Planet AI Weekly: Planet AI Weekly May 10, 2026

This week we reviewed 241 articles from 16 sources. Here is what mattered.

Agentic infrastructure dominated: NVIDIA and ServiceNow shipped a joint agent runtime, OpenAI hardened Codex for production and pushed it into the browser, and the Musk-OpenAI trial surfaced fresh detail on early equity and governance fights. Vendors filled in the plumbing with new networking fabrics, GPU capacity primitives, and inference engines.

Official Highlights

NVIDIA and ServiceNow ship a joint framework for autonomous enterprise agents that act on internal systems rather than just generate or reason. Worth a read if you are wiring agents into existing ITSM or workflow stacks. Read more

OpenAI and PwC embed agents into CFO workflows for forecasting, controls, and close processes, targeting measurable automation of existing finance stacks rather than a new model release. Read more

OpenAI documents how it runs Codex internally: sandboxing, approval gates, network policy, and agent-native telemetry. Concrete policy examples teams adopting coding agents can lift directly. Read more

Google DeepMind details where AlphaEvolve has produced measurable wins across infrastructure optimization, scientific code, and business workflows. Useful as a reference for what a Gemini-backed coding agent actually moves the needle on. Read more

NVIDIA extends Spectrum-X Ethernet with MRC for gigascale clusters and pitches it as the default scale-out fabric for large training runs. Network engineers evaluating next-gen interconnects should compare the published numbers. Read more

From the Community

NVIDIA researchers pack 30B, 23B, and 12B reasoning models into one Star Elastic checkpoint with zero-shot slicing, eliminating separate training runs per scale. Read more

OpenAI ships a Codex Chrome extension that operates inside signed-in sessions on LinkedIn, Salesforce, Gmail, and internal tools, with DevTools and page-level actions exposed. Read more

LightSeek Foundation open-sources TokenSpeed, an inference engine targeting TensorRT-LLM parity for agentic coding workloads under high tool-calling concurrency. Read more

Greg Brockman testified that his OpenAI stake exceeds $30B and recounted a 2018 confrontation with Musk over board control. Primary-source detail on the equity allocations and governance disputes now central to the trial. Read more

NVIDIA released cuda-oxide, a rustc codegen backend that compiles #[kernel]-annotated Rust functions to PTX through a Stable MIR to Pliron IR to LLVM IR pipeline, with single-source host plus device builds. Read more

NVIDIA has committed $40B to AI equity deals so far this year, deploying capital well beyond chip sales into the broader ecosystem. Read more

Cloudflare cut 1,100 roles it attributes to AI-driven efficiency gains while posting record revenue. The cuts hit support and operational roles previously needed at scale. Read more

Zyphra introduces Tensor and Sequence Parallelism, a folded strategy that reduces parameter and activation memory on the same GPU axis and reports 2.6x throughput over matched TP+SP baselines. Read more

Featured This Week

Arham Islam's guide to systematic prompting treats negative constraints, structured JSON enforcement, and multi-hypothesis verbalized sampling as production techniques rather than ad-hoc tricks. The walkthrough shows how each method cuts variance on tasks where reliability matters more than creativity, with concrete prompt patterns and failure-mode analysis. If you are pushing agents or structured pipelines into production and still iterating on prompts by feel, this is the reference to read end-to-end. The verbalized sampling section in particular is worth the click for teams currently relying on temperature sweeps to surface alternative hypotheses. Read more

Send comments, corrections, or submissions to planet-ai.net. We review every link.