
What happened
Positron's Cameron McCaskill said about two-thirds of AI infrastructure workloads are now inference versus one-third training, down from roughly 85% training in 2023.
Why it matters
With inference driving the economics, McCaskill said buyers are favoring cheap power and fast deployment, which suits Positron's commodity-memory, air-cooled servers.
WHO IT HITSBuyers of AI infrastructure — cloud providers, neoclouds and modular data center operators — now weigh power, cooling and deployment speed as much as chip performance, according to McCaskill. Inference-first chipmakers such as Positron are positioning to capture that shift.
Summaries like this, every morning.
McCaskill's figures trace a rapid reversal. In 2023, training dominated AI infrastructure workloads at roughly 85% versus 15% for inference. Today the split is about two-thirds inference to one-third training, according to McCaskill. That change matters because inference workloads have different physical requirements: they need cheap power and fast deployment more than the dense, expensive memory used in training clusters.
Positron's design reflects that. The company skipped high-bandwidth memory — which McCaskill described as very limited in supply and very expensive — in favor of commodity memory. That choice keeps a single server at about 4,500 watts, low enough for air-cooled rooms, and a full rack of Titan servers within 40 kilowatts. The company's first-generation Atlas server ran on a programmable chip; Positron plans to tape out its own silicon early next year for Titan. Oracle is bringing about 340 Atlas servers live in its cloud this quarter, and a neocloud provider has signed on for Atlas and committed to Titan.
The modular data center market fits this profile. McCaskill said operators with half-megawatt to 2.5-megawatt modular sites can deploy quickly, and that coupling those sites with Positron's technology can help inference players reach monetizable tokens within 90 to 120 days. Positron raised $875 million at a $5 billion valuation in September, positioning it to compete as inference spending shapes where chips get deployed.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
An Nvidia research team says feeding models entire novels is often expensive overkill; in many tests, periodic…

The writer told Muse only their height and that they wanted slim, cropped tops, then let it browse ZOZOTOWN in…

After a reader comment, the Search Console-based service 検索カルテ began attaching explicit start and end dates (f…

A practitioner customizing an agent called Muse had other models review documents and design proposals, found…

Shortical premiered "Inevitable" in Los Angeles last week, a fully AI series starring a digital avatar of Isra…

OpenAI published copy-ready prompts for GPT-6.2 Astra and Sol covering writing style, initiative, and coding…
