NVIDIA says its Vera Rubin NVL72 system delivered leading performance in its debut on MLPerf Inference v6.1. The company links inference economics to three levers: higher system performance generates more tokens and revenue; efficient scaling lets throughput grow proportionally as hardware is added, requiring fewer resources to serve users at scale; and continuous software optimization creates more value from infrastructure investments.
At NVIDIA’s AI Infra Summit, Ian Buck, vice president of hyperscale and high-performance computing, spoke about AI-factory efficiency. NVIDIA also highlighted its DSX platform and the energy efficiencies of optimizing tokens per watt for AI factories.
The summit drew more than 8,000 attendees, up from 3,500 last year, as NVIDIA put infrastructure performance, scaling, software optimization and energy efficiency at the center of its inference pitch.
Why it matters: For AI-factory operators, NVIDIA’s pitch ties higher system performance to more tokens and revenue, while efficient scaling means throughput can grow with fewer resources.
Wired AI says AI legislation is unlikely despite mounting concerns that models could go rogue.
The White House opposes oversight outright, according to the report.
Why it matters: The White House’s opposition means mounting concerns about models going rogue are unlikely to produce federal AI legislation anytime soon.
Anthropic is merging Claude Chat and Cowork into one product. Instead of choosing between interfaces, users can submit a task and Claude will decide whether it needs a quick answer or a larger workflow.
The update also adds Claude Docs and Claude Slides. Anthropic is initially releasing the features to Pro and Max subscribers; no broader availability date or pricing change has been published.
Why it matters: Pro and Max subscribers no longer have to choose between Claude Chat and Cowork before starting a task; Claude makes that routing decision inside one product.
AWS published four implementation guides covering agent optimization, prompt caching, document redaction, and product tagging. Amazon Bedrock AgentCore turns production traces into proposed agent-configuration changes and validates them before promotion. Bedrock prompt caching can cut input-token costs by up to 90% when applications repeatedly send the same context.
The other two guides target specific workflow bottlenecks. Bedrock Data Automation combines a custom blueprint with AWS Step Functions and AWS Lambda to detect and redact PII from scanned documents; token matching raises recall across degraded and handwritten documents. On Amazon SageMaker, AWS shows how to customize Qwen3-8B with supervised fine-tuning and reinforcement learning with verifiable rewards, then deploy it for asynchronous product-tagging inference.
The guides describe technical implementations rather than customer adoption or production volumes. They cover six prompt-caching scenarios through the Converse API, including system prompts, tool definitions, mixed TTL, tenant isolation, and LangChain integration.
Why it matters: For ML engineers running Bedrock applications, repeated context can cost up to 90% less in input tokens; teams building catalog workflows can customize Qwen3-8B and run asynchronous product-tagging inference on SageMaker.