The End of the Semiconductor Supercycle? Why Sold-Out HBM May Signal Peak Expectations
Why the 2027 “Sold-Out” Narrative May Mark Peak Expectations, Not Peak Opportunity
Executive Summary
The dominant narrative surrounding semiconductors today is straightforward: AI demand remains insatiable, high-bandwidth memory is effectively sold out for years, and the hardware supercycle still has ample room to run.
Operationally, much of that may be true.
Yet markets do not reward fundamentals in isolation. They reward the gap between reality and expectation. Once an industry’s growth trajectory becomes universally visible, the relevant question is no longer whether business conditions are strong. It is whether anything remains capable of surprising investors to the upside.
That is the increasingly difficult hurdle facing the semiconductor complex today.
Our view is not that AI demand is collapsing, nor that the semiconductor industry is entering imminent structural decline. Rather, the more nuanced risk is that the industry may remain fundamentally healthy while simultaneously entering a period in which equity returns deteriorate.
The reason is simple:
Future perfection may already be priced as present certainty.
Semiconductor supercycles rarely end because demand vanishes overnight. More often, they end when expectations outrun the economic reality of sustaining that demand.
The “Sold-Out Through 2027” Narrative May Reflect Consensus Saturation, Not Incremental Upside
Few data points have been cited more bullishly in recent months than the assertion that HBM supply is effectively committed through 2027.
Investors interpret this as evidence that the AI buildout remains in its early innings and that memory suppliers enjoy years of protected visibility.
That interpretation is not entirely wrong.
Tight supply and long-duration backlog unquestionably indicate robust demand.
But in financial markets, visibility itself has diminishing marginal value.
Future earnings only create upside when they exceed what investors have already discounted. Once an industry’s multi-year growth trajectory becomes broadly accepted and heavily modeled into consensus forecasts, visibility ceases to be a catalyst and instead becomes the baseline assumption against which all future performance is judged.
A fully visible backlog does not create upside on its own.
It merely raises the burden of proof for additional re-rating.
 |
| Figure 1. Conceptual framework illustrating divergence between base-case AI infrastructure demand growth and valuation-implied market expectations. The widening gap represents increasing execution risk required to justify prevailing semiconductor valuations. |
The AI Supply Chain Is Becoming Increasingly Vulnerable to Forecast Error
One of the least appreciated risks in the current AI infrastructure boom is how far procurement behavior has shifted away from real-time consumption and toward long-duration forecasting.
As supply constraints intensify, hyperscalers and enterprise buyers are no longer ordering against immediate deployment needs alone. They are reserving capacity years in advance through strategic agreements designed to secure future supply before competitors can do the same.
That dynamic creates the appearance of exceptional visibility.
But it also introduces a familiar late-cycle fragility: forecast risk.
The further procurement moves ahead of actual deployment, the greater the probability that future demand assumptions diverge from realized consumption.
Even modest forecasting errors can become magnified when billions of dollars of infrastructure are committed months or years before utilization is tested in production environments.
This is how hardware cycles often become unstable—not because demand disappears, but because the ordering system begins reflecting expected future demand rather than present economic reality.
The Real Constraint Is No Longer Demand — It Is the Economics of Delivering That Demand
The market’s current framework still assumes semiconductor suppliers can convert AI demand into profit with relative mechanical simplicity.
Yet semiconductor manufacturing has never functioned so linearly.
What the market increasingly underestimates is that the cost and complexity of serving AI demand are rising almost as quickly as the demand itself.
High-bandwidth memory illustrates this dynamic clearly.
Each successive generation of HBM is not merely a faster or denser version of the last. It represents a materially more difficult manufacturing challenge.
As stack counts rise and bandwidth requirements intensify, suppliers must contend with increasingly severe thermal constraints, tighter signal-integrity tolerances, more fragile packaging yields, and greater interdependence between memory, substrate, and packaging ecosystems.
The engineering burden is not rising incrementally.
 |
| Figure 2. Conceptual illustration of rising engineering, packaging, and thermal-management burden across successive HBM generations. As memory architecture scales, manufacturing complexity may increase in increasingly non-linear fashion, pressuring long-term returns on incremental capital. |
It is compounding.
That matters because investors often extrapolate semiconductor revenue growth as though it converts cleanly into proportional profit growth.
In reality, many capital-intensive industries experience periods in which revenue and even nominal earnings continue rising while economic returns quietly deteriorate beneath the surface.
China Does Not Need Technological Leadership to Pressure Industry Economics
Much of the semiconductor bull case dismisses Chinese memory expansion because Chinese firms remain behind at the leading edge.
That misses the point.
Commodity semiconductor markets are rarely disrupted by technological leadership alone. They are disrupted by incremental supply at the margin.
Chinese producers do not need to dominate HBM to alter industry economics.
They need only add sufficient lower-end DRAM and NAND capacity to pressure pricing discipline and compress profitability in commodity segments.
The more legacy memory economics weaken, the more incumbent players must rely on premium AI memory to support earnings.
That increases concentration risk across the sector.
The Market May Be Overestimating the Hardware Intensity of Future AI Demand
Much of current semiconductor valuation still assumes that future AI demand will resemble the training-driven spending wave that created the present boom.
That may prove increasingly inaccurate.
The first phase of the AI buildout was dominated by frontier model training, where brute-force compute scale mattered above all else.
Inference economics are fundamentally different.
Once AI systems move into production, enterprises optimize not for maximum model capability but for cost per token served, throughput efficiency, latency, and infrastructure ROI.
That changes deployment behavior materially.
Production deployments increasingly employ quantization, distillation, routing architectures, heterogeneous compute stacks, and model specialization to reduce premium hardware requirements per unit of AI output.
AI usage may continue expanding rapidly even as the hardware intensity of each incremental workload declines.
 |
|
Why Jevons Paradox May Still Not Save the Bull Case
Bulls correctly argue that inference efficiency improvements may trigger Jevons Paradox—where lower costs expand total usage rather than reduce aggregate demand.
That may well occur.
Cheaper inference could unlock AI deployment across every layer of enterprise and consumer software, dramatically expanding total workload volume.
But Jevons Paradox does not automatically validate current valuations.
The relevant investment question is not whether AI demand grows.
It is whether AI demand grows faster than already-discounted expectations.
If markets already price a future in which AI becomes ubiquitous across the global economy, then even explosive workload growth may merely validate current assumptions rather than exceed them.
DePIN Does Not Need to Replace Hyperscalers to Matter
DePIN’s relevance is often misunderstood.
It does not need to replace hyperscalers.
It only needs to absorb enough marginal non-premium workload demand to reduce future urgency for premium infrastructure expansion.
Even limited penetration into batch inference, overflow compute, fine-tuning, or cost-sensitive enterprise workloads introduces incremental supply into the broader compute ecosystem.
And in cyclical hardware industries, marginal supply/demand shifts often matter disproportionately.
Why “It’s Cheap on EPS” May Be the Wrong Framework
A common bullish rebuttal is that semiconductor valuations remain reasonable relative to forward earnings.
On headline multiples alone, that claim appears valid.
 |
|
At first glance, these valuations do not appear historically extreme.
But that observation may itself be misleading.
The flaw lies in assuming current forward EPS reflects sustainable through-cycle earnings rather than peak-cycle economics.
 |
| Figure 5. Peak-cycle earnings often create the illusion of undervaluation by temporarily compressing forward valuation multiples. In cyclical industries, the lowest observed P/E ratios frequently occur near peak profitability—not near trough valuation. |
A low P/E multiple does not necessarily indicate undervaluation.
It may instead indicate that the market expects future normalization.
The relevant question is not whether semiconductor stocks look cheap on current EPS.
It is whether current EPS itself reflects sustainable earnings power.
Why Historical Analogies Still Matter — Even If “This Time Is Different”
AI is not Pets.com.
It is not speculative fiber laid for nonexistent traffic.
It is not a subsidy-driven EV overbuild.
The AI infrastructure buildout is anchored in real demand, real utility, and credible long-term transformative potential.
But that does not make it immune to capital-cycle dynamics.
The lesson of history is not that every boom is identical.
It is that markets repeatedly over-extrapolate even genuine secular growth when near-term economics become unusually favorable.
The internet transformed the world.
That did not prevent internet infrastructure stocks from collapsing after 2000.
Cloud transformed enterprise software.
That did not prevent severe multiple compression in 2022.
Electric vehicles remain secular.
That did not stop battery suppliers from overearning and overbuilding during temporary peak economics.
A transformative technology can still produce cyclical overinvestment, temporary peak margins, and valuation overshoot.
The Strongest Bull Case — And Why It Still May Not Be Enough
The strongest bulls increasingly argue that investors are still underestimating the scale of AI’s ultimate economic impact.
They contend that AI’s true TAM is not global IT spending, but global labor.
They argue AGI infrastructure should be viewed less like enterprise capex and more like strategic military expenditure.
They argue passive flows and index reflexivity may extend the cycle far beyond what valuation models imply.
These arguments are not irrational.
They may all prove directionally correct.
But none resolves the central investment question:
How much of that future is already embedded in price today?
TAM Expansion Does Not Guarantee Proportional Value Capture
Even if AI ultimately attacks trillions of dollars of global labor costs rather than mere IT budgets, that does not mean semiconductor vendors capture an economically proportional share of that value.
History repeatedly shows that infrastructure providers often capture only a fraction of the value created by the ecosystems they enable.
Railroads enabled industrialization.
Most railroad equities performed disastrously.
Telecom networks enabled the internet.
Many infrastructure investors were wiped out.
Cloud enabled software transformation.
Much of the ultimate value accrued at the application layer.
Large TAM does not automatically equal large upstream equity returns.
AGI Optionality Does Not Justify Infinite Multiples
Even if AGI has civilization-scale strategic value, that does not imply every company within the AI hardware stack deserves unbounded valuation multiples.
Optionality has value.
But optionality that is universally recognized is no longer free.
Once markets broadly price in civilization-scale upside scenarios, future returns become increasingly dependent on reality exceeding not ordinary expectations—but extraordinary ones.
Passive Flows Can Delay Gravity, Not Repeal It
Passive ETF reflexivity can extend cycles.
It can amplify momentum.
It can prolong valuation overshoots far longer than fundamentals alone would justify.
But passive flows change timing—not destination.
Eventually, index weights stabilize.
Passive inflows normalize.
Marginal buyers once again become active allocators demanding returns on capital.
At that point, price must reconnect with cash flows.
Final Thesis
The semiconductor supercycle may not end with collapsing revenues, factory shutdowns, or an obvious demand shock.
It may end the way many capital-intensive booms end: with fundamentals that remain objectively healthy even as equity performance quietly deteriorates.
Growth may continue.
Earnings may continue rising.
Headlines may remain bullish.
Yet returns may still disappoint.
Because the central question is no longer whether AI semiconductor demand is robust.
That much is obvious.
The real question is whether demand can remain so extraordinary that it exceeds a market already pricing in sold-out supply years into the future.
Once 2027 production is already spoken for, 2027 is no longer upside.
It is baseline.
And when perfection becomes baseline, even greatness may fail to impress.
The history of investing is filled with moments when the narrative was fundamentally correct, yet the price paid was still too high.
One can be right about the technology, right about the adoption, and right about the macro trend—and still earn poor returns if the asset purchased already reflected all of it.
The question is not whether AI changes the world.
It likely will.
The question is whether semiconductor equities at today’s valuations still offer asymmetric upside after the market has already embraced that possibility.
That remains far less certain.
Sources
- NVIDIA FY2026 10-K Filing
- Reuters — HBM Supply / Samsung / SK hynix Coverage
- Reuters — CXMT Expansion Reports
- ASML / SK hynix Capital Expenditure Disclosures
- Akash Network Public Metrics
- io.net Public Metrics
- Aethir Public Network Statistics
Comments
Post a Comment