Tuesday, Jul 14, 2026

Top stories

Nvidia: performance-per-watt is now the metric that decides AI infra profitability

Nvidia's argument: since power is now AI infrastructure's hard constraint, token-generation capacity per watt — not raw throughput — is what determines revenue and profitability under a fixed power budget. The concrete numbers: GB300 NVL72 claims up to 25x performance-per-watt over Hopper on DeepSeek V4 Pro, 20x on GLM5.1, and 10x on Kimi K2.6 (an agentic-task model); Nvidia also says software optimization alone improved DeepSeek V4 performance up to 5x in one month, and its DSX MaxLPS power-management feature packs "40% more GPUs within the same power budget." The more interesting framing than the numbers themselves: Nvidia pushes Pareto curves over single-point benchmarks (different workloads want different optimization targets) and cites production validation from Anthropic, OpenAI, Perplexity, and CoreWeave rather than lab-only figures. Read this as a vendor argument, not independent verification — the benchmarks are Nvidia's own.

Sources & depth

Codex's encrypted sub-agent messages break local audit trails

A regression report against OpenAI's Codex CLI: PR #26210 (merged 2026-06-05) encrypted MultiAgentV2 payloads for spawn_agent/send_message/followup_task — sound privacy hardening for delivery to the recipient model — but as a side effect it also zeroes out the local, human-readable copy of what was delegated. Parent-side rollout history, trace, and debug surfaces now show ciphertext instead of the task/message text, so after the fact there's no way to answer "what did this sub-agent get asked to do" without decrypting. The reporter's source analysis traces the exact code path (InterAgentCommunication::new_encrypted() initializes plaintext content as empty) and proposes keeping encrypted delivery while adding a separate non-encrypted audit field persisted in rollout metadata — not asking for a revert, just for auditability to survive the encryption. Genuinely relevant to anyone running multi-agent CLI setups today: encrypted-by-default sub-agent comms currently mean you lose your own audit trail.

Sources & depth
Also today