AI Can Be Right and the Semiconductor Cycle Can Still Be Wrong

Statements marked with a source number, and the shaded “What we observed” boxes, report information from the sources listed at the end, with the date it refers to. Everything else is SemiVon's analysis.

Artificial intelligence is creating one of the largest infrastructure build-outs the technology industry has seen.

GPUs, HBM, advanced packaging, networking, data centers and power infrastructure are all being pulled into the expansion of AI compute. At the same time, semiconductor market growth has reached extraordinary levels, particularly in memory and AI-related logic.

This creates a tempting conclusion:

If AI demand continues growing, semiconductor growth should continue with it.

But the semiconductor industry has never operated in such a straight line.

Technology adoption may be structural. Semiconductor supply, pricing, inventory and capacity remain cyclical.

That distinction is becoming increasingly important.

The question is therefore no longer simply whether AI is a bubble.

A more useful question is:

What happens when a genuine technological revolution collides with the economics of a semiconductor cycle?

Semiconductor Growth Is Extraordinary — But the Composition Matters

But headline revenue growth alone does not tell us what kind of growth is taking place.

The distinction between units and ASPs — average selling prices — matters enormously.

This means the current semiconductor boom is real.

But it also means that:

Revenue growth is substantially stronger than underlying unit growth.

That distinction should matter to semiconductor manufacturers, distributors, OEMs, investors and policymakers.

Because unit demand and pricing power do not necessarily move together indefinitely.

Four Signals Matter More Than the Headline Growth Rate

At this stage of the semiconductor cycle, we believe four variables deserve particular attention:

ASP → Momentum → Inventory → Capacity

They describe different parts of the same mechanism.

Strong demand creates shortages.
Shortages support higher ASPs.
Higher prices and margins encourage inventory accumulation and capital expenditure.
New capacity eventually arrives.
The original shortage begins to ease.
Pricing power weakens.
Inventory becomes more difficult to clear.
The cycle turns.

None of this means that the underlying technology trend has disappeared.

It means that:

A technology can be structurally right while its semiconductor supply chain becomes cyclically wrong.

That distinction may become one of the defining questions of the AI semiconductor market.

ASP May Be the First Signal to Watch

Average selling prices deserve particular attention because the current semiconductor revenue boom has been unusually dependent on pricing.

Two months do not establish a definitive trend.

But when revenue growth is heavily amplified by ASP, the direction of ASP becomes increasingly important.

If prices begin normalising while capacity and inventory continue rising, revenue growth can decelerate much faster than end-market technology adoption.

This is why:

AI adoption and semiconductor pricing should not be treated as the same variable.

AI usage could continue expanding while individual semiconductor categories experience falling ASPs.

Memory Is Where the Tension Is Most Visible

Memory currently provides perhaps the clearest example of the difference between structural AI demand and semiconductor-cycle economics.

AI accelerators require enormous memory bandwidth.

That has made HBM strategically important and has redirected manufacturing resources toward higher-value AI memory.

But memory is not one homogeneous market.

HBM, server DRAM, PC DRAM, mobile DRAM and NAND face different supply constraints, manufacturing economics and end-market demand.

This creates the possibility of an unusual divergence:

HBM can remain constrained while conventional DRAM begins to weaken.

That is not contradictory.

Additional DRAM wafer capacity can relieve conventional memory supply before advanced packaging and HBM production constraints disappear.

The bottleneck simply moves.

For semiconductor buyers, this means that asking whether “memory is tight” is increasingly insufficient.

The better question is:

Which memory is constrained, by which manufacturing step, for how long, and what happens when that constraint moves?

AI Infrastructure and Semiconductor Cycles Follow Different Economic Logics

The semiconductor cycle traditionally follows a recognizable sequence:

Scarcity → Higher ASP → Higher Margins → CapEx → Capacity → Oversupply → Correction

AI infrastructure introduces another mechanism.

As computing capacity expands, the cost of intelligence can decline. Lower costs can make new applications economically viable. Those applications can generate additional demand for compute.

The infrastructure sequence can therefore look more like:

Capacity → Lower Cost of Intelligence → New Applications → Induced Demand → Higher Utilisation → Further Infrastructure

These two mechanisms can operate simultaneously.

This is why apparently contradictory developments may coexist.

• HBM can remain tight while conventional DRAM prices weaken.
• Leading-edge AI compute can remain structurally important while mature-node semiconductor capacity becomes excessive.
• AI infrastructure investment can continue expanding while semiconductor ASPs correct.
• A semiconductor company can remain strategically important while its margins normalize.
• And a structurally attractive company can still become a difficult investment if expectations, valuation or capacity expansion move too far ahead of economic reality.

The Semiconductor Industry Is Becoming Two Markets at Once

One of the most important developments in 2026 is the increasing separation between the AI semiconductor economy and the broader semiconductor economy.

The AI chain includes:

Advanced Logic → HBM → Advanced Packaging → Networking → Data Center Infrastructure

Meanwhile, the broader semiconductor economy continues to depend on:

Automotive → Industrial → PC → Smartphone → Consumer Electronics → Analog → MCU → Discrete & Power

These markets do not necessarily occupy the same point in the cycle.

AI-related semiconductor demand may be approaching cyclical extremes in certain categories while parts of automotive, industrial and traditional electronics are only beginning to recover from previous weakness.

The semiconductor market therefore should not be treated as one synchronized cycle.

Increasingly, we are watching multiple overlapping cycles with different constraints, pricing structures and capital requirements.

China Adds Another Layer to the Capacity Equation

The geographical composition of semiconductor investment also matters.

China has invested heavily in semiconductor manufacturing capacity, particularly across mature and increasingly advanced process technologies.

For global semiconductor buyers, this creates both opportunity and risk.

Additional capacity can improve availability, diversify sourcing and reduce dependence on individual suppliers.

But if capacity expands faster than underlying demand, pricing pressure can emerge across mature-node products including analog, MCUs, discrete devices, power semiconductors and mainstream logic.

At the same time, leading-edge AI compute, HBM and advanced packaging may remain structurally constrained.

The result could be another important divergence:

Overcapacity in one part of the semiconductor industry can coexist with scarcity in another.

For procurement teams and distributors, broad statements such as “semiconductors are in shortage” or “semiconductors are oversupplied” are therefore becoming less useful.

The real question is increasingly:

Where is the constraint moving?

What This Means for Semiconductor Buyers

For procurement and supply-chain teams, the current environment requires separating structural scarcity from cyclical scarcity.

A component experiencing allocation today is not necessarily a component that will remain scarce when new capacity becomes available.

Likewise, a falling spot price does not automatically mean that long-term demand has weakened.

The practical challenge is to distinguish four situations:

For distributors and OEMs, this distinction should influence purchasing commitments, inventory duration, alternative-source qualification and customer pricing.

Allocation is not the same as permanent scarcity.

What This Means for Semiconductor Manufacturers and Investors

The same distinction applies to capital.

High ASPs can generate exceptional cash flow and make capacity expansion appear economically obvious.

But semiconductor fabrication capacity is difficult to reverse.

Once fabs, equipment and supporting infrastructure are built, depreciation continues regardless of where ASPs move.

The important question is therefore not simply:

Where is demand growing fastest today?

It is:

Which capacity will still generate attractive returns after supply normalises?

This is particularly important in AI because structural technology adoption and cyclical semiconductor economics can point in different directions at the same time.

A correct long-term technology thesis does not automatically guarantee a correct entry price, capacity decision or inventory position.

What This Means for Governments

The same problem exists at an even larger scale for industrial policy.

Semiconductors have become strategic infrastructure.

Governments increasingly view domestic semiconductor capability through the lenses of national security, technological sovereignty and supply-chain resilience.

But strategic capacity and economic capacity are not always identical.

A government may rationally accept lower financial returns in exchange for supply security.

At the same time, subsidized capacity can create structural oversupply if multiple countries pursue similar strategies simultaneously.

The policy question therefore becomes:

Will today's semiconductor investment create strategic capability — or tomorrow's structural overcapacity?

Both outcomes are possible.

Structural Certainty Is Rising While Cyclical Certainty Is Falling

This may be the most important contradiction in the semiconductor market today.

The long-term case for AI remains powerful.

Compute demand is expanding.
AI models are becoming more capable.
Infrastructure is being built.
Advanced packaging, memory bandwidth, networking and power are becoming increasingly strategic.

But semiconductor-cycle uncertainty is also rising.

ASP momentum is changing.
Capacity is expanding.
Inventory behavior is evolving.
And enormous amounts of capital are being committed on the assumption that current demand patterns will persist.

The two conclusions are not mutually exclusive.

The structural certainty of AI can continue rising while the cyclical certainty of semiconductors declines.

AI can change the global economy.
And semiconductors can still experience another downcycle.
Both can be true.

The Questions We Are Watching

At SemiVon, we believe the next phase of the semiconductor market will be better understood through questions rather than headline forecasts:

• Which constraints are structural, and which are temporary?
• Where will scarcity migrate as new capacity comes online?
• Which semiconductor investments remain economically attractive after ASPs normalize?
• Which supply chains remain constrained even when wafer capacity increases?
• Which capacity will still matter five or ten years from now?

And perhaps most importantly:

What decisions remain robust if both the AI revolution and the next semiconductor downcycle happen at the same time?

Those questions increasingly matter not only to semiconductor manufacturers.

They matter to OEMs, procurement teams, industrial companies, investors and governments making decisions that may be difficult to reverse.

Related services

Component Sourcing → BOM Support →

Sources

1. Malcolm Penn, Founder, Chairman & CEO, Future Horizons Ltd, “September 2026 Semiconductor Monthly Report,” September 2026.

For the full report and analysis, please contact Malcolm Penn: mpenn@futurehorizons.com

This article is market analysis for general information. It is not a quotation or a commitment to supply, and it is not investment advice.

← All insights