Free A.I. Models Were Never Free | EP 101 Deedy Das

Описание: Thanks to our partners at Bright Data and the Web Data Loft for hosting this conversation in San Francisco.

Everyone says open source models are the cheap option. I asked Deedy Das to actually price it, and the answer was not the one I wanted.

Deedy is a partner at Menlo Ventures, where he co-runs the Anthology Fund. Before that he was on the founding team at Glean, and before that Google and Meta. He is also one of the very few investors I know who will sit down and walk you through the chip company he passed on that Nvidia went and bought for $20 billion.

We spent most of this conversation arguing. KIMI K3 against Opus on price. Whether frontier costs have really come down. Whether an enterprise should ever walk away from the best model and start hosting weights itself. I pushed back on most of his answers and he pushed back harder on mine.

Free models were never free. That is roughly where we landed.

In this episode we discuss:
• Why "I could vibe code that in 10 minutes" is not the threat founders think it is
• The one product metric he underwrites multiples on, and why he barely looks at ARR
• How most VCs never open the product they are funding
• Whether Anthropic and OpenRouter are a hedge against each other
• The maintenance headache at Glean that turned into OpenRouter
• What leaving frontier actually costs: a 3 to 5 year compute lock-in with 30% down
• KIMI K3 at roughly $15 per million output tokens against Opus at roughly $25
• Why frontier prices have stayed flat while everything else got cheaper
• The biggest compute crunch of all time, and who is locked out of it
• What happens when you add a million engineers worth of tokens and the top line does not move
• Rapid fire on Suno, Wispr Flow, Mercor, Goodfire and Skild: headwind or tailwind
• Pangram, AI slop, and why he will not read something an LLM wrote for you
• The chip company he passed on that Nvidia bought for $20 billion
• Why the Anthropic entry made no sense on the financials at the time
• What he tells a 19 year old deciding whether to still study computer science

Chapters:
0:00 Cold Open
1:09 Intro
2:23 Building a 10 Year Thesis When the Ground Moves
4:59 Why "I Could Vibe Code That" Misses the Point
6:36 Most VCs Never Try the Product
7:57 We Changed the Definition of ARR
8:24 The One Product Metric He Underwrites On
9:35 Anthropic and OpenRouter: Hedge or Coexist
12:15 The Glean Story That Became OpenRouter
15:58 Should Enterprises Just Run Open Weights
19:20 What KIMI K3 Actually Costs
19:47 Free Until They Can Charge for It
22:18 Frontier Prices Have Not Fallen
23:33 The Biggest Compute Crunch of All Time
24:18 Engineers Plus Tokens, Minus the ROI
27:21 Rapid Fire on the Portfolio
33:36 Pangram and the AI Slop Problem
39:01 The Groq Miss
41:19 The Anthropic Bet That Made No Sense
44:14 Should a 19 Year Old Still Study CS
48:06 The Company He Wants Walking Through the Door

Subscribe to CLIMB for more conversations with the founders, operators, and investors building the future with AI.

---

#AI #OpenSource #VentureCapital #LLM #Anthropic

CONNECT:
LinkedIn:   / jaykapoor  
X: https://x.com/jayKapoornyc

GUEST:
X: https://x.com/deedydas