Cerebras CEO, Andrew Feldman on Why Raise $1BN and Delay the IPO & Why NVIDIA’s Worried About Growth

Capital Markets and Funding Strategy

• Cerebras secured the largest raise ever in its category at the highest valuation, led by Fidelity (a critical signal to Wall Street) and including Tiger Global, Valor, and 1789.
• The capital raise provides "dry powder" to scale manufacturing, add new data centers (adding five in the US this year), and execute big ideas beyond incremental software gains.
• Cerebras intends to proceed to an IPO but prioritized this pre-IPO round to rapidly capture opportunities without distraction.
• No company has ever gone bankrupt by paying extraordinary talent too much; conversely, paying mediocre people high salaries is a primary cause of failure.
• The war for talent has reached unprecedented levels, with top scientists commanding compensation exceeding world-class athletes due to their irreplicable value generation.

Market Dynamics, Demand, and Uncertainty

• Global demand for AI compute is described as "unbelievable," with some customers requesting between 5 million and 40 million queries per second, a variance indicating profound uncertainty.
• The market operates on a "helter-skelter" pace where 6 to 12-month forecasts are impossible due to exponential growth in the number of users, frequency of use, and compute intensity per use.
• Current investment announcements (e.g., "$100 billion over five years") should be viewed as "options on the future" rather than guaranteed commitments, as the timeline and execution are often vague.
• There is a 100% certainty that the industry is underestimating demand, valuations, and the rate of new ideas entering the community.
• Risk in financial markets arises not from the assets themselves, but from investors fundamentally underestimating risk when diversification is an illusion (e.g., S&P 500 heavily weighted in "Mag Seven").

Technology, Hardware, and Depreciation

• Chip depreciation is not a fixed timeline but depends on the speed of performance gains in the next generation; if new chips are significantly faster and more power-efficient, old fully depreciated chips are retired.
• Cerebras estimates 90% of the potential performance gains in AI compute remain to be made, despite marketing claims suggesting rapid saturation.
• The primary bottleneck in chip performance is not just compute (FLOPS) but memory bandwidth; SRAM is fast but low capacity, while HBM is high capacity but slow.
• Cerebras solved the SRAM capacity limitation by developing "wafer-scale" chips (the size of a dinner plate), a feat no company had achieved successfully in 75 years of compute history.
• Training remains the harder software lift due to dependency on specific ecosystems (like CUDA), whereas inference moves are easy, often requiring only 10 keystrokes to switch hardware providers.
• The industry is moving toward using wafer-scale chips to enable "speculative decode" and reduce the physical mess of connecting thousands of standard chips.
• Large incumbents like NVIDIA are utilizing balance sheet dominance and pre-announcements of future products (e.g., B300 before B200) to manage competition and lock up demand.

Infrastructure, Energy, and Geopolitics

• The narrative that the US lacks power is false; the issue is a geographic mismatch between available power sources (West Texas natural gas, Upstate NY hydro) and population/fiber infrastructure locations.
• Nuclear power is a reasonable strategic option for nations lacking natural alternatives (geothermal, hydro), but not an absolute global necessity.
• US data center construction faces significant headwinds due to decentralized local permitting and fire ordinances, which can delay projects by 8-12 months and force design changes.
• The US lags in strategic power infrastructure planning compared to China, which utilizes centralized government policy and direct capital backstops to aggressively advance AI capabilities.
• Export controls and geopolitical tensions have limited US access to the Chinese market, though US firms previously bypassed some restrictions before Department of Commerce limitations tightened.
• Sovereignty plays a role in regional adoption (e.g., Mistral in Europe), combining local data requirements with high-performance hardware to create competitive products.

Future Outlook and Economic Impact

• Productivity gains from AI will be modest if used as replacements for existing tools (e.g., "Google replacement") but massive if the economy reorganizes around AI capabilities.
• The "Solow Paradox" suggests productivity gains occur only after a period of reorganization and new infrastructure, similar to the electrification of factories in the mid-20th century.
• A labor shortage in five years is unlikely; AI will nibble at the economy gradually rather than causing immediate, massive dislocation.
• Education and entry-level corporate roles will undergo fundamental transformation, with AI handling rote tasks (spreadsheets, summaries) to allow junior staff to focus on higher-value thinking.
• The AI silicon market will not become a monopoly; historical precedents show that dominance in one area (e.g., x86) rarely translates to dominance in all compute domains (e.g., mobile or switching).
• Data pipelines and data cleaning are currently under-invested areas where many AI projects fail, despite being critical to success.
• Cerebras generates significant revenue from the UAE (75-80% of reported figures in H1 2024), driven by early, massive adoption by entities like G42.
• Building a wafer-scale chip was a 15-month period of high burn ($6-7M/month) where 100% failure was possible, but the team persisted through systematic engineering analysis until success.