Latent Space: The AI Engineer Podcast

Episode: Anthropic, Glean & OpenRouter: How AI Moats Are Built with Deedy Das (Menlo Ventures)
Date: November 14, 2025
Host(s): Alessio (Kernel Labs), swix (Editor of Latent Space)
Guest: Deedy Das (Menlo Ventures, ex-Glean)

Overview

This episode centers on how foundational AI companies build and maintain moats, focusing on Glean (enterprise search), Anthropic (frontier AI models), and OpenRouter (model routing/integration), through the lens of engineer and investor Deedy Das. The discussion spans startup battles, AI model commoditization, infrastructure arms race, and the rise of new application and research layers in the AI ecosystem. Along the way, the hosts and Deedy dissect the challenges and opportunities facing AI-first companies—both model labs and app builders—while offering insiders’ perspectives on venture, talent, and shifting market dynamics.

Key Discussion Points & Insights

1. From "Boring" to Sexy: The Glean Enterprise Search Journey

[01:19–07:22]

• Struggle to Sexy: In 2019, enterprise search was "unsexy"—“I would say enterprise search, and it's shutting down the conversation right there.” (C, 01:38). With the rise of LLMs and ChatGPT, demand exploded, and Glean became "sexy”.
• Hard-Earned Moat: Glean’s moat isn’t about AI but the hard work building, integrating, and solving “last mile” search/data access problems for big enterprises:
  • “Now I see companies trying to tack on search. It’s not easy... The moat is just we did the hard work.” (C, 04:35).
• Market Dynamics: After ChatGPT’s debut (Dec 2022), the landscape changed overnight. Glean was quick to integrate LLMs, boosting sales and value.
• On Competition: New entrants (Claude, Slack, Salesforce) make the market hot—but Deedy feels Glean’s depth remains hard to replicate:
  • “To build a deep enterprise search system…[for] Anthropic and OpenAI…doesn’t make them that much money. The amount of effort…is big sales teams, huge FTE teams, tons and tons of customization.” (C, 07:50)

Notable Quotes:

"It's such a boring, unsexy company that became sexy later." (C, 01:38)
"The moat is just—we did the hard work." (C, 04:35)

2. The Unseen Complexity of Enterprise AI

[09:09–15:15]

• Enterprise ≠ Consumer: Glean had to invent new ranking and feedback systems, as enterprise data is sparser, constantly changing, and highly contextual.
• The Adoption Challenge: Productivity tools must feel good on day one, but search lacks the inherent virality of tools like Slack or email.
• Onboarding Tactics:
  • “We want to take over your new tab page…tell us what we need to do to earn the right to do that.” (C, 13:51)
  • Building chrome extensions that natively replace search in apps like Google Drive.

3. Anthropic’s Growth & Model Lab Moats

[15:32–27:56]

• Anthropic’s Meteoric Rise:
  • "Anthropic is the fastest growing software company of all time." (C, 15:54)
  • 0 to $100m, then to $1bn, now projected $9bn within three years—a scale unprecedented in SaaS.
• Risk, Innovation & Culture:
  • "There was a world where maybe they would have not worked at all...but the same qualities that would make them likely to fail made them have a high propensity to succeed." (C, 16:52)
  • Anthropic’s permissionless, highly-retentive culture lets researchers innovate (e.g., Claude Code, agents).
• On Competitive Positioning:
  • Despite increased diversity (OpenAI from 50% to 25% enterprise API share; Anthropic growing to 32%), model switching is rare in enterprise due to contract size/integration.
• Model Performance & Differentiation:
  • “Articulating what makes a model good is very, very difficult...benchmarks are...0.12 differences in suitebench, but what matters is retention and product fit.” (B, 28:24)
  • Anthropic stands out in workspace coding and other verticals, where quality does move the needle.

Notable Quotes:

“Anthropic is the fastest growing software company of all time... 0 to 100 in one year, 100 to a billion in one year...I couldn’t have predicted it.” (C, 15:54)
"You hire good talent and let them loose with a lot of tokens, see what they come up with." (B, 18:53)

4. AI’s Application Layer: App vs Model Moats

[30:44–37:22]

• App Layer Ecosystem: Startups like Cognition, Cursor, DevIn, etc., have built on top of Claude, creating a sustainable ecosystem of wrappers and innovative apps.
• Can Apps Out-Moat Models?
  • Deedy argues model labs have deeper moats: “It is far easier for Anthropic to go into one of the spaces of the apps than an app to go into the space of Anthropic, which makes me feel like one is more defensible than the other.”
• Classic "Means of Production" Concern: Platform risk persists—model provider can always move into the application’s territory if it becomes strategic.

Notable Quotes:

“The moat is what is the hardest to do in any part of the stack.” (C, 33:33)
"As an investor and a human with my own limited time on Earth, if Anthropic can go from $4B to $183B in two years, then everything else is a waste of time.” (B, 34:28)

5. The Anthology Fund & Modern Venture

[41:06–48:01]

• Anthology Fund Origins & Impact:
  • Menlo’s $100M Anthology fund is meant to seed Anthropic’s ecosystem, but it operates outside the corp-vc model to avoid misaligned incentives.
  • Has funded 40 companies (OpenRouter, Goodfire, Prime Intellect, etc.), with higher graduation rates than most.
• Research Bets, Uncertain Paths:
  • Investing in research-heavy teams (Prime Intellect, Goodfire, others) is risky, but potential upside is transformative.
  • “If I fast forward 10 years...what do I think is very likely to exist...If I believe there’s this team very strongly headed in that direction...maybe we can see something here.” (C, 47:05)

6. Spotlight: Startup Deep Dives

[53:14–65:41]

• OpenRouter (Alex Atallah):
  • Solves the “PLG sweet spot” of model routing/integration; beautiful product, deep technical attention, strong developer mindshare.
  • Risks: Pricing pressure as token costs decrease; converting hobbyist evaluation traffic into enterprise “stickiness.”
  • “If there can be a PLG [product-led-growth] motion in any SaaS market, the PLG motion will win.” (C, 55:40)
• Whisper (WISPR):
  • Voice transcription/dictation—focuses on zero-edit-rate experience, delighting users.
  • Faces classic “commodity” risk, but wins on retention and user love.
• StealthCo (Inception):
  • Exploring diffusion models for language (potentially cheaper/faster than transformers) and novel code reasoning. Bet on a future alternative to transformer hegemony.

Notable Quotes:

“OpenRouter is the only non-Elon company that Elon has tweeted the most about.” (C, 60:18)
“Diffusion models today are, I would say, 80-90% the quality at one-tenth the cost and latency.” (C, 64:08)

7. Moats, Markets, and Strategic Arbitrage

[69:06–73:18]

• Market Windows, Arbitrage, Roll-Ups:
  • Sometimes great execution isn’t enough if the "window" closes or market dynamics aren’t favorable.
  • “You could be the fastest runner in the world and you might not make it out of the tunnel.” (C, 68:30)
  • Roll-up and multiple arbitrage: "AI-ify" a traditional business and 50x the valuation.
  • Reflexivity in venture: “The belief that something can be true can make it true, even though it's not true at the time you believed it.” (B, 71:47)

8. Infra Arms Race & Scaling Gambles

[73:20–76:10]

• Massive Compute Buildout (Stargate, Amazon, OpenAI, Anthropic):
  • Unprecedented capital—hundreds of billions—going into data centers and GPUs.
  • Question: Is the “bitter lesson” (more compute = better AI) inevitably true?
    • “What happens if all this investment in compute doesn’t actually lead to economic gain, slash better models, slash everything else?” (C, 74:53)
  • Most of today’s spend is R&D ($5B) vs $2B for inference (B, 76:11).
  • Still, with user growth and model improvements, "enough" economic base is likely sustained.

Memorable Moments & Quotes

• On the Joy (and Future Risk) of Coding with AI:

  • “Some of the joy of coding used to really be you’re stuck on this annoyingly hard problem…and then you solve it. And that’s the muscle that you build when you improve and get better. Now…I find myself…just pulling the slot machine all day long.” (C, 78:15)
  • “It’s a cigarette for your brain because you do not think anymore when you pull that button.” (C, 82:23)

• On Cultural/Global Talent:

  • “You look at a guy like Rahul Patel who's become the CTO of Anthropic and he's not from a top university in India...he's come to a society that is quite meritocratic and he sort of worked his way up…if you work hard enough in certain environments for a long time…anything can happen.” (C, 39:31)

• On Venture Capital Reflexivity:

  • "The belief that something can be true can make it true, even though it's not true at the time you believed it." (B, 71:47)

• On Model Lab Moats:

  • "It is far easier for Anthropic to try to go into one of the spaces of the apps than an app to try to go into the space of Anthropic, which makes me feel like one is more defensible than the other." (C, 33:33)

Timestamps for Important Segments

• [01:19] - Glean’s evolution from “boring” to $7B enterprise AI darling
• [07:34] - Different flavors of competition (Claude/Anthropic vs. Glean, data APIs, rate limits)
• [09:09] - Hard problems in enterprise AI adoption & integration
• [15:32] - Anthropic’s rise and investor perspective
• [23:43] - Changing LLM API market share in enterprise
• [30:44] - Coding models & the commodity vs. premium battle
• [41:06] - Anthology Fund: model, thesis, and startup stories
• [53:14] - Deep dive: OpenRouter, its moat, and business risks
• [61:07] - Whisper, the battle for voice/language dictation
• [64:08] - Diffusion models as an emerging architecture bet
• [69:06] - Roll-ups, market dynamics, and 'reflexivity' in venture
• [73:20] - Infrastructure arms race: compute, capital, and "bitter lesson"
• [78:15] - The effect of coding AI on engineers and human cognition

Tone & Style Notes

The conversation is candid, energetic, and sprinkled with Bay Area/SF “insider” self-deprecation, market banter, and references to both technical and cultural elements in the AI world. Deedy’s perspective bridges hands-on engineering, “hard” startup experience, and now high-stakes venture. The hosts prod for specifics but also riff, making the show both accessible and relevant to practitioners, founders, and investors.

Summary Takeaways

• AI company moats are hard-won at every layer: Solving unsexy, last-mile integration or research problems remains foundational and defensible.
• Model labs vs. application layers: Models retain deeper moats due to technical/labor intensity, but a thriving application ecosystem still creates outsized opportunities.
• Market/venture “reflexivity”: Belief, timing, and capital flows can themselves be self-fulfilling—sometimes moreso than product or tech alone.
• Infra buildout is the new arms race: Billions are being poured into compute as much as models/products; "bitter lesson" holds, but strategic gambles abound.
• Coding with AI brings both superpowers and new risks: As LLMs get smarter, human craft and safety must keep up, or we risk becoming “slot machine” coders.

For deeper dives, exclusive charts, and full company list, head to latent.space.