FuriosaAI closed a $620M Series D at a $2.9B post-money valuation, the Seoul chipmaker said. South Korea's National Growth Fund put in $280M, the largest check, and U.S.-based Infinitum led $110M from international investors that also include Cisco Investments and Jericho Capital. RNGD, its inference accelerator, entered mass production in January, with about 20,000 units planned for 2026 and more than 40,000 for 2027, and proceeds go to production, software and a third-generation chip being built with Broadcom. (Business Wire)
Private Companies
General Computefundraise
General Compute raised a $100M Series A at an $800M post-money valuation, up from $60M after its May seed. Blackbird Ventures and Abstract Ventures led the round. The company builds inference capacity on chips from Nvidia rivals and rents bare-metal servers, meaning raw machines with no software layer, to inference cloud operators. It has a $400M deal to buy chips from SambaNova and has added Cerebras and Positron as suppliers. CEO Finn Puklowski said it expects to deploy 75MW in 2027. (TechCrunch)
Hyperbolicfundraise
Hyperbolic raised a $30M Series A1 led by Walden Catalyst, the firm founded by Intel CEO Lip-Bu Tan and former Samsung executive Young Sohn. Dell Technologies Capital, General Catalyst and Vista Equity Partners also invested. The company connects GPU capacity from other cloud providers through a control plane called Forge, and plans a marketplace where customers can buy, sell and resell compute. Hyperbolic says more than 250,000 developers use its platform. (Hyperbolic)
PsiQuantumlaunch
PsiQuantum unveiled its first optical circuit switch, an 8×8 demonstration device built on its silicon photonics platform for AI data centers. An optical circuit switch steers light between fibers directly rather than converting it to electrical signals. In lab testing, the company reports typical insertion loss of 1.2 dB, 40 dB of crosstalk and reconfiguration in under 3 ms, with no amplification required. Samples are with select prospective customers, and commercial products will be announced in the coming months. (PsiQuantum)
Public Markets
Microsoft reported after the close that fiscal first-quarter revenue rose 19% to $92.3B, with Azure up 48% to $33.2B. CFO Amy Hood said results “exceeded expectations across revenue, operating income, and earnings per share.” (Microsoft)
Emerging
GPU prefill, NPU decode: A paper from Junhyeok Park, Jungeun Yoon, Kyungmin Kim, Hunsoon Lee, Shinyoung Ahn and Kangho Kim, HAWK: Heterogeneity-Aware KV Cache Transfer for Efficient GPU-NPU Phase-Disaggregated LLM Inference, runs the compute-heavy prefill phase of LLM serving on GPUs and the memory-bound decode phase on NPUs, moving the KV cache, the stored attention state, between them over a direct RDMA path. On a system built from commercially available GPUs and NPUs, it reports performance within 7.88% of a GPU-only baseline under the same latency targets, with up to 1.26x the power efficiency and 1.72x the cost efficiency. The figures are the authors' own. (arXiv)
Early exit in memory: A paper from Pratham Sharma, Mohd Faisal Khan and Santosh Kumar Vishvakarma, SLICE-CiM: Selective Layer-wise Inference Control Engine with Confidence-Based Early Exit for SRAM based Digital Compute-in-Memory, adds four small control blocks to an SRAM compute-in-memory array, which does its math inside the memory cells, so a CNN can stop inference early once its output is confident. Synthesized in TSMC 65-nm at 335 MHz, the blocks occupy about 1,982 µm² after place-and-route, and the authors report 31.35 to 454.91 TOPS/W depending on precision and supply voltage. The figures come from synthesis and EDA flows on MobileNetV2 with CIFAR-100, not fabricated silicon. (arXiv)