Interview
Dylan Patel on GPT-5’s Router Moment, GPUs vs TPUs, Monetization
- NVIDIA is projected to generate over $200 billion in revenue this year and over $300 billion next year, potentially accumulating north of $100 billion in cash by year-end, while maintaining superior networking, HBM, process nodes, ramp speed, and cost efficiency compared to competitors.
- Competitors require a hardware efficiency advantage of 5x to surpass NVIDIA's trajectory, as supply chain constraints and software ecosystems often dilute this theoretical advantage to 2.5x or 50%, leading to the continued dominance of NVIDIA's supply chain over AMD and custom silicon efforts.
- OpenAI anticipates GPT-5 reducing average thinking times to five to 10 seconds from 30 seconds and will implement auto-routing to tier models by load, monetizing free users via transaction cuts on high-value queries while potentially seeing inference margins fall to 50% gross margin or less as commoditization occurs.
- Anthropic and other code providers are shifting from unlimited or weekly rate limits to hour-based limits, pushing the industry toward usage-based pricing, though enterprises may prefer flat-fee models to avoid variability while consumer subscriptions remain usage-dependent.
- AI infrastructure spend currently represents five years of revenue, suggesting potential over-investment, while hyperscalers could grow CapEx by 20% to 30% next year despite being constrained by power infrastructure availability rather than capital.
- Power infrastructure deficits in the US are a heavier constraint than capital, with 80% of GPU data center costs attributed to capital assets and 20% to land and power; investing in faster deployment is deemed more cost-effective than saving on power costs due to the value of time-to-market.
- Intel faces a high risk of bankruptcy without a significant cash infusion or a 50% reduction in workforce, as its design-to-shipping cycle takes five to six years with 14 revisions compared to the industry's three years and one to three revisions.
- TSMC may raise prices by more than the announced 3% to 10% next year, while custom silicon providers like Google, Amazon, and Meta will scale orders with Google's TPU already at 100% utilization and Amazon's Tranium expected to follow.
- AI-generated personalized ads could create a major inflection point in take rates, and software development productivity via AI theoretically holds the potential to add three trillion dollars of GDP value based on a 100% productivity increase for 30 million developers.
- Google is expected to spend $50 billion on TPU data centers next year but currently has significant capacity waiting for power; if it fails to sell TPUs externally and reorganize its software teams, it risks being surpassed by competitors in the coming years.
- Microsoft's internal model efforts are characterized as failing spectactularly on LLMs with Azure losing share to Oracle and CoreWeave, while its internal chip efforts are labeled the worst among hyperscalers and GitHub Copilot is deemed unusable.
- Apple risks losing control of the user experience to AI agents and may lose market share if it does not allocate $50 to $100 billion to infrastructure, while Meta is building data centers with temporary tents due to the potential for a superintelligence shift within five years.
- China will not gatekeep power for AI, with Chinese companies likely renting GPUs and Blackwell chips outside China where they are more cost-effective, even as they grow their AI CapEx on a percentage basis significantly more than US companies.
- AI accelerator startups are raising capital based on a disruptive technology leap to overcome NVIDIA, though if AI concentration increases, custom silicon will outperform, whereas if AI disperses via open-source models, NVIDIA is expected to remain the most valuable company for a long period.
- OpenAI may capture no more than 10% of the value it creates through current usage models, while the competitive benchmark is shifting from pure performance to the Pareto frontier between cost and performance, with an "economic release" strategy increasing tokens served without necessarily lowering prices.
- Hyperscalers and infrastructure funds like Brookfield, Blackstone, and sovereign wealth funds are not yet fully committed to AI infrastructure spending, despite the potential for AI-generated personalized ads to significantly increase purchase likelihood.