Cerebras CEO on the Future of Data Centres, Token Costs & Memory | Should US Companies Sell to China

AI Infrastructure Supply Constraints:

• There is currently a $25 billion backlog in data center construction due to an inability to build facilities fast enough to meet demand.
• Memory shortages, particularly for High Bandwidth Memory (HBM), are expected to persist for several years if demand remains high.
• HBM prices have increased 4x to 5x due to supply chain constraints involving Samsung, Micron, and Hynix.
• New semiconductor fabs require a $40 billion investment and a five-year timeline to build, creating a "step function" inability to respond quickly to demand spikes.
• Micron reports gross margins of 80% to 85% on memory production, driven by the shortage.

Market Dynamics and Competition:

• Andrew Feldman (Cerebrus CEO) characterizes the current market as "behind demand" rather than a bubble, contrasting it with the 1990s fiber optics and railroad booms where infrastructure was built ahead of demand.
• Cerebrus, founded by Feldman, completed the largest semiconductor IPO in history last week, raising over $5.5 billion with a share price surge from $185 to $311.
• Cerebrus signed a $20+ billion deal with OpenAI, a relationship that grew from an initial $1 billion agreement with G42.
• OpenAI is reportedly 1.5 to 2 generations behind the latest GPU technology (e.g., missing the Blackwell B200) because they purchased existing stock (H100s) rather than securing the newest supply.
• Feldman argues NVIDIA has over-allocated to "neoclouds" to create competitors for traditional hyperscalers, potentially creating an unhealthy dependence.
• Cerebrus claims to run the Kimi K2.6 model 6.7x faster than the next fastest GPU cloud, positioning speed as a critical differentiator with a "zero market" for slow inference.

Cost Structure and Technology Trends:

• Cerebrus avoids HBM shortages and costs by utilizing SRAM, which does not face the same supply constraints.
• The industry trend points toward a massive reduction in the cost per unit of compute over time, driven by architectural improvements delivering more tokens per dollar and per watt.
• Feldman notes that while Google's full-stack ownership (TPUs to data centers) offers a cost advantage, it historically limits market size because hardware is sold only to internal use.
• OpenAI's "brilliance" lay in recognizing exponential compute demand years in advance and securing contracts for power and hardware before others.

U.S.-China Relations and Geopolitics:

• There is a consensus in the semiconductor industry that selling leading-edge technology to China will inevitably be used by their military and government to gain an industrial advantage.
• Feldman supports keeping China out of the US ecosystem to manage the risk, viewing them as an industrial adversary, despite historical appreciation for their entrepreneurs.
• The US strategy involves using chokepoints like TSMC and ASML to limit China's access to cutting-edge lithography.
• Feldman advocates for onshore TSMC-like capabilities in the US, citing the loss of packaging expertise and the ecosystem as a strategic liability.
• He proposes a policy change: granting a 20-year exemption from local and state ordinances for TSMC and Samsung to build fabs in the US, comparing fabs to "modern pyramids" that require specialized, streamlined regulation.

Enterprise Adoption and Barriers:

• The primary inhibitor to enterprise AI adoption is currently legal and security apparatuses that lack precedent for new technologies.
• Productivity gains from AI are expected to expand engineering teams rather than shrink them, as the volume of tasks increases 50x the current capacity.
• Roles such as CIO and CSO evolved from networking and security needs; a new "AI Officer" role is emerging for AI governance.
• Data organization is a competitive moat; companies like Mayo Clinic with 30 years of disciplined data structuring hold a significant advantage over those with unorganized data.
• Europe faces a structural challenge due to a regulatory mindset ("fear, then regulate, tax") and a culture less tolerant of failure compared to Silicon Valley.

IPO Strategy and Future Outlook:

• Cerebrus timed its IPO deliberately to be the first and only pure-play AI company public before major competitors like SpaceX or OpenAI, leveraging a window where investors were shut out of private deals.
• Feldman credits the successful IPO to persistence, noting they attempted to go public a year and a half earlier but were blocked by regulatory concerns (specifically regarding the previous administration) that disappeared under the current administration.
• Feldman predicts the "limiting factor" for AI will likely be electricity supply, with the industry shifting toward multi-gigawatt data center build-outs.
• He anticipates the "good neighbor" approach to data centers—paying for own infrastructure upgrades, recycling water, and building community assets like schools—to be the standard for future local permits.

Personal and Organizational Insights:

• Cerebrus has created approximately 800 millionaires, a metric Feldman cites as a core measure of leadership success compared to his previous company's 100.
• Feldman advises that sustaining a relationship while leading a public company requires a partner with patience who understands the daily "pressure test on the soul" of entrepreneurship.
• He revealed that he spent 18 months burning $8 million per month on R&D for a technical problem that only Cerebrus has since solved, a period of doubt where the board did not apply external pressure.
• The current administration is described as "unwaveringly better for business" due to the ability to operate without obstruction, though Feldman acknowledges specific points of disagreement.