OpenAI's Jalapeño chip beats Nvidia in inference benchmarks

OpenAI unveiled benchmarks for its Broadcom-built Jalapeño inference chip, claiming 1.5–1.9x more throughput per kilowatt and up to 3.6x lower latency than Nvidia's best systems, with deployment starting late 2026. Separately, an unconfirmed report says OpenAI finished pretraining 'Bel,' a 10T+ parameter base model for Astra and GPT-6. Anthropic unified Claude's memory across chat and Cowork with user-visible editing, and Google Cloud launched Gemini Enterprise for Financial Services.

Written by AI from the sources it cites; it can be wrong — follow any citation to confirm.

14 sources2 min read

OpenAI's Jalapeño chip beats Nvidia in inference benchmarks, claims 1.9x efficiency At Hot Chips, OpenAI and Broadcom's custom inference ASIC delivered 1.5–1.9x more throughput per kilowatt and 1.7–3.6x lower latency than Nvidia's best available systems on SemiAnalysis' InferenceX benchmark. OpenAI plans small deployments by end of 2026 and larger rollout in 2027, with SemiAnalysis noting it also edges Nvidia's upcoming Vera Rubin on output tokens per megawatt. 1 2 3 4 5

OpenAI reportedly finishes 'Bel' pretraining, a 10-trillion-parameter base model An unconfirmed report from X user @synthwavedd says OpenAI completed pretraining on 'Bel,' its first full-scale frontier pretraining since GPT-4o, with over 10 trillion parameters. It is expected to serve as the base for Astra and GPT-6 successors, though OpenAI has not confirmed the model. 6 7

Anthropic unifies Claude memory across chat and Cowork with user controls Claude now remembers topics from chat when users switch to the Cowork desktop agent, and users can view, edit, or delete memories topic by topic. Sensitive subjects such as health, race, religion, and politics are not stored by default unless users opt in. 8 9 10

Google Cloud launches Gemini Enterprise for financial services and legal AI agents Google Cloud debuted the first industry-specific agentic AI offerings on Gemini Enterprise, starting with a Financial Research agent for capital markets and corporate banking plus more than 50 specialized skills, and a legal-industry solution. The financial services offering is in preview. 11 12

OpenAI launches AI Futures research team for global governance questions OpenAI's new Strategic Futures team, led by former White House policy adviser Dean Ball, launched 'AI Futures,' a research blog exploring how free societies can preserve individual rights as transformative AI decouples power from labor. The six-person team reports to Chief Strategy Officer Jason Kwon. 13

Google rolls out Gemini's interactive 3D models and simulations to Workspace Gemini's dynamic simulations, models, tables, and grids are now available to Workspace customers, Workspace Individual subscribers, and eligible users in the Gemini app. The feature, announced in April, lets users rotate and zoom into generated 3D models like a DNA double helix. 14