Cerebras CEO Andrew Feldman argues AI is not in an infrastructure bubble, claiming chip production lags far behind demand

HSHarry Stebbings@HarryStebbings

Why it is BS that we are in an AI infra bubble... "We can't build data centers fast enough to keep up with demand. We're not building ahead of demand. We're building behind demand. We are trying to keep up with demand, not the other way around." @andrewdfeldman What does no one know that everyone should know when it comes to AI infra spending? @altcap @GavinSBaker @nathanbenaich @mmurph @embirico

HSHarry Stebbings@HarryStebbings

It is BS to think we are in an AI infrastructure bubble. Dario did not get a good deal for compute with Elon. We, the US and Jensen should not be selling chips to China. This is the best podcast you will listen to this week on AI. My notes from my conversation with @andrewdfeldman, CEO @cerebras: 1. Why We Are Not in an Infrastructure Bubble and It Is Just the Start Unlike past tech bubbles that overbuilt ahead of demand, the AI infrastructure rollout is running significantly behind immediate market needs . Major chip makers face massive multi-billion-dollar backlogs because data centers cannot be built fast enough to keep pace with the exploding, real-world user demand happening today. 2. Anthropic Did Not Get a Good Deal With Elon. They Got a Deal That Was Available. Founders are often forced to take action on what is available in the market rather than what is ideal. Anthropic’s deal with Elon Musk required them to buy "down rev gear", older H100 chips rather than cutting-edge B200s, leaving them a generation and a half to two generations behind the absolute leading edge. 3. Why Jensen and NVIDIA Are Wrong to Sell Chips to China Widespread security consensus confirms that selling leading-edge tech to China means their military and government will inevitably leverage it to compete with Western industry . As an industrial adversary driving down global costs in vital sectors like solar and automotive, American chip companies should be completely comfortable selling fewer chips to protect strategic boundaries. 4. What the F**k Is Going on With the Price of Memory and Why Is It a Problem? Exploding AI demand makes High Bandwidth Memory (HBM) the most critical bottleneck right after fab space. Because only three manufacturers produce this specialized memory and cannot keep up, prices have skyrocketed, allowing suppliers to command software-like 80-85% gross margins on hardware fabrication. 5. Are Google Best Positioned to Produce the Lowest Cost Tokens and What Challenges Do They Face? Owning the full stack from data centers to chips gives hyperscalers an immense cost advantage over standard clouds paying high hardware margins. However, the historical downside is that your market is constrained strictly by your own internal demand, meaning hardware innovators must sell externally to maximize volume and lower unit costs. 6. My Biggest Advice to Entrepreneurs Scaling Their Business Your initial focus must be entirely on winning just one customer. Landing that first anchor client forces your company to build the necessary operational muscle, adjust your supply chain, and learn how to properly service a massive organization so you have the capability to keep the next ones happy. (links below)