What this covers

This conversation with Dylan Patel works through the economics of AI infrastructure. The central mechanism is that Scaling Laws are log-log, so each performance tier needs roughly ten times the compute, even as serving a given quality of model gets far cheaper. GPT-3-level serving is about 2,000 times cheaper than it was. Patel frames the risk as the next tier of spend, say $50 billion to $500 billion, rather than money already committed, and compares it to the UK railroad and fiber Overbuild Cycle. He also covers Microsoft's late-2024 retreat from building for OpenAI, which let Oracle step in, and Anthropic's climb from roughly $1 billion to $7-8 billion in revenue, almost all of it from code.

The middle of the conversation turns technical. Patel argues the bottleneck is useful data rather than model size, which is why about 40 startups now build reinforcement learning environments where models practice and are trained on successful attempts. He explains long-horizon work through note-taking outside the context window, and compares ML research to semiconductor process development, where progress depends on intuition and runs that look wasted.

The last stretch is physical and political. He argues US power is limited by supply chains, labor and regulation, not aggregate scarcity, and he points to curtailment rules in Texas and PJM, diesel engines used for on-site generation, and electrician wages that have roughly doubled. Other topics include neocloud contract structure, China's state-funded semiconductor push, Jensen Huang's argument applied to Nvidia, and Patel's contested claim that AI creates human capital rather than replacing it.

Sharpest takeaway

Dylan Patel argues the AI buildout is financed by a fragile chain of circular deals and multi-year rental contracts whose payoff depends on continued model improvement and token demand, and that AI shifts value and cost structure in software and power infrastructure.

  • Compute spending must precede revenue, and the OpenAI-Nvidia-Oracle deals spread the capex risk across balance sheets that may not be able to bear it
  • If model capability stops improving at the current scaling tier, the marginal spend has no return and overbuilding follows, as in historical railroad and fiber bubbles
  • AI cuts the cost of building software while adding large token COGS, which undermines the traditional SaaS model

The argument · threads6 threads · 100 claims
0.80

Each capability tier multiplies business value, so growth requires moving up the capability curve.

5 pointscentrality 5/5
  • Each capability tier increases the value of work a model can do disproportionately, like the gap between a 6-year-old, a 13-year-old, and a college intern, so an order of magnitude more compute can produce a drastically different business value.

    “While it may be an order of magnitude more compute, the amount of value is dramatically different.”

  • For a given model capability there is a saturation point of demand, so cheaper serving at a fixed level eventually runs out of users, and growth requires moving up the capability curve.

    “for a given model level, I think there's a saturation point of how much demand for intelligence there is.”

  • OpenAI chose to keep GPT-5 roughly the same size and cost as GPT-4o and invest in thinking modes rather than a bigger model, so it could serve more users and move them up the adoption curve, because a larger slow model would not be adopted.

    “Or do you make the model the same size, which is what they did for GPT-5? It's basically the same size as 4.0”

  • A researcher who finds a 5 percent efficiency improvement saves 5 percent of training time and 5 percent across the entire inference fleet, so a single idea can be worth far more than the researcher's salary given the scale of chip spend.

    “If they come up with the idea that's 5% more efficient, fantastic. I just saved not only 5% of my training time, I also save 5% across my entire inference fleet.”

  • Adding more researchers to frontier ML does not necessarily speed progress, because learning comes from running and interpreting many experiments with limited shared knowledge, and large compute budgets are often spent on experiments that teach little.

    “Adding more people to the problem doesn't make it faster because there's so many things you're trying.”

0.71

Control of cost, supply, and contracts determines bargaining power across the AI value chain.

6 pointscentrality 4/5
  • Nvidia has been blocked from buying major companies, such as ARM, so it can only make small startup acquisitions and must deploy excess cash via buybacks or investments in customers.

    “They weren't even allowed to buy ARM when they were a nobody.”

  • Cursor generates nearly a billion dollars of revenue but has thin margins and sends most of its gross profit to Anthropic, yet Cursor owns the user data and product interface, so power in the relationship is shared rather than one-sided.

    “Does Anthropic have all the power? The common view is yes from a lot of people. But then it's, well, Anthropic only makes the model that's generating the code.”

  • In late 2024 Microsoft pulled back from building data centers for OpenAI and relinquished exclusive compute provider status, letting Oracle take on the deal, which shifted power from Microsoft to OpenAI.

    “They pulled back because Amy Hood and whoever else at Microsoft... We're not going to build out $300 billion worth of compute for OpenAI.”

  • China has the highest average capability across the supply chain and can source most inputs domestically, while the US and Taiwan depend on a global supply chain for critical materials and equipment, so China is more insular while the US is more exposed.

    “China is more concerned with being insular than with being the best.”

  • Google has the lowest cost of goods sold for any token of any company because it owns a vertical stack on TPUs, giving it a structural advantage in serving AI at scale.

    “They have the lowest cost of goods sold for any token of any company because they have their own vertical stack on TPUs.”

  • Nvidia's gross margin is about 75 percent, meaning that for each dollar of hardware sold, most of the gross profit remains with Nvidia even when its equity stake takes part of the value.

    “Nvidia's gross margin is 75%.”

0.43

Stranded capital and credit risk concentrate on OpenAI and the companies financing it.

3 pointscentrality 3/5
  • If model improvement stalls after the industry has spent up to $500 billion on training, that spend will never earn an return, and this stranded-capital risk is what worries the financial community.

    “the moment it stops, wherever you were on the rung—if we went from $50 billion spend to $500 billion spend, well, that $500 billion spend is never going to have ROI.”

  • Neocloud and Oracle contracts with OpenAI carry greater credit risk than Microsoft-backed contracts because OpenAI lacks a strong balance sheet, so the risk of nonpayment sits with OpenAI, its suppliers, and the companies that financed them.

    “OpenAI, they don't have a balance sheet, so how can I be sure that they're actually going to pay”

  • Nebius signed a roughly $19 billion long-term contract with Microsoft, which should generate at least $6 billion of gross profit, and the market trusts Microsoft to pay more reliably than the US government, so these contracts are durable.

    “I'm signing $19 billion to Microsoft. They will pay me no matter whatever.”

0.43

AI and cheap technology drive growth through deflation and newly feasible software projects.

3 pointscentrality 3/5
  • Without AI, software development is too costly and slow to attempt many projects, so migration, database work, and automation remain unbuilt, while AI lets businesses adopt what 2020-era technology made possible.

    “Mainframe migration is something people have always wanted to do. Amazon leaving Oracle took fucking 20 years”

  • Without AI-driven rapid GDP growth, the US cannot sustain its debt and supply chain costs and risks social instability from dividing a shrinking pie, so the country needs AI to grow the pie rather than just redistribute it.

    “by do something, I mean AI has to dramatically accelerate GDP growth. Because once you start talking about dividing the pie, you're screwed.”

  • Technology is the most deflationary force in quality of life because costs fall far faster than revenues rise, yet total revenues still grow, which is the basis of semiconductors and tech in general.

    “tech is the most deflationary thing in the world ever in terms of quality of life. It gets cheaper way faster than revenues go up, but revenues still go up.”

0.29

Chat and AI agents are becoming a discovery layer that threatens existing commerce intent platforms.

3 pointscentrality 2/5
  • Meta and Google built their businesses by placing the thing a user wants to buy in front of them through recommendation systems and ads, and an AI agent that decides purchases threatens to shift that intent layer.

    “That's what has made Meta and Google such amazing companies. They figured out how to get the thing you want to purchase in front of you.”

  • More than 10 percent of Etsy's traffic already comes from GPT, and people make purchasing decisions through chat even when they complete the purchase elsewhere, which means chat is already a discovery layer for commerce.

    “More than 10% of Etsy's traffic is straight from GPT.”

  • Dylan coined the term tokenomics for the economics of tokens and hopes it becomes a search term.

    “So the thing I like to call it is tokenomics. I stumbled upon the word, actually.”

0.29

AI already advances science and physical R&D, with slow but high-value returns.

3 pointscentrality 2/5
  • AI tools worked with researchers during COVID vaccine development, and a future pandemic vaccine could be discovered even faster with AI, showing that AI's impact on science is already real even if not visibly dramatic.

    “The COVID vaccine was created with AI—AI drug discovery.”

  • Applying RL to the physical world, like testing chemistry and materials in real labs, is slow and expensive compared with digital flywheels, but it offers high value through battery chemistry, drug discovery, and semiconductor process development.

    “The flywheel is fast in digital, but the flywheel in the physical world is so slow.”

  • Semiconductor fabs run large R&D operations that produce zero direct economic value because they are learning the next process node, but that learning eventually enables volume production that makes the money.

    “And it's producing zero economic value. Besides that, it's teaching you how to do the next node, which then you can deploy at volume.”