Nvidia and SpaceX's AI unit confirmed this week that SpaceX's first Starmind satellite will run an optimized version of Nvidia's Vera Rubin NVL72 rack architecture in orbit as early as the fourth quarter of 2027, extending the same 88-core Vera CPU design that already orchestrates Grok's terrestrial AI agents into space. The move treats orbital compute not as a separate hardware track but as the same platform Nvidia already sells to every other AI factory, launched instead of built. (NVIDIA Newsroom)
Private Companies
C2i Semiconductorsother
Infineon agreed to acquire C2i Semiconductors, a Bengaluru startup founded in 2024 by veterans of Texas Instruments, National Semiconductor, and Maxim Integrated, whose software-defined multiphase controllers and smart power stages keep GPU voltage rails stable as individual accelerators now draw more than 1kW. Terms weren't disclosed; the deal is expected to close in the third quarter. As GPU power draw keeps climbing, the delivery circuitry around the chip is becoming as contested as the chip itself. (Infineon)
Groqlaunch
Nvidia's Groq 3 LPX interactive-inference accelerator entered full production at Samsung's fabs this week, the first shipped product from the roughly $20B in Groq assets Nvidia acquired last December. The chip packs 500MB of SRAM directly on-die to sidestep the memory-bandwidth limits that slow rival inference silicon; Nebius plans to bring it into its Token Factory serving platform, with Groq itself among the earliest adopters. Buying the architecture that threatened to undercut Nvidia's GPUs turned out cheaper than competing against it. (NVIDIA Newsroom)
Public Markets
Supermicro rallied after Cisco added its liquid-cooled servers to Cisco's Secure AI Factory program and Taiwan prosecutors indicted individual employees but not the company itself, clearing a governance overhang that has weighed on the stock for two years. (The Motley Fool)
Arm unveiled AGI, its first complete in-house CPU aimed at data centers — selling finished silicon to server vendors instead of only licensing designs for them to build. The stock's muted move suggests investors want shipping volume before repricing the business-model shift. (ServeTheHome)
Emerging
Cheap-memory bet: Intel detailed Crescent Island, an inference accelerator built around 480GB of LPDDR5X rather than expensive HBM, betting that memory capacity matters more than bandwidth for hosting large, long-context agentic models economically. The 350-watt card is air-cooled, a deliberate contrast to the liquid-cooling infrastructure Nvidia and AMD's HBM-heavy inference chips require — alongside Groq's SRAM-heavy approach, a second credible sign that inference-chip architecture hasn't converged on a winner yet. (StorageReview)