NVIDIA is making one of its biggest bets outside hardware: a deal tied to Poolside worth around $7 billion, with $6 billion to license its technology, plans to bring on more than a hundred employees, and an extra $1 billion investment in the company. The move could speed up NVIDIA’s open Nemotron models and confirms Jensen Huang’s push to grow the company well beyond selling AI GPUs.
NVIDIA and Poolside in 20 seconds
- NVIDIA would pay $6 billion to license Poolside’s technology, plus a $1 billion investment.
- More than 100 startup employees could join NVIDIA.
- The goal is to strengthen Nemotron and compete harder in open-weight models.
- Poolside would keep operating as an independent company.
- Framing this as a hedge against falling GPU demand is, for now, interpretation, not confirmed fact.
Poolside announced the deal to its investors, and various outlets later reported it. The licensing arrangement isn’t an acquisition: Poolside would keep its founders and part of its structure, while NVIDIA would mainly bring in talent tied to its Laguna models and gain access to the Model Factory, the infrastructure the startup uses to train and build its systems.
The number is big even by today’s AI standards, but it makes more sense given what NVIDIA is building around Nemotron and how fast Chinese open-weight models are gaining ground.
Poolside adds more than engineers to NVIDIA
Poolside focuses on AI for software development and on training models that can handle long programming tasks.
In May it introduced Laguna M.1 and Laguna XS.2, two Mixture-of-Experts models built for long-running agentic programming. Laguna M.1 has 225.8 billion total parameters, with 23.4 billion active per token, while XS.2 has 33.4 billion and activates about 3 billion per token.
More important than the parameter counts is the system Poolside built around them.
The company talks about a Model Factory, a set of tools, data, training infrastructure, and processes for creating and improving models through real software development tasks.
That knowledge lines up directly with where NVIDIA is going with Nemotron.
In August, NVIDIA introduced Nemotron 3.5 Lightning, a 30-billion-parameter Mixture-of-Experts model built for specific tasks in multi-agent systems and workloads that need to stay active a long time. NVIDIA pairs it with tools like NeMo Switchyard to route requests between models.
So NVIDIA’s interest in Poolside doesn’t look limited to buying a finished model. It would be getting technology, training processes, and specialized talent to keep building its own model family.
NVIDIA also wants to compete at the model layer
Through much of AI’s rise, NVIDIA sat in a very comfortable spot.
OpenAI, Anthropic, Meta, xAI, Microsoft, Google, and almost any lab that needed a lot of accelerated capacity would end up buying or renting NVIDIA GPUs.
In that setup, the company supplied the tools others competed with.
Nemotron starts to change that.
NVIDIA offers models developers can adapt and run on different infrastructure, but naturally they’re heavily optimized for its own hardware and software. That lets the company play in many layers at once:
GPU → networks → complete systems → CUDA software → libraries → inference → models
The Poolside move reinforces that last layer.
It also reflects a geopolitical shift in the market. Chinese labs have gained ground with open-weight models like DeepSeek and Kimi, which let companies and developers deploy systems without always depending on closed APIs.
The U.S. keeps some of the most capable proprietary models, but a big part of the competition in open-weight AI is coming from China.
NVIDIA has strong reasons to keep this layer from being dominated mostly by models designed and optimized around other technology ecosystems.
Is NVIDIA trying to become its own GPU buyer?
One striking reading of the deal suggests NVIDIA is preparing a kind of insurance against a future drop in demand.
The logic is simple.
If NVIDIA builds very competitive Nemotron models and large AI services around them, it could use more and more of its own accelerators internally. In an oversupply scenario, that internal demand would act as a semi-alternative home for its compute.
The idea isn’t impossible, but the available data don’t confirm it’s the purpose of the Poolside deal.
NVIDIA hasn’t announced plans to become a “last-resort” buyer of its own GPUs, and it isn’t stockpiling capacity to absorb a future crash.
Recent moves point the other way: the company keeps talking about very high infrastructure demand.
On August 10, NVIDIA announced deals with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to create platforms that could mobilize more than $500 billion in third-party capital for AI infrastructure. Jensen Huang says AI computing is becoming a new class of infrastructure asset.
A week later, NVIDIA also announced an agreement with SB Energy tied to Ohio’s PORTS-Pike campus. The company will provide credit support to secure an initial 4.25 GW of land, energy, and buildings, with an option for another 3.75 GW; OpenAI is listed as a prospective customer for the 8 GW of IT. NVIDIA also disclosed a $1.5 billion investment in SB Energy.
These are huge commitments, hard to square with a company publicly expecting demand to collapse soon.
The more careful read is different: NVIDIA is trying to gain more control over the AI value chain.
The $7 billion can also be read as a defense of CUDA
There’s another, less obvious reason NVIDIA might want very competitive models of its own.
Every popular model shapes the technology choices around it.
Companies pick inference engines, quantization formats, libraries, hardware, cloud platforms, and orchestration systems based on the workloads they run.
If Nemotron catches on, NVIDIA can make sure its models are optimized from day one for CUDA, TensorRT, Blackwell, Rubin, and its future platforms.
Instead of waiting for another lab to publish a model and optimizing it later, NVIDIA can co-design the model and the infrastructure.
That kind of vertical integration is an advantage companies like Apple have used for years in other markets: hardware and software evolve together.
NVIDIA now has the resources to try something similar in AI.
An increasingly odd relationship with OpenAI and Anthropic
The move also creates an awkward situation.
OpenAI and Anthropic are big buyers of accelerated infrastructure, and they’re part of the market that made NVIDIA one of the world’s most valuable companies.
But if Nemotron keeps expanding, NVIDIA starts to compete with them in certain use cases.
The difference is that, for now, it isn’t trying to copy exactly the same models.
OpenAI and Anthropic build much of their business around models people reach through products and APIs. NVIDIA is betting more openly on open-weight models other companies can deploy and customize.
That approach might even raise GPU demand.
A high-quality open model doesn’t only run in its creator’s data centers. It can be deployed across hundreds of cloud providers, companies, universities, and governments that all need compute.
From that angle, building Nemotron could be less about hedging against falling GPU sales and more about creating new reasons to buy them.
Poolside also needed a lot more compute
The relationship between the two companies shows just how much infrastructure access has come to matter.
According to information shared with investors, Poolside had tried to raise billions and even floated infrastructure built around roughly 40,000 GB300 GPUs. The startup struggled to finance that capacity fast enough.
For NVIDIA, it’s the opposite.
It has the hardware and enormous financial resources, but building top-tier models takes researchers, data, training methods, and expertise.
This deal pairs the two needs.
Poolside gets capital and keeps its independence. NVIDIA gets more than a hundred professionals and licenses technology that could save years of in-house development.
Sam Altman isn’t saying the AI boom is over
It’s also worth separating this deal from a different debate around Sam Altman’s forecasts.
OpenAI’s CEO has acknowledged the economic shift AI drives won’t necessarily arrive overnight. OpenAI compares AI’s rollout to technologies like electricity, whose effects ran deep but took years to spread through the economy.
That’s very different from saying Altman thinks AI demand will now collapse.
OpenAI’s August data show rising adoption. ChatGPT passes 1 billion users, and the company notes firms are moving from using AI just for answers to using it for full tasks and processes through agents.
Altman has also argued recently that building the most advanced models may need a more controlled pace so safety systems and society can adapt. That’s mainly about safety and governance, not a belief that the economic need for compute is fading.
The distinction matters, because it keeps two separate stories from merging into one bubble-about-to-burst narrative.
NVIDIA is becoming more than NVIDIA
The $7 billion tied to Poolside probably matters more strategically than financially.
For a company of NVIDIA’s size and cash flow, it doesn’t change the finances much. But it could change how the company sees what kind of business it wants to be in five years.
Jensen Huang started by building GPUs.
Then NVIDIA created CUDA and made a big chunk of accelerated computing depend on its software platform.
After that it added networks, complete systems like DGX and GB200 NVL72, cloud services, inference engines, and tools for building agents.
Now it’s investing billions in its own models that will use all that infrastructure.
That story is far more interesting than framing the deal as a hedge against a possible sales dip.
NVIDIA could end up making the accelerator, designing the system, providing the network, funding the data center, building the inference software, and offering the model that runs on all of it.
That much integration also carries risk. The company has to avoid competing too directly with its biggest clients, justify bigger investments, and get enough Nemotron adoption to challenge open models that change almost weekly.
Poolside provides the technology and talent to make that happen.
What the deal doesn’t show is that NVIDIA expects to run out of GPU buyers. Right now the signs point the other way: it’s using its extraordinary hardware profits to occupy more layers of the AI business before others do.
Frequently Asked Questions
How much will NVIDIA invest in Poolside?
The available information puts the total commitment around $7 billion: $6 billion for a non-exclusive technology license and another $1 billion through an investment in Poolside at a pre-money valuation of $12 billion.
Has NVIDIA bought Poolside?
No. The deal announced to investors isn’t a takeover. Poolside will keep operating independently, and its founders will stay. More than 100 employees, mainly tied to its models, could join NVIDIA.
What does NVIDIA want to do with Poolside’s technology?
A main goal is to speed up Nemotron, NVIDIA’s open-weight model family, using Poolside’s training technology and expertise.
Is NVIDIA buying its own GPUs because it fears demand will collapse?
There isn’t enough evidence for that. Its investments in models and infrastructure can be read as diversification, but NVIDIA keeps pushing hard to expand capacity and publicly says AI compute demand is still rising.
