Meta enlists tiny Korean startup to build 'one-chip-like datacenter' — CXL architecture introduced by Facebook's parent company and Panmnesia can handle almost 1000 AI GPUs per domain
THE SO WHAT
A CXL architecture that treats up to 960 accelerators as one coherent domain is a bet that memory and interconnect—not raw FLOPs—are the next AI infra bottleneck. If you’re planning large-scale training or inference, start asking cloud and colo partners about their CXL and memory pooling roadmaps, not just GPU counts.
READ THE SOURCE
MORE FROM THE WIRE
Applied AISources: Anthropic told investors it will be profitable for a second straight quarter, with 80%+ gross margins before partner revenue sharing and training costs
Model vendors showing 80%+ gross margins on adjusted bases are telling you two things at once — the core inference business can throw off software-like economics, and real profitability still depends on how you treat rev-share and training capex. For buyers, expect harder pushes toward long-term commitments and platform lock-in as these companies head into IPO mode and optimize for predictable, high-margin revenue.
Applied AIPerplexity trusts GPT-6 Astra with end-to-end systems
If Perplexity is letting Astra change code and touch production with less human oversight, the bar for "deployable autonomy" in software ops just moved. Teams still gating models to low-risk copilots should start scoping where they’d be comfortable with end-to-end execution plus spot checks, not line-by-line review.
Applied AI'That is where the machine starts winning on cost': Expert pits AMD Radeon AI PRO R9700 rig against ChatGPT and gives surprising verdict
The AMD Radeon AI PRO R9700 rig only beats cloud AI subscriptions once your monthly token volume crosses a high threshold — below that, cloud still wins on cost and flexibility. Teams with predictable, heavy workloads should start running real TCO models now, because the on-prem vs cloud crossover point is becoming a concrete budget decision, not a philosophical one.
Applied AIIt's All Fun and Games Until You Give AI Your Credit Card
Letting an AI assistant directly control your credit card turns “hallucination” into financial exposure — the failure mode is no longer a bad answer, it’s a bad transaction. If you’re building or integrating these agents, you need hard limits, human-in-the-loop checkpoints, and dispute paths designed up front, not bolted on after the first incident.