AWS is deploying Cerebras CS-3 systems in its data centers, pairing Trainium for prefill with the Wafer Scale Engine 3 for decode. Why disaggregated inference is the right architecture, and what makes a 4-trillion-transistor chip the right tool for the decode problem.
Wafer-scale inference on Bedrock and open 3T-class models both landed this month. Neither will run your robot's control loop — but together they redraw the line between what a robot computes on-board and what it can safely delegate to the cloud.
The 5 key concepts every cloud architect should know about LLM serving: PagedAttention, KV cache mechanics, continuous batching, MoE trade-offs, and real production numbers.
A practical walkthrough of two paths to working with Mistral — the managed API for fast prototyping and self-hosted deployment for full control — with real code covering prompting, model selection, function calling, RAG, and INT8 quantization.
What TFLOPS actually measures, why FP16 matters for LLMs, and why the most important GPU bottleneck for inference isn't compute at all.