Tired of ads? Enjoy an ad-free experience by signing up.
👩‍🍳 How we use AI at Tech in Asia, thoughtfully and responsibly.
🧔‍♂️ A friendly human may check it before it goes live. More news here

Broadcom launches new networking chip to challenge Nvidia

Broadcom has introduced a new networking chip, Thor Ultra, designed to connect hundreds of thousands of chips in AI computing systems.

The chip targets operators building large-scale infrastructure for AI models like those powering ChatGPT and will compete with Nvidia’s networking interface chips in data centers.

Thor Ultra follows Broadcom’s recent agreement to supply 10 gigawatts of custom chips to OpenAI starting in the second half of 2026.

Broadcom reported US$12.2 billion in AI-related revenue in fiscal 2024 and previously announced a US$10 billion unnamed customer for its custom AI data center chips.

The company’s networking chips, including Thor Ultra, are used to move data within AI data center clusters.

Engineers doubled the bandwidth in Thor Ultra compared to its predecessor and considered power and heat requirements during its design.

🔗 Source: Reuters

🧠 Food for thought

Implications, context, and why it matters.

Broadcom avoids AI fabric lock-in

  • UEC compliance puts Thor Ultra up against NVIDIA InfiniBand and Spectrum-X lock-in 1. It lets buyers pair any XPUs (a catch-all for CPUs/GPUs and custom accelerators), optics (optical transceivers for fiber) or switches to cut lock-in 1.
  • The Network Interface Card (NIC) adds features; packet-level multipathing; out-of-order delivery; selective retransmission; programmable congestion control 2. These overhaul Remote Direct Memory Access (RDMA) for bursty AI traffic, and they avoid the Go-Back-N resend that wastes bandwidth on loss 2.
  • Thor Ultra uses 50 watts, 60-67% less than Nvidia BlueField 3 Data Processing Unit (DPU) at 125-150W 2. It targets AI back-end networking (the east-west traffic between accelerators in the cluster) and skips front-end, storage offload or security 2. Broadcom projects 10-15% faster AI job completion with networks at 10-15% of total cluster cost 2.

Ethernet faces a short 800G window

  • 800G Ethernet optics will grow 60% in 2025 3. 1.6T parts land in select Nvidia and hyperscale uses in late 2025, leaving a 12-18 month window for 800G share gains 3.
  • Switch and NIC vendors need clear timing, and Thor Ultra is sampling with customers (early hardware provided for evaluation) while no ship date or design wins are public 1. Rivals that sell chips, chiplets and intellectual property (IP) can win integration revenue if they move faster than add-in card makers 2.
  • Third-party test labs and system integrators see demand for reference designs, interoperability testing and system integration as UEC-compliant AI fabrics standardize on high-speed Ethernet with Thor Ultra, and Broadcom invests $6-10 per silicon dollar 45.

Recent Broadcom developments

Stay ahead in Asia’s tech landscape

You've reached your 2 free content limit for the month. Sign up for free to read the full story.

🏄 For casual readers / 👶 Free

Basic

US$0

Free forever

Get instant access to this article and more every month

0 premium content

Unlimited news briefs

5

5 articles

Ad-free reading experience

Just US$0 per day

⌛Sign up in 20s. No payment details needed.

📖 For learners / 👍 Starter

Lite

US$4.92/month

Billed annually at US$59/year

Get instant access to this article and more every month

4

4 premium content

Unlimited news briefs & articles

Ad-free reading experience

Just US$0.17 per day

Cancel anytime

Our subscriber community includes professionals from these companies:

Stay updated on the go with our mobile app.

Get latest insights with smoother, more personalized experience through TIA mobile app.