OpenAI and NVIDIA announced a letter of intent for a strategic alliance that, if finalized, will deploy at least 10 gigawatts (GW) of NVIDIA systems dedicated to OpenAI’s next-generation AI infrastructure. The agreement— which both companies hope to finalize in the coming weeks— positions NVIDIA as preferred partner for computing and networking to support the growth of OpenAI’s so-called “AI factories” and anticipates NVIDIA will invest up to $100 billion in OpenAI gradually, as each gigawatt is brought online.
The first operational milestone is set: the first GW will be deployed in the second half of 2026 on the NVIDIA Vera Rubin platform, the foundation for the upcoming wave of ultra-scale training and inference systems that Jensen Huang has positioned as the pillar for the leap to superintelligence.
“NVIDIA and OpenAI have driven each other forward over the past decade—from the first DGX to the ChatGPT milestone,” highlighted Jensen Huang, founder and CEO of NVIDIA. “This investment and infrastructure partnership represent the next leap: deploy 10 GW to fuel the next era of intelligence.”
“Everything begins with compute,” emphasized Sam Altman, co-founder and CEO of OpenAI. “The compute infrastructure will be the backbone of the economy of the future, and we will use what we are building with NVIDIA to create new breakthroughs and put them into the hands of people and businesses at scale.”
What does “10 GW” mean in practical terms
- Magnitude: ten gigawatts of AI systems imply millions of cutting-edge GPUs, interconnected by data factories with NVLink/NVSwitch, Ethernet/InfiniBand, and high-performance storage subsystems.
- Capacity: this level of compute is designed to train and serve next-generation models (trillions of parameters), agents with reasoning, tools, and long-term memory, as well as low-latency inference services for hundreds of millions of concurrent users.
- Energy and locations: deploying 10 GW isn’t just about hardware acquisition. It requires land, substation power, permits, advanced cooling (liquid/immersion), PPA agreements, and collaboration with grid operators to avoid stressing the electrical system.
NVIDIA frames this initiative within its concept of “AI factories”: data centers designed as intelligence production plants, where data, compute, and models converge with finely optimized pipelines and orchestration.
The Platform: NVIDIA Vera Rubin
Vera Rubin is NVIDIA’s next-generation platform, designed to:
- Extreme scalability in training clusters (NVL and successors), with terabits/sec bandwidth between GPUs.
- Energy efficiency per watt trained/inferred, critical when total power reaches gigawatt levels.
- HW/SW co-optimization: CUDA libraries, TensorRT-LLM, NeMo, frameworks, and data schedulers, along with compilers and runtimes that maximize each GPU generation.
The agreement with OpenAI adds a co-design vector: both companies will sync their roadmaps for models and infrastructure software (OpenAI) and hardware/software (NVIDIA) to reduce bottlenecks and deployment times.
A Pact with Industrial and Financial Implications
- Staged investment: NVIDIA plans to invest up to $100 billion in OpenAI incrementally as each GW is deployed, aligning capex (data centers, power, equipment) with capacity ramp-up.
- Supply chain: the commitment anticipates volume agreements with foundries, HBM memory manufacturers, optical networks, PSUs, and cooling systems. The availability of high-performance HBM and optics will be critical factors.
- Time to value: the first GW in H2 2026 sets the pace; the rest will depend on power infrastructure, construction, and logistics. The maturity of OpenAI’s software stack for distributed training and large-scale inference will be key to leveraging each phase.
This move complements existing collaborations of both with a broad ecosystem (Microsoft, Oracle, SoftBank, Stargate partners), aiming to build the world’s most advanced AI infrastructure.
Why now: user traction and competitive pressure
OpenAI reports having surpassed 700 million weekly active users, with strong adoption in companies, SMBs, and developers. This volume presses the inference backend and, simultaneously, the schedule for training new models. On the other side, competition in base models, agents, and platforms (including the race for PQC, advanced multimodality, and long context) compels securing compute resources over multi-year horizons.
For NVIDIA, this agreement consolidates its role as a leading provider in the physical layer of the AI economy, reinforcing its AI factories narrative and anchoring demand for its upcoming GPU generations and networks.
Challenges Ahead: Energy, Expertise, and Sovereignty
- Energy and grid: 10 GW equals the peak capacity of several cities. It will be necessary to diversify locations, sign PPAs with renewables/nuclear, and collaborate with TSOs/DSOs to avoid bottlenecks and optimize capacity factor.
- Efficiency: software must keep pace with hardware: parallelism, schedulers, compilation, data loading, and memory management to reduce cost per token (training and inference).
- Talent: operating gigawatt-scale AI factories requires expert teams in HPC, networks, power, critical infrastructure, and MLOps at an unprecedented scale in the commercial cloud.
- Sovereignty and regulation: the geographical distribution of data centers will influence data protection, localization requirements, and advanced computing export policies.
What to watch in the next 12–24 months
- Site and power deal announcements: announcements of locations and large-scale energy agreements, including storage and network infrastructure.
- Vera Rubin milestones: tape-outs, official benchmarks, and first factories with next-generation NVL.
- Co-optimization: launches of co-designed runtimes and frameworks, improvements in latency, performance per watt, and cost per inference.
- HBM/optical network chain: signals of additional capacity and new suppliers; validations of HBM4 and 800G/1.6T networks.
Message for Clients and Developers
- Capacity in sight: the first GW in 2026 anticipates a growing service curve. Companies with massive projects (large-scale RAG, multimodal agents, massive fine-tuning) should align their product roadmaps with the capacity windows expected.
- Portability: standardizing on NVIDIA frameworks and runtimes (CUDA, TensorRT-LLM, NeMo) makes it easier to leverage the platform once it arrives, without remaking tooling.
- Cost/Price: increased compute supply may pressure marginal prices downward; it’s wise to prepare cost per token models to decide between hosted API and dedicated infrastructure depending on use case.
Conclusion
The OpenAI–NVIDIA announcement is not just about hardware procurement; it’s a multi-year industrial commitment to build gigawatt-scale AI factories. If it unfolds as planned, the first GW in H2 2026 will mark a turning point toward a compute capacity that enables new model leaps and a global service with lower latency, higher resilience, and decreasing cost per token. The challenge now isn’t whether there will be enough demand to fill them, but how to bring energy, talent, and software up to this ambitious scale.
Frequently Asked Questions
What does “10 GW” mean in AI data centers?
It is the installed electrical capacity powering thousands of racks with millions of GPUs. Practically, it defines the ceiling of compute available for training and inference of large-scale AI models.
When will the first “AI factory” go into service?
The plan indicates H2 2026 for the first gigawatt on NVIDIA Vera Rubin; deployment will proceed in phases as new sites are built and energy agreements are finalized.
What does NVIDIA’s up to $100 billion investment involve?
It is a progressive investment in OpenAI, linked to the deployment of each GW. It supports the effort in data center + power + equipment and aligns incentives between provider and client.
Does this alliance replace work with other partners (Microsoft, Oracle, etc.)?
No. According to the announcement, it complements ongoing collaborations within a broad ecosystem including Microsoft, Oracle, SoftBank, and Stargate partners, aimed at building the world’s most advanced AI infrastructure.