---
type: raw-user-note
source_type: linkedin-transcript
category: raw
domain: ai-infrastructure
status: captured
captured: 2026-05-31
published: 2026-05-29
confidence: medium
tags:
  - andrew-feldman
  - cerebras
  - ai-infrastructure
  - inference-economics
  - hbm
  - data-centers
  - neo-clouds
  - us-china
  - export-controls
source: https://www.linkedin.com/posts/andrewdfeldman_cerebras-just-went-public-in-the-largest-activity-7465386786073268224-k9AY
related_digest:
  - "[[08 Reference/2026-05-31 Andrew Feldman - 20VC AI Infrastructure Interview - Digest]]"
---

# Andrew Feldman - 20VC AI Infrastructure Interview Transcript

Source: Andrew Feldman LinkedIn post with 20VC interview transcript  
Primary URL: https://www.linkedin.com/posts/andrewdfeldman_cerebras-just-went-public-in-the-largest-activity-7465386786073268224-k9AY  
Captured from: Telegram direct transcript supplied by Noam Rosen, plus public LinkedIn source check  
Captured: 2026-05-31

## Source details

- **Interview:** Harry Stebbings / 20VC interview with Andrew Feldman, founder and CEO of [[04 Companies/Cerebras]]
- **Public source post:** Andrew Feldman LinkedIn post on Cerebras IPO and the 20VC conversation
- **Transcript status:** user-supplied long transcript; the public LinkedIn page exposes a partial transcript preview and confirms the source topic set
- **Important caveat:** the chat/runtime copy available to this agent was truncated, so this raw note preserves the source metadata, the available transcript opening, and the core user-supplied claims/themes for KB use

## Available transcript excerpt

> We can't build data centers fast enough to keep up with demand. We have a $25 billion backlog. If demand stays high, we are going to continue to see memory shortages for at least the next several years.

The opening frames the whole conversation: Feldman argues that the AI infrastructure build-out is unlike classic fiber/rail bubbles because the current constraint is not speculative oversupply. The constraint is that data centers, power, memory, and AI systems cannot be built fast enough to satisfy already-visible demand.

## User-supplied themes to preserve

- **AI infrastructure is not a normal bubble if supply is behind demand.** Feldman contrasts prior infrastructure bubbles, where capacity was built ahead of demand, with AI infrastructure where customers want capacity now and vendors are backlogged.
- **Cerebras backlog as demand evidence.** The transcript attributes a roughly **$25B backlog** to Cerebras, with similar backlog dynamics at Nvidia, AMD, and others.
- **HBM / memory shortages can persist.** Feldman argues that if demand stays high, memory shortages may last for several years because supply-chain expansion lags demand.
- **2025 is mass inference usage.** The conversation points toward inference, not just training, as the next adoption and infrastructure wave.
- **Fast inference changes agentic/coding workflows.** For hard problems and coding agents, there is no practical upper bound on how much faster users want the system to be; speed improves iteration, exploration, and accepted work.
- **Nvidia and neo-cloud dependency.** Feldman argues that Nvidia has helped create alternative buyers to the traditional hyperscalers by funding, backstopping, and overallocating to neo-clouds, creating a dependency that may become strategically unhealthy.
- **Developer token TAM.** The transcript is the source Noam wanted attached to the prior market-sizing idea: annual AI-token spend per software engineer in the **$50K-$100K** range, multiplied by **47M** software engineers, implies a multi-trillion-dollar recurring inference market. Treat this as Feldman's aggressive TAM heuristic, not a forecast.
- **Enterprise blocker: legal and security.** The transcript highlights that enterprise AI adoption is limited less by model capability than by legal, security, privacy, data-control, and workflow-risk concerns.
- **Data-center community acceptance.** AI data centers need a local license to operate: power, water, land, noise, jobs, ratepayer impacts, tax base, and community negotiation become deployment variables.
- **US-China export controls.** Feldman treats export controls as strategically important but hard to execute cleanly because demand, smuggling, substitution, and national industrial policy all interact.
- **US fab policy.** The conversation includes the need for domestic and allied semiconductor manufacturing capacity, but the KB read should separate fab resilience from the full AI infrastructure stack: chips, packaging, memory, systems, power, cooling, and data centers.
- **Europe competitiveness.** Feldman is skeptical that Europe is structurally well-positioned in frontier tech without a more competitive capital, labor, energy, and regulatory posture.
- **Application-layer ecosystem optionality.** Feldman says Cerebras may now have more room, as a public company, to acquire, invest in, or partner with companies building on top of its infrastructure. The key distinction is governance of capital: venture-backed companies and public companies have different investor mandates and different partnership tools.
- **Public-company resilience through building.** Feldman frames the delayed IPO process as a lesson in continuing to build through conditions outside management control: more customers, better roadmap execution, more sales, and more distance from competitors made Cerebras stronger when the window reopened.
- **Hard-tech failure analysis as operating discipline.** Feldman describes an 18-month period burning roughly $8M/month while the team could not solve the technical problem. The durable operating lesson is not CEO bravado; it is repeated failure analysis, learning without failing the same way twice, and extending failure time from seconds to hours.
- **Board empathy matters in hard technology.** Feldman credits a board that understood additional pressure would not solve a deep technical problem. For hard-tech governance, investors need enough empathy to distinguish avoidable execution failure from the painful learning curve of solving a problem no one else has solved.
- **Leadership cost is paid at home.** Feldman is explicit that building this kind of company is a family burden: heavy travel, repeated work failure, mood spillover at home, and relationship "debits" that need conscious credits back. This is a useful non-market signal on the real cost of frontier infrastructure companies.

## KB interpretation

This source should sit in the KB as a high-signal saved source post/interview, not as a daily brief. It connects four KB threads:

- AI infrastructure demand is constrained by deployable capacity, not by lack of interest.
- Inference economics may scale with repeated developer and agentic workflows.
- Sovereign and regional AI strategy depends on control of the recurring inference stack, not just model access.
- The physical layer of AI - data centers, power, HBM, fabs, export controls, and community acceptance - is becoming a board-level strategic constraint.
- Hard-tech leadership should be evaluated by resilience systems: investor patience, failure-analysis discipline, family/relationship cost management, and ability to keep building when public markets, regulators, or technical execution do not cooperate.

## Related notes

- [[08 Reference/2026-05-31 Andrew Feldman - 20VC AI Infrastructure Interview - Digest]]
- [[05 People/Andrew Feldman]]
- [[04 Companies/Cerebras]]
- [[03 Research/AI Infrastructure/AI Infrastructure Overview]]
- [[03 Research/Inference Economics/Inference Economics Overview]]
- [[03 Research/Sovereign AI/Sovereign AI Overview]]
- [[03 Research/Enterprise Adoption/Enterprise Adoption Overview]]
- [[06 Theses/The Inference Economics Thesis]]
- [[06 Theses/The Sovereign AI Thesis]]
- [[00 Raw/User Notes/2026-05-30 Noam Rosen - India AI Token Import Bill]]
