Japan bets $133B on an NVIDIA-built AI-robotics factory

Good morning 👋 Japan just told NVIDIA how many GPUs a national industrial policy is worth: 27,500 of them, backed by a government betting on a $133 billion robotics market it doesn't have yet.

In today's issue:

• 🔭 Japan and NVIDIA stand up a national AI-robotics factory, backed by a $133B market bet
• 🧠 Gemini Omni turns Vids into an editor you talk to, and gives you a talking avatar
• 🔬 Where robot safety filters, RAG evals, and LLM math all quietly break
• 🚀 Thira and Feathery close funding rounds for back-office and insurance AI agents
• 🛠️ Fine-tune video diffusion models for the robot-training footage nobody has yet
• 🔗 Fable 5 joins Max plans, and a million-rating study says no voice model nails emotion

🔭 THE ONE THING

🏭 Japan bets $133B on an NVIDIA-built AI-robotics factory

Japan and NVIDIA are standing up a national AI factory: 13,750 Vera CPUs, 27,500 Rubin GPUs, 140MW of capacity, built and run by Noetra Corp., a SoftBank-NEC-Sony-Honda consortium with Japan's METI underwriting it. It's the hardware backbone for METI's FRONTia project, the government's push to train multimodal foundation models for robotics and physical AI. Jensen Huang's line, "Japan invented modern manufacturing. Now it is building the AI factories that will power the next industrial revolution," is on-brand hype, but the number underneath it is real: Japan wants a 30% share of a $133 billion global AI-robotics market by 2040. Call NVIDIA's "world's first" tag what it is: an unconfirmed marketing line that Tom's Hardware pins down more precisely as the first state-tendered national AI factory. What isn't marketing is the model itself, a government picking a compute number and a vendor for an entire industrial strategy, and if it works, expect Seoul and Berlin copying the homework within the year.

🧠 MODELS & RELEASES

• 🎬 Google rolled Gemini Omni into Vids, so you can edit clips by typing what you want changed (swap the background, fix the lighting) instead of re-shooting anything, and it added personal avatars that turn a selfie plus a short voice sample into a talking stand-in for you. Locked to Pro, Ultra, and Workspace accounts, avatars stick to your own likeness and require you to be 18+, and every output gets a SynthID watermark. Google Vids updates

🔬 RESEARCH HIGHLIGHTS

• Robots don't read intent, they read words. A new probe called PRISM checks a model's hidden states instead of its output text, hunting for instructions that sound harmless but turn dangerous once a robot actually executes them. On the team's new PhysicalSafetyBench-1K it hit 99.6% accuracy with a 0.7% false-positive rate. A plain text-safety judge, by contrast, rejected two-thirds of tasks that were completely fine. If you're wiring an LLM into anything with actuators, filtering the output text isn't your safety layer. It's the wrong layer entirely. link

• Agentic search rewards documents that look useless. Retrieval has always been scored on one question: does this document answer the question in front of you. Wrong test, apparently, once an agent starts searching in multiple steps. Researchers logged 23,000+ document reads inside a ReAct agent and found almost zero correlation (Spearman rho of -0.026) between a document scoring as relevant and it actually changing the outcome. Roughly a third of the documents that mattered looked like noise to a static reader; they just handed the agent an entity name that redirected its next query. If your RAG eval pipeline scores on relevance alone, it's measuring the wrong thing for multi-hop agents. link

• A model's parts don't add up to its whole. Ask a model for estimates about subgroups of a population, then ask it for the population-level number, and the two should reconcile. Often they don't. Researchers partitioned populations into subgroups, prompted the model at each level, and found the aggregated subgroup answers frequently matched real-world data better than the model's own top-line estimate, a gap they call the "macro fallacy." The knowledge is in there somewhere. It just doesn't reliably assemble itself into one honest number, so don't trust a single aggregate estimate when you can decompose the question first. link

🚀 AI STARTUPS

• Thira raised a $21M seed led by Madrona, with FUSE joining in. The founders, Sunny Gupta and Kurt Shintaffer, are the same pair who built Apptio, and this time they're going after the back office itself: self-learning agents that plug into a company's existing systems and run operational processes off an enterprise knowledge graph. Ten design partners are testing it ahead of a fall launch. Thira raises $21M seed round

• Feathery closed a $30M Series A led by Portage Ventures, with Index, Bain Capital Ventures, and strategic checks from Allstate and Erie. It's a decisioning platform for insurance carriers, brokers, and wealth managers, already handling document processing and onboarding for 300+ firms, including Tokio Marine and Sequoia Financial. Not a flashy vertical. Real customers, though. Feathery raises $30M Series A

🛠️ TRY THIS

Fine-tune a video diffusion model for synthetic robot-training footage

Japan's new robotics AI infra needs training data nobody has enough of: real video of a gripper missing, a hallway nothing in the dataset has seen. NVIDIA's NeMo Automodel now talks directly to Diffusers checkpoints, so generating that footage got a lot less painful.

1. Pull a base checkpoint (HunyuanVideo 1.5, Wan 2.2) from the Hub and pre-encode your clips into latents with Automodel's preprocessing tool.

2. Fine-tune off the stock YAML. LoRA if the GPU budget is thin, full fine-tune if it isn't. No custom training script needed.

3. Launch with torchrun, setting FSDP2, tensor, or pipeline sharding as config flags instead of code.

4. Drop the checkpoint straight into a Diffusers pipeline and generate the clips your policy model is actually missing.

Prompt: Write a NeMo Automodel YAML config for LoRA fine-tuning HunyuanVideo 1.5 on a folder of local MP4 clips, using FSDP2 sharding across 8 GPUs with multiresolution bucketing enabled.

Worth a look

• IBM Research tested the same 417-task workload on Claude Sonnet and GPT-4.1. Sonnet came out at $79, GPT-4.1 at $155, even though GPT-4.1's advertised per-token rate is lower. If you route across models, price the actual task, not the rate card.

🔗 QUICK LINKS

• Anthropic folds Fable 5 into Max and Team Premium plans, but at half the usual rate limits, starting July 20. Source
• Hume AI ran a million-plus human ratings across 40+ voice models and found none nail paralinguistic cues like hesitation or emotion consistently, accents and background noise still trip up the "best" ones. Source

That's the issue. See you tomorrow.

Pradeep Perugu