On August 11 (Beijing time August 12), The Information, citing multiple insiders on the project, reported that NVIDIA is developing Nemotron 4, with the flagship version expected to reach at least 1 trillion parameters—roughly double the Nemotron 3 Ultra (~500B parameters) released in June. It could be ready as early as late fall 2026 (around November). The same day, NVIDIA released Nemotron 3.5 Lightning (30B total / 3B active MoE, focused on long-running AI agents) and NeMo Switchyard (an open-source model routing tool for automatic multi-model dispatch), and Dell integrated both into its Dell Deskside Agentic AI platform.
Placed in the context of NVIDIA's last three months—Jensen Huang's July public endorsement of open-weight models, multiple open-source releases in the past month, and flagship R&D resources now flowing into trillion-parameter Nemotron 4—a GPU vendor is making "building our own models" a long-term strategy.
How the Trillion-Parameter Number Adds Up
NVIDIA's open-source model trajectory:
| Release | Model | Total params | Active | Positioning | |------|------|--------|------|------| | 2024-12 | Nemotron-3 22B | 22B | 22B | Edge / mid-tier | | Mid-2025 | Nemotron-3 49B | 49B | 49B | Mid-size | | 2026-06 | Nemotron 3 Ultra | ~0.5T (500B) | — | Flagship | | 2026-08-11 | Nemotron 3.5 Lightning | 30B | 3B | Long-horizon agents | | Late 2026 (expected) | Nemotron 4 | ≥1T | — | Trillion-parameter open flagship |
But competing open-source players in the same window:
- Qwen3.8-2.4T-A95B (released open-source on 08-12, 2.4T total / 95B active)
- Moonshot Kimi K3: 2.8T / 104B
- DeepSeek V4-Pro: 1.6T / ~49B
- xAI Grok 4.6: 1.5T / 8-16
- August: internal architecture + training data finalized
- September–October: large-scale training begins
- November: possible completion / early trials
- H1 2027: GA + third-party validation
- Alibaba's Qwen series held 50%+ of global open-model downloads as of March 2026
- Chinese open-source models reached 41% of monthly Hugging Face downloads, surpassing the US (36.5%)
- On OpenRouter, DeepSeek + Qwen + MiniMax + Zhipu's combined call share rose from ~20% to near 50% in H1 2026
- No official release date; "late fall 2026" is an insider estimate via The Information
- Training data sources and license unconfirmed (Apache 2.0 vs NVIDIA Open Model License)
- Active parameters and architecture undisclosed—dense or MoE unclear
- Lightning's benchmark numbers unpublished; comparisons with Claude Haiku 4.5 / GPT-5.6 Luna pending
- Depth of Dell integration unclear (reference implementation vs bundled sale)
- Nemotron Coalition members' contribution details not disclosed
- https://www.theinformation.com (The Information original report, subscription required)
- https://www.knews.com.tw/news/3AFBF66E3423ED3EFB50AD8C2DDD0682
- https://www.163.com/dy/article/L45I4QS10519C3MG.html
- https://new.qq.com/rain/a/20260812A0B6N000
- https://k.sina.com.cn/article_7879996029_1d5af327d06801bo5u.html
- https://so.html5.qq.com/page/real/search_news?docid=70000021_7796a7bc96b24852
- https://blogs.nvidia.com
- https://x.com/karibriski
Qwen open-sourcing a Qwen-Max-class 2.4T model is seen as the key trigger. If NVIDIA didn't push Nemotron 4 to ≥1T, the open-source camp's scale would overshadow NVIDIA's ecosystem pull—leaving the "US open-source camp" without a flag-bearer.
The Information's original wording: "final training has not yet started; may be ready as early as late fall 2026 (around November)"—this is a "R&D readiness signal," not a release announcement. A realistic timeline:
So the concrete technical impact within six months is weak; the real significance is that NVIDIA has decided to do this at all.
The Business Math: Free Models + 90% Revenue from Hardware
NVIDIA's FY2026 total revenue is roughly $130B, with data centers at 90%+. This means Nemotron 4 / Nemotron 3.5 Lightning / NeMo Switchyard are not about competing with OpenAI or Anthropic for API revenue—but about pulling enterprises back into NVIDIA's ecosystem for self-hosting:
Free open-source model → customers download weights, self-host training + inference → strong hardware lock-in (GPU / DGX needed) → interconnect dependency (InfiniBand NDR / Spectrum-X for thousand-GPU clusters) → software stack lock-in (NeMo / Megatron-LM / TensorRT-LLM) → long-term hardware recoup. The open model is effectively the "giveaway product"; every layer sells hardware.
The "Nemotron Coalition" (Reflection / Cursor / Thinking Machines / Mistral and others) contributes training data and evaluation support—a "collective open R&D" posture that adds a compliance-and-diversity halo to NVIDIA's models.
Same-Day Releases: Nemotron 3.5 Lightning + NeMo Switchyard
Nemotron 3.5 Lightning (30B total / 3B active MoE) targets long-running AI agent scenarios and is claimed to beat Claude Haiku 4.5 on benchmarks like Terminal Bench 2 and SWE-Bench Verified / Pro (official numbers pending). It is a lightweight agent backend deployable on a single consumer GPU.
NeMo Switchyard (open-source routing tool) automatically assigns the best model per task, positioning NVIDIA in the model-routing track alongside Google API Gateway routing and OpenRouter's auto/cost-tier routers.
Dell's Deskside Agentic AI platform integrates both—"hardware + model + routing + software" bundled for enterprise desktop AI, leveraging Dell's channel and free models.
$28B Compute Commitment + Huang's Open Letter
Another data point from The Information: NVIDIA's multi-year cloud compute commitment has grown to $28B, roughly triple a year ago—reverse-leasing capacity from hyperscalers to train its own models. Combined with the same-month "$500B AI factory fund" (Apollo / BlackRock / Blackstone / Brookfield / Goldman Sachs / KKR, with NVIDIA providing up to 25% residual-value support), NVIDIA is positioning itself as the pricing-setter and supplier of the AI Infra era. Jensen Huang's July open letter supporting open-weight models now has concrete follow-through actions.
Defending Against the DeepSeek / Qwen / GLM Open-Source Surge
The defensive logic is concrete:
The risk: enterprises shift from NVIDIA GPUs to Chinese open models plus domestic inference stacks and chips (Ascend, Biren, Cambricon). The response: Nemotron 4 + Lightning + Switchyard + Dell integration = "US enterprises using US open models + US GPUs + US software stacks." NVIDIA has long supported Llama / Mistral / Qwen training via Inception / DL platforms—but building a top open-source flagship itself is a first.
The Real Risk: Can "Self-Disruption" Work?
Concern 1: API players fine-tuning Nemotron 4 and reselling it competes with OpenAI / Anthropic / xAI paid APIs—but the net math is positive, since AI compute demand growth exceeds private API revenue loss.
Concern 2: Training cost. A full trillion-parameter run needs 20,000+ H100s, ~3×10²⁵ FLOPs, 90–120 days—at ~$3/hr per H100, roughly $0.5–1B in R&D. Whether that pays back within a year depends on enterprise adoption. Worst case: Nemotron 4 becomes a "strategic-posture investment" rather than a commercial one. But as long as the Chinese open-source camp erodes NVIDIA's ecosystem share, the defensive spend is justified.
Limitations and Unknowns
Is Open-Source Model Building Self-Disruption for AI Infra Companies?
> Can an AI Infra company (a compute seller) building its own open-source flagship close the loop across hardware sales, model sponsorship, and training-data ecosystems?
NVIDIA's stack—Nemotron + CUDA + DGX + InfiniBand + Megatron-LM + TensorRT-LLM + NeMo Switchyard + Dell—gives away the model and monetizes everything around it. If it works, AMD (ROCm + Instinct), Cerebras / Groq / SambaNova, and Huawei Ascend / Biren / Cambricon could follow with their own models. Nemotron 4 is the first full commercial validation of "an AI Infra company entering open-source models"—whether the model itself is world #1 matters less than whether the business model holds.
For AI coding and embodied intelligence: Nemotron 4 is a next-generation foundation model designed for long-horizon agent tasks, not a chatbot. Since August, NVIDIA has placed bets on four fronts simultaneously—agent backends, agent routing, agent desktop integration, and a flagship open model. The path is fully legible.
---
Sources