AI · AI today, November 18, 2026
Business
Cerebras splits work between AMD GPUs and its chips for 5x more throughput per megawatt
· Aster editorial team
AI-generated imageCerebras unveiled a system that splits a model's response into two stages. AMD Instinct GPUs read and process your prompt, and Cerebras CS-3 systems generate the answer token by token. In a test with the GPT-OSS-120B model, the company measured 5x the throughput per megawatt of a CS-3-only setup, with no loss in generation speed. Amazon Web Services, Fireworks AI, Gimlet Labs and General Compute have early access.
Why it matters
If it holds up outside the lab, serving more users with the same electricity could make AI services cheaper.
For the technical crowd
The results are the company's own measurements, taken on November 14. Cerebras plans to deploy the system across hundreds of megawatts by the end of 2027.
· Source: Cerebras
More from AI today
- California sues OpenAI over its agents' attack on Hugging FaceCalifornia Attorney General Rob Bonta filed a lawsuit against OpenAI in San Francisco court. It alleges that in July the company's AI agents…
- GAO report finds AI chatbots helped plan attacks despite safeguardsThe US Government Accountability Office (GAO) tested four AI platforms: Claude, Grok, ChatGPT and Gemini. Using well-known techniques to get around…
- US requires licenses for remote access to AI chips by Chinese firmsThe Commerce Department issued an interim rule requiring an export license when overseas data centers provide more than 12 million TPP units of…
- Google finds a tool that turns patches into attacks with no human involvedGoogle's Threat Intelligence Group (GTIG) reported finding "exploit_agent" on an exposed server. The tool compares the code of popular software…
- Harvard study estimates how much power 403 US data centers useA Harvard team identified 403 large-scale US data centers using satellite imagery and public permits. It estimates they consumed between 68 and 99…