· Aug 25, 2026
Dylan Patel and Dwarkesh Patel discuss compute on Dwarkesh Podcast
Markdown · JSON · Original source
Published August 25, 2026, interviews Dylan Patel, founder of SemiAnalysis, about who may control and what that could mean for the economy. Dylan Patel argues that and are taking a growing share of new compute, and forecasts that, if the trend continues, they could control most of the world’s usable flops by the end of 2028. He says this depends on their ability to pay more for compute, which in turn depends on model releases, regulation and supply constraints. The conversation’s economic argument is that rising returns from AI could draw investment and borrowing toward compute infrastructure, raising costs elsewhere and making expansion harder to finance. Dylan Patel also believes the labs will devote a growing share of compute to training rather than revenue-generating inference, because he expects internal AI research to produce greater future returns. Dwarkesh Patel and Dylan Patel connect this concentration to a broader concern: if AI labor and compute gather in a few labs, economic and technological power may become concentrated there too. The speakers leave open how quickly these changes could occur and how regulation, financing and technical progress will interact.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
· Transcript
Source details
Original source date: Aug 25, 2026 (2026-08-25T15:32:57.000Z).
Date basis: Date supplied by the publisher’s feed.
Added to Escape Meridian: 2026-11-02T00:08:26.322Z. First observed: 2026-09-23T05:35:02.900Z. Retrieved: 2026-10-30T19:29:54.143Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
Text as retrieved on October 30, 2026; originally published August 25, 2026. The interview presents forecasts and estimates by its speakers; the transcript does not independently verify them. The speakers leave uncertain how regulation, financing, supply constraints and technical progress will affect the pace and extent of compute concentration. The transcript does not independently establish how the forecasts will unfold.
Open question
How much could regulation and financing constraints change the concentration of compute Patel forecasts?
Source-stated
Dylan Patel forecasts that OpenAI and Anthropic could control most of the world’s usable flops by the end of 2028 if the trend continues.
By the time you’re towards the end of 2028 — if this trend continues, and I see nothing that’s stopping it — you’ve got them just controlling most of the usable flops in the world on their own.
Source-stated
Dylan Patel says the labs may need to pay higher prices to acquire the compute required for his forecast.
They do have to start paying $25, $30, $50 million a megawatt to really gobble up 70% of the world’s compute in 2028, to get to 100 gigawatts by 2028, which is a very aggressive goal.
Source-stated
Dylan Patel believes the labs will allocate a growing share of compute to training rather than inference, because he expects internal AI research to be more valuable to them.
The labs are going to allocate less and less compute to inference over time. I think that’s very non-consensus.
Source-stated
Dwarkesh Patel and Dylan Patel discuss whether concentration of compute and AI labor could concentrate power in a small number of companies.
most “people”, in terms of work output, are concentrated within two labs who are consuming more and more of the world’s compute.
Source and contributor context
Dwarkesh Podcast · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media retrieval status: unchecked.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
AI compute
The computing resources, meaning chips, data centers and electricity, used to train or run AI models.
Dwarkesh Patel
Dwarkesh Patel hosts the Dwarkesh Podcast, asking researchers and other guests to explain and defend their ideas.
Dwarkesh Podcast
Dwarkesh Podcast features long conversations with researchers, builders and thinkers about technology, history and the future.
OpenAI
OpenAI develops AI models and products and publishes research, release announcements and evaluations.
Anthropic
Anthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Editorial provenance
Version 1. Publication ID: 2cb2deb6-98f6-8f42-5e71-373661bc0ac8.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-02T00:08:04.930Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model, second_opinion_categorical.
Observation: 768eadd3244f328001a9528a622aefe1c18f418e1a23d87a234183199f78acb3. Source text hash: c542d569e166992f55c978c8c04488e1a6d410eef5bf1a175aa6110e9a2fc039.
Revision and source-check history
Version 1 · 2026-11-02T00:08:26.322Z
Dylan Patel and Dwarkesh Patel discuss compute on Dwarkesh Podcast
Published August 25, 2026, interviews Dylan Patel, founder of SemiAnalysis, about who may control and what that could mean for the economy. Dylan Patel argues that and are taking a growing share of new compute, and forecasts that, if the trend continues, they could control most of the world’s usable flops by the end of 2028. He says this depends on their ability to pay more for compute, which in turn depends on model releases, regulation and supply constraints. The conversation’s economic argument is that rising returns from AI could draw investment and borrowing toward compute infrastructure, raising costs elsewhere and making expansion harder to finance. Dylan Patel also believes the labs will devote a growing share of compute to training rather than revenue-generating inference, because he expects internal AI research to produce greater future returns. Dwarkesh Patel and Dylan Patel connect this concentration to a broader concern: if AI labor and compute gather in a few labs, economic and technological power may become concentrated there too. The speakers leave open how quickly these changes could occur and how regulation, financing and technical progress will interact.
Original source
· Transcript
Source details
Original source date: Aug 25, 2026 (2026-08-25T15:32:57.000Z).
Date basis: Date supplied by the publisher’s feed.
Added to Escape Meridian: 2026-11-02T00:08:26.322Z. First observed: 2026-09-23T05:35:02.900Z. Retrieved: 2026-10-30T19:29:54.143Z.
Limits
Text as retrieved on October 30, 2026; originally published August 25, 2026. The interview presents forecasts and estimates by its speakers; the transcript does not independently verify them. The speakers leave uncertain how regulation, financing, supply constraints and technical progress will affect the pace and extent of compute concentration. The transcript does not independently establish how the forecasts will unfold.
Open question
How much could regulation and financing constraints change the concentration of compute Patel forecasts?
Source-stated
Dylan Patel forecasts that OpenAI and Anthropic could control most of the world’s usable flops by the end of 2028 if the trend continues.
By the time you’re towards the end of 2028 — if this trend continues, and I see nothing that’s stopping it — you’ve got them just controlling most of the usable flops in the world on their own.
Source-stated
Dylan Patel says the labs may need to pay higher prices to acquire the compute required for his forecast.
They do have to start paying $25, $30, $50 million a megawatt to really gobble up 70% of the world’s compute in 2028, to get to 100 gigawatts by 2028, which is a very aggressive goal.
Source-stated
Dylan Patel believes the labs will allocate a growing share of compute to training rather than inference, because he expects internal AI research to be more valuable to them.
The labs are going to allocate less and less compute to inference over time. I think that’s very non-consensus.
Source-stated
Dwarkesh Patel and Dylan Patel discuss whether concentration of compute and AI labor could concentrate power in a small number of companies.
most “people”, in terms of work output, are concentrated within two labs who are consuming more and more of the world’s compute.
Source and contributor context
Dwarkesh Podcast · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media retrieval status: unchecked.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
AI compute
The computing resources, meaning chips, data centers and electricity, used to train or run AI models.
Dwarkesh Patel
Dwarkesh Patel hosts the Dwarkesh Podcast, asking researchers and other guests to explain and defend their ideas.
Dwarkesh Podcast
Dwarkesh Podcast features long conversations with researchers, builders and thinkers about technology, history and the future.
OpenAI
OpenAI develops AI models and products and publishes research, release announcements and evaluations.
Anthropic
Anthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Editorial provenance
Version 1. Publication ID: 2cb2deb6-98f6-8f42-5e71-373661bc0ac8.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-02T00:08:04.930Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model, second_opinion_categorical.
Observation: 768eadd3244f328001a9528a622aefe1c18f418e1a23d87a234183199f78acb3. Source text hash: c542d569e166992f55c978c8c04488e1a6d410eef5bf1a175aa6110e9a2fc039.
passed · 2026-11-02T00:08:04.930Z · anthropic/claude-opus-5.5
Publication: 2cb2deb6-98f6-8f42-5e71-373661bc0ac8
attribution — pass: The reading correctly identifies this as one Dwarkesh Podcast interview of Dylan Patel, founder of SemiAnalysis. Forecasts are attributed to Dylan Patel, and the concentration-of-labor point is jointly attributed, which matches the transcript where Dwarkesh raises it and Dylan agrees.
Okay, I’m back with Dylan Patel, founder of SemiAnalysis.
summary — pass: The summary names the actors, keeps the 2028 forecast conditional, and covers the dependencies Dylan cites, including price, model releases, regulation and supply. It also covers the crowding-out financing argument, the non-consensus training-versus-inference view and the centralization concern. 'Flops' and 'inference' are not in the definitions list and would benefit from tooltips. Context still makes the central meaning readable. The title is generic but accurate.
if this trend continues, and I see nothing that’s stopping it
date — pass: The displayed date of August 25, 2026 matches the UTC calendar day of originalPublished, 2026-08-25T15:32:57Z. The modified date equals the publication date. This is a publisher transcript, not an archive copy. Forecasts about 2027–2029 are presented as the speakers' forecasts, and the reading uses no hindsight framing.
support — pass: All claim excerpts are exact substrings of the transcript and support their claims. The summary's points on higher compute prices, regulatory slowdown and debt-financed crowding out are grounded in the source. So is the claim that internal research is more valuable than external inference. The reading adds no unsupported benefits or costs.
This crowding-out effect is the reason it’s not YOLO 1 billion gigawatts.
media — pass: No media is selected and no verified candidates were supplied, so nothing was misidentified or invented.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
Dwarkesh Podcast features long conversations with researchers, builders and thinkers about technology, history and the future.
Explore Dwarkesh PodcastDwarkesh Patel hosts the Dwarkesh Podcast, asking researchers and other guests to explain and defend their ideas.
Explore Dwarkesh PatelThe computing resources, meaning chips, data centers and electricity, used to train or run AI models.
Definition referenceOpenAI develops AI models and products and publishes research, release announcements and evaluations.
Explore OpenAIAnthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Explore AnthropicDwarkesh Podcast features long conversations with researchers, builders and thinkers about technology, history and the future.
Explore Dwarkesh PodcastDwarkesh Patel hosts the Dwarkesh Podcast, asking researchers and other guests to explain and defend their ideas.
Explore Dwarkesh PatelThe computing resources, meaning chips, data centers and electricity, used to train or run AI models.
Definition referenceOpenAI develops AI models and products and publishes research, release announcements and evaluations.
Explore OpenAIAnthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Explore AnthropicDwarkesh Podcast features long conversations with researchers, builders and thinkers about technology, history and the future.
Explore Dwarkesh Podcast