NVIDIA to Invest in d-Matrix: The Inference Chip Race Just Got a New Playbook

The deal: folding a rival into the ecosystem

According to three people familiar with the transaction, NVIDIA plans to invest in d-Matrix, a seven-year-old startup building AI server inference chips — a direct competitor. The proposed investment is part of NVIDIA's push to make its own technology compatible with rival chip designs and pull competitors into its dominant hardware ecosystem, even as NVIDIA keeps gaining share across nearly every front.

It is NVIDIA's first investment in an AI inference chip startup since last December's $20 billion licensing deal with Groq — a signal that NVIDIA will not back just one company in this space. Three months ago, NVIDIA said it was working with d-Matrix so their chips cooperate on different parts of the same AI inference task; last month the two went further: d-Matrix's next-generation Raptor XPU will natively support NVIDIA's NVLink Fusion and use the MGX rack reference architecture.

Source: 36Kr (Chinese) · Sina Finance

The bigger picture: ecosystem building and a $1B research bet

On October 8, NVIDIA announced a $1 billion commitment over five years to frontier research in quantum computing, healthcare and energy security in the US, as part of the White House "Genesis Mission." NVIDIA's $1B is 42% of the $2.4B total pledged by the tech industry, the largest single contribution, followed by AMD ($500M), OpenAI ($200M), Anthropic and Google ($150M each).

• Groq: $20B licensing deal + founding team absorbed; GPU combined with specialized inference chips
• d-Matrix: planned investment; Raptor XPU with NVLink Fusion + MGX (Q4 2027)
• Intel / Broadcom: billions in stakes over the past 12 months to integrate GPUs with their chips
• Amazon Trainium: August proposal for NVIDIA GPUs and its own silicon in the same data center
• SambaNova: June demo of its chips interconnecting with NVIDIA GPUs for AI models

Source: Sina Finance via Toutiao (Chinese) · IT Home (Chinese)

What it means for enterprises: inference costs are falling, adoption is accelerating

Training is expensive but concentrated; inference is massive and continuous. By pulling inference rivals into its ecosystem and pushing heterogeneous integration, NVIDIA is, in effect, doing one thing: lowering the cost and deployment barrier of AI inference. For enterprise customers this means:

1. More inference supply: GPU + ASIC running inference together means more options, less single-vendor lock-in;
2. Lower unit costs: specialized inference chips are more power-efficient and cheaper than general-purpose GPUs at scale;
3. Cloud inference becomes the default: heterogeneous compute, NVLink racks and elastic scaling are exactly what cloud platforms provide on demand.

For most SMBs, building your own inference cluster is expensive and hard to operate. Putting inference on the cloud — pay-as-you-go, elastic, always current — is the realistic path from "AI applications are accelerating" to "our own business is growing."

Cloud inference vs. self-hosted inference

As inference chips enter the "compatibility era," cloud platforms will bring new compute to users faster. Instead of waiting for your own cluster to depreciate, ride the industry tailwind from the cloud.