AI compute is scarce
Demand for AI compute still exceeds available supply: some AWS customers are reportedly trying to reserve effectively all capacity. (networkworld.com) Venture capital and chip investors are shifting to alternative silicon—SiFive raised $400M at a $3.65B valuation with Nvidia among backers to build RISC‑V CPU designs aimed at AI datacenters. (parameter.io)
Artificial intelligence compute is still so tight that Amazon says some Amazon Web Services customers want to reserve nearly all available Trainium chip capacity. (networkworld.com) Compute is the raw processing power used to train and run artificial intelligence models, and cloud providers rent it out by the hour through data centers packed with chips, networking gear, and power equipment. In an April 9 shareholder letter, Amazon Chief Executive Andy Jassy said Amazon Web Services’ custom-chip business is “on fire” and said customers are trying to buy up the Trainium capacity Amazon has today. (aboutamazon.com) (networkworld.com) Amazon is pushing two in-house chip lines: Trainium for training models and Inferentia for serving them after they are built. Amazon says Inferentia2-based Inf2 instances are tuned for large language models and image generators, while Trainium chips plug into the company’s Neuron software stack for model training. (aws.amazon.com) (aboutamazon.com) Amazon has been adding capacity fast, but the build-out is measured in data centers and power feeds, not software updates. In October 2025, Amazon said Project Rainier went live with nearly 500,000 Trainium2 chips for Anthropic, one of the largest artificial intelligence compute clusters the company has built. (aboutamazon.com) The shortage is pushing investors beyond Nvidia’s graphics processors and toward other parts of the server. On April 9, SiFive said it raised $400 million in a Series G round at a $3.65 billion valuation to expand RISC-V central processing unit designs for data centers, with Nvidia, Apollo Global Management, Point72 Turion, T. Rowe Price Investment Management, Prosperity7 Ventures, and Sutter Hill Ventures among the backers. (sifive.com) RISC-V is an open chip instruction set, which is the basic rulebook that tells a processor how to understand software. SiFive said hyperscale customers want customizable central processing units in licensable intellectual-property form, and said its funding will go toward scalar, vector, and matrix designs aimed at artificial intelligence data centers. (sifive.com) That shift reflects how artificial intelligence servers are being redesigned as full systems, not just graphics-processor boxes. SiFive said central processing units are needed to coordinate “agentic” artificial intelligence workloads and manage tasks that graphics processors and accelerators do not handle efficiently, especially under tight power limits. (sifive.com) Amazon is making the same argument from the cloud side: lower-cost custom chips can widen supply even while it keeps buying Nvidia hardware. Jassy said Trainium offers better price-performance than current graphics-processing-unit alternatives, and Amazon has kept rolling out newer Trainium generations as it spends heavily on artificial intelligence infrastructure. (networkworld.com) (aboutamazon.com) The bottleneck now is not whether companies want more artificial intelligence compute, but how fast cloud operators can turn capital, electricity, and chip designs into usable capacity. The next test is whether Amazon’s custom silicon and the new money flowing into alternatives such as SiFive can add supply faster than customers reserve it. (networkworld.com) (sifive.com)
