AMD Courts Anthropic While Coding Tools Learn to Route Models

AMD sells compute by the gigawatt

AMD's announcements this week cover training, rack integration and inference. Its partnership with Anthropic covers up to 2 gigawatts of Instinct MI450-series GPUs, with AMD investing up to $5B in the buildout, so the chipmaker is helping fund a customer's capacity rather than only selling parts. At its Advancing AI event, AMD also launched Helios, a rack-scale AI systems platform positioned against Nvidia's rack-level products. On the inference side, AMD and Cerebras announced a joint solution promising ultra-low latency and high throughput, aimed at workloads where response time matters more than model size.

Hardware claims elsewhere are drawing more scrutiny. DeepSeek's earlier claim of training a model on Huawei chips now has benchmarks, and skeptics who remain unconvinced, which leaves open whether domestic Chinese silicon can match Nvidia-class training performance.

Routing replaces the single-model bet

Cursor introduced a router that sends simpler coding tasks to cheaper models and reserves expensive reasoning models for harder ones, reportedly cutting overall model costs by about 30%. Coverage of the launch asks whether expensive reasoning models are always worth invoking and treats routing logic as a core part of coding-agent design. Runway is applying the same idea to images and video with a router that picks among generative media models by task, betting that specialized models beat one in-house model in a crowded market. For teams watching inference bills, deciding which model handles each task may become as important as which models are available.

Further up the stack, Sierra acquired TakeOff, a platform for long-horizon agent tasks, folding it into its customer-facing agent product instead of building that capability itself.

Voice and image models multiply

Voice is getting more attention from assistant makers. OpenAI extended GPT-Live's full-duplex voice control to Codex and the ChatGPT desktop app, so developers can talk to coding agents in real time and interrupt naturally, while Anthropic's updated Claude voice mode brings more capable conversational interaction. Microsoft AI released MAI-Voice 2 Flash for fast speech synthesis alongside MAI-Image 2.5 Pro, adding to its first-party models; that may give it more cost and latency control than routing everything through partners.

On images, Black Forest Labs published FLUX 3, its latest generation model, as it tries to keep pace with closed frontier image models. A separate analysis argues open weights are being unbundled into specific business layers rather than pursued as a single strategy, which could explain why labs weigh releasing weights so differently.

Usage is rising faster than results

Several pieces suggest adoption is outpacing measurement and hygiene. A product-leadership essay on the AI productivity paradox notes that teams report heavy tool usage while aggregate productivity and quality metrics have moved less than expected, an argument for tracking outcomes rather than usage. Security data shows a sharp rise in exposed AI tools discoverable on the open internet, widening the attack surface for organizations that deployed quickly. For a contrast in process discipline, a long interview on Poolside describes how it organizes teams, data pipelines and iteration loops into a repeatable frontier model factory.