Anthropic's IPO prospectus, reviewed by Reuters, lists $518B of cloud, compute and infrastructure obligations over the coming years, against $4.59B of 2025 revenue. Compute and infrastructure spending reached $7.33B last year, more than triple 2024 and over half of $12.65B in operating expenses, and Anthropic ended 2025 with $20.28B in cash and an operating loss above $8B. Reuters reports the company will seek a valuation above $2T. Forward commitments run more than 100 times trailing revenue. (Reuters)
Private Companies
World Labsother
AMD agreed to acquire World Labs, the spatial-intelligence lab founded by Fei-Fei Li, in an all-stock deal valued at about $8.2B, expected to close by the end of 2026. World Labs raised $1B at a $5B valuation in February with AMD among the investors, a 64% step-up in seven months. Li becomes AMD's chief scientist, reporting to Lisa Su. AMD gains workload knowledge for its roadmap and a world-model capability Nvidia already offers through Cosmos. (AMD)
SiMa.aifundraise
SiMa.ai closed a $150M Series C at a $1.45B valuation, co-led by Fidelity Management & Research Company and Amplify, bringing total capital raised to $500M. Revenue grew four-fold from 2024 to 2025, and Bosch, Micron and Synopsys are named customers and partners. Proceeds fund its physical AI software environment and a part targeting 1,000 dense TOPS in the first half of 2028. A crossover-led round at this scale prices physical AI as the next hardware market to re-rate after the data center. (company press release)
Gimlet Labspartnership
Cerebras will supply CS-4 systems to Gimlet Labs to build 100 MW of inference capacity over one to two years, with the first Gimlet Cloud data center due later in 2026. Gimlet pairs wafer-scale processors with GPUs, routing each inference phase to the best-suited hardware, and targets up to 3,000 tokens per second. Terms were not disclosed. The deal gives Cerebras a second route to market beside its own cloud, through a neocloud built around disaggregated inference. (Quartz)
Emerging
Memory bandwidth, pooled: A new paper, BOOST: Concurrent Access to Host Memory and HBM to Accelerate LLM Inference, reads GPU HBM and host memory over the CPU-GPU interconnect at the same time, allocating pages in proportion to each tier's bandwidth instead of treating host memory as a slow overflow tier. Integrated into vLLM and tested on Grace Hopper, it lifted high-throughput serving by 31% on average and beat prefetching by 15%, with no kernel changes. The result extends what deployed superchip systems can serve before additional HBM is needed, a useful lever while HBM supply stays tight. (arXiv)