What changed
Positron AI Inc. announced that inference workloads now make up about two-thirds of AI infrastructure usage, compared to one-third for training. This shift is reshaping how companies build and deploy AI systems, with modular data centers becoming a preferred environment for inference tasks.
What this means in practice
Positron's head of sales and business development, Cameron McCaskill, explained that the change reflects a broader industry trend toward smaller, more flexible AI deployments. The company's approach uses commodity memory instead of high-bandwidth memory, reducing costs and power consumption. A single Positron server draws about 4,500 watts, and a full rack of Titan servers operates within 40 kilowatts, making them suitable for sites with access to cheap power and minimal regulatory hurdles.
Limitations and availability
Positron's first-generation Atlas server runs on a programmable chip, and the company plans to release its own silicon for the Titan server early next year. Oracle is deploying about 340 Atlas servers in its cloud this quarter, and a neocloud provider has also committed to using Atlas and Titan. The company's design choice to avoid high-bandwidth memory allows for lower power usage, but it remains to be seen how this approach compares to industry standards in performance and scalability.
