as fast as demand grows. m:…6115

The pressure is showing up in places far from Texas. Norway has implemented a moratorium on new data centers, and a major proposed data center project in the United Kingdom was terminated due to the utility's inability to provide grid access. Developers are successfully securing support for data centers by funding local community projects like hospitals and schools, which is a sign of how the negotiation has changed: power access is now bought with community investment, not just a utility contract. The dissonance between Silicon Valley optimism and Wall Street pessimism regarding AI is currently at an all-time high, and the market is pricing in a 2028 slowdown for AI chips which is likely too bearish. m:…6115

Memory, not just megawatts

Power is not the only bottleneck, and it may not be the binding one. Memory is currently the most significant component bottleneck for hyperscalers, and semiconductor supply chain tightness is expected to increase in 2027. That reframes the buildout: even where power is available, the machines themselves are constrained by memory supply. The market is currently pricing in a 2028 slowdown for AI chips, which is likely too bearish. Whether that optimism survives another year of permitting delays is the open question. m:…6115

The data center is changing shape

Inside the data center, the workload itself is changing what gets built. TensorWave, an AMD-exclusive cloud provider, says its customers are requesting more CPUs in their next AI deployments, and higher CPU-to-GPU ratios for future deployments, because AI agents require CPUs to handle tasks between model calls. Its early customers primarily used AMD GPUs for inference, and AMD's GPU memory roadmap influenced its decision to build an AMD-exclusive cloud, aiming to provide an alternative to NVIDIA's closed-source ecosystem. TensorWave is preparing to deploy AMD Helios across North America, and requires a signed customer contract before financing a new deployment, a contract-first model that ties capacity buildouts to actual demand rather than speculation. Increased memory per GPU allows for larger model weights and KV caches, which is why the memory roadmap mattered to the architecture choice. m:…0303

Vultr makes the same point with numbers. GenAI workloads typically require a 1:4-8 CPU-to-GPU ratio, while agentic AI workloads are moving towards a 1:1 ratio, because in agentic AI, CPUs are responsible for GPU orchestration, data supply, tool calls and policy checks. Agentic workflows involve continuous loops of evaluating outputs, checking for errors and calling tools, unlike GenAI workflows, which are closed, finite processes involving a simple back-and-forth between user and model. Vultr's full-stack infrastructure integrates AMD EPYC CPUs, AMD Instinct GPUs, VAST Data OS and SUSE AI Factory, and allows organizations to run agentic workloads in sovereign environments. Governance and compliance ranked as the top obstacle to greater AI maturity, cited by 64 percent of decision-makers, according to Vultr's account, which is another way of saying the buildout is gated by permission as much as by power. m:…8418

The hardware underneath is also shifting. In a preprint, researchers benchmark NVIDIA's Vera data-center CPU against AMD Turin, Intel Granite Rapids and NVIDIA Grace, finding Vera achieves the best performance in most scientific codes tested, with gains primarily attributed to high memory bandwidth, effective Spatial Multithreading and low-latency memory subsystems. Vera exhibits lower floating-point performance compared with Intel and AMD in compute-bound workloads, reflecting a design optimized for GPU-centric ecosystems. The preprint's message fits the rest of the week: in the AI buildout, memory bandwidth matters more than raw floating-point speed, because the job is feeding the accelerators, not replacing them. m:…8151