How Will AI Reasoning Reshape Global Computing?

TL;DR

AI compute demand is accelerating because usage growth and computational work per request are rising together. Jensen Huang argues that pre-training, post-training, and reasoning-time inference now form three scaling laws, while NVIDIA’s partnership with OpenAI adds direct infrastructure collaboration across chips, software, systems, and AI factories to support OpenAI’s expansion into a hyperscale company.

Transcript

I think that OpenAI is likely going to be the next multi- trillion dollar hypers scale company. Okay, [Applause] Jensen. Great to be back, of course, with my partner Clark Tang. You know, I can't believe it's Welcome to Invidia. Uh, oh, and nice glasses. Um, those actually look really good on you. The problem is now everybody's going to want you to... Read More

Key Insights

• AI compute demand is driven by three scaling laws: pre-training, post-training, and inference. Post-training lets AI practice through repeated attempts, while modern inference allocates additional computation to thinking, researching, checking ground truth, and refining an answer before producing it.
• Modern AI is a system of language models rather than a single isolated model. Multiple models can operate concurrently, use tools, conduct research, process multiple modalities, and generate content, creating much larger inference requirements than the earlier pattern of producing a one-shot response.
• OpenAI’s compute requirement reflects two compounding exponentials. Its customer base and application usage are growing, while each individual interaction is becoming more computationally intensive because models increasingly reason before answering instead of immediately generating a response.
• NVIDIA’s new OpenAI partnership supports OpenAI’s first self-built AI infrastructure. The companies plan to work directly across chips, software, systems, and complete AI factories, giving OpenAI a relationship with NVIDIA comparable to those maintained by established hyperscale operators.
• OpenAI’s self-built capacity is additive to its other infrastructure projects. NVIDIA continues supporting Microsoft Azure deployments, contracted Oracle Cloud Infrastructure projects involving OpenAI and SoftBank, and capacity associated with CoreWeave while beginning the separate direct buildout.
• OpenAI is likely to become a multi-trillion-dollar hyperscale company, according to Huang. He expects it to provide both consumer and enterprise services, and he considers NVIDIA’s opportunity to invest before that expansion a potentially strong use of capital in a field NVIDIA understands.
• Direct infrastructure ownership follows OpenAI’s growth to sufficient scale. Huang compares the intended relationship with NVIDIA to direct partnerships involving Elon Musk and xAI, Mark Zuckerberg and Meta, Sundar Pichai and Google, and Satya Nadella and Microsoft Azure.
• Wall Street expectations diverge from the infrastructure ambitions described by AI builders. The discussion cites a consensus forecast among 25 sell-side analysts that anticipates NVIDIA growth flattening to 8 percent from 2027 through 2030, despite expansive buildout plans across companies and sovereign customers.

Questions & Answers

Q: Why is AI inference compute demand increasing?

AI inference demand is increasing because models are shifting from one-shot generation to reasoning before answering. A system may think, perform research, check ground truth, learn from what it finds, and then continue reasoning before generating its response. Demand also expands because modern AI can involve several models running concurrently, sometimes using tools and processing multiple modalities rather than executing one simple model call.

Q: What are the three scaling laws for modern AI?

The three scaling laws described are pre-training, post-training, and inference. Pre-training develops the foundational model. Post-training allows AI to practice a skill through repeated attempts and reinforcement learning, integrating training with inference. The inference law concerns the computation used while answering: allowing a system to think longer, research information, and check ground truth can improve the quality of its final response.

Q: How does reasoning-time inference work?

Reasoning-time inference gives an AI system computational space to think before producing an answer. Instead of responding immediately, the system can research, compare its developing response with ground truth, learn from the information it finds, and reason further. Huang argues that longer thinking can produce better answers, making inference a scalable source of capability rather than merely the final, one-shot use of a trained model.

Q: What is NVIDIA’s partnership with OpenAI designed to build?

The partnership is designed to help OpenAI build and operate its own AI infrastructure for the first time. NVIDIA and OpenAI plan to work directly at the chip, software, systems, and AI-factory levels. The goal is to establish the kind of direct technical and purchasing relationship NVIDIA already maintains with other hyperscale operators, while supporting OpenAI’s continued expansion in consumer and enterprise AI services.

Q: Does OpenAI’s self-built infrastructure replace its cloud partnerships?

OpenAI’s self-built infrastructure is presented as additional capacity rather than a replacement for existing projects. NVIDIA expects to continue work connected to Microsoft Azure, Oracle Cloud Infrastructure, SoftBank, and CoreWeave. Huang says the new direct buildout sits on top of previously announced and contracted infrastructure, providing more capacity for OpenAI’s customer growth and the rising computational requirements of each AI interaction.

Q: Why does NVIDIA view OpenAI as an attractive investment?

Huang believes OpenAI is likely to become the next multi-trillion-dollar hyperscale company, with both consumer and enterprise services. From that perspective, investing before it reaches that scale could produce a strong return. He also says NVIDIA understands the field in which OpenAI operates. The investment opportunity is not required by the partnership, but NVIDIA considers the chance to participate financially attractive.

Q: What two exponentials are driving OpenAI’s infrastructure needs?

The first exponential is growth in customers and usage as AI quality improves, use cases expand, and applications connect to OpenAI. The second is growth in computation required for every use because models increasingly think before responding. When the number of interactions rises while each interaction also consumes more computation, the two trends compound and create a much larger infrastructure requirement.

Q: Why is OpenAI building a direct relationship with NVIDIA?

OpenAI has reached a scale at which it wants direct technical, operational, and purchasing relationships for infrastructure. Huang compares this approach with NVIDIA’s direct work with xAI, Meta, Google, and Microsoft Azure. Building directly with NVIDIA gives OpenAI collaboration across chips, software, systems, and complete AI factories, supporting its transition toward operating as a full hyperscale company with substantial internal compute capacity.

Summary & Key Takeaways

• Jensen Huang describes an AI industry moving beyond one-shot language-model responses. Modern systems combine multiple models, concurrent agents, tools, research, multimodal inputs, and generated video. Pre-training, reinforcement-driven post-training, and longer reasoning during inference collectively increase the amount of computation required to develop and operate increasingly capable AI services.

• NVIDIA’s new OpenAI partnership centers on helping OpenAI build and operate its own infrastructure through direct collaboration at the chip, software, systems, and AI-factory levels. Huang presents this work as additive to existing OpenAI capacity being developed with Microsoft Azure, Oracle Cloud Infrastructure, SoftBank, and CoreWeave rather than replacing those relationships.

• Huang expects OpenAI to become a multi-trillion-dollar hyperscale company offering consumer and enterprise services. He says its infrastructure requirements reflect two compounding exponentials: rapid customer growth and increasing computation for each use. The discussion contrasts this builder outlook with Wall Street forecasts that anticipate NVIDIA’s growth flattening beginning in 2027.