Spectrum-X

Headlines · 4

1. 2d
   Nvidia backs Upscale AI's heterogeneous compute networking push

   Upscale AI, a startup with Nvidia as a strategic investor and technology supplier, has unveiled Token Fabric, a full-stack AI networking platform meant to let data centers run accelerators from Nvidia, AMD, Intel and Chinese vendors side by side. Its in-rack scale-up layer uses Upscale's own SkyFabriX switch chip and supports the UALink and ESUN open interconnect standards, while the cross-cluster scale-out backbone is built on Nvidia's Spectrum-X Ethernet, paired with the SkyOS network OS and SkyCMD operations tooling. The report says phase one is slated for Q4 2026 with the full platform rolling out in stages during 2027, and that Upscale AI raised a $190 million extension in June 2026 led by Premji Invest with Nvidia, Temasek and Salesforce participating, bringing total funding to $500 million at a $2 billion valuation.

   36氪 人工智能 · 🔥 0
2. 4d
   Nvidia and Cisco unveil USFP optical interface with 400Gb/s per lane

   Nvidia and Cisco announced USFP (Ultra-Dense Small Form Factor Pluggable), a new optical transceiver interface for AI data centers, at the 2026 OCP Global Summit; Cisco says it is the first such interface to support 400Gb/s per lane, twice the current 200Gb/s-per-lane designs. The first modules are expected to use 16 full-duplex lanes in an octal small form factor footprint, giving a theoretical 6.4Tb/s total, with later designs expanding to 32 or more lanes and a planned 12.8Tb/s total that still requires further development and validation. The companies claim four times the bandwidth density of 200Gb/s-per-lane optics and about a 1.3x improvement in power efficiency, support for both switches and Nvidia SuperNIC adapters over RoCE, a two-step insertion design for signal integrity, and up to 100 watts of power dissipation without a dedicated internal cooling circuit; TE Connectivity and Amphenol are among partners building cages, connectors and modules.

   TechSpot · 🔥 0
3. Oct 9
   Citi: Meta Muse agents could shift CPU-GPU ratios toward 1:1

   In an Oct. 6 research note, Citi argues that always-on personal AI agents such as Meta Muse consume compute at orders of magnitude above chatbots, pushing the CPU-to-GPU ratio from 1:8 in training and 1:4 in inference toward 1:1 or higher for agentic workloads. The bank estimates 100 million Muse daily active users would require roughly 200,000–390,000 Blackwell-class GPUs, worth $7 billion–$19 billion in one-time revenue for Nvidia, and projects the global CPU total addressable market growing from $29 billion in 2025 to $300 billion by 2030. It also flags knock-on demand for DRAM, SSDs and data-centre network traffic.

   36氪 人工智能 · 🔥 0
4. Oct 8
   Nvidia-backed Upscale AI launches Token Fabric to link chips from rival vendors

   Upscale AI, a Santa Clara networking startup backed by Nvidia, launched Token Fabric, a platform built on open standards such as UALink and Ethernet for Scale-Up Networking that can put GPUs, custom AI chips and network cards from different suppliers on the same network. Reuters reported the platform combines Upscale's own scale-up switch chip, SkyFabriX, with scale-out switches based on Nvidia's Spectrum-X Ethernet chips, plus SkyOS and SkyCMD software that runs and monitors the network and can flag bottlenecks and hardware needing replacement. Upscale said the first part ships in the fourth quarter with the rest rolling out through 2027, and general availability in early 2027; CEO Barun Kar told Reuters he expects revenue in the tens of millions of dollars next year, possibly reaching the low hundreds of millions.

   The Next Web · 🔥 0

Experience and discussion from the community

Share my Spectrum-X experienceAsk about Spectrum-X