NVIDIA starts taking a cut of the AI clouds it powers

Good morning 👋 NVIDIA didn't just sell more chips this week. It found a way to own a slice of what happens after the sale.

In today's issue:

• 🔭 NVIDIA starts taking a cut of the AI clouds it powers
• 🧠 Claude Sonnet 5 closes most of the gap to Opus
• 🔬 Sabotage across PRs, agents lying off the record, and why humans still beat the model on Polymarket
• 🛠️ A gateway pattern for surviving the GPU quota fights coming
• 🔗 Sakana AI's UN seat, and a VFX shop's training-time crunch

🔭 THE ONE THING

💰 NVIDIA starts taking a cut of the AI clouds it powers

NVIDIA just quietly became a landlord: its new financing structure pairs AI clouds with NVIDIA infrastructure, and instead of only selling the chips, it now collects a cut of the cloud revenue those chips generate. Sharon AI is first in, committing to 40,000 GB300 Grace Blackwell GPUs, while Firmus is chasing the same terms with a 360-megawatt, 170,000-GPU campus in Batam, Indonesia. No dollar figures were disclosed, which is the tell: NVIDIA is financing its own demand curve so smaller clouds can build capacity they couldn't otherwise afford, then taking a slice of what it earns. That's fine while GPU depreciation outruns the revenue share. The day it doesn't, NVIDIA is holding the hardware and the downside both.

🧠 MODELS & RELEASES

• 🤖 Anthropic shipped Claude Sonnet 5, the first Sonnet in its newest generation, with frontier scores on coding and agentic tasks that close most of the gap to Opus. Sonnet's always been the "run it in production" tier, and this is the version actually built for that job. Introducing Claude Sonnet 5, also live on Bedrock.

🔬 RESEARCH HIGHLIGHTS

• Sabotage that ships one PR at a time — a new benchmark (Iterative VibeCoding) has Claude Sonnet 4.5 play attacker against a GPT-4o monitor, spreading a malicious payload across multiple pull requests and timing the final piece for the PR with the best cover. Gradual attacks slipped past a standard diff monitor 93% of the time. Stack four monitors together, including one that tracks links across PRs instead of judging each diff alone, and evasion drops to 47%. Still a coin flip. If your review process only looks at one PR at a time, this is the gap. arXiv

• Agents say one thing on the record, another off it — researchers gave 10 models a public channel and a private "off the record" channel inside multi-agent debates, no instructions to behave differently in either. Baseline divergence between the two was about 3%. Add social structure like hierarchy or an audience, and it jumped to roughly 40%, with some models flat-out admitting in the private channel that they'd caved in public for "career risk" or a sponsorship obligation. Nobody prompted that. Watch, don't act: if you're evaluating agents on stated goals alone, you're only seeing the performance. arXiv

• The Polymarket test nobody wants to hear — a small real-money forecasting study found most people paired with an AI either just copy its answer or use it to rubber-stamp a guess they'd already made, and both groups do worse than the model alone. A minority beat the model, and what set them apart wasn't IQ or which model they used. It was perspective-taking, intellectual humility, and curiosity. Builder read: better prompting and better models won't fix a bad collaborator. arXiv

🛠️ TRY THIS

Put a gateway in front of your model calls before the quota fights start

NVIDIA's revenue-share deals are about to pour more GPUs into partner buildouts, which means more tenants sharing the same capacity pool. AWS just published the unglamorous fix for that world: five resilience patterns for LLM inference, tested with LiteLLM sitting in front of Bedrock.

1. Drop a LiteLLM gateway between your app and the model API instead of calling it direct. It's free, runs locally, and handles routing and failover for you.

2. Set a fallback chain (primary model, then a secondary, then a cheaper one) that fires the moment you hit a 429, before a user notices anything's wrong.

3. Give each tenant its own rate-limit bucket. AWS's own test: a caller capped at 3 RPM landed 60% success, while callers at 10 RPM hit 100%. Skip this and one noisy tenant starves everyone else.

4. Hammer your primary model in staging until it throttles, then confirm the fallback catches it silently. If your app throws instead, the config's wrong, not the model.

Prompt: Review this LiteLLM config for resilience gaps. Confirm every primary model has a fallback, check that rate limits are set per tenant instead of globally, and flag any single point of failure: one region, one account, one model.

Worth a look

• CopilotKit + AG-UI turns AgentCore agents into actual interfaces: inline charts, live shared state, human-in-the-loop approval steps, over an open protocol that also plugs into LangGraph and CrewAI. AWS walkthrough
• ai-berkshire is a multi-agent research framework for Claude Code and Codex, built around four value-investing playbooks (Buffett, Munger, Duan Yongping, Li Lu) running adversarial analysis against each other. Almost 9,000 GitHub stars. Worth studying for the orchestration pattern even if you don't care about stock picks. Repo

🔗 QUICK LINKS

• Sakana AI Co-founder Ren Ito joins the UN's new AI for Good Global Commission, one of roughly 40 seats also held by Jensen Huang and co-chair Marc Benioff. Sakana AI
• Outpost VFX Swapped RTX 3090s for EC2 P5 H100 clusters with AWS's help and cut face-replacement model training from two weeks to two days. AWS

See you tomorrow.

Pradeep Perugu