Issue No. 46Week ending Sunday, November 15, 2026897 episodes · 3,860 write-ups
The Podcast Summary.

The best ideas from 30+ top podcasts, every week.

AI infrastructure and compute

Kush Bavaria on AI infrastructure and compute

7 quotes from 1 episode on 20VC with Harry Stebbings, each with a timestamped link to the source.

7 quotes1 episode

The short version

Kush Bavaria states that AI compute costs will climb as supply timelines stretch. Securing new capacity from providers today pushes delivery dates well into 2027 or 2028.

Most interesting insights

Buying new hardware from providers like CoreWeave pushes delivery dates out to late 2027 or early 2028.

“If you want to go to CoreWeave and buy something, you are probably quoted Q4 2027 or Q1 2028 at this point…”

Kush Bavaria, 20VC with Harry Stebbings · November 2026 · Watch at 8:23 ↗

From Why GPU Prices Are Not Dropping Over Time

Early hardware buyers incorrectly expected forward price curves for compute to slope downward.

“When we started the company, a lot of people thought the forward curve for chip prices is actually backwardated…”

Kush Bavaria, 20VC with Harry Stebbings · November 2026 · Watch at 10:23 ↗

From Why GPU Prices Are Not Dropping Over Time

Compute platforms reject highly profitable short-term deals from buyers to protect long-term supply positions.

“You have to say no to a modal. You have to say no to a fireworks. You have to say no to a foul which is willing to pay you gazillions of dollars in revenue and profit for your underlying business to like kind of look in the long term.”

Kush Bavaria, 20VC with Harry Stebbings · November 2026 · Watch at 40:35 ↗

From Kush Bavaria Expects 100 Specialized AI Inference Engines

Top talking points

  1. New hardware pushes compute prices higher

    Next-generation chip costs will climb. Nvidia Rubin hardware could reach $20 to $30 per GPU hour.

    “I would not be surprised if their Rubin costs are like $20, $30 per GPU hour.”

    Kush Bavaria, 20VC with Harry Stebbings · November 2026 · Watch at 11:07 ↗

    From Why GPU Prices Are Not Dropping Over Time

  2. Training mirrors speculative biotech research

    Founders burn cash developing models without knowing the final outcome. Inference demand tracks actual verified usage.

    “The good thing about the inference demand is it's not speculative. It is like real…”

    Kush Bavaria, 20VC with Harry Stebbings · November 2026 · Watch at 42:23 ↗

    From Kush Bavaria Expects 100 Specialized AI Inference Engines

    “Training is much more speculative, right? It is like kind of like biotech where you spend all this money developing this drug and you don't actually know if it works out or doesn't work out.”

    Kush Bavaria, 20VC with Harry Stebbings · November 2026 · Watch at 42:36 ↗

    From Kush Bavaria Expects 100 Specialized AI Inference Engines

  3. Older chips run smaller inference models

    Depreciated hardware steps down from frontier training clusters. These older chips handle lower-scale tasks.

    “I also believe that like the trip depreciation like the older chips will have a use besides being like frontier model training. They'll be used for lower scale inference, smaller model inference.”

    Kush Bavaria, 20VC with Harry Stebbings · November 2026 · Watch at 18:09 ↗

    From Kush Bavaria Expects 100 Specialized AI Inference Engines

Key takeaways from these write-ups

Why GPU Prices Are Not Dropping Over Time

  • Tech buyers assume GPU compute follows standard hardware depreciation, but forward price curves for compute are actually upward sloping.
  • CoreWeave continues to sign three-year rental contracts for older Nvidia A100 chips, proving older hardware does not quickly become worthless.

Kush Bavaria Expects 100 Specialized AI Inference Engines

  • Kush Bavaria compares frontier AI training to speculative biotech R&D, where founders burn cash without knowing if the model will produce a viable outcome.
  • Inference spending reflects verified usage because buyers only pay when someone consumes actual tokens.

How we attribute quotes. Every quote was matched against the episode transcript, so the words and the timestamp are real (we trim filler words like "um", nothing else). The name comes from our written summary of the episode, and we use it only when a separate check of the captions finds that person on the episode. YouTube gives us no voice-by-voice transcript, so open the timestamp to hear who is talking. See a wrong name? Tell us and we fix or remove it.

More

The Sunday Email

Get next Sunday's issue in your inbox.

The week's best AI, tech, business, science, and culture podcasts in one short read. Free, every Sunday.

You get both newsletters, The Throughline and The Carry. No spam. Unsubscribe with one click.