AMD's Anthropic deal, new image models, and agent cost discipline

AMD assembles a full alternative compute stack

AMD's announcements cover supply, systems and serving at once. The company announced a strategic partnership with Anthropic to deploy up to 2 gigawatts of Instinct MI450 GPUs, tied to up to $5B in compute investment. It also unveiled Helios rack-scale AI systems and launched a joint ultra-low-latency, high-throughput inference offering with Cerebras. The rack-scale focus is relevant because, as one essay argues, a GPU-hour is not fungible across vendors and configurations. Interconnect topology, memory bandwidth and cluster design all change effective throughput, so comparing AMD and Nvidia capacity on headline hourly price can mislead anyone negotiating compute contracts.

Claims about training away from Nvidia also need outside checking. DeepSeek's earlier claim of training a model on Huawei silicon has now been put through independent benchmarking, with skeptics raising methodology questions. Chip-independence claims are only as strong as results others can reproduce.

Image and voice models arrive from every direction

Generative media now has credible new options from both open and closed camps. Black Forest Labs' FLUX 3 continues its line of openly licensed diffusion models, giving developers an alternative to closed APIs. Microsoft AI introduced in-house MAI-Image 2.5 Pro and MAI-Voice 2 Flash, adding first-party models alongside its existing partnerships. Runway is declining to bet on any single model: its new router selects among underlying generative models by task. A VC essay on the open-weight unbundling argues that as open weights close the quality gap, advantage shifts toward distribution, tooling and data. Runway's orchestration layer may be an early example of that kind of tooling play.

Making coding agents cheaper and sturdier

For teams running agents in production, the most practical material concerns cost and reliability. A deep dive on prompt caching in agents explains how cache design decisions affect cost and latency as reasoning-heavy loops multiply token consumption. On the reliability side, claude-thermos, a new GitHub project, offers a way to stress-test and insulate Claude-based agent workflows. The interface is changing too: OpenAI brought GPT-Live's full-duplex voice into Codex and the ChatGPT desktop app, so developers can speak to and interrupt a coding agent in real time instead of typing prompts.

At the platform level, Sierra acquired TakeOff, a long-horizon agent platform, folding its technology into Sierra's customer-facing agent stack. For engineering leaders comparing their own ML organization with a lab built to produce frontier models, Latent Space's long profile of Poolside covers the data pipelines, eval infrastructure and organizational design behind a challenger operation.

Skeptics question the productivity story

Not every piece takes the spending at face value. SVPG argues that measured productivity gains from AI coding assistants lag expectations, pointing to heavier code review burden and declining quality downstream. A much longer essay goes further, claiming AI mania is degrading corporate and institutional decision-making as leaders defer to AI-generated analysis and hype cycles. Google, meanwhile, published research proposing a framework for measuring how AI reshapes economic activity, pushing back against simplistic productivity-stat narratives. For engineers using coding agents daily, the SVPG argument suggests review capacity, rather than generation speed, may be the constraint that determines real gains.