NVIDIA lines up $500B for AI compute, but only on paper so far

Good morning 👋 NVIDIA just got six of Wall Street's biggest names to put their reputations on a $500 billion number. None of it is signed yet.

In today's issue:

• 🔭 NVIDIA's $500B compute financing plan, MOUs not money
• 🧠 NVIDIA open-sources voice AI, Sakana swaps engines, Google gets agentic
• 🔬 Encrypted reasoning traces that leak anyway, SWE-bench's broken tests, safety harnesses that patch themselves
• 🛠️ Move GPU dev boxes onto a shared cluster before you buy more compute
• 🔗 ChatGPT Business perks, a 75% faster FinOps build, and a war on scrapers

🔭 THE ONE THING

💰 NVIDIA lines up $500B for AI compute, but only on paper so far

NVIDIA signed MOUs with six of Wall Street's biggest names, Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR, to build financing platforms aimed at mobilizing over $500 billion in third-party capital for AI infrastructure. The pitch: hyperscalers and frontier labs get compute without loading up their own balance sheets, and asset managers get a new category sized for pension and insurance money. Read past the release and it's six MOUs, not six checks: nothing is priced, nothing is committed, and the whole structure depends on final agreements nobody's signed yet. Huang, Larry Fink, and David Solomon all lined up for quotes anyway, because being first in line on the biggest financing story in AI is worth more right now than the ink itself. If even a fraction of that number actually closes, the real story isn't the compute, it's who ends up holding the risk when a data center's economics don't pencil out.

🧠 MODELS & RELEASES

• 🔊 NVIDIA open-sourced Magpie TTS Multilingual on Hugging Face: 364M parameters, 12 languages, NVIDIA Open Model License. 32ms to first audio on a B200, NIM containers if you want it production-ready. You get the whole voice pipeline in your own infra, not a rented API. link
• 🐡 Sakana AI swapped the base model under its Fugu conductor from Qwen to Google's Apache-2 Gemma 4, ran it against their own eval suite, and got matching accuracy and cost savings. The conductor layer is the product now, not whatever model sits under it. link
• 📊 Google pushed agentic features into Ads and Analytics: homepage summaries that flag what moved since your last login, and natural-language dashboard generation for Ads (Analytics gets it next). Beta, English only, so don't restructure a reporting workflow around it yet. link

🔬 RESEARCH HIGHLIGHTS

• Encrypting a model's chain-of-thought doesn't stop it from leaking, it just changes who leaks it. Providers return reasoning traces to the client as encrypted blocks instead of storing them server-side, meant to block distillation and IP theft. Turns out those blocks are interchangeable across sessions, users, and models within a provider's stack: feed a flagship model's trace into a weaker, less-guarded sibling model, and it decrypts and prints the whole thing in plaintext. No jailbreak of the strong model required. Scraping public logs this way pulled 367 PII artifacts out of 315,320 blocks. paper

• An audit found nearly 60% of SWE-bench Verified's unsolved tests are just broken. Too strict, too loose, take your pick, and some frontier models were just regurgitating gold patches they'd memorized during pretraining. SWE-Bench ProMax is the fix: 170 hand-curated refactoring tasks across seven languages, specs rewritten from scratch, tasks that touch 11 files and 260-plus lines on average. Best model on it resolves 41.2%. If your coding-agent numbers look too clean, ask which benchmark they ran. paper

• Agent safety usually lives in the model. SHE moves it into the harness and lets the harness rewrite itself. The framework splits an agent's harness into four pieces (system prompt, rule bank, safety memory, tool policy) and patches them automatically whenever a trajectory fails, diagnosing what broke and validating that the fix didn't tank performance. Against a static baseline it cut attack success rate 3.1x, and the learned rules generalized to unseen risks on AgentHarm with no extra tuning. paper

🛠️ TRY THIS

Move dev GPU boxes off individual laptops and onto a cluster before you commit to more compute

Wall Street lining up $500B for AI compute doesn't mean your team needs more of its own. Most orgs are already paying for GPU dev instances that sit idle overnight and on weekends. Amazon's new SageMaker AI Spaces add-on for EKS runs managed JupyterLab and VS Code environments on the cluster your ML team already operates, so engineers get GPU access without each one owning a dedicated box.

1. Pull a list of every standalone dev/notebook instance in your AWS account and sort by last-active date.

2. Flag anything under 15% average utilization over the past two weeks.

3. Install the SageMaker AI Spaces add-on on your existing EKS cluster and move one team's workflow onto it (browser or VS Code over SSH-over-SSM both work).

4. Re-check utilization after a week before you migrate the rest.

Prompt: List every GPU dev/notebook instance in our AWS account idle more than 14 days or under 15% average utilization, and draft a checklist for migrating the active ones onto a shared SageMaker AI Spaces environment on our EKS cluster.

Worth a look

• TestSprite shipped an official CLI, so its AI-powered test generation now runs from your terminal instead of a web dashboard, useful if you're already scripting CI around the migration above. github.com/TestSprite/testsprite-cli

🔗 QUICK LINKS

• OpenAI Premium seats are landing in ChatGPT Business, and signing up before August 20 gets a workspace $100 in credits plus higher usage caps for the heavy users on your team. Premium seats coming to ChatGPT Business

• nOps Ditched a self-managed LangChain/LangGraph stack on EKS for Bedrock AgentCore and took its Clara FinOps agent from a 10-12 month build to four. How nOps shipped FinOps agents 75% faster

• PatronView Nick Gray's server logs: 214 bot page loads for every human one, and Claude's search crawler alone hit his site 35,000 times for every visitor it actually sent him. He blocked it. PatronView

See you tomorrow.

Pradeep Perugu