In 2026, the semiconductor industry is at the intersection of the "trillion-dollar inflection point" and the "post-Moore era." Deloitte's 2026 Global Semiconductor Industry Trends Report projects that global semiconductor sales will reach USD 975 billion in 2026, with generative AI chip revenue approaching USD 500 billion, accounting for nearly half of global chip sales; however, AI chip production is less than 20 million units, representing only 0.2% of total sales volume.
A single NVIDIA B200 or Huawei Ascend 950PR can cost tens of thousands of CNY, while an ordinary MCU or analog chip might only cost a few CNY or even a few cents. This extreme divergence between volume and price has pushed the semiconductor industry beyond the old analytical framework: if revenue is increasingly determined by a very small number of AI chips and the capital expenditures of a few cloud providers, is the semiconductor industry still a classic cyclical industry? When computing infrastructure is completed, who will take over?
I. Beyond AI, Is There Growth? — There Is Growth, But It Is Merely a "Trickle-Down Effect"
To answer this question, we must first look beyond AI. Since AI accounts for half of the revenue, the other half determines whether the industry has real breadth.
Gartner data shows that global non-memory semiconductor revenue was approximately USD 589 billion in 2025 and is expected to increase to USD 717.9 billion in 2026, a year-on-year growth of about 21.9%. JPMorgan's metrics excluding memory also show a growth of about 37% in 2026. The conclusion is consistent: revenue from non-memory, non-AI core chips is indeed growing.
However, upon closer examination, the growth is highly differentiated.
Categories benefiting from the AI spillover effect are performing strongly. The CPU is a typical representative. AI inference workloads have changed the computing power ratio in data centers; in the past, AI servers had about 8 GPUs for every 1 CPU, which has now been compressed to 4:1 and is even trending towards 1:1. AI agents require continuous scheduling and task decomposition, unexpectedly activating the demand for general-purpose processors. Intel has even had to shift some consumer-grade capacity back to server CPUs. Power Management ICs (PMICs), networking chips, and optical interconnects have also surged due to data center expansion.
Traditional consumer electronics-related categories, on the other hand, are much milder. Omdia's quarterly tracking shows that in Q1 2026, semiconductor revenue excluding memory grew by only slightly over 2% quarter-on-quarter, with MCUs, discrete devices, and optoelectronics even experiencing seasonal quarter-on-quarter declines. Analog chips are expected to grow by about 10% for the full year, while sensors and optoelectronic devices will grow by only about 3%. These are the cyclically sensitive semiconductors traditionally linked to mobile phone, PC, and industrial demand.
In other words, the growth beyond AI is largely a "trickle-down effect" of AI capital expenditures, rather than a comprehensive recovery in end-market demand.
To verify this, we cannot just look at revenue; we must look at consumer semiconductors—the "thermometer" of the traditional cycle. If the end market is truly recovering, its shipment volume should rebound; if not, the shipment volume will expose the truth first.
II. Have Consumer Semiconductors Truly Recovered? — Volume Is Falling, Revenue Is "Propped Up"
Consumer semiconductors best illustrate where the industry actually stands after excluding AI.
Looking at shipment volume first, the situation is worse than expected. IDC forecasts that global smartphone shipments will reach approximately 1 billion units in 2026, a year-on-year decrease of 16.7%, marking the largest annual decline in the industry's history; Counterpoint forecasts a 14.3% decrease. The PC segment is equally dismal, with Goldman Sachs forecasting a 14% decrease and Omdia forecasting a 12% decrease.
However, consumer chip revenue has not fallen synchronously. The Consumer Integrated Circuits market is expected to reach approximately USD 184.7 billion in 2026, a year-on-year growth of about 7.6%; the Electronic Consumer Device Semiconductor market will increase from USD 106.75 billion in 2025 to USD 118.69 billion in 2026, a growth of about 11.2%.
Volume has fallen by double digits, yet revenue is still growing; the contradiction stems from prices and per-unit value. Omdia points out that although smartphone unit shipments are expected to remain relatively stable, semiconductor revenue will grow due to higher memory pricing, significantly driving up the overall bill of materials cost. Citi forecasts that the average selling prices of DRAM and flash memory will increase by 88% and 74% respectively in 2026. Mobile phone manufacturers have raised their average selling prices by 27.6%, and consumers are paying for it, ultimately increasing the revenue flowing to memory chip manufacturers.
A deeper change is the steady increase in "semiconductor content per device." AI smartphones and AI PCs are driving upgrades in SoC (System on Chip), NPU, and LPDDR specifications, with the unit price and integration level of high-end SoCs, RF front-ends, and OLED drivers steadily improving. The growth of consumer semiconductors is essentially the high-value transformation of existing devices, rather than an explosion of new devices.
The complete logical chain is:
AI infrastructure → squeezing memory capacity → surging memory prices → soaring bill of materials costs for consumer electronics → manufacturers raising prices/high-end transformation → increased per-device semiconductor value → "growth" in consumer chip revenue.
Therefore, as smartphone and PC shipments decrease, the "volume" of consumer chips is indeed shrinking; the "revenue" figure is temporarily propped up by rising memory prices and increased per-device value. This is not a spring for consumer semiconductors, but a false boom under the "blood-sucking" effect of AI infrastructure.
III. Two Worlds, How Big Is the Temperature Difference? — AI at Full Capacity Locking in Production, Consumer Passively Raising Prices
If the semiconductor industry is split into two worlds, the contrast will be clearer.
The AI infrastructure world: full capacity, customers locking in production, advance payments. Leading foundries' capacity utilization rates are near full capacity, and the memory sector's capacity utilization rate remains above 95%, with some companies' contract liabilities and new orders hitting recent highs. JPMorgan data shows that in August 2026, global semiconductor sales increased by 23% month-on-month, but the driving force comes entirely from pricing—the composite average price rose by 24% month-on-month, while shipment volume actually fell by 1% month-on-month. The price increase is not due to a comprehensive boom in demand, but because the supply side is squeezed by AI orders.
The consumer world: plunging shipments, high inventory, passive price increases. The inventory of A-share listed semiconductor companies grew by 32.24% year-on-year in Q1 2026, but this is more about manufacturers actively stocking up rather than end-market pull. The inventory and days of inventory outstanding for overseas leading CPU, analog, and power device manufacturers remained high in Q2 2026. The consumer electronics and automotive markets even saw a decline in demand in the first half of 2026, with foundry revenue growth relying on AI-related orders to "fill the gap."
This temperature difference reveals a key fact: the prosperity of AI infrastructure is "sucking the blood" of the normal cyclical rhythm of consumer semiconductors. Foundry capacity is occupied by AI orders, and general-purpose and specialized memory capacity is squeezed. SMIC (Semiconductor Manufacturing International Corporation) also stated that its logic circuit and MCU production lines are already at full capacity and cannot directly switch capacity to produce specialized memory. The supply side of consumer chips is compressed, but the demand side has not seen a strong recovery—this is a false tension driven by supply, not a real boom driven by demand.
IV. Is This Still a Semiconductor Cycle? — From the Classic Cycle to the "Capital Expenditure Super Cycle"
The driving force of the traditional semiconductor cycle is supply-demand mismatch: consumer electronics demand explosion → foundry expansion → overcapacity → price plunge → clearing → next round. Its rhythm is determined by decentralized end markets such as mobile phones, PCs, and industrial applications. The fluctuations are relatively smooth and traceable.
For the current semiconductor growth, the driving force has switched to the capital expenditure decisions of a few hyperscale cloud providers. Global AI infrastructure spending is expected to reach USD 450 billion in 2026, with inference computing power accounting for over 70% for the first time. Memory chip prices are soaring—DRAM is up 4.4 times year-on-year, and NAND is up 9 times—essentially not because everyone is replacing their phones, but because Microsoft, Google, Meta, ByteDance, etc., are rushing to buy HBM and DDR5 "at all costs."
This can no longer be described using the traditional "cycle"; a more accurate term is the "capital expenditure super cycle." Its volatility may be greater, not smaller, than the traditional cycle. The traditional cycle is smoothed by the decentralized decisions of tens of millions of consumers, whereas the AI infrastructure cycle depends on the budget sheets of fewer than ten companies. When these companies contract simultaneously, the downward slope will be much steeper than the mild recession of the consumer electronics cycle.
Omdia's analysis also points out that the 2026 semiconductor revenue growth is driven by "highly concentrated AI-related demand," rather than the "broad consumer behavior or industrial production trends" that historically influenced the market. If memory and logic ICs (Integrated Circuits) are excluded, the overall semiconductor revenue growth rate will plummet from 30.7% to only 8%. This 8% "base growth" is the real state of consumer semiconductors and traditional industrial semiconductors—mild and slow, and a significant portion of it is still "propped up" by rising memory prices.
So, is it a cycle? Yes, but not a healthy one. At the bottom of a healthy cycle, consumer demand will naturally bottom out and rebound, driving the recovery of the entire industry. The current "recovery," however, is a shot in the arm injected from the top by AI capital expenditures, with consumer semiconductors only passively benefiting from specification upgrades and capacity crowding-out effects. More dangerously, the shipment volume foundation of consumer chips is collapsing. When the memory price increase cycle ends, consumer chips will face a double blow of "falling volume and falling prices."
V. Will It Plunge Rapidly After Computing Infrastructure Is Completed? — No Crash, But a Painful Clearing
In the short term, there are no signs of a collapse in 2026-2027. The HBM capacity gap reaches 50%-60%, advanced packaging capacity is at full load, and TSMC's 2nm/3nm capacity is fully booked by NVIDIA, AMD, and Broadcom, with supply bottlenecks lasting at least until 2027. Hyperscalers have unusually high order visibility, and some have already signed long-term supply agreements. As long as the AI arms race does not see a collective brake, semiconductor revenue data will continue to look good.
The medium-term risk lies in "training peaking and whether inference can take over." The current revenue core is the arms race of training clusters—the more GPUs one has, the stronger one's model might be. Once the capabilities of large models converge and training demand shifts from "infinite expansion" to "on-demand iteration," the pace of GPU procurement will slow down significantly. At that time, what will support demand is inference, i.e., the actual usage volume of AI applications. Inference demand is more decentralized and relies more on real commercial closed loops. The shift from chatbots to autonomous agents may create sustained computing demand, but this requires time to verify.
Long-term structural risks cannot be ignored. The semiconductor industry has never had a precedent of "not falling after being built." After fiber optic infrastructure peaked in 2000, optical communication chips experienced years of capacity clearing. The logic of AI computing infrastructure is similar: when data centers, power, cooling, and chip deployment reach a stage of saturation, and AI application revenue fails to keep pace with the rhythm of capital expenditures, the "AI revenue" of the semiconductor industry will face a drastic repricing.
Deloitte also admits that the current growth "masks significant structural differences," with the industry highly dependent on data center AI chips, and this concentration itself is a risk. Once the growth rate of capital expenditures by hyperscalers slows down—it doesn't need to decline, just slow down—the price support for memory chips will loosen. And memory is precisely the largest contributor to semiconductor growth in 2026, accounting for 54% of industry revenue.
Consumer semiconductors will not "take over"; they will be "hit." Looking at current data, the annual increment of consumer semiconductors is roughly in the range of USD 10-20 billion, while the annual increment of AI-related chips, especially memory and advanced logic, is in the range of hundreds of billions of USD—memory chips alone expanded from about USD 220 billion in 2025 to USD 837 billion in 2026 in just one year. This gap in magnitude means that even if consumer semiconductors maintain a 10% growth rate, it will take many years to make up for the gap left by the ebb of AI.
More importantly, the "foundation" of consumer semiconductors themselves has already become weak. With smartphone and PC shipments falling by double digits, the "volume" of consumer chips is shrinking; the "price" support brought by rising memory prices will also disappear after the AI ebb. At that time, consumer semiconductors will face a more difficult environment: the capacity expanded by foundries during the AI boom needs to find new outlets, while consumer demand itself is insufficient to absorb this capacity.
Of course, the semiconductor industry will not "crash rapidly," but it will undergo a painful structural clearing. After the AI infrastructure is completed, the industry will not return to the past "classic cycle," but will enter a dual-track state: AI-related high-value chip categories, such as GPUs, HBM, and Advanced Packaging, will experience severe fluctuations and consolidation; traditional consumer/industrial semiconductor categories will return to their familiar, milder cyclical rhythm. The fluctuation logic of the two tracks is completely different, and analyzing them using the same "semiconductor cycle" framework is itself a kind of intellectual laziness.
The real suspense lies in whether AI inference can grow into a sufficiently large and sustainable source of revenue before training capital expenditures peak. If it can, the semiconductor industry will complete the paradigm shift from "cycle" to "growth"; if it cannot, the trillion-dollar spectacle of 2026 will become a textbook case of the next classic cycle peak.