If you think the semiconductor rally has run its course, Nomura’s equity research team has a message for you: you’re not looking hard enough. In a detailed note published Wednesday, analysts led by David Wong laid out a case that the chip boom isn’t just alive — it’s about to get a second wind from bottlenecks the market has completely ignored.
We’re talking about the stuff that doesn’t make headlines. Not GPUs, not CPUs, not the flashy AI accelerators everyone’s been chasing. No, Nomura’s pointing at the boring, unsexy, absolutely critical components that are about to become the next great choke point in the global supply chain. And the hyperscalers — Amazon, Microsoft, Google, Meta — are going to keep throwing money at the problem, because they have no choice.
The Bottlenecks Nobody’s Talking About
Nomura’s team identifies three specific areas where supply is about to get squeezed: advanced packaging, silicon photonics, and high-bandwidth memory (HBM) interconnects. These aren’t the headline-grabbing pieces of silicon that power ChatGPT. They’re the glue that holds the entire AI infrastructure together.
“The market is focused on logic chips and foundry capacity. But the real constraints are in the layers below — the substrate, the interconnects, the packaging. That’s where we see the biggest mismatches between demand and supply over the next 18 months.” — Nomura Research Note
Take advanced packaging. TSMC’s CoWoS (Chip-on-Wafer-on-Substrate) capacity is already booked solid through 2026. Every hyperscaler wants it for their AI accelerators. But the bottleneck isn’t just TSMC’s fabs — it’s the substrate suppliers, the equipment makers, the chemical companies that provide the raw materials. One Japanese substrate maker, Ibiden, has been ramping capacity for two years and still can’t keep up. That’s not going to change overnight.
Hyperscalers: Trapped in a Spending Spiral
The real kicker is that the hyperscalers have no off-ramp. Their entire business models — cloud revenue, advertising algorithms, search, recommendation engines — now depend on AI inferencing at scale. If they stop building, they lose. So they keep spending. Nomura estimates capital expenditure from the top four hyperscalers will hit $250 billion in 2026, up 30% year-over-year. And a growing chunk of that is going to these overlooked bottlenecks.
“They can’t afford to wait,” Wong wrote. “If they don’t secure packaging capacity now, they’ll be locked out for quarters. The lead times are stretching to 12–18 months for some advanced packaging services. That’s a lifetime in AI.”
This creates a self-reinforcing cycle: more spending drives up demand for these niche components, which drives up prices, which forces more spending. It’s a beautiful thing for the companies that sit at those choke points.
The Winners Nobody’s Betting On
Nomura flags several names that could benefit from this dynamic. Among them: Amkor Technology (advanced packaging), Lumentum Holdings (silicon photonics), and SK Hynix (HBM). But the most interesting call is on a company most investors have never heard of: Disco Corporation, a Japanese toolmaker that produces dicing saws and grinders used in packaging. If packaging capacity needs to double, Disco’s tools are essential. And they’re the market leader with 70% share.
“Disco is the pick-and-shovel play nobody’s talking about,” the note says. “Every new packaging line needs their equipment. They’re basically a toll booth on the AI highway.”
Disco’s stock is already up 45% this year, but Nomura thinks it has room to run as orders accelerate. The same logic applies to other niche suppliers: BE Semiconductor (die attach), Onto Innovation (inspection), and even some chemical companies like Entegris (specialty materials).
Why This Rally Is Different
Every chip cycle has its skeptics. The usual argument is that capacity additions will eventually catch up, margins will compress, and the cycle will turn. But this time feels different. The demand isn’t coming from a single product cycle or a consumer fad. It’s coming from a structural shift in computing — the move from CPUs to accelerators, from local processing to cloud inferencing, from dumb data centers to AI factories.
And the bottlenecks aren’t going away because they’re in areas that require massive capital investment and years of lead time. You can’t just flip a switch and add CoWoS capacity. You need to build factories, qualify processes, and train engineers. That takes two to three years minimum.
Meanwhile, the hyperscalers are racing to outspend each other. Microsoft just announced a $50 billion data center expansion in Texas. Google is building a new campus in Nebraska. Amazon is reportedly in talks to buy a nuclear power plant to run its AI workloads. The arms race is real, and it’s only getting hotter.
The Risks Nobody’s Discussing
Of course, there are risks. Geopolitical tensions could disrupt supply chains — especially if Taiwan becomes a flashpoint. The US export controls on advanced chips to China are creating uncertainty for companies with exposure to both markets. And there’s always the chance that AI demand doesn’t materialize as expected, though that seems increasingly unlikely.
But Nomura’s team argues that the market has systematically underpriced these bottlenecks. “Investors are still thinking in terms of the old cycle — peaks and troughs driven by PC and smartphone demand,” Wong wrote. “This cycle is different. The drivers are secular, not cyclical. The bottlenecks are structural, not temporary.”
That’s a bold claim. But if Nomura is right, the chip rally has a lot further to run — and the biggest gains will come from the companies nobody’s watching.
One detail that sticks with me: Disco Corporation’s CEO recently said in an earnings call that their order backlog now stretches to 2028. Think about that. A company that makes saws for cutting silicon wafers is booked out for four years. That’s not a cycle. That’s a supercycle.
So forget the GPU hype for a minute. Start looking at the glue. That’s where the real money is made.
