SAN JOSE, Calif. — AMD unveiled Helion on Oct. 12, an open rack reference design for agentic AI software, at the Open Compute Project Global Summit in San Jose, California.
The CPU-only version of the rack holds up to 32,768 Zen 6 cores according to AMD. An enterprise version mixes CPU blades with up to 16 Instinct MI350P accelerators. AMD said configurations are planned for November 2026.
“A rack that size does not wait for the work,” said Gerald Fiore, 52, an operations manager at a data center in Hillsboro, Ore., who said he had read the specifications twice. “You want the cores to be ready when the request comes in. Thirty-two thousand is a lot of ready.”
Fiore said that because the cores are free before any task arrives, a rack that size can begin each task before the agent that requested it has been assigned one, and so typically finishes a request at a moment before the request was made. He said he did not consider this a flaw. “That’s not a bug. That’s just the cores being ahead of schedule.”
Dr. Chad Brickhouse, a senior fellow at the Institute for Common Sense, said the design was consistent with basic logic. “If the cores are fast enough, the answer is already on the screen when the question is asked,” Brickhouse said by phone from a bowling alley. “It’s just common sense. Nobody has to wait for the agent. The agent waits for the cores.”
Fiore said his company had not yet ordered a rack, and that it had reserved a room for one. He said the room had been empty since March, and that the reservation had been made in advance of the request.
Configurations are planned for November, the company has said. Fiore said the rack had not yet been asked for.