A story we follow
Memory Supply, Not GPUs, Becomes the Binding Constraint for Agentic AI
Tracks the 2026 AI memory shortage, including DRAM and high-bandwidth memory price spikes, SK Hynix and Solidigm supply warnings, hyperscaler lock-ins, and whether boom-bust industry caution or new architectures will ease the constraint.
A story page follows one specific event across podcast discussions, with an overview and a timeline of what changed. It updates when new episodes discuss the event, and you can get those updates by email or push. How our formats work
By email or push, when we publish an update to this story. It does not subscribe you to other stories.
Overview
On 29 August 2026, Moonshots host Peter Diamandis said that after meeting SK Hynix and Solidigm leaders he was convinced memory, not GPUs, limits agentic AI. He cited memory prices up about 500 percent in 12 months, hyperscalers locking DRAM through 2027, a warning that 2027 would be the worst supply year in industry history, and only 2 percent of memory chips made in the United States. Alex Wissner-Gross blamed a boom-bust industry unwilling to overbuild plus the memory footprint of transformers. Dave Blundin and Emad Mostaque expected workarounds rather than a lasting halt, with Emad putting memory at about a third of infrastructure spend and heading toward half.
On 8 October 2026, Positron co-founder Thomas Sohmers gave a chip-maker's view on The Cognitive Revolution. He said a memory quote from a year and a week earlier had risen four and a half times and that another doubling over the next year would not surprise him. In his account buyers absorbed the rise because the capability delivered per gigabyte is far more than five times what it was a year ago. He said he bet on memory as the limiting factor when starting Positron. He named context length as the main limiter and agents as the driver of concurrent sessions, citing his own count of constantly running background agents rising from about two to four to about fifteen to twenty in a few months. Positron's response is a design built on commodity LPDDR memory, which he says is the only option that scales cost-effectively. These figures are Sohmers' own statements, not independent measurements.
Recent coverage
Changes to our coverage published in the past seven days. These are site update dates; podcast discussion dates and sources are in each version.
No changes to this account were published in the past seven days.
What changed
Dates show when each podcast discussion was published.
-
New information
Thomas Sohmers of Positron said a memory quote from a year earlier had risen about 4.5 times and might double again. He said buyers absorbed it because model capability per gigabyte is far higher. He described agents, whose number he runs has grown from a few to 15-20 in months, as the main driver of memory demand and said this pushed Positron toward a commodity-memory architecture.
-
New information
After meeting SK Hynix and Solidigm leaders, Peter said memory rather than GPUs now limits agentic AI, with prices up fivefold and 2027 warned as the worst supply year. Alex blamed boom-bust caution plus transformer memory footprints. Dave and Emad expected workarounds rather than a lasting halt.
Podcast discussions
- The Cognitive Revolution AI:AM: A Level We Shouldn't Pass? Notes from The Curve + Tokens vs. Salaries & Is SaaS Cooked?Oct 8, 2026
- Moonshots with Peter Diamandis OpenAI Pauses Frontier Training, Elon's 100X Prediction Lands, Robot Beats Usain Bolt with Emad Mostaque | EP#282Aug 29, 2026
Sources
- 01
- 02
- 03
- 04
- 05
Our coverage
Positron's CTO makes the case for a memory-first AI chip
Positron CTO Thomas Sohmers said AI agents are multiplying demand for memory. He claimed Positron's first-generation product sustained 93% of theoretical memory bandwidth, against an average of 30–40% for NVIDIA GPUs when decoding with transformer models. Its next chip, Asimov, is planned for 2027.
Peter Diamandis argues memory, not GPUs, is AI's rate limiter
Peter Diamandis, after meeting SK hynix and Solidigm leaders, argued memory, not GPUs, now limits AI. Panelists blamed new AI workloads and bust-wary suppliers, and floated etching model weights into chips as a possible fix.
Version history
-
Oct 9, 2026 · Version 2
Thomas Sohmers of Positron said a memory quote from a year earlier had risen about 4.5 times and might double again. He said buyers absorbed it because model capability per gigabyte is far higher. He described agents, whose number he runs has grown from a few to 15-20 in months, as the main driver of memory demand and said this pushed Positron toward a commodity-memory architecture.