CoreWeave completed the industry-first bring-up and full rack-level validation of the Vera Rubin NVL72 on June 1 — Dell delivered the rack on May 31 on Dell PowerEdge XE9812 servers with Micron liquid-cooled NVMe SSDs, and CoreWeave cleared production validation the following day. The NVL72 packs 72 Rubin GPUs and 36 Vera CPUs per rack over NVLink 6 fabric at 260 TB/s, with 1.6 Tb/s of backend bandwidth per GPU. Nvidia's claimed performance versus Blackwell: 10x inference per watt, 75% fewer GPUs required for equivalent workloads, and one-tenth the cost per million tokens.
The NVL72's liquid cooling and rack-scale architecture required two pieces of internal tooling: Valvey, a programmable per-rack valve assembly enabling software-defined liquid cooling with real-time monitoring, and Racky, a unified rack control appliance that aggregates power, cooling, and environmental sensors across the system. CoreWeave also deployed multi-rail networking with Nvidia BlueField-4 DPUs for tenant isolation. The operational complexity of running Vera Rubin at scale is meaningful, and CoreWeave is the first organization to have worked through it.
CoreWeave is the leading indicator for the broader Vera Rubin rollout — AWS, Google Cloud, Microsoft, Oracle, Lambda, Nebius, and Nscale are all expected to deploy in H2 2026. The tooling CoreWeave built to manage cooling, power, and networking at rack scale will likely become the operational baseline the rest of the industry references. Jane Street is named as an early customer: "Their ability to deliver highly performant clusters with full observability gives us confidence to partner on Vera Rubin."
CoreWeave's full announcement has the technical architecture detail on Valvey, Racky, and the rack-level validation process.
Groq is raising $650 million from existing investors Disruptive and Infinitium to build a pivot business after Nvidia's $20 billion LPU architecture license deal transferred Groq's core inference chip IP and most engineering talent to Nvidia. Under new CEO Simon Edwards, the remaining team is building "Groq 2.0" as a dedicated AI inference neocloud — GroqCloud. The overnight GTC announcement adds direct competitive context: Nvidia confirmed Groq 3 LPX ships Q3 on Samsung 4nm, leveraging Groq's own licensed architecture against the market the original company was building. The $650M buys time to establish GroqCloud as a software-differentiated inference services business before Nvidia's full-stack LPX-equipped platform is in wide deployment.
Red Hat OpenShift AI powered by Rebellions ATOM NPUs went generally available on May 27 — the first enterprise Kubernetes AI inference platform to certify a non-Nvidia accelerator as a production-grade deployment target, shipping certified vLLM container images for OpenShift AI. Rebellions' Rebel100 NPU is in mass production at Samsung Foundry (64 PFLOPS FP8, 600W). Named discussion partners include Saudi Aramco's Humain subsidiary and Mistral AI. The Red Hat GA converts Rebellions from a competitive proof point into a production sales motion inside Fortune 500 enterprise environments that standardize on OpenShift — the enterprise distribution unlock that inference chip challengers require to reach meaningful volume outside hyperscaler pilots.
DriveNets closed a $410 million Series D today led by Bessemer Venture Partners and Atreides Management, with AMD as a strategic investor, D1 Capital Partners, Pitango, and new entrant Red Dot Capital Partners. Valuation: $8.5 billion. Total funding: approximately $1 billion. The company is cash-flow positive with over $1 billion in backlog. DriveNets builds disaggregated networking software — separating the control plane from commodity switching hardware — and has extended this model to AI cluster fabric, offering lossless Ethernet fabric optimized for AMD Instinct GPU clusters. AMD's strategic investment is the structural signal: DriveNets is becoming the networking layer for the AMD-native AI cluster alternative to Nvidia's NVLink ecosystem. AT&T separately committed $650 million in a secondary transaction in mid-2025, providing early-stage liquidity.
RTX Spark — Nvidia's first laptop and desktop SoC, co-designed with MediaTek — was also unveiled overnight. It pairs a 20-core Grace CPU (Arm Cortex-X925/A725) with a Blackwell RTX GPU (6,144 CUDA cores, 5th-gen Tensor Cores with FP4), up to 128GB LPDDR5X unified memory, and 600 GB/s GPU-to-CPU bandwidth via NVLink-C2C. Claimed AI performance: 1 petaflop. Ships fall 2026 in slim laptops and compact desktops from ASUS, Dell, HP, Lenovo, Microsoft Surface, and MSI. Nvidia is entering the PC SoC market directly, competing against Apple M-series and Qualcomm Snapdragon X Elite — and using the same NVLink-C2C interconnect architecture that runs from Vera Rubin down to a laptop chip.
ChangXin Memory Technologies (CXMT), China's leading DRAM manufacturer, received Shanghai Stock Exchange approval on May 27 to list on the STAR Market, targeting a raise of approximately $4.2–$5.5 billion — China's largest chip IPO since 2022. Q1 2026 revenue: 50.8 billion yuan, up 7x year-over-year; net profit up approximately 17x. CXMT is actively developing HBM capability, with IPO proceeds earmarked for DRAM capacity expansion, advanced manufacturing technology, and HBM development. A successful CXMT HBM ramp introduces China's first domestically manufactured HBM into competition with SK Hynix, Micron, and Samsung by 2027–2028 — the first credible threat to the oligopoly pricing structure that currently drives HBM valuations. Watch CXMT's stated HBM yield roadmap in its IPO prospectus.