Qualcomm Cloud AI 100 Ultra accelerator card detailed
Archived — this story has rotated out of today’s deck. It is kept here in full.
The gist
Qualcomm's Cloud AI 100 Ultra packs 128GB of memory at 548GB/s into a 150W inference card.
The 2023-era card drew renewed scrutiny over whether its specs still hold up.
Background
Qualcomm Technologies is the company whose data center portfolio includes the Cloud AI 100 Ultra, an AI inference card (hardware built to run trained models rather than train them). The card was announced in November 2023 and became available in 2024, so it is not new; a November 2026 Hacker News thread about its specs renewed attention on it. Qualcomm positions the Ultra at the top of its inference lineup alongside the AI 80 Ultra, AI 100 Pro and AI 100 Standard, and within its broader Dragonfly data center roadmap.
How it unfolded
- Nov 2023Qualcomm announced the Cloud AI 100 Ultra.
- 2024The Cloud AI 100 Ultra became available.
- Oct 2024Qualcomm began shipping the AI 100 Ultra, a PCI-Express card interlinking four AI 100 chips on a single package, alongside geared-down AI 80 versions.
- Mar 10, 2026Advantech and Qualcomm launched the SKY-641E3 4U edge server powered by Cloud AI 100 Ultra accelerator cards.
- Nov 4, 2026The card reached the Hacker News front page under the title "150W AI Accelerator Card with 128GB at 548GB/s".
- Nov 5, 2026Coverage described the card as Qualcomm's newest cloud AI inference card, with up to 870 TOPS INT8, 288 TFLOPS FP16, 150W TDP, 64 AI cores, 576 MB on-die SRAM, 128 GB LPDDR4x at 548 GB/s and PCIe Gen4 x16.
Key terms
- TDP
- The maximum heat a computer component generates and its cooling system must dissipate during normal non-turbo operation.
Source: Wikipedia
Who’s saying what
- Official
- Qualcomm calls the Cloud AI 100 Ultra a performance- and cost-optimized inference solution, programmable across workloads from computer vision to natural language processing, and says a single 150W card can run models of about 100 billion parameters and two cards about 175 billion.
- Skeptics
- Hacker News commenter varispeed wrote that the specs already seem outdated and "seems meh," while commenter pearl_panel5 asked how memory bandwidth holds up under sustained inference loads versus spec-sheet numbers.
- Public
- A daily.dev summary rated community reaction mixed (20% positive, 55% mixed, 25% skeptical): some see the memory-bandwidth-for-capacity tradeoff and low power draw as reasonable for prefill-heavy or capacity-constrained inference, others find the specs underwhelming versus newer GPUs and Apple's M-series.
Still unverified
The 870 TOPS INT8 figure is second-hand: explainx.ai said it could not retrieve Qualcomm's own product brief. Pricing and availability remain unclear.