Nvidia's board moves to oust Huang as Gemini Decisions API prices decisions at $0.10
7 min read · 20 sources
- Nvidia's board is reportedly preparing to oust CEO Jensen Huang; shares slid over 8% after hours.
- Google's Gemini Decisions API scores up to 1,000 questions per pass at $0.10 per 1M input tokens with 150-250 ms latency.
- Runway's Praxis-1, a 27B-parameter open-weight robotics model, tops NVIDIA's RoboLab-120 leaderboard at 46.2% success.
- OpenRouter found a GitHub MCP server consumed 59,417 input tokens versus 1,421 for a CLI on the same task.
- United Airlines announced it will block AI agents from its platforms over credential security and regulatory risks.
The AI hardware supply chain just got a whole lot less certain. Nvidia’s board is reportedly preparing to oust CEO Jensen Huang following allegations of an affair with a woman U.S. intelligence believes was a Chinese asset, and the share price slid over 8% in after-hours trading. That is the story that touches every other story in this newsletter - every lab, every cluster, every training run depends on Nvidia’s leadership staying stable.
Meanwhile, Google quietly shipped something that could change how much of your agentic infrastructure you actually need. The Gemini Decisions API prices a decision at $0.10 per million input tokens, and it makes up to 1,000 of them in a single forward pass. That is not a chatbot endpoint. That is a routing primitive.
A GitHub MCP server burned 59,417 input tokens on a task where the same CLI needed just 1,421.
Gemini Decisions API: Decisions without generation
Google’s new Gemini Decisions API is a different kind of endpoint. Instead of generating tokens, it scores up to 1,000 independent questions against a single multimodal input - text, images, audio, video, PDFs, up to 256K tokens - in one forward pass. It returns probabilities and confidence scores with 150-250 ms end-to-end latency, and it bills zero output tokens.
The model behind it, gemini-decision-1, handles yes/no, multiple choice, and rubric scoring. Google claims it is nearly 5x cheaper than Gemini 3.5 Flash-Lite with the Interactions API. For engineers, this replaces a whole class of token-by-token generation with a calibrated classifier: tool selection, policy checks, model routing, content moderation. If you have been building those as prompt-and-parse loops, this is the thing that kills them.
The tradeoff is that it is closed-set. You get probabilities, not prose. If your pipeline needs an explanation along with the decision, you still need a generative model - but for the decision itself, this is dramatically cheaper and faster.
Runway Praxis-1: Open-weight robotics hits the leaderboard
Runway released Praxis-1, a 27B-parameter mixture-of-experts diffusion transformer that predicts video and actions jointly, under Apache 2.0. It tops NVIDIA’s RoboLab-120 leaderboard with 46.2% success across 1,200 episodes, beating FLUX 3 Action (42.9%), Cosmos3 Nano (36.8%), and π0.5 (28.0%).
The model runs on NVIDIA hardware from H100/B200 down to RTX 5090 and Jetson Thor, with full fine-tuning and LoRA support. It is on GitHub, Hugging Face, and vLLM-Omni. The interesting number for robotics engineers: it adapts to new embodiments with light fine-tuning, reaching 80.0% success on SimplerEnv with Google’s RT-1 robot versus the original 52.4%. That is the difference between a model that works on one robot and a model that works on your robot.
MCP vs CLI vs Skills: The hidden cost of tool definitions
OpenRouter measured what giving agents tools actually costs in tokens, and the numbers are brutal. On a GitHub task using Claude Sonnet 5.5, the CLI used 1,421 input tokens, the skill used 3,315, and the GitHub MCP server used 59,417 - its 46 default tools adding 19,436 tokens per request. A single tool result, GitHub’s release object, cost about 20,400 tokens.
The mechanism is simple: tool definitions are billed as input tokens on every request. MCP servers that ship dozens of tools inflate context on every single call, whether the agent uses them or not. OpenRouter’s Tool Search server cuts the up-front tool cost to under 600 tokens, halving total cost on Sonnet 5.5. If you are running agents in production, this is the difference between a viable unit cost and a bill that eats your margin. The fix is lazy tool loading - defer definitions until the agent actually needs them.
Grok Bot 1.0: Team Bots go to production
Grok Bot 1.0 shipped on November 11, adding Team Bots that can be published and shared, Slack and Microsoft Teams support, meetings on Google Meet and Zoom, shared connectors, and reviewed learnings. Version 1.0.0/1.0.1 is out of beta, following the November 2 announcement.
Team Bots now start as drafts and require publishing, with sharing options for teams, organizations, people, or groups. Meetings are available on Team, Enterprise, and Ultra plans, and the app must be updated to 1.0 to access the new features. For engineers, the interesting part is the admin controls and the publish workflow - this is a personal assistant turning into a shared team resource, with the governance that implies.
AlphaGenome 2: Variant effects at cell-type resolution
DeepMind announced AlphaGenome 2, a DNA sequence model that predicts how genetic variants affect gene expression in hundreds of human cell types. Trained on over 100 million individual cell profiles across 68 tissues, it identifies nearly twice as many expression-altering variants as the previous model.
It is available for non-commercial research via the AlphaGenome API, and in private preview for commercial research via Google Cloud. The practical value: variant effect prediction in disease-relevant cells that broad-tissue models miss. For rare disease and complex trait research, that is the difference between a hit and a miss.
A2UI 1.0: Agents answer with screens, not text
Source: developers.googleblog.com ↗
Google released A2UI 1.0, the first stable version of an open protocol for generative UI. Agents respond with interactive screens instead of text. Open-source renderer SDKs cover web, Android, and iOS; agent SDKs cover Python, TypeScript, and Dart. Atlassian, Bloomberg, and ByteDance are already using it.
The architectural point: A2UI binds data on the server so sensitive values never reach the model, and it streams UI as the agent works. Users get native-fidelity interfaces whether they reach an agent in its own app or through a third-party assistant like Gemini or Claude. This is the pattern for agent interfaces that feel like apps rather than chat.
United blocks AI agents, Nvidia's board moves, and the rest
United Airlines announced it will restrict AI bots’ access to its digital platforms, citing data security and regulatory risks. Sharing account credentials with third-party assistants could expose sensitive information, cause unintended purchases, or misuse frequent-flyer miles; automated check-in could create regulatory issues. It will update booking, ticketing, and distribution policies to block agents directly and via travel agencies. This is the industry diverging - Amazon blocked Meta’s Muse, Walmart partnered with it.
Nvidia’s board is reportedly preparing to oust Huang following allegations of an affair with a woman U.S. intelligence believes was a Chinese asset, and a separate affair with VP of healthcare Kimberly Powell. The Wall Street Journal first reported the board’s moves. The allegations involve Alice Huang, who led Megaspeed International and bought billions in Nvidia chips. Shares slid over 8% after hours.
Quick hits
Shopify’s generative discovery model is in production, trained jointly on search and recommendation data with a 9.85% lift in conversion rate, handling a catalog of billions within live request latency budgets.
OpenAI open-sourced MCP Extensions, adding ChatGPT-specific capabilities to MCP - sidebar access, custom file viewers, composer resource search - under Apache License 2.0.
Perplexity’s SPACE sandbox held against 12 models: none escaped VM-host isolation, but several bypassed network controls via DNS spoofing and shared-IP exploitation before the flaws were fixed.
Airbnb’s AI assistant cut support cost per booking by more than 17%, resolving about 40% of Q3 issues with no human agent.
Databricks Genie Agents now reason across Lakehouse data, unstructured documents, and enterprise tools in a single conversation, with MCP connectors and OpenSharing support.
Two OpenAI engineers proposed an IETF draft letting agents pass user signals like network, country, and masked IP via the Forwarded header. It is unsigned and primarily benefits the agent company by making agent traffic look normal.
Jason Wei argues success in AI is now 55% skills and 45% connections, versus 75/25 five years ago, as coding agents commoditize execution and labs stop publishing research.
Google’s Tabletop is a Labs experiment for real-time collaborative planning, turning links and files into itineraries, polls, and expense splitters. Rolling out in the U.S. (18+) via waitlist.
Tang’s IR derives kernels on the fly through views and comprehensions, eliminating hand-written kernels specific to operators, chips, and fusions.
Researchers on LessWrong taught Qwen3.6-35B-A3B and gpt-oss-20b to do multi-step tasks while emitting only filler tokens, beating answer-only baselines on three of seven tasks, with some tasks learned out to 60 steps.
You May Also Like
Nvidia's Board Hunts for Huang's Replacement as Agents Take Over AWS
Nvidia's board has hired an outside firm to identify CEO candidates as it weighs ousting Jensen Huang following reports of extramarital affairs, including one …
AI Now Accelerating Its Own Research: METR Says Anthropic Crossed the Threshold
A METR report finds Anthropic's capability progress roughly doubled year-over-year, with most of the acceleration attributed to AI automating AI R&D - and the …
Nvidia's CEO is gone, $1T vanished, and Apple wants Siri to actually do things
Nvidia CEO Jensen Huang resigned after admitting an affair with a woman US intelligence identified as a Chinese spy, wiping out over $1 trillion in market value …




