NVIDIA Vera CPU, Anthropic Plaid Finance Tool, and Box Nemotron Debut

Enterprise artificial intelligence infrastructure is expanding across custom processor architecture, consumer financial analysis, and multi-document retrieval. NVIDIA announced its Vera CPU, an architecture engineered specifically for autonomous AI agents that reason, invoke external software tools, and coordinate multi-step workflows across cloud clusters.

In a 100-megawatt data center comparison cited by NVIDIA, the Vera CPU delivers 1.6x more agents per watt than the fastest x86 processor and achieves 2x faster agent execution speed compared with the most energy-efficient x86 CPU. Early infrastructure partners reported substantial performance gains during benchmark testing. Perplexity executed agentic coding workloads 1.5x faster on Vera than on its fastest competing processor and plans to deploy Vera in production environments. DeepInfra recorded up to 2.2x the speed of the best x86 processor on agentic orchestration benchmarks. In database and analytics workloads, ClickHouse scaled near-linearly across 176 cores to claim the top spot on the ClickBench leaderboard, while DuckLabs reached roughly 1.5x the median TPC-H SF1000 query performance of an Intel Xeon 6 baseline. Sandbox providers also measured rapid provisioning: Daytona brought 1,000 isolated agent sandboxes online in 10.8 seconds—compared with 88.4 seconds on an x86 platform—and sustained 12.7 jobs per second across 2,000 concurrent sandboxes, while Modal achieved up to 2x sandbox throughput per core.

While hardware makers optimize server-side agent execution, Anthropic launched a beta Personal Finance tool for Claude tailored to Pro, Max, and Ultra subscribers in the United States. Built on read-only financial connections powered by Plaid, the feature allows users to link bank and credit card accounts so Claude can inspect the institution name, account name and type, the last 4 digits of the account number, current balances, credit limits, and up to 2 years of transaction history. Subscribers can prompt Claude to categorize monthly spending, flag duplicate charges or rising recurring bills, and explain annual percentage rates. Anthropic emphasized that Claude receives strictly read-only visibility, never accesses banking passwords, cannot transfer money or alter account settings, and does not train its foundation models on Personal Finance conversations. The release follows OpenAI's May integration of Plaid into ChatGPT and expands Anthropic's domain offerings alongside Claude for Legal and Claude for Financial Services.

Enterprise document management is undergoing a parallel shift toward agentic search. At its Thursday developer conference, Box unveiled Box Content Retrieval, a retrieval architecture integrating NVIDIA's open-weight Nemotron Embed models, which topped the RTEB retrieval benchmark in July. Scheduled to roll out early next year to Box Enterprise Advance customers, the service consolidates 5 distinct search technologies into 1 unified application programming interface. Box Chief Executive Officer Aaron Levie noted that returning targeted document chunks spares AI agents from ingesting entire files, while NVIDIA Vice President of Generative AI Software Kari Briski stated that the integration is expected to answer enterprise queries 4 times faster with 20% fewer errors and 90% fewer tokens consumed.