Battle for the AI Data Center: Deep Dive on the Semiconductor Supercycle

Summary

Michael Marks and Bernstein’s Stacy Rasgon dissect why the current AI-driven semiconductor surge looks unlike prior cycles. Rasgon argues it is primarily a demand cycle, not a classic inventory or supply cycle, and says the scale and speed of the move across memory, wafers, equipment, networking, power semis, and even CPUs is unprecedented. He highlights HBM as a core AI bottleneck, explains why Moore’s Law no longer delivers cheaper transistors, and describes how chiplets and advanced packaging are becoming necessary to push performance forward. The conversation also turns to business economics: training creates chip demand, but inference is where revenue is realized, with agentic coding and low-latency workloads emerging as important monetization paths. Rasgon is bullish on the cycle, skeptical that it is a bubble yet, and warns that power availability and grid buildout may ultimately constrain the next leg.

Notes

• Rasgon says this may be the first real semiconductor supercycle he has personally seen.
• The current AI shortage is defined by compute scarcity, not model scarcity.
• Marks frames hyperscaler AI capex as massive, with the four largest cloud companies near $600 billion of spend this year.
• Rasgon distinguishes supply, inventory, product, and demand cycles in semiconductors.
• He argues the present episode is a demand cycle.
• The speed and magnitude of the AI ramp are unusually large.
• The rally has broadened beyond AI names into much of semiconductors.
• Demand moved from accelerators to memory, tools, networking, optical, power semis, and now CPUs.
• He says the memory cycle is likely the strongest ever seen.
• He contrasts that with the prior memory trough, which he says was the worst since the tech bubble.
• Memory cycles can still be brutal, with prior crashes including bankruptcies in the tech bubble era.
• HBM is the special DRAM format powering AI systems.
• HBM is stacked DRAM packaged together for AI accelerators.
• HBM consumes far more silicon area per gigabyte than standard DRAM.
• Stacking lowers yield, which is part of the HBM scarcity problem.
• Moore’s Law no longer delivers cheaper transistors in the old way.
• The industry is getting performance and power gains, but at rising cost.
• Physical die-size limits are forcing architectural changes.
• Blackwell is described as two GPUs combined with multiple memory dies on an interposer.
• Nvidia’s pricing power remains exceptional, reflected in roughly 75% gross margins.
• Hyperscaler capex is still rising despite already enormous spending.
• Anthropic is presented as evidence that AI revenue can ramp very fast.
• Customers are reserving packaging and co-packaging capacity as part of the bottleneck chain.
• Hyperscalers are not new to custom silicon; Google and Amazon have done it for years.
• Google’s TPU program is already many generations deep.
• Apple’s vertical integration went as far as migrating its PCs off Intel x86 to custom ARM silicon.
• Vertical integration is partly about bottleneck avoidance and supply security.
• The industry is effectively sold out across multiple layers.
• AI demand has hit one capacity ceiling after another in the stack.
• Raw materials such as substrates and T-glass are emerging constraints.
• The industry has maintained pricing power through the cycle.
• Between 2019 and 2024, semiconductor units were flat while ASPs rose 50%.
• February industry revenue was up roughly 80% year over year, mostly because of memory pricing.
• The sector is already on a run rate above $1 trillion.
• Chinese memory makers are the most important new entrants.
• Those Chinese firms are constrained by U.S. regulatory limits.
• Marks says the decisive AI question is whether inference can generate durable revenue.
• Training itself does not monetize; inference does.
• Agentic inference is emerging as the next step beyond basic chatbots.
• Agentic coding appears to be the clearest early AI revenue source.
• AI spend can exceed the labor cost of the workers it is meant to replace.
• Inference uses GPUs, but it also consumes substantial CPU capacity.
• Token economics are uneven; not all tokens have the same value.
• Low-latency tokens can monetize at much higher rates than ordinary workload tokens.
• The chip-equipment market is strong, but factory capacity is the near-term limiter.
• He rejects the near-term AI bubble narrative because physical expansion limits are still binding.
• The WFE market is concentrated in five major vendors with 70%+ share.
• Chinese equipment vendors are gaining share because U.S. firms face export restrictions.
• Broadcom’s AI business is scaling fast enough to imply more than $100 billion next year.
• Broadcom’s role is driven by custom workload optimization and negotiation leverage versus Nvidia.
• ASICs trade flexibility for lower total cost of ownership on stable workloads.
• ASICs are already a meaningful share of AI shipments and could grow to roughly a quarter or more of a much larger market.
• Intel’s foundry strategy is strategically necessary, but execution was weak under Pat Gelsinger.
• Lip-Bu Tan is portrayed as a better fit because he understands customers, foundries, and industry relationships.
• Intel’s balance sheet has improved materially thanks to outside investments.
• Panther Lake looks promising, but yields are still not good.
• Intel’s server competitiveness is still years away.
• Strong server CPU demand is even helping Intel move previously written-off inventory.
• AI buildout may eventually be constrained by the power grid, not just chip supply.
• On-site generation and even restarts like Three Mile Island are part of the power discussion.
• Startup formation in semiconductors is picking up again after years of sparse VC activity.