A rival joins Micron in the $1 trillion club as one bank argues AI is actually underhyped
The trillion-dollar badge has long been the preserve of platform titans and one or two extraordinary chipmakers. Now, one of Micron’s chief memory rivals has vaulted into that club as well, a milestone that says as much about the changing character of artificial intelligence as it does about the companies themselves. For all the talk of GPUs, the market is finally pricing the other half of the AI equation: memory and bandwidth. And at least one major Wall Street bank believes the rally still isn’t fully reflecting what’s coming, calling AI not overhyped but underhyped.
What it means when memory is worth a trillion
For years, memory makers were treated as cyclical suppliers—highly leveraged to booms and busts, prone to price wars, and valued at a discount to logic peers. That mental model is breaking. High-bandwidth memory (HBM) has become the indispensable partner to accelerated computing. Without enough fast memory stacked inches from the compute die, even the most advanced AI processors idle.
Two implications follow:
– Value shifts toward bandwidth. The effective performance of an AI system is bounded by how quickly it can move and reuse data. That makes HBM capacity, power efficiency, yields, and packaging prowess strategic moats rather than interchangeable parts.
– Pricing power returns. With HBM demand outrunning supply and multi-quarter qualification cycles, the industry’s triopoly is exercising discipline. Long-term agreements, better mix, and rising layer counts translate into structurally higher margins than in prior commodity cycles.
Micron’s ascent—and now its rival’s—reflects this redesign of the stack. The market is acknowledging that memory is no longer a volume game but a performance product where design wins, yields, and packaging expertise command premium multiples.
Why a top bank says AI is underhyped
Against this backdrop, a leading investment bank argues investors are still underestimating AI’s arc. The thesis rests on five underappreciated vectors:
1) The memory ceiling hasn’t been priced in. Street models focus on GPU unit growth but often haircut the corollary explosion in HBM bits per GPU. Each new accelerator generation has raised HBM stacks, capacity per stack, and bandwidth per pin. As architectures push toward even larger context windows and massive multimodal models, that curve is steepening.
2) Inference will be bigger, sooner. Training grabs headlines, but inference—where models answer questions, generate content, and run continuously in production—drives recurring demand. Longer sequence lengths and retrieval-augmented pipelines push memory footprints higher, both in the cloud and, increasingly, in on-prem and edge deployments.
3) The build-out is moving sideways and down the stack. Hyperscalers are just the tip. Sovereign AI, industry-specific private models, and enterprise data estates are triggering a second wave of capex. That spills into networking, optical interconnect, advanced packaging, cooling, and, critically, memory supply chains.
4) Software maturity is multiplying hardware utility. Compiler advances, sparsity, quantization, and memory-efficient attention aren’t substitutes for bandwidth; they are enablers that let systems consume even more memory at higher utilization. Better software increases the return on every additional HBM stack.
5) Productivity is compounding, not linear. Early pilots show order-of-magnitude task acceleration across coding, customer support, design, and analytics. As workflows reconfigure, the bank argues GDP and earnings sensitivity will look more like the smartphone plus cloud eras combined rather than a single-cycle server refresh.
Together, those drivers support a longer runway for elevated capex and mix-rich memory sales than current consensus embeds.
The supply side: physics, packaging, and power
The underhyped view also leans on real-world constraints that help sustain pricing and profitability:
– HBM is hard. Moving from 8- to 12- and 16‑high stacks, thinning dies, and increasing TSV density while maintaining thermals and yields is a technical gauntlet. Only a handful of companies can execute at scale.
– Capacity expansion has long lead times. Cleanroom space, lithography tools, bumping, and 2.5D/3D packaging each represent bottlenecks with multi-quarter debottlenecking cycles.
– Power and cooling are gating factors. Data center power availability, liquid cooling footprints, and facility upgrades constrain how quickly hyperscalers can absorb new compute, smoothing demand waves rather than amplifying them.
Valuation math: is a trillion justified?
Skeptics will point to memory’s history: when supply catches up, prices crack. The counterargument is that AI has lifted memory from a commodity to a performance co-processor, advantaging leaders with:
– Long-term agreements that stabilize volumes and pricing
– Mix that shifts to HBM and higher-value managed NAND for AI-optimized storage
– Operating leverage as yields and utilization rise on next-gen nodes
– Ecosystem entrenchment via joint qualifications with top accelerator vendors
If margins prove sustainably higher, multiples can rerate. Layer on multi-year bit growth driven by larger models, inference at scale, and edge AI, and the earnings power to support trillion-dollar valuations becomes less fanciful.
The competitive chessboard
The memory race is moving on several fronts:
– HBM4 and beyond. Next nodes will demand even tighter logic-memory co-design, with broader adoption of advanced interposers and potential silicon bridges to push bandwidth further.
– Processing-in-memory experiments. Early efforts to move simple compute closer to memory won’t replace accelerators but could offload bandwidth-heavy primitives.
– CXL-enabled memory pooling. Composable memory can increase utilization and total system memory footprints, benefiting suppliers that master both HBM and DDR/CXL stacks.
– Vertical integration pressure. Accelerator vendors and hyperscalers are deepening co-development with memory partners, embedding long-term supply visibility—and raising the qualification bar for challengers.
Risks that could spoil the party
No thesis is bulletproof. Key risks include:
– Overbuild and price compression if supply ramps faster than expected
– Architectural shifts that reduce memory intensity per unit of compute
– Geopolitical frictions and export controls that fragment demand and complicate supply chains
– Power and permitting delays that defer data center expansions
– Execution missteps in yields, packaging, or node transitions
What to watch next
– HBM contract pricing into the next cycle
– The pace of 12‑high and 16‑high stack qualifications tied to new accelerators
– Data center power procurement and liquid cooling deployments
– The diffusion of AI beyond hyperscalers into sovereign and enterprise workloads
– Evidence of durable margin uplift versus past memory cycles
The bottom line
A rival’s arrival alongside Micron in the $1 trillion club signals a re-rating of memory from cyclical component to strategic performance layer in AI. If the underhyped camp is right, the industry is still early in a multi-year transition where bandwidth, capacity, and power efficiency are as decisive as raw compute. That shift doesn’t erase risk, but it does expand the set of companies with durable claims on AI economics—and explains why investors are starting to price memory like the keystone it has become.
This article is for informational purposes and does not constitute investment advice.
