Why AI Is Creating the Biggest Semiconductor Boom in History
A handful of AI accelerators now earn about half the industry's revenue. Here is where that money lands along the semiconductor supply chain, and why the memory cycle is the part to watch.
Summary
- WSTS expects global chip sales to jump from about $792 billion in 2025 to $1.51 trillion in 2026, and memory accounts for most of that increase.
- The value is concentrated: Deloitte estimates AI chips are about 0.2% of units but roughly half of 2026 revenue, and it pools at a few chokepoints in the supply chain.
- Our read is that the demand is real but the pricing is cyclical, so investors should separate companies paid for scarcity from companies paid for capability.
For forty years the chip business followed one rhythm. Consumers bought more phones and laptops, fabs added capacity, prices fell, and the cycle turned. A chip was a commodity input, and the winners were the ones who made it cheapest.
AI broke that rhythm. The growth now comes from data centres buying accelerators and the memory that feeds them, in quantities that are small by unit count and enormous by value. That shift explains almost every headline number in the sector this year.
The size of the jump
The World Semiconductor Trade Statistics body put 2025 sales at $791.7 billion. Its May 2026 forecast has 2026 at $1.51 trillion, growth of around 90% in a single year, with memory alone forecast above $800 billion. For an industry that took decades to reach its first half-trillion, that is an extraordinary forecast.
Two cautions belong next to it. First, it is a forecast, and WSTS revised its 2026 figure upward by more than half a trillion dollars in six months, which tells you how fast the ground is moving. Second, much of the rise is price rather than volume. TrendForce reports that conventional DRAM contract prices rose roughly 93% to 98% quarter on quarter in the first quarter of 2026, as memory makers moved wafers to AI products.
A market that doubles on price can halve on price.
Value, not volume
Deloitte's 2026 outlook puts the concentration plainly: generative AI chips will be fewer than 20 million units, about 0.2% of all chips shipped, yet close to half of industry revenue. No earlier cycle looked like this. The PC and smartphone booms spread money across thousands of parts; this one routes it through a few.
Nvidia is the clearest case. Its data centre segment earned $193.7 billion in the fiscal year to January 2026, up 68%. TSMC, which manufactures most of those accelerators, reported first quarter 2026 revenue of $35.9 billion with a net margin of 50.5%. Those are pricing-power margins, the kind a manufacturer only earns when customers have nowhere else to go.
Where the money pools in the supply chain
Follow a single AI accelerator from design to data centre and the profit collects at the stages that are hardest to copy. The table below is our map of those stages and why each one holds its price.
Where AI chip value concentrates along the supply chain
| Stage | Who controls it | Why it holds pricing power |
|---|---|---|
| Accelerator design | Nvidia, with AMD and hyperscaler in-house chips | Software ecosystem and a multi-year product lead |
| Leading-edge fabrication | TSMC | Industry estimates put its share of the most advanced chips near 90% |
| High bandwidth memory | SK hynix, Samsung, Micron | Three suppliers, hard stacking yields, capacity sold ahead |
| Advanced packaging | TSMC CoWoS, primarily | Capacity added over years, not quarters |
| Lithography tools | ASML | Sole maker of EUV machines |
Memory and packaging are the stages most readers underrate. An accelerator is useless without stacked memory beside it and a package that joins the two, and both are supply-limited in their own right. We cover those constraints in detail in the hidden bottlenecks behind AI chips.
Memory is the swing factor
If you want one line item to watch, make it memory. Logic chips such as accelerators are designed years ahead and sold on long contracts. Memory is sold closer to spot, so its price reflects this quarter's balance of supply and demand, and that balance has swung harder than anything else in the industry.
The mechanics are simple. A wafer that goes into high bandwidth memory yields far fewer bits than one making ordinary DRAM, so every shift toward AI memory tightens supply for phones, laptops and servers. Buyers of those products then pay more, and by mid-2026 TrendForce was reporting that PC and smartphone buyers were reaching their affordability limit. The same logic runs in reverse: when new memory fabs come online, prices can fall as fast as they rose.
For anyone reading a memory maker's results this year, the useful question is how much of the margin comes from price and how much from mix. Mix improvements, such as a larger share of high bandwidth memory, tend to stick. Price gains on commodity DRAM usually do not.
The buyers are changing too
The largest cloud companies are no longer content to buy every accelerator from one vendor. Google, Amazon, Microsoft and Meta all design custom AI chips, partly to control cost and partly to secure supply. That is a real threat to the current leader's share over time, but it does not change where the chips are made. Custom silicon still goes to the same leading-edge foundry, still needs stacked memory, and still competes for the same packaging lines.
So the design layer may become more contested while the manufacturing layers stay concentrated. For an investor, that argues for paying more attention to the manufacturing chokepoints than to any single design winner.
Why this cycle may not end like the others
Every previous chip boom ended in oversupply. Fabs ordered in the good years came online in the bad ones, prices collapsed, and margins went with them. The bull case says AI spending is infrastructure, like power grids or telecom networks, so demand will keep absorbing new capacity for years.
We think the truth sits between the two positions. Demand for compute looks durable, because the buyers are the largest and best-funded companies in the world and they are building for multi-year programmes. Pricing looks cyclical, because memory has always been a commodity business and capacity is already being added. The companies that earn on capability, such as design ecosystems and leading-edge process technology, should hold up better than those earning on a temporary shortage.
Separate the companies paid for scarcity from the companies paid for capability.
The risks that can break the story
- Export controls. US restrictions on advanced chips and tools to China have already removed a large market for leading suppliers, and the rules keep changing.
- Taiwan concentration. Most leading-edge output comes from one island, and new fabs in the US, Japan and Europe will add only a fraction of that capacity this decade.
- Customer concentration. A few hyperscalers buy most AI hardware. If one slows capital spending, suppliers feel it in the next quarter.
- Replacement cost. A competing leading-edge fab takes the better part of a decade and tens of billions of dollars, which protects incumbents but also slows any relief from shortages.
What this means for Indian investors and founders
India will not build accelerators or leading-edge fabs in this cycle, and pretending otherwise leads to poor capital decisions. The openings are in the layers around the chokepoints: chip design services, assembly and test, electronics manufacturing, specialty materials and the utilities a fab needs. India's semiconductor opportunity sets out which of those layers the country is already competing in.
- Investors: when a listed or private company says it is an AI beneficiary, ask whether its revenue depends on memory prices or on a capability a customer cannot buy elsewhere. Price-driven earnings deserve a cycle-adjusted multiple.
- Founders: build for the part of the chain where a shortage is structural, such as packaging, test and qualified supply to fabs, rather than the part where it is a price spike.
- Both: model a downside where memory prices fall by half. If the plan survives that, it can survive this cycle.
If you are sizing an entry into this sector, our market intelligence team builds exactly this kind of value-chain map before capital is committed, and our deep-tech practice works with hardware founders on the raise that follows.
How useful was this article?
One tap. It tells us what to write more of.
About the author
SRF Capital Studio
The next one
Get what we publish next, by email.
Working notes on raising, borrowing, protecting, growing and structuring capital in India. One email a week at most, and you can leave any time.
We use your address only to send this. See our privacy policy.
We store your address to send you these emails and nothing else. See our privacy policy.
Related reading
