Ecosystem / Other AI Chip Providers
Ecosystem
Other AI Chip Providers
NVIDIA, Intel and AMD are the main merchant platforms, but they are not the whole market. Custom accelerators, networking silicon and new inference chips from other providers shape what you can buy, rent and plan for.
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.
Providers in the landscape
- Broadcom Designs custom AI accelerators, often called XPUs, with large cloud customers, and supplies Ethernet switching and networking silicon used in AI clusters. Most buyers meet Broadcom through cluster networking or a cloud provider custom chip rather than buying its accelerators directly.
- Marvell Custom silicon and data-centre connectivity, including the optical and electrical interconnect that links accelerators, racks and sites.
- Qualcomm A newer entrant in data-centre AI inference, with its AI200 and AI250 accelerators aimed at rack-scale inference. Check software support and availability before planning around it.
- Cerebras Wafer-scale AI systems, offered as hardware and as a cloud service, focused on very large models and fast inference.
- Groq Inference-focused accelerators and a hosted inference service, aimed at low-latency serving.
- Cloud-built silicon Google TPUs, AWS Trainium and Inferentia, and Microsoft Maia are mostly reached through the owning cloud. Availability to outside buyers is changing, so check current terms.
- Arm-based CPUs Increasingly used as host processors alongside accelerators in AI systems and cloud instances.
What it means for buyers
- Check the software stack first. A faster chip does not help if your frameworks, drivers and tooling do not run well on it. Software support often decides the choice.
- Availability follows the channel. Many of these chips reach buyers through cloud services or specific system builders, not off-the-shelf. That shapes lead times and your rent or own decision.
- Networking is a decision in its own right. Switching and interconnect silicon affects cluster performance as much as the accelerator. See networking infrastructure.
How Embedica helps
- Neutral by design. Our detailed platform advice covers NVIDIA, Intel and AMD, where we can compare options in depth. We track other providers so you can weigh them, and we do not claim partnerships with any provider listed here.
- Workload first. We start from what you need to run and work back to the silicon, rather than starting from a vendor.
Common questions
Should I consider accelerators other than GPUs?
Sometimes. Custom accelerators and inference chips can suit steady, well-defined workloads, but software maturity and availability matter as much as raw speed. We compare options against your actual workload.
Can I buy TPUs or Trainium chips directly?
Cloud-built chips such as Google TPUs and AWS Trainium are mostly reached through the owning cloud, though availability is changing. For owned hardware, most buyers still choose from NVIDIA, Intel and AMD platforms. See the platform comparison.
Do these providers face the same export controls?
Advanced AI chips from any provider can be export controlled, depending on classification, destination and end user. See our export controls page.
Next steps
Related: NVIDIA · Intel · AMD · Memory and Storage Suppliers
Weighing accelerators beyond the big three?
