Check out this week’s episode of The Compute 100 Podcast with guest, Naveen Rao, CEO of Unconventional AI! The conversation covers Naveen’s incredible background, starting his third company, the parallels he believes computing should have with nature, and his motivations for building Unconventional AI. Give it a listen below!
Akamai signed a seven-year, $11.6B cloud agreement to run inference workloads for Anthropic, with an option to expand to roughly $20B, and took on Anthropic as an equity holder in return — a warrant for up to 5% of Akamai's stock, vesting as the contract scales. Akamai will spend about $5.5B through 2028 building out the infrastructure, and revenue does not begin until the second half of 2027, a multi-year bet that a legacy content-delivery network can be re-platformed for distributed AI inference. It is the clearest signal yet that frontier labs are diversifying compute suppliers beyond hyperscalers and neoclouds to any operator with distributable capacity, paying with warrants rather than cash to lock in supply. (Akamai)
Private Companies
Crusoecustomer
Crusoe will run production inference for Thinking Machines Lab's open models — including its Inkling models and GLM 5.2 and 5.3 — on a dedicated cluster of NVIDIA HGX B200 systems networked with Quantum-2 InfiniBand, under a new $65M annual contract. The deal pushes Crusoe's Managed Inference product past $100M in contracted annual recurring revenue less than a year after launch, evidence that GPU clouds built for training are finding a second growth line in dedicated, SLA-backed inference capacity for model developers who don't want to run their own stack. (Crusoe)
Emerging
Quantum interconnects, without matching photons: A new paper, "Single-atom-based asynchronous photonic interconnect for scalable modular quantum computing," from researchers at Israel's Weizmann Institute of Science and Austria's Institute of Science and Technology, proposes a memory-assisted entanglement scheme: a single trapped atom heralds and stores entanglement from one quantum processor while it waits for a matching photon from a second, removing the requirement that both photons be indistinguishable — a persistent bottleneck for linear-optics interconnects. The scheme scales linearly rather than quadratically with photon-arrival probability, an orders-of-magnitude gain in entanglement rate that bears directly on whether photonic links can support the module counts that fault-tolerant, modular quantum computing will eventually require. (arXiv)