Qualcomm data centre AI
Qualcomm AI200 and AI250: Inference Racks
Qualcomm has moved from phone chips into data-centre AI with the AI200 and AI250, inference-only accelerators sold as liquid-cooled racks that trade memory bandwidth for very large, lower-cost memory. Here is what is announced, what is claimed, and what to verify.
Export controls: AI chips, servers, networking and memory can need an export licence depending on destination, end user and end use under US, UK and EU rules. We screen every project for this. See what applies.
What Qualcomm is offering
- Inference only. The AI200 and AI250 are accelerator cards sold in pre-configured racks for running models, not for training them.
- Memory capacity first. The AI200 carries 768 GB of LPDDR memory per card. LPDDR costs less and uses less power than HBM but has much lower bandwidth, so the bet is that many inference workloads are limited by capacity, not bandwidth.
- A new memory architecture. The AI250 uses a near-memory computing design, with Qualcomm claiming more than ten times the effective memory bandwidth and lower power. No figures were given, and analysts say the claim needs scrutiny.
- Racks. Both target rack-level power of about 160 kW with direct liquid cooling, PCIe for scale-up and Ethernet for scale-out, and confidential computing support.
Timing and customers
- Availability. Qualcomm announced AI200 commercial availability for 2026 and AI250 for 2027. These dates come from October 2025 announcements, so confirm what has shipped before you plan around them.
- First customer. HUMAIN in Saudi Arabia announced it is targeting 200 megawatts of AI200 and AI250 racks starting in 2026. Qualcomm has not named other customers in the material we read.
- Earlier products. The Cloud AI 100 family is the prior work these racks build on.
Side by side
| Cloud AI 100Earlier generation | AI200Inference rack, 2026 | AI250Inference rack, 2027 | |
|---|---|---|---|
| Status | Prior work, specifications not covered in the sources we read | Commercial availability announced for 2026 | Announced for 2027 |
| Memory | Not covered | 768 GB of LPDDR per card | Near-memory architecture, with a claim of more than 10x effective bandwidth (vendor claim) |
| Rack | Not covered | Direct liquid cooling, about 160 kW per rack | Same rack envelope, about 160 kW |
| Fabric | Not covered | PCIe scale-up, Ethernet scale-out, confidential computing | PCIe scale-up, Ethernet scale-out |
| First named customer | Not covered | HUMAIN (Saudi Arabia), targeting 200 MW from 2026 | HUMAIN, same programme |
Most figures are Qualcomm claims from October 2025 and were not independently measured.
Software
- Hosting models. Qualcomm describes a hyperscaler-grade software stack with one-click deployment of Hugging Face models through its Efficient Transformers Library and AI Inference Suite, and support for disaggregated serving.
- The CUDA question. NVIDIA’s ecosystem is a large moat. Check that your models, serving engine and tooling run well on Qualcomm’s stack before you commit. See the platform comparison.
- Keeping it secure. Runtime, firmware and serving stack need patching and review under the maintain stage of our agent security and defence practice.
What it means for your data centre
- Power and cooling. At about 160 kW per rack and with direct liquid cooling, these racks need the same facility readiness as the largest GPU systems. See the rack power and cooling guide.
- Where it could fit. Large-model inference where capacity and cost per token matter more than peak speed, and where a second source to NVIDIA and AMD is valuable.
- Sovereign programmes. The HUMAIN announcement shows regional and sovereign buyers are an early market. Check the export position for your destination.
Where this fits in the stack
Common questions
What is the difference between AI200 and AI250?
The AI200 is the 2026 card with 768 GB of LPDDR. The AI250, announced for 2027, adds a near-memory architecture that Qualcomm says raises effective memory bandwidth more than ten times.
Can they train models?
No. Qualcomm positions them for inference only.
Is LPDDR a problem?
It has lower bandwidth than HBM, so token generation speed can be lower on some models, but it offers much more capacity per card at lower cost and power. Test your own models.
Do export controls apply?
Advanced AI hardware can need a licence depending on destination, end user and end use. We screen every project.
Sources and dating
Figures are as of 18 November 2026, from October 2025 coverage by The Register and NAND Research. We did not find newer sources, so availability may have changed. Some figures are vendor claims or preliminary and some differ between sources, so we confirm against current datasheets when we specify a build. Qualcomm and Cloud AI are trademarks of Qualcomm Incorporated. Embedica is independent and is not affiliated with or endorsed by any of the companies named.
Next steps
Related: Other Chip Providers · NVIDIA Data Centre GPUs · AMD Instinct GPUs · Memory Suppliers · The AI Application Layer
Weighing an inference alternative
Test it against your workload
