The Semiconductor Supercycle Is One Product Category Wearing a Market Costume

The Number Everyone Is Misreading

Global semiconductor sales hit $110.5 billion in April 2026, up 11% from March and 93.9% above April 2025's $56.9 billion. The WSTS Spring 2026 forecast endorsed by SIA projects annual global sales will reach $1.5 trillion in 2026, with 2027 projected to exceed $1.9 trillion. Those are genuine records. They're also misleading without context.

The 93.9% year-over-year growth implies broad strength across consumer electronics, automotive, industrial, and communications silicon. That's not what's happening.

The Concentration Problem

High-value AI chips now drive roughly half of total revenue while representing less than 0.2% of total unit volume. As AI chips boom, chips for automotive, computers, smartphones, and non-datacenter communications are seeing relatively slower growth.

Half the revenue. Two tenths of one percent of units. This isn't a market recovery—it's a pricing event concentrated in advanced-node logic and memory deployed exclusively in AI infrastructure. The average selling price gap between a high-end AI accelerator and commodity chips tells the entire story.

The SIA-Deloitte report released June 1 extends this picture forward. The report projects annual revenue for semiconductors used in AI data centers could reach $1.2 trillion by 2028—a nearly tenfold increase over four years that would surpass total global semiconductor sales from 2025 across all end uses by more than 50%. For perspective, global semiconductor sales hit $791.7 billion in 2025, an increase of 25.6% compared to 2024's $630.5 billion. The $1.2 trillion AI datacenter figure alone would dwarf that baseline.

What's Actually Driving This

LLM training runs and inference serving require dense compute clusters built around high-bandwidth memory and specialized accelerators. A single AI server rack contains more than 4,500 packaged chips. Chips account for more than 95% of a leading AI server rack's content value and more than 50% of the total capital expenditures required for building and operating an AI data center.

That density, combined with aggressive pricing power from dominant accelerator vendors and sole-source access to advanced foundry capacity, means a small number of racks generates enormous revenue. Unit counts stay low because each unit costs tens of thousands of dollars and supply is deliberately constrained by fab capacity at cutting-edge nodes.

To meet global demand for new AI applications, government and industry will invest over $4 trillion in new data center infrastructure through 2028, of which up to $2.8 trillion will be spent on semiconductors. This spending isn't dispersed. It flows through a narrow channel: TSMC's N3 and N5 nodes, NVIDIA's H-series and Blackwell architectures, and the HBM supply chain anchored at Samsung and SK Hynix.

A leading AI chip manufacturer has secured about 800,000 wafers for its main chip in 2026 and produces about 20 chips per wafer, suggesting approximately 16 million chips in total. Sixteen million units generating revenues that constitute a meaningful fraction of global semiconductor sales. That's extreme ASP concentration in action.

Who Gets Starved

The inverse of datacenter concentration is allocation starvation everywhere else. Fab capacity at advanced nodes can't flex dramatically within 12 months. When TSMC's N3 and N5 lines run at full AI accelerator production, wafer starts for everything else fight for scraps. Automotive, industrial, and consumer programs needing trailing-edge nodes face a different squeeze: capital, engineering talent, and equipment procurement migrate upstream.

The SiC market is experiencing an overcapacity downturn, with utilization rates dropping to around 50% for upstream processes and 70% for device lines as of 2025. This correction, driven by the 2019-2024 capex boom and a slowdown in automotive EV demand, is expected to persist until 2027-2028. That's the other side of the barbell: while AI chip fabs run flat-out, power semiconductor fabs built for automotive electrification sit half-empty.

Analog and legacy-node vendors face a worse problem. They can't reprice into the AI tailwind because their chips aren't in that supply chain. Yet the "supercycle" narrative sets customer and investor expectations that bear no relation to actual demand for their silicon.

The Projection Caveat Operators Need

The $1.2 trillion AI chip revenue figure by 2028 isn't a floor—it's a straight-line projection of current hyperscaler capex. 2027 and 2028 could diverge sharply from current expectations. If monetization of AI takes longer or generates lower returns than anticipated, data center projects could be canceled or postponed, with adverse impacts on chip sales.

The miss scenario isn't implausible: macro contraction, a plateau in foundation model scaling, or faster-than-expected shifts from centralized training to distributed edge inference would all compress ASPs and reduce datacenter chip demand. None of that would help the commodity chip market—it would just remove the single prop holding up the headline number.

Consider this: if hyperscalers shift capex from large training clusters toward edge inference at lower ASPs, the revenue curve bends well before unit volume changes. That's the risk in a market where 0.2% of units carry 50% of revenue.

What to Watch

1. TSMC N3/N5 utilization disclosures in quarterly earnings through Q3 2026. Softening fill rates on leading nodes signal peaking datacenter demand or forward-pulled allocation.

2. Whether AI datacenter chips maintain 50% revenue share into H1 2027 or normalize as legacy-node recovery broadens. SIA monthly data is the clearest signal.

3. Inference chip ASP trajectories. Monitor NVIDIA, AMD, and custom silicon from Google, Amazon, and Microsoft for shifts favoring inference efficiency over training performance at lower price points.

4. New foundry capacity targeting AI chips outside TSMC. Intel Foundry, Samsung yield improvements, and any challengers to the incumbent duopoly represent the structural relief valve.

5. SiC and analog vendor guidance revisions in Q2 2026 earnings. If those vendors lower outlooks while headline semiconductor numbers hold steady, the divergence isn't just real—it's accelerating.