Microsoft posted a Maia 300 architecture write-up on Monday: a third-generation custom inference chip in a 96-accelerator rack at 167 kW, with a 12,288-accelerator Ethernet superpod aimed at long-context decode and tool use. xAI put Grok, Cursor, and X Premium on one plan, The X Subscription, from $8 to $200 a month with a shared usage pool. NVIDIA contributed Vera Rubin rack and tray designs to the Open Compute Project through MGX 1.2 and opened pieces of DSX OS. A France ecosystem roundup names Bull's 2028 Vera Rubin Ultra plan in Angers, a first satellite on orbit this month, and Twake on Nemotron for French public agencies.
In the news
Azure Maia 300
Microsoft's Azure Blog published a Maia 300 architecture post on October 12. Maia 300 is the third-generation custom inference accelerator, co-designed as silicon, rack, network, and datacenter for cloud-scale agentic work: multi-step execution, long-context decode, and tool use.
The rack holds 96 accelerators across 24 four-accelerator trays at 167 kW, twice the density of Maia 200. A two-tier Ethernet fabric builds a superpod of up to 12,288 accelerators, and prefill and decode workers stay on that same scale-up path rather than bouncing through a separate host network.
Microsoft says compute sits on TSMC N2P, with memory and I/O on separate dies. Versus Maia 200, the company quotes 36 percent more compute FLOPs, 2.5 times the HBM bandwidth, 33 percent more HBM capacity, and about twice the HBM bandwidth per watt, with eight HBM4 stacks, 288 GB, and 18 TB/s on each accelerator. The dataflow layer, SDLA, adds NUMA across chiplets, enhanced MXFP4 and sub-3-bit storage, and hardware type conversion. An accelerator can drive up to 750 GB/s to any peer in the superpod when only a subset of traffic goes cross-rack. Those performance ratios are Microsoft's.1
The X Subscription
xAI posted The X Subscription on October 12. Grok (chat, voice, Imagine image and video), Grok Bot, Cursor, and X Premium sit on one bill.
Paid plans are Plus at $8 a month, Premium at $30, Super at $100, and Ultra at $200, with a Free tier below them. Usage across Grok and Cursor shares a pool that refreshes weekly. xAI pitches Premium as SuperGrok plus Cursor Pro plus X Premium+, previously $90 a month combined, and says each new tier delivers more usage per dollar than the plan it replaces. Those comparisons are the company's.
New customers can start now, with a global rollout over the next few days. Existing SuperGrok, Cursor, and X Premium subscribers keep their price and features unless they switch; the switch is optional, prorated, and appears in billing settings when it is available. Old plans are closed to new sign-ups.
Team seats match the $30, $100, and $200 prices and add SSO, admin controls, consolidated billing, and SOC 2. Third-party models in Cursor remain on Premium and above, billed at API list prices through top-ups. This is a bundle, not a new Grok model. Grok 4.8 has not shipped.2
NVIDIA at OCP
NVIDIA posted from the Open Compute Project Global Summit in San Jose on October 12. It contributed Vera Rubin rack and compute-tray designs to OCP through the open MGX architecture.
MGX 1.2 covers rack-scale blind-mate liquid cooling, a liquid-cooled 5,000 A busbar, native 19-inch support, fanless liquid-cooled trays, and racks up to 250 kW. A separate Vera CPU rack spec is meant for tool calling, orchestration, and sandboxed execution beside Vera Rubin NVL72 training and inference.
NVIDIA also opened DSX OS components for switch config, cluster readiness tests, rack management, and bare-metal onboarding that it says can take hours instead of days. With AMD, Google, and Meta it published Open Data Center Specification v0.7. Google, Meta, Microsoft, and NVIDIA published a grid-resilience spec that calls for restoring near-full load in under two seconds. The tokens-per-watt and onboarding-time claims are NVIDIA's.3
NVIDIA in France
NVIDIA's Europe post from October 12 is factory and sovereign-stack news around GTC Berlin. Bull will extend European AI server production to Vera Rubin Ultra NVL72, planned for 2028, with trays in Europe and rack integration at the Angers factory that opened October 2. That is a production plan, not a ship this week.
Altair-1, first of a claimed $1 billion, 50-satellite constellation, reached orbit this month. Loft Orbital integrated it and Orbitworks built it in Abu Dhabi, flying Jetson AGX Orin modules that run Mistral models for Earth observation and autonomous operations.
LINAGORA and Docaposte's CANUT agreement names Twake.ai, built on Nemotron, as the referenced sovereign workplace for more than 2,500 French public organizations. Jensen Huang's GTC Berlin keynote is still Wednesday, October 21, at 11:00 CEST.4
