Beyond Nvidia: The hidden winners of the AI investment boom

Capital moves in a day. Capacity takes years. The gap between those two facts is where the money in the AI build-out is actually being made.

That gives an investor a usable test. Does a company own a constraint that capital cannot quickly relieve? A large power transformer now carries a 128-week lead time. A grid connection in the United States takes a median of more than five years. Extreme ultraviolet lithography has exactly one supplier in the world. Software capacity, by contrast, can be added in an afternoon.

Five layers of the AI chain pass that test today. They became constraints in sequence, and the profit pool has moved with them.

1. Custom silicon

The hyperscalers want alternatives to merchant GPUs, and they are designing their own. Broadcom builds them. It reported $10.8 billion of AI semiconductor revenue in the quarter ended 3 May 2026, up 143% year on year, and guided the current quarter to $16 billion, growth of more than 200%. Consolidated revenue for the quarter was $22.2 billion, which makes AI silicon close to half of everything the company now sells.


Fabrication can be bought. A custom accelerator is co-designed with a single customer, shaped around that company's models and its own data centres over several years, and the customer then writes its software to the silicon. Switching suppliers means starting again from the beginning. The relationship is the asset.

2. Memory

High-bandwidth memory stopped being a commodity and became a gating input. Micron reported revenue of $41.46 billion in the quarter ended 28 May 2026, against $9.30 billion a year earlier.

The number that matters more sits in Micron's strategic customer agreements, which carry $22 billion of total commitments, of which roughly $18 billion is cash already deposited. Customers are paying upfront, years ahead, to reserve supply.

Memory is the most cyclical corner of semiconductors, and it has always been priced accordingly. Micron trades at under 7 times forward earnings, the multiple a market assigns to profits it expects to vanish. What is changing is the contract structure underneath those profits. Volume and price are now locked for years rather than quarters, which converts a spot-price business into a contracted-revenue one. Predictable earnings have always commanded higher multiples than volatile ones. If the contracts hold, the return in memory is likely to come from the multiple rather than from the earnings.

3. Manufacturing

Money can fund a fab. Money cannot fund the two decades of process knowledge that make one work, which is why this constraint has held the longest.

ASML is the sole supplier of extreme ultraviolet lithography, the machines without which no leading-edge chip gets made. It closed 2025 with a backlog of €38.8 billion, having booked €13.2 billion in a single quarter, of which €7.4 billion was EUV. First-quarter 2026 sales were €8.8 billion at a 53% gross margin.

TSMC is the customer that matters. It turned over $35.90 billion in the March 2026 quarter, up 40.6% year on year, at a gross margin of 66.2%. Nodes at 7 nanometres and below accounted for 74% of its wafer revenue. Contract manufacturing does not earn two-thirds gross margins in a competitive market.

4. Networking

A cluster of accelerators is useless without the fabric that connects them, and this is the constraint most investors have not yet priced. Nvidia's own networking revenue reached $14.8 billion in the quarter ended April 2026, up 199% year on year, growing faster than its GPUs.

The mechanism is physical. Beyond roughly a metre at modern speeds, copper stops working, so once a cluster outgrows a single rack the network has to go optical. Arista posted revenue of $2.71 billion in the March 2026 quarter, up 35%. Astera Labs grew 93% to $308 million at a 76% gross margin. Nobody pays those margins for a part they can buy elsewhere.

5. Power and the physical layer

In PJM, the grid serving the largest concentration of American data centres, capacity for the 2026/27 delivery year cleared at $329.17 per megawatt-day, hitting the regulated price cap in every zone. The following auction cleared at the cap again, at $333.44.

GE Vernova booked $2.4 billion of electrification equipment orders serving data centres in the first quarter of 2026, more than in all of 2025. Its gas turbine backlog and slot reservations reached 100 GW, on a total backlog of $163 billion. A gas turbine ordered today, on current lead times, would not generate power until 2031.

Vertiv, which cools this equipment, grew revenue 30% to $2.65 billion in the same quarter. This is the one constraint in the chain that money cannot shorten.

The valuation test

A constraint is only worth owning if it is not already in the price. On S&P Global data as of 14 July, Broadcom trades at 24.4 times forward earnings against a PEG ratio of 0.53. Vertiv trades at 44 times, with a PEG of 1.36. Both own a real bottleneck. One of them is being handed to you cheaply.

The multiple on its own settles nothing. Run it against the growth rate, then look hard at what that growth rate assumes. A low PEG, built on estimates that quietly assume the bottleneck is permanent, measures the forecast rather than the business. Three questions do most of the work here: How long would it take to relieve this constraint? Who is spending money to relieve it? And how much of that relief is already priced in?

What can an Indian investor actually own?

The companies that own these constraints are listed where the industries were built: lithography in the Netherlands, foundry in Taiwan, memory in Korea, custom silicon and grid equipment in the United States. That is a fact of industrial history rather than a verdict on any market. It also means a portfolio held entirely at home does not reach these layers, however well it is constructed.

Until 10 July, owning SK hynix, the largest supplier of high-bandwidth memory, meant trading in Seoul. That day it listed American depositary shares on Nasdaq, raising $26.5 billion in the largest US listing by a foreign company on record. The memory layer is now reachable through the same account as everything else on this list.

Indian investors have been moving in this direction already. Equity and debt remittances under the Liberalised Remittance Scheme rose 17% year on year in April 2026, even as total outward remittances fell.

Access is the easy part now. Find the constraint. Then check whether anyone has priced it.

(Subho Moulik is Founder & CEO at Appreciate)

(Disclaimer: Recommendations, suggestions, views and opinions given by the experts are their own. These do not represent the views of The Economic Times)