OpenAI presented benchmark results for Jalapeño at Hot Chips 2026, claiming its Broadcom-co-developed 700W inference ASIC delivers 1.5x to 1.9x more throughput per kilowatt and 1.7x to 3.6x lower end-to-end latency than Nvidia GB200 and GB300 systems.

• Six HBM4 stacks power the chip, which was designed around minimizing data movement and communication delays for modern LLM workloads, with limited deployment expected in OpenAI data centers by end of 2026 and broader rollout in 2027.
• The vertical integration play is explicit: OpenAI built chip, memory, network, software, and rack-scale system together, breaking the dependency on Nvidia merchant silicon margins that has constrained inference economics.

Google Cloud today shipped Gemini Enterprise for Financial Services and Gemini Enterprise for Legal simultaneously, purpose-built agentic platforms for capital markets, corporate banking, and law firm workflows.

• The legal product was developed directly with Cleary Gottlieb, Freshfields, Weil, and Williams & Connolly, and targets billable-hour automation rather than generic chat, signaling Google is cutting out legal tech middleware vendors.
• The financial services product includes a Google-managed Financial Research agent, 50-plus specialized skills, enterprise data connectors, and a third-party agent ecosystem built on the Gemini Enterprise platform.

Anthropic shipped unified memory across Claude chat and Claude Cowork on August 25, making context bidirectional and persistent across both surfaces.

• Memory is on by default for Free, Pro, and Max plans; Team and Enterprise orgs default to off, and users can edit or delete any remembered topic under Settings > Memory.
• Sensitive topics such as health and beliefs are excluded from memory unless the user explicitly enables them, a privacy control that reflects enterprise compliance pressure Anthropic has faced since Cowork launched.

GitHub's Copilot coding agent can now ingest Microsoft Teams conversation history as context for a coding task, eliminating the manual step of re-explaining decisions already made in chat.

• The integration combines what engineers have discussed in Teams with repository and issue data, giving the agent a fuller picture of intent before it writes a single line of code.
• The practical unlock is for distributed teams where product decisions live in chat threads: the agent can now act on those decisions without a human translating them into a ticket.

Salesforce measured production agent activity across 400 businesses and found the average number of agents per organization grew from five to thirteen in fifteen months, with 70% of customer-service sessions now completing without a human.

• Agent build time fell 53% over the same period, while escalation rates held steady — the metric Salesforce uses to distinguish genuine deflection from contacts being routed to harder-to-find channels.
• The 400-company sample represents production deployments, not pilots, making this one of the more credible data points on enterprise agent scale yet published.

Paxos reported that its Hoplites AI coding agents, operating inside Slack, now account for 15% of all merged pull requests as of July 2026, while also handling incident response automation and enabling non-engineers to contribute to internal tooling.

• Agents run non-critical engineering tasks autonomously within Slack channels, cutting pull request cycle times without requiring engineers to context-switch to a separate coding environment.
• Non-engineer contributions to internal tools increased as a direct result, a workforce expansion effect that goes beyond the productivity narrative typically attached to coding agents.