Astera Labs has announced a third generation of its Leo CXL memory controller line, and the new top-end part is built to plug memory directly into the network fabric that connects AI accelerators, not just into a host CPU.
Leo 2 doubles bandwidth for CPU-attached memory
The existing Leo 2 E-series and P-series controllers, aimed at CPU-attached memory expansion, move from PCIe Gen5/CXL 2.x to PCIe Gen6/CXL 3.2, the newest shipping revision of both standards and the one the first compatible host CPUs are launching with. Astera Labs doubled the memory channels per controller from two to four to keep pace with the extra PCIe bandwidth, which in turn doubles both bandwidth and the memory each expander can carry: up to 768GB of DDR4 or 4TB of DDR5 behind a single expander. CXL 3.x also adds multi-level switching and direct peer-to-peer memory access between expanders.
A specific use case behind the update is keeping older DDR4 in service. New DDR5 is expensive, and high-capacity DDR5 more so, while DDR4 pulled from retired servers is cheap. ServeTheHome’s writeup details the mechanism: a hardware test engine runs multiple patterns across each DIMM and layers software ECC on top of the DIMM’s own hardware ECC, so a DDR4 module that develops a fault can be corrected and kept in the pool instead of pulled from service, a real risk with memory that has often sat in service for half a decade already.
Leo X moves memory onto the accelerator fabric itself
The new Leo X-series is Astera Labs’ bigger move: a fabric-attached controller that pairs with the company’s own Scorpio PCIe switch silicon so a memory expander can join the network fabric behind a rack of AI accelerators directly, rather than routing through a host CPU. The company hasn’t named which accelerator fabrics it supports beyond the PCIe-based Scorpio switching, saying only that it’s targeting hyperscalers running a mix of vendor and homegrown accelerators. Leo X can sit either outside a rack in the scale-out network or inside a node in the tighter scale-up domain, and the first-generation part ships with four onboard DIMM slots supporting both DDR4 and DDR5.
The target application is offloading key-value cache, the memory-hungry state that accumulates as an AI model works through a long context or an agent chains several reasoning steps. Astera Labs says its own testing improved time to first token by as much as 62% and overall token throughput by 22%. That’s a vendor benchmark rather than an independent result, but the direction fits where the rest of the industry is headed: NVIDIA’s own Context Memory Storage push is aimed at the same decode-phase bottleneck. All three new Leo controllers are sampling with hyperscaler customers now, per Astera Labs, not yet in general availability.
AWS open-sources an inbox-style manager for AI agents
Separately, AWS’s own announcement introduces Pizza Bot, an open-source tool that manages AI agent tasks as email-style threads instead of a live chat window. Finished tasks land in an Unread category with a full writeup; anything needing a decision goes to Action. The team built it internally at Amazon for agent uses outside of coding, where the recurring need was background work that only interrupts the user when it’s done or needs a decision, closer to how a delegated colleague reports back than how a chat session behaves.
Pizza Bot runs on-device for Windows, macOS, and Linux, stores its threads locally in SQLite, and doesn’t send anything off the device without permission. It supports Ollama for locally run open-weight models alongside hosted providers, and can drop in Claude Code-compatible .mcp.json files and skills directly. Out of the box it can read, write, edit, and search files, delegate to specialist agents, run a sandboxed JavaScript interpreter, and automate a browser. AWS is explicit that this is open-source code released for outside use, not an officially supported product: there’s no AWS SLA behind it.