Kazakhstan Just Bought a Seat at NVIDIA's Table

Good morning 👋 Jensen Huang flew to Kazakhstan this week and came home having cleared a 45,000-GPU export license through Washington. The ribbon-cutting was the least interesting part.

In today's issue:

• 🔭 Kazakhstan Just Bought a Seat at NVIDIA's Table
• 🧠 Models & Releases: a cheaper branch for agents, a Polly refresh, RAG you don't have to build
• 🔬 Research Highlights: planet-scale embeddings for two cents on the compute dollar
• 🛠️ Try This: audit your single-cloud GPU dependency before it's not your call anymore
• 🔗 Quick Links: bank earnings, a SPAC pitch, and an 8-K that's just a cover sheet

🔭 THE ONE THING

🛰️ Kazakhstan Just Bought a Seat at NVIDIA's Table

Jensen Huang made his first-ever trip to Kazakhstan this week, and he didn't come empty-handed: local firm Firebird walked away with a U.S. export license for 45,000 Grace Blackwell GPUs, Kazakhtelecom is already running 400 H200s and 480 B300s in Astana and Almaty, and NVIDIA (with TCV) put money into Higgsfield, the AI video startup whose engineers sit in Kazakhstan. Huang's headline engagement wasn't a photo op; it was a fireside chat with Deputy PM Zhaslan Madiyev, who wants AI that speaks Kazakh, not a localized English model wearing a different flag. Read past the ribbon-cutting and the real story is supply chain, not diplomacy. A 45,000-GPU export license doesn't clear Washington on its own merits. Someone decided Kazakhstan's compute buildout serves U.S. interests in Central Asia, and that calculus is worth more attention than the Astana press tour it bought.

🧠 MODELS & RELEASES

• 🧪 Microsoft trained agents on ThinkingBox by grading the database's final state instead of the tool calls along the way. Fine-tuning Qwen3.8-27B against those checks took reliable task completion (pass on all 20 tries) from 38 tasks to 107 out of 187, and cut the cost per dependable task 67%, down to $8. ThinkingBox training
• ⚡ OpenAI took the Decisions API to general availability. It's a stripped-down call built on GPT-6 Luna for bounded questions, classify this, route that, pick an action, and runs up to 10x faster than a normal Responses API call. $0.10 per million input tokens, nothing for output. If your agent is spinning up a full chat completion just to branch, this is the cheaper lane for that. Decisions API
• 🎙️ Amazon Polly's generative engine got a drop-in upgrade. 14 new voices, Hindi and Finnish now generally available, 57 generative voices total across 25 locales. Same API calls, same pricing. Polly upgrade
• 📚 IskayPet, a pet-retail chain spanning Spain, Portugal, and France, built a four-language knowledge assistant on Amazon Bedrock Managed Knowledge Bases. No RAG pipeline to run yourself. Answer time dropped from 3-5 minutes to about 7 seconds. Bedrock case study

🔬 RESEARCH HIGHLIGHTS

• Ai2's OlmoEarth v1.3 turns satellite imagery into ready-made embeddings for the whole planet: 10-meter resolution, built from Sentinel-1 radar and Sentinel-2/Landsat optical data, compressed to 128 dimensions you can just query instead of training your own model. It beats AlphaEarth Foundations and Tessera v2 on 20 of 27 benchmarks across six continents, and this version trained in 18% of the compute the original OlmoEarth needed, for roughly $2,720 in H100 time. Weights, data, and a browser explorer are all open. Hugging Face
• Sensorimotor Alignment proposes a cheap, offline score for whether a robot model's internal representations actually line up with how motion works in the physical world. The appeal is obvious: screen candidate vision-language-action backbones before burning GPU hours fine-tuning and running expensive closed-loop robot tests. Early results show the score tracks real task success across model families, though the paper doesn't publish hard correlation numbers yet. Treat it as a filter, not a verdict. arXiv
• Pass@k optimization for thought pruning tackles the quality-versus-diversity tradeoff in test-time reasoning by optimizing for it directly, instead of juggling two separate heuristics. Across eight open-weight reasoning models, it beat both quality-only and diversity-only pruning, with the biggest gains when keeping a moderate number of candidates alive. Worth swapping in if you're running test-time scaling in production. arXiv

🛠️ TRY THIS

Audit your single-cloud dependency before someone else's compute comes online

Kazakhstan going from "we bought GPUs" to full-stack deployment is the tell: regional capacity is stopping being a rounding error. More places to run inference means more leverage for you, if you know where you're locked in.

1. Pull 30 days of inference/GPU spend by provider. One number, not a dashboard.

2. Pick the single workload that breaks hardest if that provider has an outage or a price hike.

3. Find an open-weight model that covers 80% of that workload's job.

4. Stand it up on a second provider (or bare metal) and clock the latency and cost gap. Decide if the gap's worth closing now or just worth knowing about.

Prompt: List every AI inference workload I run, then for each one tell me: (1) which provider it's locked to, (2) whether an open-weight model could handle 80% of the job, and (3) what it'd take to get a backup running on a second provider by Friday.

🔗 QUICK LINKS

• Primis Financial posted Q3 numbers: net interest margin up 21bps to 3.66%, stock still sitting at 0.89x book despite it. SEC filing
• First Citizens BancShares paired Q3 earnings with a fresh $1.2B buyback authorization for 2027, fresh off swallowing 138 branches from BMO. SEC filing
• K&F Growth Acquisition Corp. II pushed out more SPAC paperwork on its David's Bridal spinoff, Pearl, which drops "AI trends" into the pitch without saying what that means yet. SEC filing
• AITX says its RAD robotics unit landed orders "across three product lines." The 8-K itself is just a cover sheet. The specifics live in a press release nobody's linked yet. SEC filing
• Data443 launched Convaa.ai, an on-prem AI collaboration platform for regulated shops that don't want their data leaving the building, built to run on GPU capacity most enterprises already own and barely use. SEC filing
• Mattel added EdgePoint partner Lee Matheson to its board under a new cooperation deal with the toymaker's largest shareholder. No AI in sight, just a boardroom seat changing hands. SEC filing

Keep an eye on who's buying the compute, not just who's shipping the model.

Pradeep Perugu