Global Semiconductor Industry Structure and the AI Cycle: 2022–2026

The two cyclesOverview

The value chain consists of the following:

The major product differentiator before the AI boom was generally the ability to manufacture chips at increasingly smaller process nodes. As a result, competition was heavily focused on process technologies such as 7nm, 5nm, and 3nm.

The competitive structure has become more complicated with AI. Performance now depends more heavily on the combined system of:

This means that stronger demand for NVIDIA and other AI accelerators can sequentially influence demand for advanced foundry capacity, HBM, advanced packaging, networking equipment, semiconductor manufacturing equipment, and eventually the physical infrastructure required to operate AI data centers.

However, the quantitative analysis shows that these effects do not occur simultaneously. Different stages of the semiconductor value chain respond at different speeds.

The two cyclesTwo Semiconductor Cycles

To test the idea that AI-related and traditional semiconductor companies are currently operating under different cycles, semiconductor companies were separated into two groups.

The AI-exposed basket consists of NVIDIA, Broadcom, TSMC, SK hynix, Micron, ASML, Applied Materials, and Lam Research.

The traditional semiconductor basket consists of Qualcomm, NXP, STMicroelectronics, onsemi, Infineon, and MediaTek.

Median values are used for the main comparison to reduce the effect of unusually high or low results from individual companies.

Revenue Growth

The divergence did not immediately appear when ChatGPT was released.

During the first half of 2023, the median AI-exposed company was still experiencing declining revenue. Median YoY revenue growth was approximately -4.5% in Q1 2023 and -13.5% in Q2 2023.

This is important because it shows that the initial AI boom was still concentrated in a relatively small number of companies, particularly NVIDIA, while memory and semiconductor manufacturing companies were still affected by the traditional semiconductor downturn.

The change became more visible from late 2023.

Median AI-basket revenue growth increased to approximately:

• Q4 2023+9.9%
• Q1 2024+23.4%
• Q2 2024+38.0%
• Q4 2024+44.0%
• Q2 2026+40.4%

At the same time, the traditional semiconductor basket remained weak through most of 2024. Median revenue growth was approximately -7.8% in Q2 2024 and -8.7% in Q4 2024.

The traditional cycle eventually began recovering during 2025–2026, reaching median revenue growth of approximately +12.4% by Q2 2026.

Therefore, the data supports the existence of two semiconductor cycles. However, the important timing difference is that the divergence became broad and persistent during late 2023 and 2024, rather than immediately after the beginning of the Generative AI boom.

The two cyclesProfitability Shows an Even Larger Structural Difference

Gross margins provide another way to observe the separation.

At the beginning of 2022, the two groups were economically very similar. Median gross margin was approximately:

• AI basket48.1%
• Traditional basket48.2%

By Q2 2026, this had changed significantly:

• AI basket67.0%
• Traditional basket41.3%

The difference therefore expanded from almost zero to approximately 25.6 percentage points.

The change was gradual.

The AI basket’s median margin remained around the high-40% range through much of 2022–2024 before increasing significantly during 2025 and 2026. In contrast, the traditional semiconductor group’s margin gradually declined from the high-40% range toward the low-40% range.

This suggests that the distinction between the two cycles is not limited to revenue growth.

AI exposure is increasingly affecting:

pricing power, product mix, utilization, and profitability.

The result is particularly important because it remains visible even when median values are used, meaning the difference is not simply the result of NVIDIA’s unusually high margins.

Compute layerFabless: The Starting Point of the AI Boom

Major fabless semiconductor companies include:

One important change since 2023 is the increasing role of custom ASICs.

During the initial stage of the AI boom, AI computing was heavily concentrated around NVIDIA GPUs. However, hyperscalers have increasingly developed their own processors, including Google TPU, Amazon Trainium, Meta MTIA, and Microsoft Maia.

Therefore, the likely future market structure is not necessarily NVIDIA being replaced. Instead, AI compute is developing toward a combination of:

The financial data shows how unusually large the initial NVIDIA acceleration was.

Using Q1 2022 as an index of 100, NVIDIA Data Center revenue increased to approximately:

• Q3 2023317
• Q2 2024693
• Q2 20251,199
• Q2 20262,307

In dollar terms, the quarterly series increased from approximately $3.3 billion in Q1 2022 to $75.2 billion by Q2 2026 in the aligned dataset.

This increase is substantially larger than the growth observed at later stages of the semiconductor supply chain.

The result supports the argument that the AI semiconductor cycle first became visible at the compute layer, before becoming equally visible across the broader industry.

Foundry layerFoundry: One of the Most Significant Chokepoints

Fabless companies can design chips, but they still require foundries to physically manufacture them.

The major global foundry companies include:

AI accelerators place particularly high demands on foundries because customers require more than general wafer production.

They increasingly require:

• advanced nodes such as 3nm and eventually 2nm
• high manufacturing yields
• large semiconductor dies
• HBM integration
• advanced packaging capability

TSMC currently holds the strongest position in providing these capabilities together at large scale.

Quantitative Evidence from TSMC HPC

To estimate the expansion of high-performance-computing manufacturing activity, TSMC’s total quarterly revenue was multiplied by its reported HPC revenue share.

This creates a TSMC HPC Revenue Proxy.

The proxy should not be interpreted as pure AI revenue because HPC also includes non-AI high-performance computing products. However, it provides a useful measure of the manufacturing layer most closely exposed to advanced AI computing.

The proxy increased from approximately $7.2 billion in Q1 2022 to $26.5 billion in Q2 2026.

Using Q1 2022 = 100, the index reached approximately 369 by Q2 2026.

The most interesting result is the lead-lag relationship with NVIDIA.

Using quarterly YoY growth rates, the strongest positive correlation between NVIDIA Data Center revenue growth and the TSMC HPC proxy occurs when TSMC is lagged by approximately three quarters:

Correlation ≈ 0.78 at +3 quarters

This does not establish causality. However, it is consistent with the idea that a rapid increase in AI-compute demand is followed several quarters later by broader growth in advanced foundry activity.

Memory layerMemory: AI Demand Overlaid on a Traditional Commodity Cycle

Memory represents one of the most important examples of why the AI semiconductor cycle cannot be analyzed independently from the traditional semiconductor cycle.

AI accelerators require large quantities of HBM because traditional DRAM cannot provide sufficient bandwidth for high-performance AI workloads.

The major suppliers are:

However, the financial data shows that memory did not immediately move upward with NVIDIA.

The combined SK hynix + Micron revenue index, with Q1 2022 = 100, fell to approximately 43 by Q1 2023.

This reflects the severe memory downturn occurring at the same time that AI accelerator demand was beginning to accelerate.

The memory proxy then recovered:

• Q2 2024105
• Q2 2025141
• Q2 2026528

This is one of the strongest examples of the two-cycle structure.

AI demand was increasing while the traditional memory cycle was initially collapsing. Only later, as HBM demand, DRAM pricing, and tighter supply conditions combined, did the financial effect become visible across the memory companies.

Unlike the NVIDIA-to-TSMC relationship, the analysis does not show a strong simple lead-lag correlation between NVIDIA revenue growth and total memory revenue growth.

This is reasonable because total memory revenue is influenced by multiple factors:

• HBM demand
• DRAM and NAND pricing
• inventory corrections
• capacity reductions
• PC and smartphone demand
• product mix

Therefore, NVIDIA revenue alone is not sufficient to explain the memory cycle.

This means HBM-specific disclosures from SK hynix, Micron, and Samsung should be used together with the aggregate memory financial series rather than attempting to treat total memory revenue as a direct function of GPU sales.

Packaging layerAdvanced Packaging

Connecting GPUs and multiple HBM stacks requires technologies including:

• 2.5D packaging
• CoWoS
• silicon interposers
• chiplets
• 3D stacking

The major players include TSMC, ASE, Amkor, and Intel.

The broad ASE + Amkor revenue proxy does not show the same explosive growth observed in GPU, HBM, or TSMC HPC revenue.

This should not be interpreted as evidence that advanced packaging demand is weak.

ASE and Amkor generate substantial revenue from non-AI packaging activities, while an important portion of AI advanced packaging is performed directly by TSMC through technologies such as CoWoS.

Therefore, total OSAT company revenue is too broad to function as a clean AI advanced-packaging indicator.

For this section, company disclosures on CoWoS capacity, 2.5D/3D capacity, and AI/HPC packaging should therefore remain more important than the broad revenue proxy.

Equipment layerSemiconductor Equipment

The major semiconductor-equipment companies include:

The equipment layer grows more slowly than the compute layer because equipment purchases occur after semiconductor manufacturers decide to expand or upgrade manufacturing capacity.

The combined equipment revenue index increased from:

• Q1 2022100
• Q2 2024109
• Q2 2025131
• Q2 2026157

This is much slower than NVIDIA Data Center revenue, but the lead-lag analysis provides important evidence that equipment demand follows manufacturing activity.

The strongest positive correlations are:

• TSMC HPC growth → Equipment growth+1 quarter, r ≈ 0.89
• Memory growth → Equipment growth+2 quarters, r ≈ 0.92

The memory relationship remains similarly strong at +3 quarters.

In addition, NVIDIA Data Center growth shows its strongest positive relationship with equipment revenue approximately four quarters later, with a correlation of approximately 0.71.

These results are consistent with the following transmission mechanism:

The lag becomes longer as the analysis moves further upstream in the semiconductor manufacturing process.

TransmissionQuantitative View of the AI Transmission Cycle

Using Q1 2022 = 100 allows the major layers to be compared despite very different absolute revenue levels.

By Q2 2026:

The result shows a clear difference in the magnitude and timing of growth.

NVIDIA’s Data Center business expanded much faster than every other stage.

TSMC HPC activity followed with substantial but smaller growth.

Memory initially declined sharply, then accelerated dramatically later.

Equipment increased last and at a slower rate.

Together, these figures provide stronger support for a sequential transmission model than for the idea that AI immediately lifted the entire semiconductor industry.

TransmissionRevising the Lead-Lag Thesis

The original interpretation suggested that AI demand might move through the semiconductor value chain with a lag of approximately one to three years.

The quarterly data suggests the financial transmission is generally faster than that, at least for the period currently available.

The strongest positive relationships observed are:

Therefore, the observable financial transmission appears to operate more closely on a one-quarter to one-year scale, rather than one-to-three years.

However, the sample is short. Depending on the lag, there are only approximately 10–14 paired quarterly observations.

The correlations should therefore be interpreted as descriptive evidence rather than causal estimates.

Demand layerHyperscaler CAPEX: An Important but Different Relationship

One result does not support the initial hypothesis.

Aggregate hyperscaler CAPEX does not positively lead NVIDIA Data Center revenue growth in the available dataset.

Across the tested 0–4-quarter lags, the correlations are negative.

This does not mean hyperscaler spending is unrelated to NVIDIA demand.

Instead, it likely reflects differences in timing and measurement.

Semiconductor orders, GPU deliveries, construction spending, networking investment, buildings, power infrastructure, and accounting recognition of CAPEX do not occur at the same time.

In fact, the financial series is more consistent with compute demand accelerating before the full infrastructure CAPEX response becomes visible.

Hyperscaler CAPEX itself still increased significantly, from approximately $71.3 billion in Q1 2022 to $234.7 billion in Q2 2026, or approximately 3.3 times the Q1 2022 level.

Therefore, hyperscaler CAPEX remains strong evidence that AI has become a major capital-investment cycle, but it should not be presented as a clean short-term leading indicator for NVIDIA revenue.

Year by year2022: The Traditional Semiconductor Cycle Turned Down

The macroeconomic environment in 2022 was relatively unfavorable for semiconductors.

Inflation increased, central banks raised interest rates rapidly, and demand for PCs, smartphones, and consumer electronics weakened after strong purchases during COVID.

Inventory corrections began across the semiconductor industry.

At approximately the same time, Generative AI began emerging as a new source of semiconductor demand.

Therefore, two different forces began moving in opposite directions:

Year by year2023: AI Growth While Much of the Industry Was Still Weak

The quantitative data shows this divergence clearly.

NVIDIA Data Center revenue accelerated significantly during 2023, particularly from the second half of the year.

However, the median AI-exposed semiconductor company was still experiencing negative revenue growth during Q1 and Q2 2023 because memory and manufacturing-related companies remained in a downturn.

The memory proxy reached its lowest level around this period.

Therefore:

AI Boom ≠ Immediate Semiconductor-Wide Boom

The initial AI boom was concentrated primarily in compute.

Year by year2024: The AI Cycle Broadened

During 2024, the financial effect became visible across more companies.

Median revenue growth for the AI basket increased from approximately 23% in Q1 to 44% by Q4.

TSMC’s HPC proxy accelerated, memory recovered from its trough, and semiconductor equipment revenue began improving later in the year.

Meanwhile, the traditional semiconductor basket remained in negative median revenue growth during all four quarters of 2024.

This is the period where the two-cycle structure became most clearly visible.

Year by year2025–2026: The AI Cycle Became Broad-Based

By 2025–2026, the AI cycle was no longer concentrated primarily around NVIDIA.

Foundry, memory, equipment, networking, and infrastructure investment were increasingly participating.

By Q2 2026:

• Median AI-basket revenue growth40.4%
• Median traditional-basket revenue growth12.4%

and:

• Median AI-basket gross margin67.0%
• Median traditional-basket gross margin41.3%

The traditional semiconductor industry was also recovering, meaning the market was no longer simply an AI boom occurring against a universal semiconductor recession.

Instead, both cycles were moving upward, but with very different growth rates and economics.

ConclusionsCurrent Interpretation of the Semiconductor Cycle

The evidence supports the following sequence:

This sequence is more accurate than treating the semiconductor industry as one synchronized cycle.

There are also important exceptions.

Memory follows its own pricing and inventory cycle, while hyperscaler CAPEX accounting does not cleanly lead GPU revenue.

Therefore, the AI semiconductor cycle should be understood as an overlay on top of existing semiconductor cycles, rather than a complete replacement of them.

ConclusionsCurrent Major Chokepoints

The major chokepoints can be summarized as:

The chokepoint has gradually moved through the value chain.

In the early stage, the most visible limitation was AI compute availability.

As production expanded, constraints moved toward HBM, advanced packaging, and advanced-node capacity.

As semiconductor supply continues expanding, physical data-center infrastructure and electricity may become increasingly important limitations.

ConclusionsFramework for Analyzing the Industry

Instead of simply asking whether the semiconductor industry is doing well, each layer should be observed independently.

The major structural conclusion can therefore be summarized as follows:

The AI semiconductor cycle began as a highly concentrated compute boom. From late 2023 and through 2024, the financial effect spread into advanced foundry activity. Memory recovered later as AI-related HBM demand interacted with the traditional memory cycle. Semiconductor equipment followed manufacturing activity with an additional lag. By 2025–2026, this process had developed into a broader semiconductor and infrastructure investment cycle, while the traditional electronics semiconductor cycle recovered at a significantly slower pace.

The important point is therefore not simply that AI is causing semiconductor growth.

The more useful conclusion is that AI demand is being transmitted through different semiconductor layers at different speeds, while each layer continues to be influenced by its own traditional industry cycle.