AI Data-Center Chips Could Hit $1.2 Trillion by 2028 — Nearly 10x in Four Years

What "10x in four years" really tells you

Here's the deal: a report from the SIA with Deloitte threw out a heavy number. Annual revenue from chips in AI data centers could top $1.2 trillion by 2028 — nearly tenfold from today. Ten times in four years. That's not just "the market will grow"; it's closer to a signal that the whole industry is being pulled in one direction.

The second number is even more striking: semiconductors make up 95% of an AI data server rack's value. In old general-purpose servers, chips were one piece; in an AI rack, they're effectively almost all of the value. A single rack reportedly holds more than 4,500 packaged chips. AI infrastructure = a block of silicon, proven in numbers.

Read this as pure optimism and you miss the point. Some analysts read the $1.2 trillion as a supply-chain warning, not a market forecast — because the bigger question is whether that many chips can actually be made. Today let's unpack where the number comes from, who benefits, and what traps are hiding underneath.

What the report flags — revenue, share, and investment

First, revenue explosion. The SIA and Deloitte see AI-data-center chip revenue exceeding $1.2 trillion by 2028 — nearly 10x the last four years, and more than 50% above total global semiconductor sales across all uses in 2025. One slice — "AI data-center chips" — becomes a bigger market than the entire past semiconductor industry.

Second, 95% of rack value is chips. An AI server rack holds 4,500+ packaged chips, and 95%+ of its value is semiconductors — GPUs, HBM memory, networking chips, power semis, the full stack. Chips also account for more than 50% of total capex to build and run a data center. The AI-infra race is, at bottom, a race to secure chips.

Third, astronomical infrastructure spend. The report sees government and industry investing $4 trillion-plus in new data-center infrastructure through 2028, with up to $2.8 trillion going to semiconductors. At that scale it's not a one- or two-company story; it's national industrial policy and capital moving together.

What each side gets — chipmakers, clouds, and tools/materials

Chipmakers are the most direct winners. Nvidia (GPUs), the memory trio (HBM), TSMC/Samsung/Intel (foundry), and power and networking chip makers — with nearly all rack value coming from chips, demand pours their way. Tight-supply parts like HBM gain pricing power especially. Direct good news for Korea's memory industry.

Clouds and hyperscalers (Amazon, Google, Microsoft) face a double-edged sword. Exploding AI demand means big revenue, but also fierce chip competition and capex burden — which is exactly why they build their own chips (Trainium, TPU, Maia) to avoid sole dependence on Nvidia. The report's "chips are 50%+ of capex" lands on them as cost pressure.

Tools and materials firms are quiet winners too. Printing that many chips needs more lithography gear, packaging, materials, and fabs themselves. Equipment makers like ASML and materials/back-end firms scale with it. The three interests meet at one point — AI infra = a chip-demand explosion — and that's the core of this forecast. But not everyone gets to smile.

Echoes of the past — the boom-and-bust of silicon supercycles

Semiconductors are infamously cyclical. Demand explodes, everyone races to add capacity, supply overshoots, prices collapse — repeat. The 2021–2022 memory supercycle is the classic: COVID demand spiked prices, then expansion plus a demand slowdown snapped it back fast. Success hinged on "does demand actually persist."

What's claimed to be different this time is that demand is "structural" — not a temporary surge but AI becoming the infrastructure of every industry, laying down chip demand for the long haul. The $1.2 trillion forecast rests on that premise. If true, it differs in kind from past short supercycles; if wrong, the shadow of over-investment looms again.

That's why some read the number as a warning. Actually printing $2.8 trillion of chips needs fabs, power, materials, and people all in place. If any one becomes a bottleneck, the forecast stays just a number. Power and HBM supply are already tight, so the report is closer to an alarm asking "are we ready to make this much" than "the market is this big."

Competitor counter-play — in-house chips and supply diversification

The whales aren't sitting still. Hyperscalers grow in-house AI chips to cut Nvidia dependence — Amazon Trainium, Google TPU, Microsoft Maia. If Nvidia takes a big share of a $1.2 trillion market, customers rationally build their own card to cut that cost.

Governments are countering too. If chips are 95% of AI infra, the chip supply chain is security. Massive subsidies for domestic production in the US, EU, Japan, and Korea fit here — "can you make chips" ties directly to "can you survive the AI race." Government money is a big slice of the report's $4 trillion infrastructure outlook.

For latecomers and smaller firms, the calculus sharpens. Even with a $1.2 trillion market open, getting in needs chips, power, and capital — barriers keep rising. Conversely, selling the "picks and shovels" (tools, materials, power) opens new opportunity. Selling jeans in a gold rush is viable again.

So what actually changes

If you work in semiconductors or memory, this is a direct signal. If AI-data-center demand is structural, work and investment flow your way. Cycle risk always exists, though, so keep checking "does demand truly persist."

If you're an investor or market watcher, look at "bottlenecks" more than the $1.2 trillion. Which of power, HBM, or foundry capacity jams decides whether it's realized. Tracking the weak link in the supply chain is more useful than the headline number.

If you're a general user or company, no immediate change. But infrastructure at this scale means AI services get more ubiquitous and their cost structure shifts — and social costs like power and environmental burden ride along. Worth remembering.

One step further — the bottlenecks of an era where "chips are power"

What $1.2 trillion really points to isn't market size but the question "can we make this much?" The history of semiconductors is the history of bottlenecks. However much demand explodes, if any one of fabs (foundry capacity), power, materials, or people jams, that demand stays just a number. The two tightest spots in AI-chip demand now are HBM memory and power. HBM is effectively governed by Korea's three memory makers, but capacity takes time to add and lags demand. Power is worse — the electricity data centers draw is hitting grid limits in some regions.

Why it matters: the formula "AI infra = chips + power" creates a new power structure. Where oil-holding nations once held sway, in the AI era it's those who make chips and supply power. That's why the US, EU, Japan, and Korea pour tens of billions in subsidies into domestic production while staking everything on expanding power infrastructure. Printing $1.2 trillion of chips needs that many fabs and power plants at once, and both take years to build. So the forecast is closer to an alarm asking "are we ready to make this much" than "the market is big."

Still, the over-investment shadow can't be ignored. If everyone believes "AI demand is forever" and races to expand, at some point supply overshoots and prices collapse — the semiconductor cycle's fate. The $1.2 trillion rests on "demand persists structurally"; if AI investment fervor cools once, that premise shakes. The data centers and fabs being built astronomically now could become huge excess capacity if a demand slowdown lines up a few years out. So rather than trusting the number, it's safer to doubt both demand persistence and supply feasibility at once.

There's also how this reshapes the chip market's structure. If 95% of an AI rack is chips, chipmakers move beyond parts suppliers to become core infrastructure of the AI economy. That's why Nvidia printed a market cap rivaling big tech. At the same time, as hyperscalers cut dependence with in-house chips, the line between "chip buyers" and "chip makers" blurs. Amazon, Google, and Microsoft jumping into chip design is a survival strategy to control cost in a $1.2 trillion market.

Finally, for Korea it's a double-edged opportunity. As a memory and foundry power, an AI-chip demand explosion is a direct boon, but Korea also sits at the center of the HBM supply bottleneck, so the responsibility is large. If power infrastructure and talent training can't keep pace, a boom could turn into a burden. In the end, the report hands Korea the homework that "to seize the opportunity, you must grow supply capacity at the same time." $1.2 trillion is not a promise but a conditional possibility — and who fills the conditions first is the real contest.

🥄 Three Things You're Probably Wondering

— So what does this mean for me? Almost no direct impact. But infrastructure this large signals AI services getting more common, with chips and power at the center. Worth tracking if you follow that industry or invest.

— $1.2 trillion — will it really happen? Too early to say. It's a forecast, not a given. If any bottleneck — power, HBM, foundry capacity — jams, the number shrinks. That's why some read it as a supply-chain warning, not a market forecast.

— Does Nvidia just take all of it? No. Nvidia's share is big, but hyperscalers are diversifying with in-house chips, and demand spreads to memory, tools, and materials. It's less one winner taking all, more the whole ecosystem growing.

References

• New Report Finds Semiconductors Account for 95% of an AI Data Server Rack's Value — SIA
• SIA: AI data center chips could hit $1.2 trillion — Electronics360
• SIA-Deloitte Report Puts Chips at Center of AI Buildout — HostingJournalist
• 2026 Semiconductor Industry Outlook — Deloitte Insights
• Semiconductor Revenue from AI Could Hit $1.2 Trillion Soon — Electronics For You

Numbers and criteria are as of announcement and may change.