How Do Three AI Scaling Laws Drive AI Factories?

TL;DR

AI compute demand is expanding through three scaling laws: pre-training, post-training, and thinking inference. Post-training repeatedly practices skills through inference, while modern inference researches, uses tools, and reasons before answering. NVIDIA’s proposed work with OpenAI adds a 10-gigawatt self-built infrastructure program to existing Microsoft Azure, Oracle Cloud Infrastructure, and CoreWeave projects.

Transcript

Jensen great to be back of course with my partner Clark Tang you know I can't believe it's been welcome to NVIDIA haha oh and nice glasses um this actually look really good on you the problem is now everybody's gonna want you to wear them all the time they're gonna say where are the red glasses I can vouch for that so it's been over a year since we... Read More

Key Insights

• AI development is governed by three scaling laws: pre-training, post-training, and inference. This framework expands the earlier focus on pre-training by recognizing that skill practice after training and additional computation during answer generation can also improve model quality.
• Post-training is AI practicing a skill until it gets the result right. It tries multiple approaches and uses inference within reinforcement learning, which means training and inference become integrated rather than functioning as entirely separate workloads.
• Thinking inference is a multi-step process rather than a one-shot response. An AI can reason, conduct research, check ground truth, learn from what it finds, continue thinking, and only then generate its final answer.
• Agentic AI is a system of language models rather than a single language model. Multiple models can operate concurrently, use tools, perform research, and process multiple modalities, substantially increasing the computation required to complete useful tasks.
• OpenAI’s compute demand is shaped by two compounding exponentials. Its customer and application usage is growing as AI quality and use cases improve, while every individual use requires more computation because the system increasingly thinks before answering.
• The 10-gigawatt partnership is intended to help OpenAI build self-operated AI infrastructure. NVIDIA would work directly with OpenAI at the chip, software, systems, and AI-factory levels, adding capacity beyond previously announced infrastructure programs.
• Direct infrastructure relationships become practical when an AI company reaches hyperscale. Huang compares OpenAI’s desired relationship with NVIDIA to the direct working and purchasing relationships maintained by xAI, Meta, Microsoft, and Google.
• Accelerated computing is presented as the successor to general-purpose computing. The interview connects this transition with NVIDIA’s annual Hopper, Blackwell, and Rubin cadence, AI-factory construction, sovereign AI, large computing clusters, and constraints involving power and energy.

Questions & Answers

Q: What are Jensen Huang’s three AI scaling laws?

The three scaling laws are pre-training, post-training, and inference. Pre-training remains the foundation-building stage. Post-training lets an AI practice skills repeatedly, trying different approaches through reinforcement learning until performance improves. Modern inference adds thinking time, research, tool use, and ground-truth checks before an answer is generated, so greater inference computation can produce a higher-quality result.

Q: How does post-training improve an AI model’s skills?

Post-training improves skills by allowing an AI to practice repeatedly and explore many possible ways to solve a task. The process uses inference as part of reinforcement learning, evaluating attempts and continuing until the system gets the result right. This arrangement integrates training and inference, creating significant inference demand even though the activity is described as a training stage.

Q: Why does thinking inference require more computing power?

Thinking inference requires more computing power because the model no longer generates an answer immediately in a single pass. It can think, research relevant information, check ground truth, learn from those checks, and then reason further before responding. The longer and more extensively it performs these steps, the more inference computation it consumes and the better its answer may become.

Q: Why are agentic AI systems more computationally demanding?

Agentic AI systems are more computationally demanding because they consist of systems of language models rather than one isolated model. Several models may run concurrently, with some using tools, others conducting research, and the overall system handling multiple modalities. These coordinated activities create many inference operations for a single user request and expand demand beyond traditional one-shot generation.

Q: What is NVIDIA’s 10-gigawatt partnership with OpenAI?

The partnership is a plan for NVIDIA to help OpenAI develop its own self-built AI infrastructure at a scale of 10 gigawatts. NVIDIA would collaborate directly across chips, software, systems, and complete AI factories. Huang says this capacity would be additive to existing OpenAI-related work involving Microsoft Azure, Oracle Cloud Infrastructure, and CoreWeave rather than replacing those projects.

Q: Why does OpenAI want to build its own AI infrastructure?

OpenAI has reached a scale where it wants direct working and purchasing relationships for infrastructure, similar to relationships NVIDIA maintains with xAI, Meta, Microsoft, and Google. Self-building gives OpenAI a full-stack operating model suited to a hyperscale company. Huang expects it will probably consume the capacity itself because both usage and computation per use are increasing rapidly.

Q: What two exponentials are increasing OpenAI’s compute needs?

The first exponential is customer and application usage growth, supported by improving AI quality, better use cases, and connections between OpenAI and many applications. The second is the computation required for every use, because inference increasingly thinks before answering instead of producing a one-shot response. When these two growth patterns compound, total infrastructure requirements rise especially quickly.

Q: Why does Jensen Huang expect AI factories to keep expanding?

Huang’s argument begins with the claim that general-purpose computing is over and that future computing will be accelerated and AI-based. Expansion is also supported by three simultaneous scaling laws, agentic systems that run multiple models and tools, growing customer usage, and more computation for each answer. The broader discussion identifies power and energy constraints as important limits on this build-out.

Summary & Key Takeaways

• Jensen Huang describes three scaling laws that now shape AI development. Pre-training builds foundational capabilities, post-training improves skills through repeated practice and reinforcement learning, and thinking inference spends additional computation researching and reasoning before answering. Training and inference consequently become increasingly integrated rather than remaining clearly separated stages.

• OpenAI’s infrastructure requirements are driven by two compounding exponentials: rapid customer growth and rising computation for every use. NVIDIA plans to help OpenAI build self-operated infrastructure directly across chips, software, systems, and complete AI factories, supplementing projects already associated with Microsoft Azure, Oracle Cloud Infrastructure, and CoreWeave.

• Huang argues that general-purpose computing is ending and that accelerated computing and AI computing represent the future. The broader interview connects annual NVIDIA platform transitions from Hopper to Blackwell to Rubin with sovereign AI, massive clusters, energy constraints, talent policy, global competition, and AI factories as potential contributors to global GDP.