Bernstein's star analyst Stacy Rasgon believes that with AI infrastructure investment surging towards 4.4% of US GDP, the semiconductor industry is experiencing an unprecedented, genuine "supercycle."
On June 21st, the Tech Surge Deep Tech Podcast, which focuses on emerging technology frontiers, released a new interview transcript featuring an in-depth conversation between Celesta Capital founding managing partner Michael Marks and Bernstein's renowned chip analyst Stacy Rasgon.
During the nearly hour-long dialogue, the two delved into the semiconductor revenue growth driven by AI, the transition from AI training to inference, capacity bottlenecks across the supply chain, and the sustainability of future industry growth.
Unlike most Wall Street analysts, Rasgon holds a Ph.D. from MIT and has a pure engineering background, which makes him place greater emphasis on established physical laws and capital flows.
Rasgon clearly stated that the semiconductor industry is currently experiencing the largest demand explosion he has witnessed in his career. Last year, the industry's total revenue surpassed $800 billion, and this year it is racing towards a scale of $1.3 trillion.
Rasgon remarked in the interview: "Throughout my entire career, I've always heard the term 'supercycle.' And this might be the first real one I've seen. The only thing we're hearing now is that no one has enough compute."
Rasgon emphasized that the current market focus is shifting from "model training" to "AI inference," which is the core for achieving commercial monetization. Simultaneously, capacity bottlenecks are spreading comprehensively from GPUs to HBM memory, semiconductor equipment, and even power supply.
In the future, custom chips (ASICs) represented by Broadcom and GPUs from NVIDIA will coexist long-term in the expanding incremental market, jointly absorbing this seemingly bottomless demand for compute power.
The Supply Chain's Whack-a-Mole Game: The Entire Industry is Being Forcibly Pulled by AI
As the bottomless pit of AI compute demand opens up, the market is exhibiting a peculiar "whack-a-mole" effect—capacity bottlenecks are erupting one after another along the industry chain.
Rasgon detailed this phenomenon: "Everything is being dragged by this insatiable demand for AI compute. I've never seen anything of this scale in my career. It went from accelerators to memory, to semiconductor manufacturing equipment, to networking and optics, to power semiconductors, and now even CPUs are in shortage."
Taking memory as an example, the industry is experiencing its strongest upcycle ever, with prices doubling every quarter. The core driver behind this is HBM (High Bandwidth Memory). Rasgon revealed a key data point: "In the silicon area of an AI chip, over 85% might be HBM."
More critical is the "trade ratio" issue. He said: "Due to yield loss from stacking technology and the occupation of logic die space, manufacturing 1GB of HBM requires roughly 4 times the silicon area of standard DRAM."
This means that even if wafer fabs expand capacity frantically, the actual incremental increase in memory capacity (bits) remains highly constrained.
This extreme demand is even benefiting weaker players unexpectedly. Discussing Intel's server CPU business, Rasgon pointed out bluntly that current server demand is so strong that Intel is even seeing margin upside: "Demand is so strong that they are even selling inventory that was previously written off, sitting in a corner of a warehouse like garbage. Customers' attitude now is: 'We don't care, we want it, please sell it to us.'"
The Inflection Point Arrives: "You Can't Make Any Money Training Models"
Despite hundreds of billions of dollars flowing in, the biggest market concern is: Is this growth sustainable? Where is the actual potential?
Rasgon points directly to "Inference" as the breakthrough. He emphasizes that a huge amount of capital was previously used for large model training, but that is not the endgame for commercialization. Rasgon stated: "You can't make any money training models... You have to be able to use the model. That's inference."
This shift is already reflected in the staggering data from startups. Rasgon cited data in the interview, noting that companies like Anthropic are showing a vertical rise in their annualized revenue run rates: "$90 billion in December last year, $140 billion in January this year, and recently (April) it has reached $300 billion."
Furthermore, with NVIDIA's recent acquisition of Groq, the segmented demand in the inference market is becoming prominent. Rasgon pointed out that not all data "tokens" hold equal value. For specific inference tasks requiring extremely low latency and fast response, custom chips or dedicated inference architectures often offer better economics than general-purpose GPUs.
Custom ASICs and GPUs Are Not a Zero-Sum Game
Against the backdrop of exploding inference demand, the momentum of custom chips (ASICs) is challenging the absolute dominance of GPUs. Broadcom has become the biggest beneficiary of this trend.
Rasgon, mentioning Broadcom, said: "Before all this started, Broadcom said semiconductors were a mature industry with mid-single-digit growth. But now everything has exploded. They say next year they think they can do $100 billion in AI revenue."
Why are major cloud service providers so intent on developing their own ASICs? Rasgon believes it's not just for performance optimization but also to have bargaining chips when facing NVIDIA's gross margins as high as 75%. Rasgon said: "At least when you sit at the negotiating table with Jensen Huang to discuss next year's contract, you want to have some cards in your pocket."
But Rasgon emphasized this is not a game of one replacing the other. If ASICs capture a larger share, it's because the entire pie is getting bigger. For large, stable, internally developed workloads, ASICs can provide lower total cost of ownership; but if the model architecture changes, the programmability advantage of GPUs is irreplaceable. Rasgon believes: "The right pain point is: Is the opportunity in front of us still getting bigger? If it's big enough, they will both thrive; if it's not, everyone is screwed."
The Ultimate Future Ceiling: The Grid Might Not Hold Up
When asked about risks the market might be overlooking, Rasgon shifted the focus from code and silicon back to the physical infrastructure of the real world—electricity.
Currently, cloud giants' capital expenditure this year has reached $600 billion. If future infrastructure spending develops according to NVIDIA's forecast of $3 to $4 trillion annually, humanity's existing energy systems will face collapse.
Rasgon shared a calculation model he previously built: "Do we even have enough electricity to do this? The grid might not hold. US power capacity needs to grow about 5% annually over the next decade. In the eyes of power equipment analysts, a 5% annual growth rate is simply unachievable."
This means the next wave of AI innovation and bottleneck breakthroughs will inevitably fall on areas like energy generation, cooling, and nuclear power. As he has always believed: "Never underestimate human ingenuity. If there's money to be made, engineers will find a way."
Overall, as long as AI demand does not plummet off a cliff, the "supercycle" for the entire semiconductor industry chain will continue. The focus of capital markets must closely follow these "capacity bottlenecks" that are constantly shifting across various segments.
Comments