The Circuit EP 158: Arm 2.0, Memory Freak Out, and AI Models
36 snips
Mar 30, 2026 A deep dive into ARM's bold move from licensing to selling CPUs and what that means for the CPU market. Discussion of agent-native versus cloud-native chips and whether ARM will rival x86 players. Exploration of memory market worries and the TurboQuant reaction. Reaction to leaks about Mythos and rapid shifts in AI model leadership.
AI Snips
Chapters
Transcript
Episode notes
Buying Arm Chips Lets Hyperscalers Focus On Accelerators
- Offering Arm-designed merchant chips may let hyperscalers focus R&D on AI accelerators while buying agent CPUs instead of building them.
- Ben suggests this build-vs-buy shift could change hyperscaler decisions for agentic infrastructure, particularly for companies that won't prioritize CPUs.
Arm Roadmap Implies Hundreds Of Cores Per CPU
- Arm's roadmap shows rapid scaling: announced chips with ~136 cores now, roadmap items implying 250+ and 500-core CPUs in future generations.
- Jay notes that 500-core CPUs blur lines between CPU and GPU topologies and indicate aggressive core-count scaling for agentic workloads.
Arm Shows Big TAMs Beyond CPUs Which Raises Questions
- Arm included accelerator and robotics TAMs in its presentation, prompting speculation about entering accelerators despite no public accelerator IP roadmap.
- Ben and Jay note Arm likely needs partners for accelerators and left mobile ambitions open-ended, signaling future optional expansion.