Formed via an intent in September 2025, the new strategic partnership between Openai and Nvidia is designed for both Openai's next-generation computing infrastructure, which is expected to be deployed early in the second half of 2026.
At a high level:
- The target scale is over 10 gigawatts (GW) of deployed computational capacity realized through NVIDIA systems (including millions of GPUs).
- The first phase (1 GW) is scheduled for the second half of 2026 and is built on the upcoming Vera Rubin platform.
- Nvidia will gradually invest up to $100 billion in Openai, subject to gradual capacity deployment.
- The initial $10 billion investment from NVIDIA is linked to the implementation of the system's first gigawatt decisive purchase agreement.
- Equity stakes that Nvidia acquires are described as non-voting/non-controlled. That is, it provides in-game financial skins without governance control.
From a strategic perspective, linking investments to capacity deployments can help open locking capital and hardware on the long horizon, mitigating supply chains and financing risks. This type of gradually fixed commitment, as it is frequently cited as a binding constraint to progression models, gives Openai a predictable growth pathway (says that, at least in theory, accurate economic terms and risk sharing are not fully revealed.
The press conference statement highlights that millions of GPUs will ultimately be involved and co-optimizing Nvidia's hardware using Openai's software/stack will become a key feature of collaboration.
Importantly, the deal also fits Openai's broader strategy of diversifying infrastructure partnerships beyond a single cloud provider. Microsoft remains a central supporter and collaborator, but this Nvidia partnership will further expand Openai's computing base and complement other announced partnerships (such as Oracle, Softbank, Stargate).
The transaction also illustrates NVIDIA's strategic change. Become a strategic investor for anchor customers who directly drive GPU demand, not just chip vendors. This alignment strengthens the connection between Nvidia's roadmap, software ecosystem, and actual termination deployments.
In an interview with CNBC, Nvidia CEO Jensen Huang characterized the initiative.
“This is the largest AI infrastructure project in history. The partnership is to build an AI infrastructure that allows AI to travel from the lab to the world.”
Nvidia's press documents also add that the new AI infrastructure will provide “billion times more computing power” than the first DGX system provided to Openai in 2016.
Openai's critical architecture upgrade
Nvidia's Vera Rubin family and the new NVL144 CPX system represent an intentional architectural pivot to ultra-density rack-scale platforms optimized for extremely long context windows, multimodal workloads and generated videos. Nvidia's Rubin CPX announcement frames the platform surrounding the “Million-Token” inference/Prefill use cases, highlighting extremely large on-rack memory and extreme cross-rack bandwidth. The VeraRubinNVL144 CPX Rack is advertised as offering about 8 ExaFlops of ~100 TB of ~100 TB of Agg, ~100 TB of about 8 ExaFlops. Overall performance uplift Nvidia says it is about 7.5 times more than the previous GB300 NVL72 system.
Technically, the Rubin family also introduces a decomposition approach to long contextual inference. NVIDIA ships both multiple Rubin CPX (reported with ~30 PFLOPS NVFP4 with 128 GB of GDDR7 per socket in some configurations) and high-band wide-rubin GPUS (both genent/contation/in phs in firfy ph, both large-scale HBM interpolation). It can be mapped to the most suitable silicon. This split is intended to provide much more effective context length and throughput when combined with software optimization, along with the next-generation NVLink/NVSwitch fabric, along with the new NVLINK-144 switches and NICS referenced in reinforced silicon photonics and platform materials.
In the case of OpenAI, the practical meaning is a co-design of hardware + software scale. Rubin-ERA GPUS, NVL144 NVLINK/NVSWITCH fabric, and accelerated networks should require alignment and performance adjustments for OpenAI training and inference pipelines and CUDA/SDK roadmap. Public coverage and Nvidia materials explicitly invoke joint optimization and rack-scale system designs aimed at 1 million token workloads.
There are operational and deployment results that are worth flagging. The NVL144 CPX rack is located for production cargo late 2026 (NVIDIA public timetable). This coincides with the timing of the first phase of 1 Golden Week announced at Nvidia – Openai Loi. Scaling OpenAI and NVIDIA capacity types may require a distributed deployment across multiple campuses and providers. Multiple independent reporting outlets and system-level analyses highlight the nature of the platform's rack-scale and practical limitations (power supply, cooling, site procurement). Treat the “distributed campus” statements as inferences based on information from platform design and industry power/site reality, rather than lines from NDAs or decisive filings.
