NVIDIA Turns AI Compute Into a $500B Wall Street Asset Class

Good morning 👋 The $500 billion NVIDIA teased yesterday just got six real balance sheets attached to it. Whether "asset class" holds up past Huang's own pitch deck is the question worth sitting with.

In today's issue:

• 🔭 NVIDIA turns its $500B financing plan into six actual balance sheets
• 🧠 Cyber-defense models on Bedrock, AMIE watches a patient instead of just reading about one, a construction foundation model that beat the general-purpose approach
• 🔬 Sparse autoencoders' bad day
• 🏛️ The paperwork behind putting Claude in a regulated enterprise, Brussels lines up drone rules
• 🛠️ Score an AI compute financing deal like a credit analyst
• 🔗 A robot that rams speed bumps on purpose, a CIDP patient walking again, and a stack of small-cap AI filings

🔭 THE ONE THING

🏭 NVIDIA Turns AI Compute Into a $500B Wall Street Asset Class

NVIDIA locked in six of Wall Street's biggest balance sheets, Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR, to mobilize north of $500 billion for AI compute financing. Jensen Huang's pitch to CNBC: this is the first time chips have become an investable asset class, and the pricing backs him up: Blackwell B200 now goes for $5.30 to $7.05 an hour. That's the kind of number you want if you're trying to convince pension funds that GPUs are yield-bearing infrastructure, not a depreciating expense. One thing to sit with: Huang wrote that framing himself, on NVIDIA's own blog, and a company financing the buyers of its own chips is the exact circular-financing setup Michael Burry has been warning about for a year. The capital is real. Whether "asset class" is more than a sales pitch is still NVIDIA's bet to prove.

🧠 MODELS & RELEASES

• 🛡️ Daybreak Red & Blue OpenAI's specialized cyber-defense models are live on Amazon Bedrock for eligible customers, with zero-operator access enforced at the chip so vulnerability data never passes through a human reviewer. source
• 🩺 AMIE Google's medical research model ran real-time clinical video consultations in a first-of-its-kind study, still simulated patients, not a live rollout, but a real step past text Q&A toward a model conducting the visual exam itself. source
• 🏗️ Ishigaki-IDS ONESTRUCTION built a foundation model for construction and BIM workflows with AWS's GenAIIC team, synthetic data, a three-stage training pipeline, verifiable rewards. Evidence a narrow custom model can outperform prompting a general one for a niche domain. source
• 📸 Pixieset took its AI alt-text feature from concept to production in four months on Amazon Bedrock, then took sixteen months to get 35% of its photographer users onto it, automating the SEO grunt work and leaving the creative call to the humans. Notable because photographers are about the most AI-skeptical audience there is. source

🔬 RESEARCH HIGHLIGHTS

• Sparse autoencoders, interpretability's favorite feature-finding tool, don't track human concepts any better than the plain embeddings they're supposed to improve on. The paper tests whether SAE latent sets recover human category boundaries more faithfully than dense representations. They don't. They track the model's own internal structure instead. Worse, semantic edits a human would barely register can scramble the SAE's active feature set entirely. If your interpretability story assumes SAE features compose like a bag of concepts, this is the paper that says check that assumption. paper

🏛️ POLICY & REGULATION

• AWS shipped a reference deployment for its Claude Apps Gateway: a self-hosted governance layer sitting between Claude Code or Desktop and Bedrock. No new rule got written today. What did ship is the plumbing (audit logging, cost controls, access policy) that gets a Claude rollout past legal at a regulated enterprise. Deploying Claude apps gateway for AWS
• EU Commission opened registration for the first D-TECT Forum on November 11, its coordination venue for drone and counter-drone tech across the bloc. Nothing binds yet. This is Brussels lining up the industry players before it decides what rules the sector actually gets. Join the Inaugural D-TECT Forum

🛠️ TRY THIS

Score an AI compute deal like a credit analyst, not a headline reader

1. Pull the raw terms from the filing or press release: capacity (MW/GPUs), contract length, off-take counterparty, financing structure.

2. Feed the terms to an LLM and ask for the revenue-per-GPU-hour needed to clear a target IRR, given the depreciation schedule.

3. Ask it to name the three assumptions most likely to break the return: counterparty credit, power cost inflation, GPU residual value at contract end.

4. Sanity-check the output against a deal you already understand (a CoreWeave or Crusoe financing) before you trust the number.

Prompt: You're a credit analyst underwriting an AI compute financing deal. Given these terms: [paste capacity, contract length, off-take counterparty, financing structure], calculate the revenue-per-GPU-hour needed to hit a 15% IRR, list the three assumptions most likely to break that return, and flag whether the off-take counterparty looks investment-grade. Output as a one-page memo.

🔗 QUICK LINKS

• AITX ROAMEO, the company's autonomous patrol robot, went viral for repeatedly ramming a speed bump. The official line in an SEC exhibit: deliberate throttle logic that ramps up force until it clears the obstacle, not a stumble. Judge the video yourself. 8-K exhibit
• Bioxytran says its first CIDP patient on PLICID, a galectin antagonist the company designed with AI, regained use of his fingers and can walk several miles again. One patient isn't a trial, but it's the kind of result that gets a real one funded. filing
• Eva Defense, Eva Live's drone unit, signed Boumarang to build a jam-resistant UAS that keeps flying when GPS and radio get denied. $1.2M, paid out at $100K a month through July 2027. 8-K
• Rocket One got accepted into Seagate's partner program. Modest news for a company selling itself as an AI-infrastructure and radiation-tolerant computing play, worth watching for what it does with the access rather than the acceptance itself. 8-K
• Aura Consolidated Group upsized the private placement funding its buyout of Qoria, the AI-driven student-safety software maker, from $75M to $100M. The S-1 with the deal paperwork is out. S-1 exhibit

That's the issue. Back tomorrow.

Pradeep Perugu