Is there a meaningful ceiling on how much compute the economy can productively absorb?

Jonathan Ross · published Sep 29, 2025

AI product economics differ fundamentally from SaaS because you can buy quality with compute on a per-query basis, so demand for compute is effectively unbounded

In SaaS quality is fixed by what engineers built; in AI you can run two instances of a prompt and pick the better answer, or give more compute to more valuable customers — so competitors keep spending, which is why token bills nearly match revenue

25:17 20VC: OpenAI and Anthropic Will Build Their Own Chips | NVIDIA Will Be Worth $10TRN | How to Solve the Energy Required for AI... Nuclear | Why China is Behind the US in the Race for AGI with Jonathan Ross, Groq Founder

Jonathan Ross · published Sep 29, 2025

Demand for inference compute is highly elastic — halving the price roughly doubles the volume customers buy

Customers aren't spend-conscious; they spend as much as they make because whatever they spend increases the quality of the output

69:26 20VC: OpenAI and Anthropic Will Build Their Own Chips | NVIDIA Will Be Worth $10TRN | How to Solve the Energy Required for AI... Nuclear | Why China is Behind the US in the Race for AGI with Jonathan Ross, Groq Founder

Harry Stebbings · published Oct 26, 2026

Demand for inference is effectively infinite, and recognizing this a year earlier would have led to very different investment decisions

He underestimated the depth of each layer of the stack, particularly inference demand

Scope: framed as a personal lesson learned

17:01 20VC: Satya Nadella on Open vs Closed Models: Who Wins | Whether to Trust Frontier Model Providers | Why Companies Need to Own Their Own Intelligence | The Importance of Verticalisation: From Chips to Models and More

Harry Stebbings · published Nov 2, 2026

Demand for inference is effectively infinite and some of the biggest companies in the world are being built in that space

0:25 20VC: Who Wins the Inference Market | Has Open-Source LLM Inference Peaked? | What Happens to the Cost of Compute | The Real Limits of Post-Training and RL with Erik Bernhardsson, Modal CEO

Jonathan Ross · published Sep 29, 2025

There is no limit to the amount of compute the economy can absorb — unlike the industrial revolution, adding compute directly strengthens the economy without needing complementary machinery to be built first

In the industrial revolution more oil was useless without building cars and machines to consume it; with AI, doubling compute immediately doubles users and improves model quality

Scope: better models can also do more with the same compute

42:01 20VC: OpenAI and Anthropic Will Build Their Own Chips | NVIDIA Will Be Worth $10TRN | How to Solve the Energy Required for AI... Nuclear | Why China is Behind the US in the Race for AGI with Jonathan Ross, Groq Founder

Erik Bernhardsson · published Nov 2, 2026

We are very early in AI compute demand — there is enormous latent demand and it is unclear whether compute can be overbuilt at all

So much latent demand exists for these products that supply may never get ahead of it

Scope: eventually capital constraints (rising rates, data center build speed) will bind; he explicitly says he doesn't know whether overbuild is possible

22:41 20VC: Who Wins the Inference Market | Has Open-Source LLM Inference Peaked? | What Happens to the Cost of Compute | The Real Limits of Post-Training and RL with Erik Bernhardsson, Modal CEO

Jonathan Ross · published Sep 29, 2025

AI is unprecedented because it lets the economy add labor by producing more compute, something that has never happened before in economic history

Producing more compute and better AI directly adds labor supply to the economy

0:00 Compute growth acts like adding labor with no historical precedent

Harry Stebbings · published Nov 2, 2026

You cannot overbuild AI compute capacity in the next five years

23:06 Cannot overbuild ai compute capacity within the next five years

Erik Bernhardsson · published Nov 2, 2026 · hedged

The real risk to compute demand is that many heavy GPU-consuming companies don't yet make revenue, and the lag before they monetize could be long enough that the market runs up against it

Companies like Instinct clearly can make a lot of money eventually, but they have to actually start doing so, and that lag is a market dynamic

23:09 Revenue monetization lag not compute appetite is the real risk to continued demand