After HBM: where the next AI bottleneck is forming

Memory was the squeeze of 2026. The constraint is already moving outward from the chip: to packaging, optics, storage and power. Here's the order, the numbers, and what could break it.

In the HBM explainer, the takeaway was simple: AI chips are only as fast as the memory feeding them, and HBM became the bottleneck. But bottlenecks don't stay put. Fix one and the pressure moves to the next-weakest link. Watching where it moves next is one of the most useful habits in this trade.

My read: the constraint is moving outward from the chip, in roughly this order.

1. Advanced packaging: bonding it all together

HBM only works when it's mounted millimeters from the GPU on a shared silicon base, using TSMC's CoWoS packaging. That step has become as hard and as scarce as making the chips themselves.

TSMC has been racing to scale CoWoS from roughly 35,000 wafers a month at the end of 2024 to around 130,000 by the end of 2026, close to a fourfold increase.
Nvidia has reportedly reserved around 60% of that capacity for 2026.
Even so, industry estimates put the supply-demand gap at about 20% in mid-2026, narrowing to around 10% by year end.
Demand is projected to roughly double again in 2027, from about 1.4 million wafers to about 2.7 million, as bigger chips with more HBM stacks need bigger packages.

Packaging is also pulling scarce materials with it, especially the ABF substrates these packages are built on. This is the Fab & Packaging layer of the Pyramid: TSMC, plus outsourced packagers like Amkor and ASE picking up overflow.

2. Optics: connecting thousands of chips

Modern models train and run across thousands of GPUs acting as one machine, so the links between them matter as much as the chips. Copper is hitting its limits on distance, speed and power, and the industry is shifting to light.

The next step is co-packaged optics (CPO): building the optical engines directly into the switch package instead of using plug-in transceivers. It cuts power and failure points.
It's early. CPO is estimated at only about 0.5% of optical modules in AI data centers in 2026, but forecasts put it near 35% by 2030.
Nvidia made $2 billion equity investments in each of Coherent and Lumentum to secure laser and optical supply, a strong signal of where it expects the constraint to be.

This is the Networking & Optics layer: Coherent and Lumentum for lasers and optics, Broadcom and Marvell for switching and CPO, Credo for the copper links that still dominate short distances.

3. Storage: memory's next tier

As AI shifts from training to inference (running models for real users), systems need to hold enormous amounts of working context. That spills beyond HBM and DRAM into flash storage.

Citi estimated that each of Nvidia's Rubin-generation systems could need over 1,000 TB of additional SSD storage for inference context, pushing AI's share of global NAND demand from under 3% in 2026 to over 9% in 2027.
Enterprise SSD contract prices were projected to rise roughly 235% cumulatively during 2026.

This is why flash names like SanDisk and Western Digital joined the Memory & Storage layer. Storage used to be the boring part of a server. For inference, it's becoming part of the memory system.

4. Power and cooling: the slowest fix

This is the hardest constraint because it moves on utility and construction timelines, not chip timelines. As the supercycle post covered, US data center power demand is expected to roughly double between 2025 and 2027, while large transformers carry multi-year lead times.

Inside the building, every new rack generation runs hotter, which is forcing the shift from air to liquid cooling. That's the Cooling & Electrical layer (Vertiv, Eaton, Modine) and the Power & Energy layer (GE Vernova, Constellation, Vistra).

The thesis

The scarcest link sets the price, and the scarce link keeps moving further from the chip. Compute was the constraint of 2023 and 2024. Memory became the constraint of 2026. Packaging and optics are tight now and scaling fast, which means their squeeze may ease first. Power is the one that can't be fixed quickly, which makes it the longest-lasting bottleneck in the stack.

For investors, the useful question isn't "who benefits from AI?" Almost everyone in the stack does when spending rises. It's "who owns the constraint, and for how long?"

What could break this

Capacity catches up. Packaging and memory are expanding aggressively. If supply catches demand in 2027, pricing power fades and these become ordinary cyclical businesses again.
Technology timelines slip. CPO is hard to manufacture at scale, and there has been public debate over whether it's running behind expectations. Delays push revenue out for the optics names.
Efficiency gains. Better models and software that do more with less hardware could reduce how much of everything is needed.
Spending slows. Every bottleneck here depends on hyperscaler capex staying high. If 2027 budgets flatten, the whole stack feels it, starting with the most expensive and most cyclical parts.

Sources

Not investment advice. I may hold positions in companies mentioned.