20VC: Cerebras CEO on the Future of Data Centres, Token Costs and Memory | We are Not in an Infra Bubble & Dario Got a Bad Deal with Elon for Compute | Should US Companies Sell to China & Why Most Layoffs are AI Washed with Andrew Feldman
20VC: Cerebras CEO on the Future of Data Centres, Token Costs and Memory | We are Not in an Infra Bubble & Dario Got a Bad Deal with Elon for Compute | Should US Companies Sell to China & Why Most Layoffs are AI Washed with Andrew Feldman
Podcast1 hr 1 min
Listen to Episode
Note: AI-generated summary based on third-party content. Not financial advice. Read more.
Quick Insights

Investors should consider Cerebras Systems (CBRS) as a high-conviction "pure-play" AI hardware alternative to NVIDIA, leveraging its $25 billion backlog and a 15x speed advantage in complex AI tasks. Because Cerebras utilizes SRAM etched directly into its chips, it is uniquely insulated from the global HBM memory shortages and packaging bottlenecks currently capping competitors' revenue. For those seeking to capitalize on supply chain constraints, memory manufacturers like Micron (MU) maintain massive pricing power and software-like margins that are expected to persist for several years due to fab capacity limits. The broader AI infrastructure trade remains robust, with a strategic shift toward companies solving the energy bottleneck as data centers scale from megawatts to multi-gigawatt requirements. To mitigate geopolitical risk, focus on firms involved in the onshoring of semiconductor fabrication and packaging to protect against vulnerabilities in the TSMC and Samsung supply chains.

Detailed Analysis

Cerebras Systems (CBRS)

• Cerebras recently completed the largest semiconductor IPO ever, with the stock price rising from $185 to $311 on its debut. • The company currently has a $25 billion backlog, indicating that demand for their AI chips is significantly outstripping supply. • Architectural Advantage: Cerebras chips are reportedly 15x faster than competitors for certain tasks. They recently demonstrated running the Kimi K2.6 model 6.7x faster than the next fastest GPU cloud. • Supply Chain Resilience: Unlike NVIDIA, Cerebras does not use HBM (High Bandwidth Memory) or CoWoS (Chip on Wafer on Substrate) packaging. They use SRAM, which is etched directly into the chip, insulating them from the current global memory shortages and packaging bottlenecks. • Customer Concentration: The company recently signed a deal worth over $20 billion with G42, moving beyond their previous concentration with OpenAI.

Takeaways

Pure-Play AI Investment: Cerebras is positioned as the only "pure-play" AI hardware peer to NVIDIA in the public markets. • Speed as a Moat: For "agentic flows" and complex coding tasks, speed is the primary value proposition. Investors should watch if Cerebras can maintain its 15x performance gap as NVIDIA releases new architectures (Blackwell/B200). • Margin Protection: Because Cerebras avoids expensive third-party memory (HBM), they may maintain more stable gross margins compared to competitors who are currently paying "software-like" margins to memory makers like Micron.


NVIDIA (NVDA)

• NVIDIA is described as having a strategy of funding and "over-allocating" compute to "neoclouds" (specialized AI cloud providers) to create competition for traditional hyperscalers like AWS and Azure. • Backlog Issues: Like the rest of the industry, NVIDIA is struggling with a massive backlog due to data center construction speeds and memory shortages. • Product Lifecycle: The transcript notes that Elon Musk’s recent compute purchases were "downrev gear" (H100s), placing him 1.5 to 2 generations behind the current cutting-edge B200 (Blackwell) chips.

Takeaways

Market Dominance vs. Dependence: While NVIDIA remains the market leader, the CEO of Cerebras suggests NVIDIA has created a "dependence" in the neocloud market that may not be healthy long-term. • Supply Chain Risk: NVIDIA is highly dependent on HBM (from Samsung, Micron, Hynix) and CoWoS packaging from TSMC. Any further shortages in these specific components directly cap NVIDIA’s revenue potential.


Memory Manufacturers (MU, Samsung, SK Hynix)

• There is an "extraordinary growth in demand" for HBM (High Bandwidth Memory), which is essential for most GPUs. • Pricing Power: Companies like Micron (MU) are achieving 80-85% gross margins on memory, which are typically seen only in software companies. • Shortage Duration: Memory shortages are expected to persist for the "next several years" because increasing capacity requires building new $40 billion fabs, a process that takes roughly five years.

Takeaways

Bullish Outlook for Memory: As long as AI demand stays high, memory makers hold significant pricing power. • Step-Function Growth: Investors should realize that supply cannot be added incrementally; it arrives in "steps" as new factories come online, meaning supply-demand imbalances will likely remain "lumpy."


AI Infrastructure & Data Centers

Not a Bubble: The speaker argues we are not in an infra bubble because, unlike the fiber-optic bubble of the 90s, current infrastructure is being built behind demand rather than ahead of it. • The Energy Bottleneck: The industry is moving from 20-megawatt facilities to multi-gigawatt buildouts. The core business of AI is now described as "turning electricity into intelligence." • The "Lawyer" Bottleneck: The biggest inhibitor to enterprise AI adoption is not technology, but security and legal departments that are risk-averse and lack precedent for AI contracting.

Takeaways

Energy is the Limit: Investment opportunities may lie in companies that can solve the power problem (e.g., Crusoe, SoftBank Power, or nuclear/grid upgrades). • Data Organization Advantage: Companies with decades of organized data (e.g., Mayo Clinic, GlaxoSmithKline) have a massive head start in implementing AI compared to those with "messy" data. • AI-Washing in Layoffs: Most recent tech layoffs are attributed to "AI washing"—companies correcting for over-hiring during COVID rather than AI actually replacing workers yet.


Geopolitical Insights: US vs. China

Export Controls: The speaker supports the ban on selling cutting-edge chips to China, labeling them an "industrial adversary." • Chinese Model Strength: Chinese open-source models (e.g., Kimi K2, DeepSeek, GLM) are described as "extraordinarily good," though slightly behind US closed-source frontier models. • Onshoring: There is a critical need to bring TSMC and Samsung fabrication and packaging capabilities to the US to protect against Taiwan’s vulnerability.

Takeaways

Strategic Choke Points: The US maintains an advantage through control of ASML (lithography) and TSMC (fabrication). • Regulatory Risk in Europe: Europe is viewed as being in a "be afraid of it, regulate it, tax it" mentality, which is stifling innovation compared to the US and China.

Ask about this postAnswers are grounded in this post's content.
Episode Description
Andrew Feldman is the co-founder and CEO of Cerebras Systems. This month, Cerebras went public achieving a market cap of $70BN, the largest semiconductor IPO in history. Cerebras has a massive commercial backlog with a monumental, multi-year $20 billion compute agreement from OpenAI. AGENDA:  05:58 - Why we are not in an infrastructure bubble and it is just the start 08:00 - Sam Altman's superpower is his ability to forecast capex spend. 08:58 - Anthropic did not get a good deal with Elon. They got a deal that was available.  10:39 - What is going on with the price of memory and why is it a problem? 16:40 - Are Google best positioned to produce tokens and what challenges do they face? 19:23 - Is Coreweave dramatically undervalued or overvalued? 24:34 - My biggest advice to entrepreneurs scaling their business  30:13 - Why most of the layoffs are AI-washed and 33:41 - What will we spend on tokens for software engineers in five years? 34:48 - Why does the role of HR change so significantly in the world of AI? 35:36 - Why lawyers are the biggest inhibitor of enterprise AI adoption 39:20 - Why Jensen and Nvidia are wrong to sell chips to China 42:49 - What needs to change in the U.S. to build a strategic asset in chips? 51:00 - Should Cerebras invest in companies building on top of their platform; as Nvidia is? 53:28 - Nothing changed when Cerebras IPO'd but I did make 800 millionaires.
About The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch
The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch

The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch

By Harry Stebbings

The Twenty Minute VC (20VC) interviews the world's greatest venture capitalists with prior guests including Sequoia's Doug Leone and Benchmark's Bill Gurley. Once per week, 20VC Host, Harry Stebbings is also joined by one of the great founders of our time with prior founder episodes from Spotify's Daniel Ek, Linkedin's Reid Hoffman, and Snowflake's Frank Slootman. If you would like to see more of The Twenty Minute VC (20VC), head to www.20vc.com for more information on the podcast, show notes, resources and more.