DRAM Oligopoly Threatens AI Future Amid 700% Price Surge

Original author: P Equity Research

DeepChain TechFlow

Overview: P Equity Research presents a rarely acknowledged view: the three memory giants—Samsung, SK Hynix, and Micron—are pushing the AI capital expenditure cycle toward a breaking point through price hikes. DRAM contract prices are nearing a 700% year-over-year increase, and memory is projected to account for 40% of cloud providers' capital expenditures by 2027. The author predicts the turning point will arrive in mid-2027, significantly earlier than the market’s general expectation of 2030. A contrarian analysis of the memory cycle.

SK Hynix (000660.KS), Micron (000660.KS), Micron (000660.KS), Micron (MU), and Samsung ($005930.KS) control the DRAM market, accounting for a combined 89% share, with Samsung alone holding 38%. This is an oligopoly alliance.

Chart source: Counterpoint Research

These DRAM manufacturers took advantage of the supply-demand imbalance, raising prices quarter after quarter to an alarming level.

It's simple logic: to manufacture advanced chips, you need DRAM.

Let’s take a quick detour to explain how DRAM becomes HBM.

Stacking DRAM dies layer by layer and connecting them with TSVs (through-silicon vias) creates HBM.

Chart source: SemiAnalysis

In ordinary DRAM chips, data must travel to the edge of the silicon wafer to find wiring. HBM is different: manufacturers use lasers and chemical etching to create thousands of micron-sized holes right in the center of the silicon wafer, which are then filled with copper—these are called TSVs. They act like vertical shafts, piercing straight through the entire chip.

Between each layer of DRAM, tens of thousands of tiny solder balls called microbumps are placed. After the entire stack is heated, the microbumps melt, connecting the TSVs of the upper and lower layers to form a continuous, ultra-high-speed vertical data pathway.

This is the entire process of DRAM becoming HBM.

Chart source: Bloomberg

Hash rate requires more advanced chips, and the number of HBM layers is increasing accordingly. HBM3 has 12 layers, while HBM4 aims for 16 layers. More layers mean higher bandwidth and greater capacity—that’s the direction.

Returning to DRAM demand: the stronger the chip, the more memory it requires, and the memory market is becoming increasingly tight.

These manufacturers could easily live like kings with a 60% gross margin, yet they’re still pushing harder—I believe they’re voluntarily sacrificing their AI capital expenditure cycle to achieve higher profits.

No one yet knows when the gross profit will peak. This is one of the reasons I wrote this article.

It is certain that the price will continue to rise for the remainder of the calendar year 2026 (CY26). DRAM contract prices have already approached a 700% year-over-year increase.

Chart source: Morgan Stanley

Micron, Samsung, and SK Hynix have delayed large-scale capacity expansion until 2024 to 2025. All three have previously experienced cycles of boom and bust—after price increases, when demand declines and supply surges, prices collapse.

Chart source: Morgan Stanley

I don't blame them for taking so long; there are two reasons:

In the past, expanding production lowered memory gross margins; enduring a longer expenditure cycle increases visibility into demand.

The problem is that they now hold global pricing power, sufficient to strangle the entire capital expenditure cycle—a point that hasn’t received enough attention.

Memory is expected to account for 30% of hyperscaler capital expenditures in calendar year 2026, rising to 36.2% in 2027.

Chart source: SemiAnalysis

I believe even these estimates are too low, as memory prices have consistently surpassed forecasts. I predict memory will account for 40% in CY27.

Take ALETHEIA CAPITAL as an example:

We now expect the average selling price (ASP) of server DRAM to increase by another 30% in the third fiscal quarter of 2026 (up from our previous forecast of 10% to 15%); a further increase of 10% to 15% is possible in the fourth fiscal quarter (consistent with our prior outlook). We anticipate the ASP of HBM to double year-over-year in 2027.

Chart source: ALETHEIA CAPITAL

They even project that the value of memory within AI hardware will rise from just over 40% in 2025 to more than 70% by 2027, with some memory-intensive racks exceeding 90%.

Chart source: Company financial reports, P Equity Research

Samsung and Micron's gross margins may reach highs of over 70%, while SK Hynix could reach the mid-80s. This situation may persist through 2027 and extend into 2028.

Micron CEO Sanjay Mehrotra said in a Bloomberg interview that meaningful new capacity will not come online until 2028.

Video: https://x.com/MilkRoadAI/status/2066231053749006634/video/1

Wait until 2028?

Memory costs may not peak until 2028, forcing cloud providers with already tight free cash flow (FCF) to adjust spending to offset rising memory expenses.

Chart source: Tom's Hardware

Microsoft increased its capital expenditures by $25 billion to address rising memory and chip prices. $25 billion.

Other cloud providers haven't provided specific figures directly tied to memory costs, but their wording is similar or indirectly acknowledges it:

Meta said, "Component prices are higher this year, especially memory"; Microsoft said, "Component prices are higher"; Amazon said, "Memory prices have surged due to supply constraints and strong industry-wide demand."

Chart source: EPOCH AI

No matter who you ask, memory has become a cost threat to everyone. It accounted for 64% of total component costs in the fourth quarter and is likely to exceed 70% by the end of 2026.

What can the cloud providers do? Nothing at all. Even long-term agreements (LTAs) can't save them.

In short, cloud providers are facing a sharp rise in memory costs because they need to buy both HBM and memory modules. HBM consumes three times the production capacity of standard server memory. Factories are urgently shifting equipment to produce HBM, causing a collapse in the supply of standard server memory and driving prices up sharply.

LTA also imposes a hard cap on the quantity that can be purchased at the discounted rate. The AI boom arrived too quickly, and cloud providers exhausted their contract allocations almost instantly. Any additional demand must be met at current market prices.

Chart source: TrendForce

Cloud providers have no choice but to sign new LTAs with memory manufacturers. These contracts are no longer for one year but for three to five years, as chipmakers aim to lock in prices early to hedge against rapid DRAM price increases. Compounding the issue, these LTAs lock in older memory that won’t be widely adopted in the future. Moving from HBM3 to HBM4 will push prices even higher.

Cloud providers remain in a passive position, with pricing power firmly held by this alliance.

Cloud providers have no choice but to continuously issue more equity and bonds. Google and Meta (possibly hinting at upcoming issuance) are doing it, and Amazon may soon follow suit.

Free cash flow is rapidly depleting, with cloud providers pouring 98% of their operating cash flow into capital expenditures—the highest level since the dot-com bubble.

Chart source: Goldman Sachs

Meanwhile, Morgan Stanley forecasts that capital expenditures will remain strong in 2027, at approximately $1.1 trillion.

Chart source: Morgan Stanley

Calculate it: About 40% of this money, roughly $440 billion, will go to memory. This is nearly equivalent to the total capital expenditure for the entire year of 2025.

Two things make me uneasy:

First, equity and debt financing in the market are already sending negative signals to participants—cash is hitting bottom, and price-to-sales and free cash flow multiples are soaring.

Second, cost pressures may cause capital expenditure growth to slow, or even halt earlier than expected. Based on my estimates, by mid-2027, earnings calls will begin to hint at a pause.

I believe the second point will hit memory manufacturers by the end of 2026, much earlier than many anticipate.

From now on, the top issue repeatedly emphasized on earnings calls has been component prices—especially memory—and how they are squeezing budget allocations. I don’t believe cloud providers will ignore this and continue ramping up capital expenditures without concern.

This is just my opinion.

AMD, NVIDIA, and Google are already moving toward memory optimization.

The CPU-side SOCAMM DRAM for NVIDIA’s next-generation Rubin NVL72 rack may be reduced from approximately 55TB per rack to around 28TB, nearly halved. This makes sense, as the BOM for the VR200 shows that memory costs have increased by 435% compared to the GB300.

Chart source: Morgan Stanley

SOCAMM is not HBM, but the cost pressure driving cost-saving solutions is the same—whether AMD uses MEXT for memory pooling (making flash behave like DRAM) or simply eliminates SOCAMM DRAM altogether.

Chip manufacturers really have no choice: they’ve already paid for HBM—adding the cost of SOCAMM? That hurts. They’re caught between two fronts.

Finally, let's discuss the periodicity of memory.

I disagree with the statement that "memory no longer has cycles."

Even if everything I said is wrong, and capital expenditures truly remain strong for a decade, you will still inevitably encounter the cycle of boom and bust. Those who refute me must assume that memory demand grows every year and that cloud providers’ spending never enters a cycle—this is simply impossible.

Chart source: SEMI

These manufacturers are betting on continued growth in capital expenditure (which seems unlikely at this pace) and sustained strong demand for memory (which in turn depends on ongoing growth in capital expenditure).

My projection is that DRAM pricing will begin to peak in 2027:

SK Hynix has a gross margin of approximately 80%; Micron has a gross margin of approximately 78% to 80%; Samsung has a gross margin of approximately 70% to 75%.

The price curve flattens around February or March, while capacity remains tight. Then, around mid-2027, you'll sense signs of slowing, or even a pause, in capital expenditure growth.

I believe most memory stocks are beginning to give back their gains at this level, as investors are already pricing in the upcoming margin contraction.

By 2028, more capacity comes online (supply remains tight), but demand expectations weaken, causing gross margins to steadily decline to just over 60%. From 2028 to 2030, capacity continues to come online, easing supply constraints, while capital expenditures show no meaningful increase. I predict a true crash occurs during this phase, with significant stock price gains retracing from the end of 2027.

Everyone believes memory will remain strong until the end of 2030; my prediction is that gross margins will begin to contract from mid-2027, and many memory stocks' gains will reverse.

Then again, if cloud providers in 2027 say that capital expenditures in 2028 will be significantly stronger, my article will have been written in vain, and I’ll look foolish. Time will tell whether I’m right or wrong, but I believe memory is about to follow this path.

I’m not as optimistic about memory as others are, for just a few reasons:

Memory manufacturers are too greedy on margins; I believe memory still follows a cycle—the "acyclical theory" bets entirely on capital expenditures never entering a cycle; chipmakers seeking ways to save on memory costs themselves prove they are already tired of high expenses; CFOs' cash is nearly 100% consumed by capital expenditures, with memory accounting for 40% of 2027 costs—further bond issuances or equity offerings are no longer sustainable.

The only good outcome would be a sudden flood of supply dumping in the market, crashing the memory prices of all three manufacturers. In that case, the same capital expenditure would yield significantly more output.