The Semiconductor Cycle Split in Two. 2026 outlook Not investment advice

For decades, people talked about "the semiconductor cycle" as if it were one thing.

Demand rose, fabs filled up, prices rose, capacity expanded, inventory built, demand slowed, prices fell, and the cycle reset. That model still exists, but in 2026 it is no longer enough. AI has pulled the semiconductor market apart. Some segments are supply constrained and capacity hungry. Others are recovering unevenly or being squeezed by component inflation from the same shortages. The industry is approaching $1 trillion, but the growth is concentrated.

The semiconductor cycle did not disappear. AI split it in two.

---

Section 01 What the 2022 SemiAnalysis note was really about

The 2022 SemiAnalysis launch note for Transistor Radio is the historical hook for this essay^[1]. It announced the joint podcast with Fabricated Knowledge and described its scope: semiconductor news, market outlooks, subsegments, and opinions that did not always become full written articles. It framed semiconductors as the world's most important industry. The piece was not a forecast model. It was a signal that semiconductor outlooks had become too important and too fragmented to ignore.

In 2022, the industry needed more outlook commentary after pandemic-era shortages. In 2026, the outlook matters more because the market is no longer one cycle. The 2022 launch note implicitly bet that semiconductor markets would split into stories that needed more space than a single article. That bet was right.

---

Section 02 The headline number, approaching $1 trillion

WSTS forecasts the global semiconductor market to grow by more than 25% in 2026 and reach about $975B, with memory and logic both expected to grow more than 30% year over year^[2]. SIA reported 2025 global semiconductor sales of $791.7B, up 25.6% year over year, with logic the largest category at $301.9B and memory second at $223.1B^[3].

The headline number hides the real story.

The market is clearly growing. But the headline hides concentration. Logic and memory are exactly where AI demand hits hardest: compute and data movement. The headline is almost $1 trillion. The real story is where the growth is concentrated.

---

Section 03 AI is not buying GPUs, it is buying the whole stack

AI demand pulls on many layers at once: AI accelerators, leading-edge logic, HBM, CoWoS, EUV exposure time, advanced substrates, networking silicon, custom ASICs, datacenter CPUs, power management, memory capacity, packaging capacity, and cloud capex. AI is not one product category. It is an infrastructure buildout.

AI is not just buying GPUs. It is buying the whole stack.

---

Section 04 Nvidia is the demand signal

Nvidia is not the whole semiconductor market, but it is the clearest demand signal for the AI infrastructure side of the split. Nvidia reported FY2026 revenue of $215.9B, up 65% year over year, with data center revenue up 68% year over year, framing the growth as driven by accelerated computing and the AI platform shift^[4]. Nvidia's 10-K context adds long manufacturing lead times, premium pricing, deposits, long-term supply agreements, and capacity commitments as ways the company secures future supply^[5].

Nvidia is the visible part of the AI infrastructure supercycle. The scale shows that AI infrastructure is no longer an early-cycle experiment; it is reshaping foundry capacity, memory demand, networking, and packaging across the industry.

---

Section 05 TSMC became the foundry proof

TSMC used to be understood largely through the iPhone cycle. In 2026, it is better understood through AI/HPC. TSMC's 2025 Form 20-F reports HPC at 58% of net revenue, with smartphones at 29%, and shows 7 nm and below at 74% of wafer revenue. TSMC plans large 2026 capex, primarily focused on 2 nm and 3 nm capacity expansion alongside advanced packaging build-out^[6].

The center of gravity moved from smartphone seasonality to AI/HPC capacity. The foundry is no longer led by the same cycle that defined the 2010s.

---

Section 06 ASML shows the equipment bottleneck

ASML's Q4 2025 results reported Q4 net bookings of EUR 13.2B, EUV bookings of EUR 7.4B, and a year-end backlog above EUR 38B, with 2026 net sales guidance of EUR 34B to EUR 39B. ASML attributed stronger medium-term customer capacity plans mainly to sustainable AI-related demand^[7].

EUV is not a side variable. Advanced AI chips need leading-edge lithography. If EUV tool supply, source power, uptime, or customer capacity plans are constrained, the AI compute buildout is constrained. AI demand reaches all the way upstream into lithography.

---

Section 07 Packaging became part of the market outlook

TSMC's 2026 Technology Symposium framed the packaging-as-allocation story explicitly. TSMC is producing 5.5-reticle-size CoWoS, plans 14-reticle-size CoWoS in 2028, and that 14-reticle version is expected to integrate around 10 large compute dies and 20 HBM stacks. TSMC is also advancing SoIC and co-packaged optics^[8].

The semiconductor outlook cannot be written only as wafer supply. Packaging capacity is now part of AI compute capacity. CoWoS, HBM attach, and package-scale systems are central to the AI roadmap. AI scaling moved from the die into the package.

---

Section 08 HBM became the memory bottleneck

Micron announced that it had completed price and volume agreements for its entire calendar 2026 HBM supply, including HBM4^[9]. Memory and logic are the two categories driving the WSTS 2026 forecast, both expected to grow more than 30%^[2].

HBM is not ordinary DRAM in the AI cycle. It is a gating input for accelerators. If HBM supply is tight, accelerator shipments and system builds are constrained. Memory stopped being background capacity. HBM became strategic capacity.

---

Section 09 The traditional electronics cycle is different

Smartphones, PCs, autos, industrial, and analog are not experiencing the same cycle as AI infrastructure. Some are recovering, some are still digesting inventory, and some are being squeezed by memory price increases. IDC analysis of the global memory shortage says it is affecting both the smartphone and PC markets, with DRAM and NAND supply growth running below historical norms and 2026 PC shipments potentially declining sharply because of memory shortages and component-price pressure^[10].

The AI cycle is capacity hungry. The traditional electronics cycle is price sensitive.

AI datacenters can pay for scarce memory. PC and smartphone OEMs are much more price sensitive. That means AI can cause semiconductor revenue growth while hurting device unit growth elsewhere. AI datacenters are creating a memory supercycle. Traditional devices are paying for it.

---

Section 10 The two-cycle framework

The cleanest way to read 2026 is to separate the two cycles and look at each on its own terms.

The AI cycle is not replacing the old semiconductor cycle. It is sitting on top of it and distorting it. Both cycles can be true at once.

---

Section 11 Why the split matters for every company

The same market can look amazing or weak depending on where a company sits. Beneficiaries cluster in the AI infrastructure stack: leading-edge foundries, AI accelerator vendors, HBM suppliers, lithography and process equipment, advanced packaging, networking silicon, hyperscaler custom silicon providers, and power and thermal infrastructure. Pressured segments cluster in the traditional electronics cycle: price-sensitive PC and smartphone OEMs, low-margin consumer electronics, analog and industrial names still digesting inventory, and companies without AI/HPC exposure or dependent on older cycles.

There is no single semiconductor outlook anymore. There are positions inside the bottleneck stack.

---

Section 12 The bottleneck stack

The new semiconductor cycle is best read as a stack of bottlenecks. A shortage in any single layer can slow the entire AI buildout. That is why the market can grow rapidly and still feel constrained.

The new semiconductor cycle is a stack of bottlenecks.

---

Section 13 What people get wrong about the 2026 outlook

The weak interpretation of 2026 is that semiconductors are booming. The better interpretation is that AI infrastructure is booming, and that boom is repricing the rest of the semiconductor stack. Revenue growth can coexist with device shipment weakness. HBM shortages can help memory suppliers and hurt PC OEMs. CoWoS shortages can constrain AI accelerators even when wafers exist. EUV bookings can rise because AI customers are reserving future capacity, not because every chip segment is suddenly healthy. The headline number is less important than the constraint map.

---

Section 14 Risks and limits

The two-cycle framing relies on a mix of WSTS and SIA market data, company filings, and IDC reporting. It is worth being explicit about where the case can break.

The point is not that the whole market goes up forever. The point is that AI split the cycle and changed which bottlenecks matter.

---

Section 15 Final verdict

In 2022, the industry needed more semiconductor outlook commentary because the market was becoming too important and too complex for simple headlines. In 2026, that complexity is the story. The market is approaching $1 trillion. But it is not one smooth boom. AI infrastructure is pulling logic, HBM, EUV, CoWoS, memory, and cloud capex into a supercycle. Traditional electronics are recovering unevenly and sometimes being squeezed by the same shortages AI created. The industry is no longer one cycle. It is a stack of bottlenecks.

The semiconductor cycle did not disappear. AI split it in two.

---

Section 16 Evidence ledger and source notes