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

Market Trends & Demand - The AI market is experiencing unbelievable demand, but future direction remains uncertain [1] - Customers are requesting between 5 million and 40 million queries per second, highlighting demand uncertainty [1] - The announcements of deals should be considered as options on the future due to the unknown environment [1] - The growth rate of inference is exponential, driven by more users, increased frequency of use, and more complex tasks [6] Financial & Investment - A $1 billion raise was completed at the highest valuation with premier investors like Fidelity [1] - The funding will be used to expand manufacturing, add new data centers, and pursue big ideas in AI [1] - Nvidia's investment of $100 billion into OpenAI is difficult to analyze beyond Nvidia trying to secure a portion of OpenAI's demand [2] - The company's revenue is concentrated in the UAE, accounting for 75-80% of revenues in the first half of 2024, due to large orders consuming manufacturing capacity [11] Technology & Infrastructure - The industry is seeing performance improvements of around 2-25x per meaningful generation move in chips [5] - The company chose wafer scale to overcome the limitations of SRAMM by using a huge amount of silicon area [5] - There is a shortage of data center capacity, despite significant investment commitments [8] - Building data centers is not for everyone, and the best can build them for $8 million per megawatt [9] Competition & Strategy - Large companies tend to use their balance sheet more and technology less as they worry about growth [2] - Nvidia's high margins (potentially 78% or higher) are driving interest in building alternative solutions [10] - The company believes that sovereignty, combined with fast hardware, makes Mistral's product compelling [10] Challenges & Bottlenecks - There are fundamental limitations in AI expertise, with not enough AI practitioners and data scientists being produced [8] - TSMC's ability to build fabs quickly enough is limited, impacting chip supply [8] - Power infrastructure and permitting processes at the local level can interfere with projects and set them back [10]