Has the AI Boom Permanently Broken the Semiconductor Cycle?

 

Filed under: Tech & AI | Market Outlook

The Semiconductor Cycle: Oscillating Between Boom and Bust

The semiconductor industry has historically oscillated between extreme boom and bust cycles. When demand surged, manufacturers expanded their fabrication plants (fabs). Years later, this inevitably led to oversupply. As prices collapsed, companies slashed capital expenditures (CapEx). These investment cuts eventually triggered the next supply shortage, causing prices to rebound and restarting the cycle.

Historically, personal computers and smartphones formed the foundation of this cycle. When PC and smartphone sales accelerated, device manufacturers ramped up their memory orders. Industry leaders like Samsung Electronics, SK Hynix, and Micron Technology enjoyed rising prices and subsequently increased their CapEx.

However, bringing new production capacity online entails a significant lead time. Eventually, a massive wave of new supply would hit the market simultaneously. Concurrently, inventory would build up in clients' warehouses, leading to a sharp drop in new orders. Memory prices would plummet, dragging semiconductor corporate profits down with them.

In short, the classic boom-and-bust loop looked like this:

Demand growth → Price increases → CapEx expansion → Oversupply → Price collapse → Investment cuts → Supply shortage

To illustrate, over the past five fiscal years, Micron’s average selling prices (ASPs) for DRAM have swung wildly from annual increases in the 40% range to annual declines exceeding 40%. Similarly, NAND flash prices have fluctuated from 30% gains to devastating 50% declines.

Consequently, semiconductor equities have been highly cyclical. Share prices did not wait for earnings to recover; they often began pricing in a rebound before memory prices even reached their absolute bottom. Conversely, at peak earnings, the market was usually already discounting the impending oversupply phase.

Recently, however, this predictable pattern has started to break down.

How HBM Disrupted the Cycle

High Bandwidth Memory (HBM) in the artificial intelligence (AI) era operates quite differently from traditional memory. Conventional commodity DRAM was produced in massive volumes with standardized specifications, forcing companies to compete primarily on price.

HBM, conversely, is a highly complex component that vertically stacks multiple DRAM dies, connecting them via Through-Silicon Vias (TSVs) and advanced packaging techniques. Because of this architecture, producing a given capacity of HBM consumes significantly more wafers than standard DRAM. As HBM production scales up, fewer wafers remain available for conventional DRAM manufacturing. Furthermore, stacking multiple layers makes yield management and quality control exponentially more difficult.

Despite these hurdles, corporate demand for HBM is surging. In fact, companies require massive, unprecedented quantities. This megatrend is fueled by aggressive AI data center investments from mega-cap tech giants like Microsoft, Meta, Alphabet, and Amazon.

The numbers reflect this surge:

• NVIDIA’s data center revenue for the first quarter of fiscal 2027 rose 92% year-over-year to $75.2 billion.
• Micron reported its data center revenue exceeded $25 billion in the third quarter of fiscal 2026, translating to more than $100 billion on an annualized basis.
• Samsung Electronics also appears poised to report the world’s highest operating profit in the sector.

HBM is currently a severe bottleneck on its own, but its production has concurrently driven a shortage of conventional DRAM. Crucially, this current supply deficit feels different from historical shortages. While fabs are actively expanding, an increasing number of market forecasts suggest that supply and demand may not invert even after new capacity comes online—a stark departure from previous expansion cycles.

In other words, the AI-driven structural demand for HBM may have permanently altered the traditional semiconductor cycle.

The Shifting Balance of Power

As a direct result, the balance of power between memory suppliers and their buyers has shifted. Historically, inevitable periods of oversupply made it easy for customers to pivot between competing suppliers. During those cyclical troughs, buyers frequently drove memory prices lower, making tense price negotiations a routine occurrence every quarter or half-year.

Today, however, the market dynamics are entirely inverted. Global technology clients are now making highly unusual proposals—including offering direct financial support for fab investments—simply to secure SK Hynix’s future production capacity. Industry reports even suggest that supply is so constrained, there is literally zero additional volume available for allocation to certain clients.

The fundamental nature of supply contracts is also evolving. Short-term negotiations are increasingly giving way to binding long-term agreements (LTAs). Reports indicate that Samsung Electronics has already locked in highly binding contracts with key customers.

This is a significant structural shift. It implies that the extreme volatility characteristic of the old memory cycle may be fading. Instead of blindly building fabs on the expectation of future orders, memory makers are now deploying CapEx only after securing firm, long-term demand commitments from clients.

We are witnessing a parallel trend in the foundry sector. TSMC is aggressively expanding its 3-nanometer production capacity to meet AI demand, with new fabs in Taiwan, the United States, and Japan expected to come online shortly. Despite this expansion, TSMC management notes that client demand remains so robust that the entire supply chain is struggling to keep pace.

Of course, it remains premature to declare a complete and permanent reversal of industry pricing power. While constrained supply has undeniably bolstered the negotiating leverage of semiconductor manufacturers, the customer base is increasingly concentrated among NVIDIA and a handful of mega-cap tech companies. If memory suppliers become overly reliant on a few dominant buyers, this concentration risk could quickly become a critical weakness in future price negotiations.

Has the Semiconductor Cycle Really Disappeared?

Let’s start with the bottom line: The semiconductor cycle has not disappeared. Rather, its duration and amplitude are fundamentally changing.

AI data center buildouts have permanently altered the structural architecture of the memory industry. Even as manufacturers race to boost output, the process of constructing new fabs, installing advanced lithography equipment, and stabilizing commercial yields requires significant lead time.

Micron has explicitly forecasted that DRAM and NAND shortages will persist beyond 2027. The company noted that even if supply gradually improves throughout 2028, it remains highly uncertain when it will fully align with market demand. At a minimum, this points toward a structurally longer upcycle than investors are historically accustomed to.

However, long-term contracts will not permanently eradicate the cycle either. The current supply shortage could evaporate far sooner than the market expects if:

• The return on investment (ROI) for AI falls short of lofty expectations, prompting major tech companies to slash CapEx.
• A massive wave of new memory capacity comes online simultaneously across 2027 and 2028.
• HBM production yields improve faster than anticipated.

Investors must also monitor bottlenecks elsewhere in the ecosystem—namely, electrical power and investment capital. While AI demand currently appears almost limitless, grid capacity and corporate balance sheets are not bottomless wells. Furthermore, robust current earnings are heavily priced into equity valuations. Blockbuster earnings no longer guarantee an automatic, sharp rally in semiconductor stocks. The critical question for markets is the absolute durability of these earnings.

Bottom Line

HBM did not kill the semiconductor cycle. Instead, it attached a high-performance engine to the traditional cyclical framework.

Historically, marginal capacity increases could rapidly generate excess memory inventory. Today, the industry is weathering chronic, structural shortages. This is because modern AI data centers are acting as massive sinks—absorbing not just HBM, but conventional DRAM, enterprise SSDs, advanced node foundry capacity, and advanced packaging volumes. This paradigm shift explains the emergence of long-term contracts and advance payment negotiations, as desperate clients rush to secure future production.

These market dynamics are entirely unprecedented. Yet, fundamental economics dictates that supply shortages cannot last forever. Elevated prices eventually attract capital investment, and that investment ultimately translates into fresh supply. The current friction lies in the market's divergent views on exactly when that supply wave will hit.

The most accurate assessment of the current landscape is this: The semiconductor cycle is not broken. It has merely elongated, and what used to be a single, unified cycle has fractured into multiple distinct sub-cycles.

Moving forward, investors must look past the simplistic narrative of a broad "semiconductor boom." The focus must shift to monitoring three key metrics:

• Whether AI CapEx by major tech companies sustains its growth trajectory.
• Whether binding HBM contracts translate to tangible price stability.
• Exactly how fast new capacity ramps up in 2027 and 2028.

This time really may be different. But we must remember a timeless market truth: the most dangerous phase in the semiconductor cycle is the exact moment everyone believes "this time is different" forever.