Compute Desk raised $45M led by Lightspeed Venture Partners to build markets for pricing, trading and hedging GPU compute. Menlo Ventures, Hedosophia and Motive Partners also invested. The company publishes price indexes for Nvidia GPU capacity, such as DESKB300 and DESKH100, built from private transaction data shared by neoclouds and distributed through Bloomberg. Exchange partners Nodal Exchange, Architect Financial and Liquid Compute are seeking regulatory approval to list compute futures and options on those indexes, and its Compute Clear product is designed to let clearinghouses settle those contracts in deliverable compute (Compute Desk)
Private Companies
Argentum AIpartnership
Argentum AI signed a 20-year, $630M agreement with Kyndryl for an initial 10MW at Kyndryl's Tier III data center in Boulder, Colorado. The GPU cloud provider is deploying Nvidia B300 systems there for AI inference, and deployment is underway. An unnamed enterprise AI inference company is the site's anchor tenant. Argentum develops, finances and runs bare-metal GPU clusters for model builders and enterprises, and both companies said they intend to expand the arrangement (PR Newswire)
Public Markets
Nvidia reported after the close that fiscal third-quarter revenue rose 103% to $115.8B, with data center up 110% to $107.6B. It guided to $130.0B, assuming no China data center compute revenue, in its first report since Jensen Huang resigned (Nvidia press release)
Cerebras put disaggregated inference into production, with AMD Instinct GPUs processing prompts and CS-3 systems generating tokens. It reports 5x throughput per megawatt versus CS-3 alone at the same speed. AWS, Fireworks AI and Gimlet Labs have early access (Cerebras)
Emerging
FPGAs for LLM decode: A paper from Zhiheng Chen, Yifan Zhang, Yu-An Chou, Ye Qiao and Sitao Huang, Chimera: A Configurable LLM Decode Accelerator Architecture with Compute-Memory-Floorplan Co-Design for HBM-Equipped Multi-Die FPGAs, targets single-request decoding, which is limited by memory bandwidth and leaves GPUs underused. It routes HBM access to compute on every die of the FPGA and keeps mixture-of-experts weights that exceed HBM capacity in a second storage tier. Across ten designs built on an AMD Alveo U280, including an 80B-parameter mixture-of-experts model, it averages 1.6x the token throughput of prior FPGA designs and 3.0x and 1.5x the energy efficiency of an RTX 3090 and an A100 (arXiv)