NVIDIA gave away a frontier-class AI model on August 11, 2026, and investors are treating it as a GPU sales pitch. Nemotron 3.5 Lightning, a 30-billion-parameter mixture-of-experts model, shipped with open weights, open training data and a permissive commercial license, according to NVIDIA’s developer blog. Anyone can download it, modify it and run it in production without paying NVIDIA a cent. The catch, if you can call it that, is where the model runs best: NVIDIA’s own hardware, from a GeForce RTX card on a desktop to a DGX Spark box built around the company’s GB10 chip.
The release landed two weeks before NVIDIA posted a fiscal second-quarter that quieted almost every doubt about slowing AI infrastructure spending. Data center revenue hit $89 billion, up 117% year-over-year, according to CNBC’s coverage of the launch and earnings cycle. The timing is not a coincidence. NVIDIA’s open-source strategy and its hardware revenue are now the same story, and Nemotron 3.5 Lightning is the clearest example yet of how the company plans to keep it that way.
What NVIDIA Actually Shipped on August 11
Nemotron 3.5 Lightning is a 30-billion-parameter mixture-of-experts (MoE) model with only 3 billion parameters active per token, a design choice that keeps inference costs down while preserving the reasoning capacity of a larger network. NVIDIA built it for what the company calls “always-on agents”: narrow, repetitive AI tasks that run continuously rather than a single chat session. The model supports context windows up to 1 million tokens, enough to hold a full codebase or a stack of legal contracts in memory during a single agent run.
NVIDIA published the model in BF16, FP8 and NVFP4 formats, alongside the base checkpoint, post-trained variants and the full training recipe. That last part matters more than it sounds. Most “open” AI releases hand over weights and stop there. NVIDIA’s NVIDIA-NeMo Nemotron repository on GitHub states plainly: “Weights, data, and recipes are released under OpenMDW-1.1.” That license lets a company retrain, fine-tune or redistribute the model commercially without crediting or paying NVIDIA anything.
NVIDIA says the model delivers up to 4x faster output than comparably sized systems, a claim aimed squarely at enterprises running high-volume, low-latency agent workloads such as customer support triage or code review bots. Whether that speed figure holds up against independent benchmarks is still an open question. No third-party lab had published a peer-reviewed comparison as of late August, and reporting from MarkTechPost and other outlets echoes NVIDIA’s own figures without independent verification.
The Single-GPU Pitch: Why Deployment Hardware Is the Real Story
Strip away the marketing language and Nemotron 3.5 Lightning’s headline feature is simple: it fits on one GPU. NVIDIA lists single-GPU deployment on one DGX Spark system (built on the GB10 superchip) or one H100 as the supported configurations, according to MarkTechPost’s coverage of the launch. NVIDIA’s own developer blog frames it as making “capable agentic AI accessible on local systems including NVIDIA Jetson, GeForce RTX 5090, and DGX Spark.”
That framing does two things at once. It genuinely lowers the barrier for a small team or solo developer to run a serious agentic model without renting a data-center GPU cluster. And it keeps every one of those local deployments tied to NVIDIA silicon, whether that’s a $2,000 RTX 5090 card or a DGX Spark desktop box priced well above consumer GPUs. Nobody is running Nemotron 3.5 Lightning efficiently on an AMD Instinct accelerator or a Google TPU pod, at least not with NVIDIA’s own tuned recipes and NVFP4 quantization kernels.
A developer wanting to try the model locally can pull it directly from Hugging Face and serve it with a standard inference stack:
huggingface-cli download nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B
vllm serve nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B
--quantization nvfp4
--max-model-len 1000000The model is also listed on OpenRouter’s free tier, giving developers a hosted option before they commit to buying hardware.
NeMo Switchyard: The Router NVIDIA Launched Alongside It
NVIDIA didn’t release Nemotron 3.5 Lightning in isolation. The same day, the company shipped NeMo Switchyard, an open-source routing library that sends each incoming prompt to whichever model handles it best, weighing cost, speed and task complexity. Coverage from NVIDIA’s official blog describes Switchyard as infrastructure for enterprises running mixed fleets of models, some large and cloud-hosted, some small and local like Nemotron 3.5 Lightning.
The pitch to security-conscious industries is specific: keep proprietary data on local hardware, route only non-sensitive queries to larger cloud models, and use Switchyard to manage that split automatically. For a cybersecurity firm or a legal research shop, that argument does real work. It also, again, nudges enterprise buyers toward standing up more NVIDIA-based infrastructure rather than less.
Why Give Away a 30B Model for Free
NVIDIA’s logic isn’t hidden. Reporting on the launch has consistently tied the open-source release back to one idea: broader access to capable AI models expands the pool of people running AI workloads, and more workloads means more GPU sales, whether that’s a single RTX card or a rack of H100s. The company has applied the same playbook across the Nemotron Speech and Nemotron Safety model families released earlier in 2026, each one free, each one tuned to run best on NVIDIA hardware.
It’s a bet that software margin matters less than hardware volume. Meta gives away Llama for a related reason: it wants the industry standardizing on infrastructure Meta controls or influences. NVIDIA’s version swaps the infrastructure layer from social data to compute. Give developers a genuinely useful model for nothing, and the GPU purchase that follows is the actual transaction.
The Nemotron Family: From a 30B Assistant to a Trillion-Parameter Ambition
Nemotron 3.5 Lightning didn’t appear out of nowhere. NVIDIA introduced the Nemotron 3 family in December 2025, starting with Nemotron 3 Nano available immediately and Nemotron 3 Super and Ultra following in the first half of 2026. Nemotron 3 Ultra eventually landed at 550 billion parameters in its Ultra-550B-A55B configuration, shipped with BF16 and NVFP4 checkpoints, post-trained variants and full training datasets.
Lightning sits at the compact end of that range by design, a 30B model meant for constant, low-cost agent work rather than frontier reasoning. But NVIDIA isn’t stopping at 550 billion parameters. The company is already developing Nemotron 4, a family expected to reach at least 1 trillion parameters in its largest configuration, roughly double Nemotron 3 Ultra, aimed at matching the world’s best open-source models head-on rather than just competing with them from the mid-tier.
The Nemotron Coalition and the Mistral AI Partnership
In March 2026, NVIDIA formed the Nemotron Coalition, a group of AI labs collaborating on an open frontier model trained on NVIDIA’s DGX Cloud infrastructure. The coalition’s first project pairs NVIDIA with Mistral AI on an open base model that’s expected to feed directly into Nemotron 4. The arrangement ties open-model development to DGX Cloud GPU capacity from the start, meaning the training runs themselves already generate NVIDIA hardware demand well before any model ships.
NVIDIA has also struck a multibillion-dollar agreement with startup Poolside to build a separate open-weight AI system, extending the same playbook beyond its own labs. Every one of these partnerships shares a structure: NVIDIA supplies compute and, often, engineering resources, and the resulting model gets released openly, expanding the base of developers who build on NVIDIA-optimized tooling.
How Nemotron 3.5 Lightning Compares to Other Open Models
Direct, apples-to-apples benchmark scores against Llama, Qwen, DeepSeek and GPT-OSS were not available from any outlet at launch, and no serious report should pretend otherwise. What is documented is how the models differ in architecture, licensing and deployment footprint, which matters just as much to a team deciding what to run.
| Model | Total / Active Params | Context Window | License | Minimum Deployment |
|---|---|---|---|---|
| NVIDIA Nemotron 3.5 Lightning | 30B / 3B active (MoE) | 1,000,000 tokens | OpenMDW-1.1 (permissive) | 1x H100 or 1x DGX Spark |
| NVIDIA Nemotron 3 Ultra | 550B / 55B active (MoE) | Not disclosed at 1M tier | NVIDIA Open Model License | Multi-GPU data center node |
| NVIDIA Nemotron 4 (in development) | Up to 1T+ (largest variant) | Not yet published | Expected open license | Multi-node cluster |
| Llama 4 (Meta) | Varies by variant | Up to 10M tokens (Scout) | Llama Community License | Multi-GPU, model-dependent |
| DeepSeek V3.2 | 671B / 37B active (MoE) | 128,000 tokens | MIT License | Multi-GPU data center node |
The comparison that stands out isn’t parameter count, it’s who can run the thing at all. Nemotron 3.5 Lightning is the only model in that table that a developer can genuinely run on a single consumer-adjacent GPU today. That’s the entire point of the release: NVIDIA is optimizing for reach among smaller developers and edge deployments, not for topping a leaderboard against DeepSeek V3.2 or Llama 4’s largest configurations.
Early Enterprise Adopters Are Already Customizing It
NVIDIA named specific early adopters testing and customizing Nemotron 3.5 Lightning for their own agentic workloads: CrowdStrike, CodeRabbit and Harvey. That’s a telling trio. CrowdStrike works in cybersecurity, where keeping model weights and training data on-premises is close to a hard requirement. CodeRabbit builds automated code review agents, exactly the kind of high-volume, always-on task Nemotron 3.5 Lightning was designed for. Harvey serves the legal industry, another sector where data never leaves the building if the client has any say in it.
None of those three are hobbyist users. They’re enterprise accounts that, once committed to a local Nemotron deployment, become long-term buyers of the GPUs that run it, whether that’s H100s already in a data center or a fresh DGX Spark order. NVIDIA doesn’t need Nemotron 3.5 Lightning to be free for goodwill. It needs it to be sticky.
The Earnings Backdrop: A Record Quarter Two Weeks Later
NVIDIA reported fiscal second-quarter results on August 26, and the numbers gave the open-source-for-hardware thesis some real evidence. Total revenue reached $96.2 billion, up 106% year-over-year and above the roughly $92.17 billion analysts had penciled in. Data center revenue alone hit $89.0 billion, up 117% year-over-year and 18% sequentially, accounting for about 92% of total sales.
| Metric | Q2 FY2026 (ended July 26, 2026) | Prior Quarter | Year-over-Year Change |
|---|---|---|---|
| Total revenue | $96.2 billion | $81.5 billion (implied) | +106% |
| Data center revenue | $89.0 billion | $75.2 billion | +117% |
| AI clouds/industrial/enterprise (ACIE) spend | $40.3 billion | Not separately disclosed | +138% |
| Non-GAAP EPS | $2.22 | Not disclosed here | Above $2.06–$2.09 estimate |
| Q3 FY2026 revenue guidance | ~$91 billion (+/- 2%) | — | Above prior Street consensus |
The AI clouds, industrial and enterprise segment, the bucket that captures exactly the kind of customer NVIDIA is courting with Nemotron 3.5 Lightning, grew 138% year-over-year to $40.3 billion. That’s demand from companies that aren’t hyperscalers, the same audience the Nemotron open-model strategy targets directly.
Wall Street’s Read on the Strategy
NVIDIA shares traded in a wide band through August, climbing to $227.88 on August 17 before easing back toward $214 as investors waited on the August 26 report, and settling around $206.89 to $210.38 in the days immediately after earnings. Market capitalization sat between roughly $5.1 trillion and $5.3 trillion through the month, keeping NVIDIA in position as the world’s most valuable public company. The stock was up about 12.6% year-to-date heading into the print.
NVIDIA has also told customers to expect price hikes exceeding 15% on some AI products, citing strong demand and limited supply for AI chips, a detail that undercuts any narrative that giving away software is somehow softening NVIDIA’s pricing power on hardware. If anything, the two moves reinforce each other: free models expand the buyer pool, tight supply keeps the price per chip high.
Historical Context: NVIDIA’s Slow Pivot to Open Models
NVIDIA wasn’t always an open-model company. For most of the last decade it sold hardware and left model development to labs like OpenAI, Google DeepMind and, eventually, Meta and DeepSeek. That changed once open-weight models from Meta’s Llama family and China’s DeepSeek proved they could match or beat proprietary systems on many practical tasks while running on hardware the buyer already owned.
NVIDIA’s answer was to become a model publisher itself, but one whose models are consistently tuned around its own quantization formats (NVFP4 in particular) and its own inference stack. The first Nemotron models targeted narrow enterprise use cases. By the time Nemotron 3 launched in December 2025, NVIDIA had a full roadmap running from small, locally deployable models up to a 550-billion-parameter Ultra tier. Nemotron 3.5 Lightning and the coming Nemotron 4 extend that roadmap in both directions, smaller and faster on one end, larger and more capable on the other.
What This Means for GPU Supply and Pricing Through 2027
Every additional developer or enterprise that adopts Nemotron 3.5 Lightning for local agent work is, in effect, a new GPU customer, even if a small one. Multiply that by CrowdStrike-scale enterprises rolling out fleets of DGX Spark boxes and the demand curve NVIDIA is chasing becomes clearer. Analysts covering the August earnings cycle noted that AI infrastructure demand from CoreWeave, Super Micro Computer and similar buyers remains strong enough that NVIDIA is raising prices rather than cutting them, even as it hands out free software.
That combination, rising list prices alongside free model releases, is unusual outside the chip industry. It only works if the free software genuinely expands the addressable market for the hardware rather than just redistributing existing demand. NVIDIA’s bet, backed by a quarter that beat estimates by roughly $4 billion in data center revenue, is that it does.
Competitive Pressure: AMD and the Open-Model Arms Race
AMD has its own open accelerator push, and the company’s Instinct MI-series chips remain the most credible alternative to NVIDIA’s data center GPUs on raw specs. What AMD doesn’t have is a matching open-model ecosystem tuned specifically to its own hardware at NVIDIA’s scale. Readers comparing the raw chip specs side by side can see how close the two companies have gotten on paper in our RX 9070 XT vs RTX 5070 Ti breakdown, but Nemotron shifts the competition from silicon benchmarks to software gravity.
By giving developers a genuinely capable free model that happens to run best on NVIDIA chips, the company builds a software moat around its hardware lead that’s harder for AMD to counter with price cuts alone. It’s the same dynamic playing out in the next-generation accelerator race between NVIDIA’s Rubin platform, AMD’s Helios and Microsoft’s Maia 300, where compute performance is only half the argument and ecosystem lock-in is the other half.
Predictions: Where This Goes Next
- Nemotron 4 arrives with real benchmark data. Once NVIDIA publishes its trillion-parameter-class model, expect independent labs to finally produce head-to-head scores against Llama and DeepSeek, filling the gap that exists today.
- More enterprise names join the CrowdStrike-Harvey-CodeRabbit cohort. Regulated industries (healthcare, finance) are the most likely next adopters given the on-premises data argument Switchyard is built to support.
- DGX Spark sees a sales bump tied directly to Nemotron. Expect NVIDIA to cite Nemotron adoption explicitly in its next two earnings calls as a driver of edge and workstation GPU demand.
- AMD responds with its own open-weight release tuned for Instinct hardware. The Nemotron playbook is easy to describe and hard to ignore, and AMD has the incentive to copy the free-model-for-hardware-demand structure.
- Price hikes on AI chips continue into 2027 even as more free software ships, since the two trends reinforce rather than offset each other as long as data center orders keep beating estimates.
Frequently Asked Questions
What is NVIDIA Nemotron 3.5 Lightning?
It’s a 30-billion-parameter, open-weight mixture-of-experts AI model NVIDIA released on August 11, 2026, built for always-on AI agents and designed to run on a single GPU, such as an H100 or a DGX Spark system.
Is Nemotron 3.5 Lightning really free?
Yes. It’s released under the OpenMDW-1.1 license, which lets developers use, modify, distribute and commercially deploy the model and any derivatives without paying NVIDIA or crediting the company.
What hardware do I need to run Nemotron 3.5 Lightning?
NVIDIA lists single-GPU deployment on one DGX Spark system (GB10 superchip) or one H100 as the supported configurations. The developer blog also references GeForce RTX 5090 and NVIDIA Jetson as local deployment targets.
How does Nemotron 3.5 Lightning compare to Llama or DeepSeek?
No independently verified benchmark comparison was available at launch. What’s documented is architecture and deployment footprint: Nemotron 3.5 Lightning is smaller and runs on a single consumer-adjacent GPU, while DeepSeek V3.2 (671B total parameters) and Llama 4’s larger variants require multi-GPU data center setups.
What is NeMo Switchyard?
It’s an open-source model routing library NVIDIA released alongside Nemotron 3.5 Lightning that automatically sends each prompt to the most appropriate model based on cost, speed and task complexity, letting enterprises mix local and cloud models in one pipeline.
Why is NVIDIA giving away an AI model instead of selling it?
The stated logic, echoed across coverage of the launch, is that wider access to capable open models expands the number of people and companies running AI workloads, and those workloads still need NVIDIA GPUs to train, fine-tune and serve, both locally and in the cloud.
How did NVIDIA’s stock react around the Nemotron launch?
NVIDIA shares traded between roughly $206 and $228 through August 2026, with market capitalization ranging from about $5.1 trillion to $5.3 trillion, and the launch came two weeks ahead of a fiscal second quarter that beat consensus revenue estimates by several billion dollars.
What is Nemotron 4 and when is it coming?
Nemotron 4 is NVIDIA’s next open-model family in development, expected to reach at least 1 trillion parameters in its largest configuration, roughly double the 550-billion-parameter Nemotron 3 Ultra. No confirmed release date had been published as of late August 2026.
Related Coverage
- Nvidia Rubin vs AMD Helios vs Microsoft Maia 300: AI Chip Race Hits $100B [2026]
- Nvidia Rubin’s 6 Chips Take On Google TPU 8i/8t [2026]
- Nvidia Skips New Gaming GPUs, First Time Since 1999 [2026]
- DeepSeek V4 vs R1 vs V3.2: Peak Prices Surge 355% [2026]
- Cerebras CS-4 Claims 30x Faster AI Inference Than GPUs [2026]
- RunPod vs Lambda vs Vast.ai: 63% H100 Price Gap [2026]
- South Korea Releases 1.56T-Token AI Training Data [2026]
For ongoing coverage of AI accelerators, model releases and the chip market driving them, see our AI chips 2026 hub.