NVIDIA gives d-Matrix a route into its AI racks with Raptor inference chips

NVIDIA is giving d-Matrix’s next-generation Raptor inference chips a route into its AI-factory racks. The multi-year collaboration will integrate Raptor XPUs into NVIDIA’s MGX architecture through NVLink Fusion, targeting the part of AI serving where generating tokens quickly matters most.

NVIDIA Watch analysis

What happened

The Futurum Group reports that the planned rack will combine 144 Raptor XPUs with NVIDIA Vera CPUs, NVLink switches, BlueField-4 DPUs, ConnectX-9 SuperNICs and Spectrum-X networking. The design splits the work: GPUs handle compute-heavy prefill, while d-Matrix hardware tackles latency-sensitive decode.

Raptor is built around stacked 3D DRAM and is expected to tape out before the end of 2026. Initial availability in NVIDIA MGX racks is expected in Q4 2027, so this is an architectural direction with a sizeable calendar attached, not a product you can order this afternoon.

Key findings

• More memory for larger models
  The proposed 144-XPU rack is described with 2.3TB of 3D-DRAM capacity, aimed at frontier-scale contexts.
• A split inference pipeline
  NVIDIA’s design pairs GPU prefill with specialist decode hardware to target faster interactive responses.
• Projected speed, not a benchmark result
  The report cites roughly 1,000 tokens per second per user on 3-trillion-parameter-class models at 1M context, but Raptor is pre-silicon.
• NVIDIA keeps the plumbing
  d-Matrix gains access to the MGX ecosystem while NVIDIA supplies the rack, interconnect, networking and software platform around it.
• The commercial clock is ticking
  Q4 2027 availability leaves tape-out, manufacturing, thermal management and customer validation as substantial hurdles.

Why it matters

AI inference is becoming a hardware-matching exercise rather than a simple GPU shopping list. NVIDIA already has specialist silicon in its orbit for fast decoding; bringing d-Matrix’s DRAM-heavy approach into MGX suggests a portfolio segmented by model size and context demands.

The practical promise is capacity. The report contrasts Raptor’s projected 2.3TB of 3D DRAM with the much smaller memory pools typical of SRAM-based designs, potentially making very large models and long contexts easier to serve. The important qualifier is that the figures remain projections from a chip that has not taped out.

Our read

This is a strategically important partnership, but not yet a performance verdict. NVIDIA is turning its rack platform into a meeting place for different inference architectures, and d-Matrix is trading some independence for access to a vast deployment ecosystem. Watch the silicon, the first named customer and the price per useful token. The brochure can wait in the lobby.

What to watch

• Whether Raptor tapes out before the end of 2026.
• Whether Q4 2027 MGX availability arrives on schedule.
• Named hyperscaler or frontier-lab deployments.
• Production measurements for latency, energy use and cost per token.

Discussion spark: Will heterogeneous racks with specialist decode chips become the sensible default for large AI services, or will simpler all-GPU systems remain cheaper and easier to operate?

Independent WittyWires tracker for public updates about NVIDIA. Not affiliated with or endorsed by NVIDIA; this is not an official account.