Bernstein Senior Analyst: The First True Chip Super Cycle Is Coming, with the "Bottleneck" Being the Wealth Creator

June 21 — Bernstein senior semiconductor analyst Stacy Rasgon, an 18-year industry veteran with an MIT PhD and engineering background, is calling this the first true semiconductor supercycle he’s witnessed in his career. Rasgon shared eye-opening data: the semiconductor industry hit $800 billion in total revenue last year and is now on track to surge to $1.3 trillion in 2024. Every subsector — from accelerators and memory to semiconductor equipment, network optics, power chips, and even CPUs — is grappling with unprecedented demand. “The only widespread consensus right now is that no one has enough computing power,” Rasgon noted. Take memory, for example: high-bandwidth memory (HBM) used in AI chips makes up over 85% of the chip’s total silicon area, and producing 1GB of HBM requires roughly four times more silicon than standard DRAM. Even as fabs ramp production at full tilt, capacity gains remain tight — a supply-demand mismatch that’s even benefiting Intel. Its previously written-off inventory is now flying off shelves, with customers saying, “We don’t care — just ship it.” The industry’s core focus is shifting from AI model training to inference, the step that drives real commercial value: training models doesn’t generate revenue, but using them does. Anthropic reported annualized revenue skyrocketed from roughly $9 billion last December to $30 billion this April — a near-vertical surge. On the AI chip landscape, competition between custom ASICs (led by Broadcom) and Nvidia’s GPUs isn’t zero-sum, Rasgon explained. “The right question is whether the opportunity is big enough for everyone to grow. If it’s large, both parties will thrive.” Broadcom forecasts AI revenue will hit $100 billion next year; its ASICs currently make up 10% to low teens of the AI chip market, a share that could rise to 25-30% long-term — but ASICs won’t fully replace GPUs. For inference startups like Groq, which was recently acquired by Nvidia, Rasgon cited Jensen Huang’s view: “Not all tokens are equal. Low-latency tokens are more valuable, and GPUs aren’t the best fit for every task.” When asked about the industry’s most underrated risk, Rasgon shifted focus from silicon to the physical world: electricity. If Nvidia’s projected $3 trillion to $4 trillion AI infrastructure buildout comes to fruition, the U.S. power grid would need to expand by 5% annually — a pace power analysts call nearly impossible. That points to upcoming bottlenecks in energy generation, cooling, and nuclear power. But Rasgon added: “Never underestimate human creativity. Engineers will find a way when there’s enough profit on the line.” For Intel, new CEO Chen Liwu has adopted a pragmatic, low-expectations strategy. The company’s new 18A chip process is beating yield projections, while investments from the U.S. government and Nvidia have eased prior concerns about its balance sheet. Rasgon concluded: as long as AI demand doesn’t collapse, the full semiconductor supply chain supercycle will continue. Investors need to keep a close eye on capacity bottlenecks affecting every stage of the industry.