Greedy DRAM oligarchs are strangling the future of AI
- Core thesis: The big three memory makers (Samsung, SK Hynix, Micron) are leveraging their oligopoly to continuously drive up DRAM prices, which will cause their share of cloud capital expenditures to rise to approximately 40% by 2027. The author predicts that high memory costs will compel cloud providers to begin cutting capital expenditures in mid-2027, triggering a cyclical turning point in memory prices, significantly earlier than the market consensus of 2030.
- Key elements:
- The DRAM market is monopolized by three giants, holding a combined 89% market share (Samsung at 38%). DRAM contract prices have surged nearly 700% year-over-year, with manufacturers enjoying gross margins of 60% or even higher, yet they continue to push prices up in pursuit of greater profits.
- Memory costs are becoming a major burden for cloud providers: expected to account for 30% of cloud capital expenditure in 2026 and rising to 36.2%-40% by 2027. Microsoft has already added $25 billion in extra capital expenditure, while Amazon, Meta, and others have acknowledged memory as a cost threat.
- Cloud providers have weak pricing power, and long-term agreements (LTAs) cannot control costs. New capacity additions (such as HBM) are slow to ramp; Micron's CEO indicated that meaningful new capacity will not come online until 2028, leaving supply tight in the near term.
- Cloud providers' free cash flow is under pressure, with capital expenditure as a percentage of operating cash flow reaching 98%, the highest level since the internet bubble. They are forced to raise funds through debt issuance or equity offerings, which is seen as a negative signal.
- AI chip companies have already begun taking steps to save on memory costs. For example, NVIDIA's next-generation Rubin rack may halve SOCAMM DRAM usage, while AMD is exploring memory pooling technology, providing side evidence that cost pressure can no longer be ignored.
- The author predicts the DRAM price cycle turning point in mid-2027: by then, cloud providers may slow or even pause capital expenditures. Memory makers' gross margins will contract from peak levels (e.g., SK Hynix falling from 80% to over 60%), and stock prices will decline in advance.
Original Author: P Equity Research
Original Translation: TechFlow
Introduction: P Equity Research presents a seldom-discussed thesis: The memory Big Three (Samsung, SK Hynix, Micron) are driving the AI CapEx cycle toward a breaking point through price hikes. DRAM contract prices are approaching a 700% year-over-year increase, and memory is expected to account for 40% of cloud CapEx by 2027. The author predicts the inflection point will arrive in mid-2027, much earlier than the market's consensus of 2030. A contrarian analysis of the memory cycle.
The Big Three Control 89% of the DRAM Market
SK Hynix (000660.KS), Micron (MU), and Samsung (005930.KS) dominate the DRAM market, holding a combined 89% share, with Samsung alone accounting for 38%. This is an oligopoly.
Image Source: Counterpoint Research
These DRAM manufacturers have capitalized on a supply shortage, raising prices quarter after quarter to alarming levels.
The logic is simple: to build advanced chips, you need DRAM.
How DRAM Becomes HBM
Let's take a brief detour to explain how DRAM becomes HBM.
Stacking DRAM dies layer by layer and connecting them with TSVs (Through-Silicon Vias) creates HBM.
Image Source: SemiAnalysis
In a standard DRAM chip, data must travel to the edge of the silicon die to find the connecting wires. HBM is different. Manufacturers use lasers and chemical etching to carve thousands of micron-sized holes directly through the center of the silicon die, filling them with copper to create TSVs. These act like vertical shafts, running straight through the entire chip.
Between each layer of DRAM, thousands of tiny solder balls called microbumps are placed. When the entire stack is heated, these balls melt, connecting the TSVs of the upper and lower layers to form a continuous, ultra-high-speed vertical data highway.
This is the complete process of turning DRAM into HBM.
Image Source: Bloomberg
As computing power demands more advanced chips, the number of HBM layers also increases. HBM3 has 12 layers, while HBM4 will have 16 layers. More layers mean higher bandwidth and greater capacity—this is the trend.
Returning to the DRAM demand issue: stronger chips need more memory, tightening the memory market further.
My Discontent with These Manufacturers: 60% Gross Margin Isn't Enough?
These manufacturers could live like kings on a 60% gross margin, yet they keep pushing for more. I believe they are actively sacrificing the AI CapEx cycle for higher profits.
So far, no one can clearly say when gross margins will peak. This is one reason I'm writing this.
What is certain is that prices will continue to rise for the remainder of calendar year 2026 (CY26). DRAM contract prices are already approaching a 700% year-over-year increase.
Image Source: Morgan Stanley
Micron, Samsung, and SK Hynix only began planning large-scale capacity expansions between 2024 and 2025. These companies have all experienced boom-and-bust cycles before—following price hikes, demand recedes, supply becomes excessive, and prices collapse.
Image Source: Morgan Stanley
I don't blame them for taking so long. Two reasons: past expansions have compressed memory gross margins, and prolonging the expenditure cycle provides greater demand visibility.
The problem is, they now hold the world's pricing power, enough to choke the entire CapEx cycle, a point that hasn't received enough attention.
Memory Will Account for 30% of Cloud CapEx in 2026; I Bet 40% in 2027
Memory is expected to account for 30% of hyperscaler CapEx in calendar year 2026, rising to 36.2% in 2027.
Image Source: SemiAnalysis
I believe even these estimates are low because memory prices have consistently beaten forecasts. I predict memory's share of CapEx will reach 40% in CY27.
Take ALETHEIA CAPITAL as an example:
"We now expect server DRAM ASP to jump another 30% in the fiscal third quarter of 2026 (previously expected 10% to 15%); likely another 10% to 15% in the fiscal fourth quarter (consistent with prior expectations). For HBM, we expect ASP to double year-over-year in 2027."
Image Source: ALETHEIA CAPITAL
They even predict that memory's content value in AI hardware will rise from just over 40% in 2025 to over 70% in 2027, exceeding 90% in some memory-intensive racks.
Image Source: Company Financials, P Equity Research
The gross margins of Samsung and Micron could peak in the 70% range, with SK Hynix reaching the mid-80% range. This situation could persist until 2027 and extend into 2028.
Micron CEO Sanjay Mehrotra stated in a Bloomberg interview that meaningful new capacity would not come online until 2028.
Video: https://x.com/MilkRoadAI/status/2066231053749006634/video/1
Wait until 2028?
Memory costs might not peak until 2028. Cloud providers, whose free cash flow (FCF) is already tight, will have to adjust their spending to hedge against rising memory costs.
Microsoft Spent an Extra $25 Billion on Memory and Chips
Image Source: Tom's Hardware
In response to memory and chip price hikes, Microsoft raised its capital expenditure by $25 billion. $25 billion.
Other cloud providers haven't given specific figures directly tied to memory costs, but their language is similar or implicitly acknowledges the pressure:
Meta stated, "Component prices are higher this year, especially memory." Microsoft said, "Component prices are higher." Amazon noted, "Memory prices have soared due to supply constraints and strong industry-wide demand."
Image Source: EPOCH AI
No matter who you ask, memory has become a cost threat for everyone. It accounted for 64% of total component costs in the fourth quarter and will likely exceed 70% by the end of 2026.
What can cloud providers do? Nothing. Even long-term agreements (LTAs) can't save them.
Simply put, cloud providers face skyrocketing memory costs because they need to buy both HBM and memory modules. HBM consumes three times the fab capacity of standard server memory. Factories are frantically converting equipment to produce HBM, causing the supply of standard server memory to collapse and prices to surge.
LTAs also have hard limits on the quantity available at discounted prices. The AI boom happened so fast that cloud providers exhausted their contract quotas almost instantly. Any additional demand must be purchased at current market prices.
Image Source: TrendForce
Cloud providers have no choice but to sign new LTAs with memory manufacturers. These contracts are no longer for one year but three to five years, as chip makers rush to lock in terms and hedge against rapid DRAM price increases. The problem is, these LTAs lock in older memory that won't see widespread adoption in the future. Transitioning from HBM3 to HBM4 will cause prices to jump again.
Cloud providers remain in a passive position, with pricing power firmly held by this oligopoly.
Free Cash Flow Hits Bottom: 98% of Operating Cash Flow Eaten by CapEx
Cloud providers have no choice but to continuously issue equity and debt. Google and Meta (hinting at a potential issuance?) are doing it, and Amazon may soon follow.
Free cash flow is rapidly drying up. Cloud providers are pouring 98% of their operating cash flow into capital expenditures. This is the highest level since the dot-com bubble.
Image Source: Goldman Sachs
Meanwhile, Morgan Stanley predicts CapEx in 2027 will remain strong, around $1.1 trillion.
Image Source: Morgan Stanley
Do the math: about 40% of this amount would go to memory, roughly $440 billion. This is almost equal to the total CapEx for the entire year of 2025.
Two things concern me:
First, the equity and debt financing in the market is already sending negative signals to participants—cash is running dry, and the multiples of price-to-sales and free cash flow are off the charts.
Second, cost pressures could slow down or even halt CapEx growth earlier than expected. Based on my estimate, earnings calls around mid-2027 will begin to show signs of pulling back.
I believe this second point will confront memory manufacturers by the end of 2026, much sooner than many expect.
From here on, the number one issue repeatedly emphasized on earnings calls will be component prices—especially memory—and how they squeeze spending budgets. I don't think cloud providers will ignore this and continue to increase CapEx without concern.
That's just my opinion.
Chip Makers Are Already Finding Ways to Save on Memory
AMD, Nvidia, and Google are already moving toward memory optimization.
Nvidia's next-generation Rubin NVL72 rack might cut the CPU-side SOCAMM DRAM from roughly 55TB per rack to about 28TB, nearly halving it. This makes sense because the bill of materials for the VR200 shows memory costs have increased 435% compared to the GB300.
Image Source: Morgan Stanley
SOCAMM isn't HBM, but the cost pressure driving the need for savings is the same—whether it's AMD using MEXT for memory pooling (making flash memory behave like DRAM) or directly cutting SOCAMM DRAM.
In fact, chip makers have even fewer options: they're already paying for HBM, and stacking SOCAMM costs on top of that? Painful. They're being squeezed from both sides.
Memory is Still Cyclical, with an Inflection Point in Mid-2027
Finally, let's talk about the cyclical nature of memory.
I disagree with the statement that "memory is no longer cyclical."
Even if I'm completely wrong and CapEx can remain strong for a decade, you will naturally hit a boom-and-bust cycle. Those who argue against me would have to assume that memory demand grows year after year and that cloud provider spending never enters a cycle—which is impossible.
Image Source: SEMI
These manufacturers, aggressively expanding capacity, are betting on continuous CapEx growth (unlikely under current trends) and sustained strong memory demand (which itself relies on continuous CapEx growth).
My thesis is
