Nscale's S-1 shows a $44.6B bet from Anthropic and a $1B check from Nvidia

Good morning 👋 Nvidia just wrote a $1 billion check into a company whose biggest contract is buying Nvidia chips. Nscale's S-1 makes that circularity official, and it isn't even the strangest number in the filing.

In today's issue:

• 🔭 Nscale's S-1: a $44.6B Anthropic bet and a circular Nvidia check
• 🧠 AWS ships a GPU-aware router that cuts inference latency without touching your code
• 🔬 Reproducible agent bugs, cheaper decoding, and a cut-rate path to diffusion models
• 🏛️ Anthropic and Accenture put $2B behind embedded AI safety evaluators
• 🛠️ Benchmark your vector store before you commit to one
• 🔗 Kimi K3 lands on Bedrock, GPT-6 Astra powers Hex, and OpenAI's burn rate leaks

🔭 THE ONE THING

📄 Nscale's S-1 shows a $44.6B bet from Anthropic and a $1B check from Nvidia

Nscale filed to list on the NYSE as NSCL yesterday, and the numbers inside are wild: $140.6 million in first-half revenue, up 1,252% from a year ago, against a $1.02 billion net loss. The engine behind that growth is a GPU Services Agreement with Anthropic worth up to $44.6 billion, signed August 25 to build out the Monarch Compute Campus in West Virginia. Three days before filing, Nscale locked in $3.1 billion in convertible notes, a billion of it from Nvidia itself, the chip supplier effectively financing the customer that buys its own chips. Press pegs the IPO at a $30 billion valuation, which is the conversion cap on those notes, not a market test. One customer accounting for tens of billions in contracted revenue is either the best backlog in AI infrastructure or the single point of failure that sinks this stock the day Anthropic renegotiates.

🧠 MODELS & RELEASES

• ⚡ AWS shipped SageMaker HyperPod Inference Gateway, a GPU-aware router for EKS that reads live GPU signals and sends each inference request to whichever pod can actually serve it fastest. No changes to your model servers or client code required, and AWS says it cuts first-token latency up to 82%. Builder read: if you're running multi-model inference on EKS and eating queuing tax today, this is worth a pilot. Amazon SageMaker HyperPod Inference Gateway

🔬 RESEARCH HIGHLIGHTS

• Chronicle goes after a real agent-debugging headache: you patch a bug, rerun the failing trajectory, and nondeterminism means it doesn't fail the same way twice. It records a run's nondeterministic boundaries (model calls, tool reads) as immutable envelopes, then replays some of them from the record while your new code runs live for the rest, turning a one-off incident into an actual regression test. Recording overhead is 23 microseconds per boundary crossing, and in mutation testing it caught every injected bug that let an unsafe action slip through; a baseline that just stubs every boundary caught none. arxiv.org/abs/2609.20625v1

• On-Demand Attention trains a small "recall head" that predicts, mid-decode, whether the model actually needs to reread its full history or can get away with local attention. Full-attention decoding pays that cost at every single step, whether the history helps the next token or not. ODA cuts those global reads substantially on Qwen and Gemma (hybrid-attention backbones included) while keeping most of the accuracy, and it's already wired into vLLM as conditional GPU execution, so the speedup shows up in real decoding, not just a paper benchmark. arxiv.org/abs/2609.20734v1

• dQwen3.5 takes the cheap route to a diffusion language model: adapt a pretrained autoregressive one instead of training from scratch. The catch is that current AR models like Qwen3.5 interleave attention with RNN layers, and RNNs are causal by construction, not easy to bidirectionalize. The team adapted it anyway at 0.8B, 2B, 4B, and 9B, and the hybrid backbone hit a given training loss in about half the tokens of a full-attention baseline. arxiv.org/abs/2609.20751v1

🏛️ POLICY & REGULATION

• Anthropic / Accenture cut the first named embedded-evaluator deal: $1B+ pledged by each side over five years for a team with employee-level access inside Anthropic, watching training runs and deployment calls as they happen. It's the first real follow-through on the "pace the frontier" essay Amodei published last week. Anthropic is funding Accenture's side directly, since no independent mechanism for this exists yet, and calls the arrangement non-exclusive, so METR could be next in line. Watch, don't act: the test is whether "employee-level access" survives an actual release deadline. Anthropic

🛠️ TRY THIS

Benchmark your vector store before you lock into one, using AWS's own numbers

1. Pull your Bedrock Knowledge Base's real query pattern (top-k, filter usage, docs ingested per day) from CloudWatch, not the sample workload in the blog post.

2. Run AWS's three benchmark scenarios against OpenSearch, Aurora PostgreSQL with pgvector, and S3 Vectors using that data.

3. Score by cost per million vectors at your current scale and at 10x. Latency numbers alone will steer you wrong once volume grows.

4. Pick S3 Vectors if you can tolerate slower recall for the discount. Pick OpenSearch if you need sub-100ms lookups under load.

Prompt: Given my RAG workload (doc count, queries/day, latency budget), compare Amazon OpenSearch, Aurora pgvector, and S3 Vectors using AWS's Bedrock Knowledge Bases benchmark framework. Show projected cost per million vectors at current scale and at 10x.

Worth a look

• Octop (TencentCloud) A self-hosted, multi-user, multi-agent assistant, 4.1k stars on GitHub. Infra economics are the story today (see Nscale's S-1); if you're weighing a hosted agent platform against running your own, this is a credible fork-and-own option.

🔗 QUICK LINKS

• Moonshot AI Kimi K3 landed on Amazon Bedrock: 1M-token context, native vision, prompt caching built in. Open-weight and enterprise-distributed in the same move. link
• OpenAI / Hex GPT-6 Astra now drives Hex's data agents, turning raw analysis into visual reports nobody has to reformat before sharing. link
• Sakana AI Stood up a new research group, the Frontier Intelligence Group. No roadmap yet, just a name and a headcount plan. link
• OpenAI Published a six-pillar youth-safety blueprint for Australia. Regulatory cover, ahead of the regulation. link
• Vantora (formerly UP.Labs) took $100M from Silversmith Capital Partners, its first outside check ever, and is steering its startup-builder model at physical AI for industrial partners. link
• OpenAI Per FT reporting on a leaked deck: negative free cash flow of $278B projected from 2026 to 2030, with revenue climbing from $36B this year to $350B by 2030. The growth number is the pitch. The FCF number is the risk. link

Watch what happens the day Nscale's biggest customer asks to renegotiate. That's the whole thesis in one clause.

Pradeep Perugu