Bernstein Analyst: The Chip Super Cycle Is Real — Bottleneck Players Are the Core Investment Opportunity
Claire Weston
Bernstein chip analyst Stacy Rasgon says semiconductors are in their first true super cycle — revenue jumping from $800 billion last year toward $1.3 trillion this year — with capacity bottlenecks cascading down the supply chain and power infrastructure emerging as the ultimate hard constraint.
Why call this a "super cycle" and not just a strong upturn?
Semiconductor revenue broke $800 billion last year and is heading toward $1.3 trillion this year. This means → the industry is growing over 60% in a single year, far beyond normal cyclical swings.
Rasgon calls it the first true super cycle he has seen in his career. His case rests not on sentiment but on a physical constraint: nobody has enough compute.
In plain terms = past chip cycles were driven by inventory rhythms. This time AI is consuming every chip available, and supply cannot keep up no matter how fast it expands.
Where are the capacity bottlenecks, and why do they play "whack-a-mole"?
Bottlenecks are cascading down the supply chain: GPU accelerators → HBM high-bandwidth memory → semiconductor equipment → networking and optics → power semiconductors → CPUs — the entire chain is being dragged by AI compute demand.
Of all the silicon area in a single AI chip's components, HBM — high-speed memory made by stacking multiple memory dies — may account for over 85%. This means → the real production ceiling for AI chips is memory, not logic.
Manufacturing the same capacity of HBM requires roughly four times the silicon area of standard DRAM. Even with aggressive fab expansion, actual new memory capacity remains very limited.
In plain terms = it is like whack-a-mole — solve the GPU shortage and HBM runs out; ramp HBM and equipment and power cannot keep up. The bottleneck never disappears; it just moves.
How strong is demand — strong enough to lift even weak players?
Rasgon says Intel's server CPU business is benefiting unexpectedly: "They even sold inventory they had previously written off and left sitting in a warehouse corner like garbage. The customers' attitude was: we don't care, we'll take it."
This reflects a market no longer in "pick-your-brand" mode — buyers are taking whatever is available.
Why is the inference market the real battlefield for monetization?
Rasgon anchors the super cycle's durability to inference. Massive capital went into training large models, but training alone cannot be monetized directly — "you have to be able to use the model, and that's inference."
Anthropic's annualized revenue run rate is rising almost vertically: $9 billion in December, $14 billion in January, $30 billion by April. This means → AI application-side revenue is materializing now; inference demand is fact, not forecast.
Custom ASICs — chips designed for a single type of workload — and Nvidia GPUs are not a zero-sum fight. ASICs lower total cost of ownership for large, stable workloads, but when model architectures change, GPUs' programmability is irreplaceable.
Cloud providers pursue in-house ASICs partly for leverage against Nvidia's 75% gross margin. In plain terms = the custom chip may never fully replace Nvidia, but it gives you "a card in your pocket" when you sit across from Jensen Huang to negotiate next year's contract.
Why is Broadcom the biggest beneficiary?
Broadcom expects AI revenue to reach $100 billion next year. This reflects custom-ASIC demand exploding far faster than even the industry itself expected.
In plain terms = just before this cycle started, Broadcom itself classified semiconductors as a mature industry with only mid-single-digit growth. Now it is one of the fastest-growing companies in the AI chip wave.
Could power become the ultimate ceiling for this super cycle?
Rasgon's model shows that if AI infrastructure spending scales to Nvidia's projected $3–4 trillion per year, U.S. power capacity would need to grow by roughly 5% annually over the next decade.
Power-equipment analysts say that growth rate is "simply not achievable." This means → power is not just "the next bottleneck" — it is a hard constraint that could cap the entire super cycle.
Rasgon's conclusion: as long as AI demand does not collapse off a cliff, the super cycle will persist. The next breakthrough will inevitably fall in power generation, cooling, and nuclear energy.
In plain terms = whoever solves the power constraint first controls the ultimate ceiling of this super cycle — and the investment lens must follow wherever the bottleneck moves.
Content is for reference only, not financial advice.