20VC: Cerebras CEO on the Future of Data Centres, Token Costs...

Full Title

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

Summary

This episode features Andrew Feldman, CEO of Cerebras, discussing the intense demand for AI infrastructure, the challenges in scaling data centers, and the future of chip technology. Feldman emphasizes that the current infrastructure build-out is lagging behind demand, contrasting it with historical bubbles, and highlights the strategic implications of the AI race, particularly concerning US-China relations and the evolution of the workforce.

Key Points

• The AI infrastructure build-out is currently behind demand, unlike past infrastructure bubbles where construction outpaced need, indicating a fundamental difference in the current market dynamic.
• The immense demand for AI compute is creating shortages and driving up costs, particularly for memory components like HBM, which are critical for GPUs.
• NVIDIA's strategy may have inadvertently fostered dependence among newer cloud providers ("neoclads"), potentially creating an unhealthy market dynamic.
• Deals for AI compute, such as OpenAI's with Elon Musk, may not always be optimal, with companies sometimes forced to accept less favorable terms due to urgent demand.
• The increasing usefulness of AI models, particularly since 2025, is driving exponential demand for compute as AI moves from novelty to everyday utility across demographics.
• Hyperscalers like AWS and Azure offer broad, integrated solutions that are valuable for many enterprises, but a segment of the market may prefer more specialized, cost-effective compute solutions without the extra software layers.
• Cerebras is advantaged in the current market due to its use of SRAM (which has no shortage) and its chip manufacturing process (5nm), avoiding the constraints and higher costs faced by others reliant on more constrained nodes or HBM memory.
• The cost per unit of compute has historically decreased significantly due to industry-wide design improvements, a trend expected to continue.
• Google's full-stack ownership of TPUs, data centers, and networking could lead to cost efficiencies, though historical precedent suggests challenges when a company is its sole customer for hardware.
• Large-scale AI adoption in enterprises is significantly hindered by security apparatuses and lawyers who are risk-averse and tend to slow down new technology implementation.
• The long-term success of AI depends on the continued advancement of models and the ability to efficiently scale infrastructure, with energy availability being a potential core bottleneck.
• Building data centers faces inherent delays and complexities due to construction, supply chain issues, and municipal permitting, making rapid scaling a significant challenge.
• The industry's past poor engagement with local communities has contributed to negative perceptions of data centers, but better community engagement and corporate responsibility can foster acceptance.
• Layoffs attributed to AI are often "AI-washed" excuses for broader issues like overhiring during COVID, with true AI-driven job displacement only beginning to emerge as the technology impacts productivity more broadly.
• The US faces significant policy challenges in long-term infrastructure development and incentivizing domestic manufacturing, particularly in critical areas like semiconductor fabrication.
• The US should focus on onshoring advanced chip manufacturing capabilities and the surrounding ecosystem to counter industrial adversaries and maintain strategic independence.
• The hesitancy of European markets to embrace entrepreneurship and new technologies, due to a culture of regulation and risk aversion, is a significant inhibitor to innovation compared to regions like Silicon Valley.
• The IPO of Cerebras was a deliberate strategic move to be the first pure-play AI infrastructure company in the public market, capitalizing on investor demand for this sector.
• Entrepreneurial persistence, continuous innovation, and building a strong business are key to navigating external challenges and achieving favorable outcomes, such as a successful IPO.
• The pressure on CEOs is immense, and having a supportive partner who understands the entrepreneurial journey is crucial for personal and professional resilience.

Conclusion

The demand for AI infrastructure is outstripping supply, creating significant shortages and cost pressures, a dynamic that differentiates it from past infrastructure bubbles.

The future of AI development and deployment hinges on solving complex challenges related to scaling infrastructure, managing energy resources, and navigating regulatory and societal adoption hurdles.

Entrepreneurial resilience, continuous innovation, and strong partnerships are essential for success in the rapidly evolving and demanding field of AI and advanced computing.

Discussion Topics

• How will the current demand for AI infrastructure shape the geopolitical landscape and global technological competition in the coming decade?
• What are the most significant ethical considerations and societal impacts we should anticipate as AI becomes more deeply integrated into various industries and daily life?
• Beyond technical capabilities, what human qualities and leadership styles will be most critical for navigating the transformative period driven by AI and advanced computing?

Key Terms

HBM

High Bandwidth Memory, a type of RAM used in high-performance GPUs to provide fast data access.

Hyperscalers

Large cloud computing providers like Amazon Web Services (AWS), Microsoft Azure, and Google Cloud, capable of operating at a massive scale.

Neoclads

A term used in the transcript to refer to newer, emerging cloud providers, possibly distinct from traditional hyperscalers.

SRAM

Static Random-Access Memory, a type of semiconductor memory that uses latches to store each bit of data, known for its speed and stability.

TPUs

Tensor Processing Units, Google's custom ASICs designed to accelerate machine learning workloads.

CFIUS

The Committee on Foreign Investment in the United States, an interagency body that reviews proposed or pending transactions involving foreign investment in the United States for national security concerns.

Timeline