BG2: NVIDIA OpenAI Future of Compute
Category: Expert Interviews · Duration: 104 min · ▶ Watch
Speakers: Brad Gerstner and Clark Tang · Jensen Huang
Segments (13)
- 00:00:00 · Introduction and Recap
- The hosts welcome Jensen Huang and recap his previous prediction that AI inference would grow by a billion-fold.
- 00:01:47 · The Three Scaling Laws of AI
- Jensen explains the evolution from one to three scaling laws driving compute demand: pre-training, post-training, and test-time scaling (reasoning).
- 02:46 · Nvidia’s Partnership with OpenAI
- Discussion of Nvidia’s strategic partnership and planned $100 billion investment in OpenAI to build out their AI infrastructure.
- 03:53 · The Rationale for Investment and Hyperscale
- Jensen justifies the OpenAI investment by predicting it will become a multi-trillion dollar hyperscale company, similar to Nvidia’s work with Microsoft and Oracle.
- 05:45 · The Disconnect with Wall Street
- Brad highlights the massive gap between the exponential growth Nvidia sees and the flat revenue forecasts from Wall Street analysts.
- 07:03 · Three Pillars of AI Compute Demand
- Jensen outlines three fundamental drivers for AI compute: the shift from general-purpose to accelerated, the migration of existing hyperscale workloads, and the augmentation of human intelligence.
- 08:17 · Why There Won’t Be a Compute Glut
- Jensen argues that demand continues to outstrip supply due to two compounding exponentials: user growth and increasing compute-per-user from reasoning models.
- 09:18 · The Annual Release Cycle and Extreme Co-Design
- Jensen explains Nvidia’s one-year product cadence is necessary to keep pace with demand and is enabled by ‘extreme co-design’ across the entire stack.
- 10:18 · Total Cost of Operation (TCO) vs. Chip Cost
- The argument is made that even if a competitor’s chip were free, Nvidia’s system offers a better TCO due to superior performance-per-watt.
- 11:30 · Sovereign AI and the Global AI Race
- The discussion shifts to the geopolitical importance of AI, with Jensen arguing every nation needs its own sovereign AI capability.
- 12:31 · Reindustrializing America and the Future of Work
- Jensen praises the administration’s focus on reindustrializing America and using AI as an equalizer to close the technology divide.
- 13:31 · The Future of Jobs
- Jensen argues that AI will eliminate tasks, not jobs, and that human ingenuity will create new ideas and new work, leading to economic growth.
- 14:00 · The Exponential Future
- The hosts reflect on the exponential rate of change, with Jensen advising to ‘get on the train’ now rather than trying to predict its exact destination.
Specific Prices (7)
| Timestamp | Item | Value | Context |
|---|---|---|---|
| 03:27 | Nvidia’s planned investment in OpenAI | $100 billion | Nvidia intends to invest up to $100 billion in OpenAI as part of the Stargate data center buildout. |
| 03:40 | Potential revenue for Nvidia from OpenAI’s 10 gigawatt buildout | $400 billion | If OpenAI used Nvidia for its 10 gigawatt build, it could result in up to $400 billion in revenue for Nvidia. |
| 07:03 | Value of the human intelligence economy | $50 trillion | Jensen estimates the part of the world’s GDP represented by human intelligence that can be augmented by AI. |
| 07:03 | Cost of an employee vs. augmenting AI | $100,000 vs $10,000 | Jensen illustrates the ROI of augmenting a $100k employee with a $10k AI to double their productivity. |
| 11:19 | AI Factory ROI example | $3M cost for $30M revenue | A slide shows that a $3M investment in a GB200 NVL72 system can generate $30M in token revenue. |
| 11:29 | AI Factory ROI example (Free compute) | $1M cost for $8M revenue | A slide shows that even with ‘free’ GPUs (representing 1/4 performance), the non-GPU costs of $1M only generate $8M in revenue, showing a lower ROI. |
| 17:27 | H1-B Visa Fee | $100,000 | The administration’s proposed new fee for an H1-B visa. |
Memory Facts (2)
- [08:35] Huawei is developing in-house high-bandwidth memory.
- N/A
- [42:55] AI requires long-term and short-term memory, with intense KV cache processing.
- N/A
Bottleneck Claims (3)
- [08:17] The primary bottleneck for AI buildout is not GPU supply, but the ability to build data centers.
- Evidence: The real constraints are securing land, power, and shell for the data centers. Nvidia can build chips in response to demand signals, but the physical infrastructure takes longer.
- [10:18] The cost of the AI factory is dominated by infrastructure (land, power, shell), not the compute hardware itself.
- Evidence: A chart shows that facility capex (land, power, shell) is 39% of the total cost of a 1GW data center, nearly equal to the GPU capex at 43%.
- [12:31] Restricting immigration of top talent is an existential threat to the US’s AI leadership.
- Evidence: The ‘American Dream’ brand is a singular advantage that attracts the world’s best talent. Damaging this brand or making it harder for talent to come and stay is a self-inflicted wound.
Predictions (4)
- [00:03, Unspecified] OpenAI will become the next multi-trillion dollar hyperscale company.
- [01:34, Unspecified] AI inference compute will increase by a billion-fold.
- [38:04, End of 21st century] In the 21st century, we will have 20,000 years of progress, not 100.
- [39:31, 5 years] In the next 5 years, AI and mechatronics/robotics will fuse, and AIs will be wandering around us.
Key Technologies (5)
- AI Scaling Laws: Three fundamental principles driving exponential compute demand: Pre-Training (learning from data), Post-Training (practicing skills via reinforcement learning), and Test-Time Scaling (multi-step reas
- Agentic Systems: A system of multiple language models working concurrently, using tools and performing research to generate a comprehensive answer.
- Extreme Co-design: Nvidia’s strategy of optimizing the entire AI stack simultaneously—from the algorithm and software to multiple interconnected chips (CPU, GPU, networking) and the data center system—to achieve perform
- NVLink Fusion: An open standard that allows third-party CPUs (like Intel’s) to be coherently connected to Nvidia GPUs, creating a powerful, fused ecosystem.
- CPX Chip: A new specialized processor announced by Nvidia for context processing and diffusion video generation, designed to handle specific, intensive workloads within the AI data center.
Companies Mentioned (18)
OpenAI · Nvidia · Meta · Google · Microsoft Azure · Oracle (OCI) · CoreWeave · SoftBank · Intel · Alibaba · Huawei · ByteDance · Amazon · Databricks · Snowflake · Cisco · Nortel · Anthropic
Notable Quotes (8)
I think that OpenAI is likely going to be the next multi-trillion dollar hyperscale company. — Jensen Huang @ 00:00
It’s going to one billion X. — Jensen Huang @ 01:34
This is the industrial revolution. — Jensen Huang @ 01:57
The ultimate extreme co-design. Nobody’s ever co-designed at this level before. — Jensen Huang @ 09:18
They could literally price them at zero and you would still buy an Nvidia system because the total cost of operating that system… is still a better bet. — Brad Gerstner @ 10:33
The concept that AI comes along and therefore there’s going to be a mass destruction of jobs starts with the premise that we have no more ideas. — Jensen Huang @ 13:31
If you have a train that’s about to get faster and faster and go exponential, the only thing that you really need to do is get on it. — Jensen Huang @ 14:13
Nobody needs atomic bombs. Everybody needs AI. — Jensen Huang @ 59:54
Key Topics
AI Infrastructure Economics · Exponential Growth of AI Compute · Nvidia's Strategic Partnerships (OpenAI, Intel, etc.) · Total Cost of Operation (TCO) vs. Chip Cost · The Three Scaling Laws of AI · Sovereign AI and Geopolitics · The Future of Jobs and the American Dream · Nvidia's Competitive Moat and Annual Cadence · US-China Tech Competition · Immigration Policy for Tech Talent
Takeaways
- The demand for AI compute is driven by three compounding exponentials: pre-training, post-training, and reasoning, leading to a potential billion-fold increase in inference demand.
- The true cost of an AI factory is not the chips, but the total cost of operation (TCO), where power and infrastructure are dominant. Superior performance-per-watt is the key to maximizing ROI.
- Nvidia has shifted to a one-year product cadence and a strategy of ‘extreme co-design’—optimizing the entire data center stack at once—to maintain its performance lead.
- OpenAI is on a trajectory to become a multi-trillion dollar hyperscale company, justifying Nvidia’s massive strategic investment in their infrastructure.
- The AI buildout is not a bubble but a fundamental re-platforming of the entire global IT industry, moving from general-purpose to accelerated computing.
- Sovereign AI is becoming a national priority for every country, as they need to own their AI infrastructure to protect their culture, data, and economic future.
- Attracting and retaining the world’s best talent is critical for US leadership in the global AI race; policies that hinder this are a significant risk.
- AI will augment human intelligence and increase productivity, which will create more economic growth and new jobs, rather than causing mass unemployment.