Nvidia Opens Up Nemotron 3.5 Lightning, a 30B MoE Built for Always-On Agents
Databricks Hits ~$190B Valuation as Data Infrastructure Becomes the AI Moat.
Nvidia Opens Up Nemotron 3.5 Lightning, a 30B MoE Built for Always-On Agents
Nvidia dropped Nemotron 3.5 Lightning 30B on August 11 — a hybrid Mamba+Transformer MoE with only 3B active parameters, licensed under OpenMDW-1.1 for full commercial use [1]. It ships with a new model routing library and is explicitly positioned for always-on agent deployments and post-training customization, not one-off inference [1][2].
The efficiency angle is the point. With just 3B active params doing the work of a much larger model, Nemotron 3.5 is cheap enough to run continuously — which is exactly the profile you want for background agents monitoring systems, triaging tickets, or orchestrating other tools 24/7. It's available on Hugging Face with quantized variants and speculative decoding checkpoints already live [2][3].
X commentary is focused on the openness of the release — training data transparency included — and how well it slots into existing agent harnesses [3]. This is Nvidia continuing its strategy of commoditizing the model layer while it owns the compute underneath. Expect more labs to route mundane, high-frequency agent tasks to small open models like this one and save the frontier models for genuinely hard problems.
Databricks Hits ~$190B Valuation as Data Infrastructure Becomes the AI Moat
Databricks announced a strategic funding round on July 16 at a $188B valuation (some reports put it closer to $190B), with annualized revenue in the $5.4–7B range and 65–80% year-over-year growth [1][2]. Notably, the company turned free-cash-flow positive in 2025 — this isn't a growth-at-all-costs story, it's a profitable one raising money anyway because the AI infrastructure land grab isn't over.
The money is earmarked for AI products and data/AI infrastructure, doubling down on the idea that the bottleneck for enterprise AI isn't model quality anymore — it's whether your data is clean, governed, and queryable by an agent in the first place [1]. Industry watchers see the valuation as a proxy for how much capital is chasing the "plumbing" layer of AI, not just the model layer [2].
For any company that's spent the last two years hearing "just plug in GPT-5" and ignoring their own data mess, this is the wake-up call: the winners of this AI cycle are increasingly the ones who fixed their data foundations before the agents arrived.
What This Means For Your Business
Three stories, one thread: the industry is quietly re-organizing around orchestration, not generation. Grok 4.6's whole pitch is surviving long agentic trajectories without falling apart. Nemotron 3.5 is built to be a cheap, always-on cog in a bigger agent system. Databricks is raising nearly $200B specifically to make sure your data is in shape for agents to actually use. None of these stories are really about "AI writes better code" anymore — they're about AI systems that plan, retry, self-correct, and run continuously, with humans setting direction rather than typing syntax.
If you're a Nordic company still evaluating AI on "does the model write good code" or "does the chatbot answer well," you're benchmarking last year's question. The real question now is: can our data support an agent that runs unattended for hours, and do we have the judgment layer — the taste, the guardrails, the escalation rules — to trust it when it does? Code is genuinely becoming free, as SpaceXAI and Nvidia are proving by shipping capable models at falling cost. What's expensive, and what Databricks investors are betting on, is the infrastructure and judgment required to point these systems at the right problems and trust their output.
This is exactly the terrain Up North AI operates in — voice AI, data tooling, orchestration — and today's news is a validation lap, not a surprise. The gap between companies that adopt agentic AI and those that just bought a chatbot license is going to become the defining competitive gap of 2026-2027.
Key takeaway: The AI race has quietly moved from "whose model is smartest" to "whose data and judgment can support an agent running unsupervised" — and that's a much harder, much more valuable problem to solve.
Sources
- https://x.ai/news/grok-4-6
- https://thenewstack.io/grok-4-6-agent-training/
- https://venturebeat.com/technology/spacexai-debuts-grok-4-6-overtaking-kimi-k3s-performance-and-matching-gpt-5-6-sol-for-worlds-third-best-on-artificial-analysis
- https://build.nvidia.com/nvidia/nemotron-3.5-lightning-30b-a3b/modelcard
- https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16
- https://ollama.com/library/nemotron-3.5-lightning
- https://www.databricks.com/company/newsroom/press-releases/databricks-raising-strategic-round-funding-188-billion-valuation
- https://memeburn.com/ai-global-funding-statistics-2026/
Stay ahead of AI
No spam. Unsubscribe anytime.
Want to go deeper?
Reading the news is one thing. Exploring the frontier is another. See what we're building.