Nvidia spent the past 48 hours doing something it rarely does: giving away the playbook. On October 8, 2026, longtime AI analyst Michael Parekh used his daily newsletter AI: Reset to Zero to lay out what he calls Nvidia’s open-source “AI Factory” push, detailing how the chipmaker is funding rivals, buying distribution, and publishing its own models to keep pace with China’s open-weight momentum. The dispatch, numbered AI-RTZ #1233 (Part 1), arrived the same week trade outlet shattered.io reported that building a single one-megawatt AI factory now costs roughly $60 million, a figure that frames just how much is riding on Nvidia’s bet that open models, not closed ones, will decide who controls the next decade of AI infrastructure.
The stakes are not abstract. Chinese labs have spent 2026 shipping open-weight models at a pace US developers are struggling to match, and Nvidia CEO Jensen Huang has spent the summer publicly arguing that banning them would hurt America more than help it. This article unpacks what Nvidia actually announced, what it is spending, and what it means for every company that runs AI workloads on Nvidia silicon.
What Nvidia Just Announced
According to Parekh’s AI-RTZ #1233 dispatch, titled “Nvidia the Kingmaker” in its opening section, Nvidia is running what amounts to a six-lever strategy to keep the open-weight AI ecosystem anchored to its chips. The newsletter describes how the company funds open-model startups, convenes industry coalitions, lobbies for open-weight-friendly policy, allies with security initiatives, owns distribution infrastructure, and builds its own competing open models in parallel. The timing lines up with a separate NVIDIA DSX and Omniverse DSX blueprint push, which Nvidia’s own newsroom describes as an open playbook for building gigawatt-scale AI factories, validated at a new AI Factory Research Center at a Digital Realty site in Manassas, Virginia.
Nvidia says the Omniverse DSX blueprint scales from 100 megawatts up to multiple gigawatts and is designed to help operators tap into roughly 100 gigawatts of underutilized grid capacity using Max-Q power optimization, which Nvidia claims can lift GPU throughput by as much as 30%. That is the physical infrastructure half of the story. The software half, per Parekh’s reporting, centers on who actually trains and distributes the open models that run inside those factories, and this is where the China comparison gets sharp.
The $2 Billion Reflection Bet and the Arrival of Beam
The clearest example of Nvidia’s funding lever is its roughly $2 billion investment in Reflection AI, a New York lab that priced its latest round at a $25 billion pre-money valuation, according to AI-RTZ #1233. Reflection used that capital to ship Beam, its first open-weight model, a launch tech-insider.org covered in detail in our piece on Reflection AI’s Beam and its 4x compute efficiency claim, and in our earlier look at Reflection’s $25 billion bet against China’s open models.
Parekh’s newsletter reports that Beam needs roughly 3 to 4 times less computing power than comparable open models to perform equivalent reasoning tasks, and that while it still trails the US frontier, it closes most of the gap with China’s strongest open-weight releases, landing near Alibaba’s Qwen3.8-Max on capability. Reflection CEO Misha Laskin framed the competitive backdrop bluntly in comments cited by the newsletter: Western labs chasing frontier-scale open models “don’t really have very good options today” if they want to match what Chinese developers are shipping on a comparable budget.
Nemotron 3: Nvidia’s Own Open-Weight Answer
Funding outside labs is only one lever. Nvidia is also building its own open models under the Nemotron 3 family. The smallest member, Nemotron 3 Nano, shipped on Hugging Face and runs a hybrid architecture combining 23 Mamba-2 layers, 23 mixture-of-experts layers, and 6 attention layers, for 30 billion total parameters with only 3.5 billion active per token, according to Nvidia’s own model card. It supports a 1-million-token context window and ships under the Nvidia Open Model License, which permits commercial use and modification without a licensing fee.
Shattered.io reports that larger Nemotron 3 models, including a version Nvidia calls Nemotron 3 Ultra, are still coming in the first half of 2026 and will stay under the 1-trillion-parameter mark, deliberately smaller than the multi-trillion-parameter models Chinese labs are now shipping. Nvidia reportedly spends around $6 billion annually on the Nemotron program, treating it less as a product line and more as a reference implementation that keeps developers building on CUDA and Nvidia GPUs regardless of which lab’s weights they ultimately deploy.
China’s Open-Weight Lead, By the Numbers
The scale gap between Chinese and Western open releases is the detail driving most of the urgency in Nvidia’s strategy. Moonshot’s Kimi K3 reportedly weighs in at 2.8 trillion parameters, Alibaba’s Qwen3.8-Max sits around 2.4 trillion, and Alibaba is already signaling plans for “super-scale” successors in the 5 to 10 trillion parameter range. Z.ai’s GLM-5.2 landed in June 2026, adding another entrant to a crowded field tech-insider.org has tracked elsewhere, including our comparison of DeepSeek’s V4.1 Flash against Claude Haiku 5.5 and Gemini 3.8 Flash.
| Model | Developer | Region | Reported Scale | Notable Claim |
|---|---|---|---|---|
| Beam | Reflection AI | United States | Not fully disclosed | 3-4x less compute for comparable reasoning |
| Qwen3.8-Max | Alibaba | China | 2.4 trillion parameters | Frontier-class open weights |
| Kimi K3 | Moonshot AI | China | 2.8 trillion parameters | Largest widely cited open release |
| GLM-5.2 | Z.ai | China | Not fully disclosed | June 2026 release |
| Inkling | Thinking Machines Lab | United States | Under 1 trillion parameters | US frontier open-weight entrant |
| Nemotron 3 Ultra | Nvidia | United States | Under 1 trillion parameters | Reference model for CUDA ecosystem |
According to AI-RTZ #1233, Nvidia helped form a Nemotron Coalition of eight labs back in March 2026, specifically to coordinate open-weight releases that stay competitive with Chinese scale without each lab bearing the full training cost alone. A separate Open Secure AI Alliance is reportedly working the policy side, lobbying regulators to treat open weights as a security asset rather than a liability.
The $60 Million Megawatt: Pricing the AI Factory
Shattered.io’s reporting puts a hard number on the infrastructure side of this race: roughly $60 million to build out a single megawatt of AI factory capacity. That figure matters because Nvidia’s Omniverse DSX blueprint is explicitly designed for multi-gigawatt facilities, meaning the capital commitment behind a single DSX-validated site can run into the billions before a single chip ships. Our earlier coverage of the Nvidia DGX Spark 64GB launch amid surging memory prices touched on the same underlying pressure: memory and power are now bigger bottlenecks for AI factory buildouts than GPU allocation itself.
Shattered.io also reports that Huang’s forecast commitment to roughly double chip output has locked up around 37% of available HBM supply, a constraint that touches every hyperscaler and neocloud trying to stand up new capacity in 2026 and 2027. That squeeze is part of why Nvidia’s open-source strategy matters commercially as much as ideologically: every open model that ships, whether it is Beam, Nemotron, or a Chinese release, still needs GPUs to run inference, which keeps demand flowing back to Nvidia regardless of which lab wins the open-weight race.
Owning Distribution: The Hugging Face and Poolside Deals
Funding and building models only gets Nvidia so far if developers go elsewhere to find and deploy them. That is reportedly why Nvidia moved to own distribution infrastructure directly, acquiring Hugging Face for about $13 billion and separately acquiring the Poolside team and licensing its technology for roughly $6 billion, a combined $20 billion commitment according to AI-RTZ #1233. Hugging Face is the default hosting and discovery layer for open-weight models worldwide, including most of the Chinese releases named above, so controlling it gives Nvidia visibility into exactly which models developers are pulling and how they are being deployed.
| Commitment | Reported Amount | What It Buys Nvidia |
|---|---|---|
| Reflection AI investment | $2 billion | Stake in a $25B open-weight lab, access to Beam |
| Hugging Face acquisition | $13 billion | Default distribution layer for open models |
| Poolside team + licensing | $6 billion | Coding-model talent and technology |
| Nemotron program | $6 billion / year | In-house open-weight reference models |
| SpaceX Colossus 2 compute deal | $6.3 billion through 2029 | Long-term training capacity commitment |
| Nebius compute agreement | $1 billion | Additional neocloud training capacity |
Layered on top of those deals, Nvidia’s compute agreements with SpaceX for its Colossus 2 data center (reported at $6.3 billion running through 2029) and a separate $1 billion arrangement with neocloud provider Nebius show the company locking in training capacity years in advance, regardless of which lab’s models ultimately run on that hardware.
Jensen Huang’s Message to Washington
Huang has been unusually direct about why he thinks banning Chinese open models would backfire. In a July 2026 interview, he praised Chinese open releases directly and argued that strong open-source models should be adopted rather than blocked, a stance Fortune reported put him at odds with parts of Washington’s China-hawk contingent. When Nvidia released Nemotron 3.5 Lightning the following month, Huang used the launch to reinforce the same point, tying free, open models directly to future hardware and chip demand, as CNBC covered.
Huang made the same argument to a wider audience in his first-ever post on X, where he shared an open letter signed by roughly 20 companies arguing that open weights make the AI ecosystem safer and more secure, not less. The subtext across all three statements is consistent: Nvidia profits either way, since open models (American or Chinese) still need to run somewhere, and Nvidia GPUs remain the default place to run them.
Market Impact: Compute Deals, HBM, and Enterprise Buyers
For enterprise buyers, the practical effect of this strategy is day-zero optimization. Shattered.io reports that Nvidia’s software stack, including the open-source Dynamo 1.0 serving layer, now ships with tuned support for Chinese models like DeepSeek V4 Flash and Qwen3.8 the same day they land on Hugging Face, alongside Nvidia’s own Nemotron releases. For IT buyers choosing between Nvidia-optimized Chinese open weights and closed Western alternatives, the calculus increasingly comes down to enterprise security posture rather than raw capability, since security and compliance concerns still steer most regulated-industry API spending toward Western vendors even when Chinese open models benchmark competitively.
South Korea’s Shinsegae Group is reportedly piloting a sovereign AI factory build using Nvidia’s DSX blueprint, an early signal that national and corporate buyers outside the US and China are treating Nvidia’s open blueprint, rather than a specific model, as the safer long-term bet. That is consistent with how tech-insider.org covered the parallel story of a California CEO charged in a $300 million Nvidia GPU smuggling case, which underscored just how much demand exists for Nvidia silicon outside approved export channels, open-source strategy or not.
Competitive Landscape: Nvidia’s Open Bet vs the Field
Nvidia is not the only chipmaker watching the open-weight race reshape demand. Chinese silicon vendors are riding the same wave of Chinese open-model adoption. Our coverage of Biren Technology’s GPU revenue surging nearly 2,000% in 2026 shows domestic Chinese GPU demand climbing in lockstep with the open models trained and served on them, even as Biren posts a net loss while scaling. AMD, meanwhile, has leaned on its own open software stack to court the same developers Nvidia is courting with Nemotron, a dynamic our piece on who actually makes GPUs besides Nvidia and AMD breaks down in detail.
What distinguishes Nvidia’s approach from AMD’s or from a pure hardware play is the breadth of the six-lever strategy itself. Few chipmakers can afford to simultaneously fund a $25 billion startup, acquire a $13 billion distribution platform, bankroll an in-house model family, and lock in multi-billion-dollar compute deals. That breadth is Nvidia’s moat as much as its chips are.
Historical Context: From Closed CUDA to Open Weights
Nvidia’s turn toward open models is a notable shift for a company that built its dominance on a closed software moat. CUDA has been proprietary since its 2006 introduction, and for most of the 2010s Nvidia’s strategy was to lock developers into that stack rather than give software away. The open-weight era, kicked off in earnest by Meta’s original Llama releases and accelerated by Chinese labs like DeepSeek shipping competitive open models at a fraction of Western training budgets, forced a rethink. Once developers could get frontier-adjacent capability for free, the scarce resource stopped being the model and became the compute to run it, which is exactly the resource Nvidia controls. Huang’s public embrace of open weights throughout mid-2026 reads less like a philosophical conversion and more like a company recognizing where its actual leverage sits.
Predictions: Where Nvidia’s Open-Source Bet Goes Next
Based on the trajectory Parekh’s newsletter and shattered.io both describe, a few outcomes look likely over the next two to three quarters:
- Nvidia will keep Nemotron models deliberately under the trillion-parameter mark, competing on efficiency and day-zero tooling rather than raw scale against Chinese super-scale releases.
- Expect at least one more Nvidia-backed open-model startup announcement, following the Reflection AI pattern, as Nvidia continues funding external labs rather than relying solely on in-house Nemotron development.
- HBM supply constraints, with roughly 37% already committed against Huang’s 2x forecast, will keep pushing AI factory costs higher before any meaningful relief arrives.
- Policy pressure in Washington over Chinese open models will intensify rather than fade, putting Nvidia in the position of repeatedly defending open-weight adoption in public forums.
- More sovereign and enterprise buyers, following Shinsegae’s lead, will pilot Omniverse DSX-based builds specifically to avoid being locked into any single model provider, Chinese or American.
What Enterprise Developers Should Watch
For engineering teams deciding where to build, the near-term signal is less about which model wins and more about which models get day-zero Nvidia tooling support, since that determines real-world deployment cost regardless of benchmark scores. Dynamo 1.0’s open-source serving layer and the Nvidia Open Model License attached to Nemotron releases both lower the switching cost between open models, which is precisely the point: Nvidia wants teams comfortable moving between Beam, Nemotron, Qwen, or DeepSeek without rewriting their serving infrastructure, because every one of those paths still runs through Nvidia silicon.
Expert Voices on Nvidia’s Open-Source Push
Jensen Huang has made the clearest public case for why Nvidia is betting on openness rather than treating it as a threat. Speaking to Axios in July 2026, Huang said plainly: “These Chinese models are excellent,” adding that “Open-source models that are excellent should be used” (Axios, also reported by Fortune).
A month later, discussing the Nemotron 3.5 Lightning release, Huang distilled his commercial logic into two short lines, as covered by CNBC: “Free AI should be great for hardware. Free AI should be great for chips.”
On the policy side, Huang framed open weights as a strategic asset rather than a liability in the open letter he shared on his first-ever post on X, arguing that “Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty” (as reported by CNBC). Taken together, the three statements show a consistent through-line: Huang is betting that openness grows the market Nvidia sells into faster than it erodes Nvidia’s share of that market.
Related
- Reflection AI Beam: 501B Params, 4x Less Compute [2026]
- Reflection AI’s $25B Bet Takes On China’s Open Models [2026]
- Nvidia DGX Spark 64GB Launches at $4,999 as Memory Prices Surge [2026]
- Biren GPU Revenue Soars 1,998%, Posts $56M Loss [2026]
- Who Makes GPUs Besides Nvidia, AMD? 90% Share Gap [2026]
Frequently Asked Questions
What is Nvidia’s “AI Factory” open-source strategy?
Per Michael Parekh’s AI-RTZ #1233 newsletter, it is a six-part approach: Nvidia funds open-model startups like Reflection AI, builds its own open models under the Nemotron brand, convenes industry coalitions, lobbies for open-weight-friendly policy, partners with security initiatives, and owns distribution through its Hugging Face acquisition.
How much has Nvidia invested in Reflection AI?
AI-RTZ #1233 reports a roughly $2 billion Nvidia investment in Reflection AI, which priced its latest funding round at a $25 billion pre-money valuation before releasing its first open-weight model, Beam.
What is Beam, and how does it compare to Chinese open models?
Beam is Reflection AI’s first open-weight model. According to the newsletter’s reporting, it needs roughly 3 to 4 times less compute than comparable open models for equivalent reasoning performance, trails the US frontier, but approaches the capability level of Alibaba’s Qwen3.8-Max, one of China’s leading open releases.
Why did Nvidia buy Hugging Face?
Hugging Face is the primary hosting and discovery platform for open-weight AI models worldwide. Reports place the acquisition at roughly $13 billion, giving Nvidia direct visibility into and influence over how open models, including Chinese releases, get distributed and deployed.
How much does it cost to build an AI factory in 2026?
Shattered.io reports a cost of roughly $60 million per megawatt of AI factory capacity. Since Nvidia’s Omniverse DSX blueprint is designed to scale from 100 megawatts to multiple gigawatts, a single large facility can represent a multi-billion-dollar capital commitment.
Does Nvidia sell different chips to China?
Nvidia has stated it has no LPU sales in the China market and no China-specific LPU product on its roadmap, according to company statements cited by shattered.io. Export rules continue to shape which Nvidia products are available in which markets.
What is the Nemotron Coalition?
According to AI-RTZ #1233, it is a group of eight labs that Nvidia helped form in March 2026 to coordinate open-weight model releases, pooling resources so no single lab has to bear the full cost of matching Chinese open-model scale alone.
Will Nvidia’s Nemotron models ever match China’s largest open models in size?
Not based on current reporting. Nemotron 3 Ultra is expected to stay under 1 trillion parameters even as Chinese labs like Moonshot and Alibaba push past 2.4 trillion and plan 5-to-10-trillion-parameter successors, suggesting Nvidia is competing on efficiency and tooling rather than raw scale.