My resentment towards memory manufacturers: Samsung is choking the AI capital expenditure cycle.

Author: P Equity Research

Translation: Deep潮 TechFlow

Deep潮 Guide: P Equity Research presents a rarely acknowledged judgment: the memory giants (Samsung, SK Hynix, Micron) are using price hikes to push the AI capital expenditure cycle toward a rupture. DRAM contract prices are approaching 700% year-over-year, and memory will account for 40% of cloud providers' capital spending by 2027. The author predicts the inflection point will arrive in mid-2027, much earlier than the market generally expects in 2030. An contrarian memory cycle projection.

The three giants capture 89% of the DRAM market

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

Chart source: Counterpoint Research

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

The logic is simple: to make advanced chips, you need DRAM.

How DRAM becomes HBM

Let's take a detour and briefly explain how DRAM turns into HBM.

Stacked DRAM chips are connected with TSVs (through-silicon vias) in the middle, forming HBM.

Chart source: SemiAnalysis

In regular DRAM chips, data must travel to the edge of the silicon to find wiring. HBM differs: manufacturers laser and chemically etch thousands of micron-scale holes into the silicon die, fill them with copper—these are TSVs. They act like vertical shafts piercing through the entire chip.

Between each layer of DRAM, thousands of microbump solder balls are placed. When the stack is heated, the solder balls melt, connecting the TSVs of the upper and lower layers, forming a continuous, ultra-high-speed vertical data highway.

This is the full process of transforming DRAM into HBM.

Chart source: Bloomberg

Computing power requires more advanced chips, so the number of HBM layers also increases. HBM3 has 12 layers, and HBM4 aims for 16 layers. More layers mean higher bandwidth and larger capacity—that's the direction.

Returning to the demand for DRAM: as chips become more powerful, more memory is needed, and the memory market becomes increasingly tight.

My dissatisfaction with these manufacturers: 60% gross margin still isn't enough

These companies could comfortably enjoy kingly days with 60% gross margins, yet they keep pushing higher. I believe they are actively sacrificing the AI capital expenditure cycle in exchange for higher profits.

So far, no one can say when their margins will peak. That’s one reason I’m writing this.

What’s certain is that in the remaining time of 2026 (CY26), prices will continue to rise. DRAM contract prices are already approaching 700% YoY.

Chart source: Morgan Stanley

Micron, Samsung, SK Hynix only started planning large-scale capacity expansion in 2024-2025. These companies have all experienced cycles of boom and bust—prices rise, then demand recedes, oversupply causes prices to collapse.

Chart source: Morgan Stanley

I don’t blame them for dragging it out; there are two reasons:

Past capacity expansions suppressed memory margins; waiting longer during the expenditure cycle provides higher demand visibility.

But the problem is, they now hold the global pricing power, enough to choke the entire capital expenditure cycle, and not enough people are paying attention to this.

Memory will account for 30% of hyperscalers’ capital expenditure in 2026, I bet it will reach 40% in 2027

Memory is expected to account for over 30% of hyperscalers’ (large-scale cloud providers) capital expenditure in 2026, rising to 36.2% in 2027.

Chart source: SemiAnalysis

I believe these estimates are even too low, as memory prices have been defying predictions all along. I forecast that in CY27, memory’s share will reach 40%.

Taking ALETHEIA CAPITAL as an example:

“We now expect the average selling price (ASP) of server DRAM to jump another 30% in Q3 2026 (previously expected 10-15%); in Q4, it may rise another 10-15% (consistent with prior expectations). We expect HBM ASP to double YoY in 2027.”

Chart source: ALETHEIA CAPITAL

They even predict that the content value of memory in AI hardware will rise from just over 40% in 2025 to over 70% in 2027, with some memory-intensive cabinets exceeding 90%.

Chart source: Company financial reports, P Equity Research

Samsung and Micron’s gross margins could reach over 70%, while SK Hynix might hover around 80%. This situation could persist until 2027 and extend into 2028.

Micron CEO Sanjay Mehrotra said in an interview with Bloomberg that meaningful new capacity won’t come online until 2028.

Video:

Waiting until 2028?

Memory costs might peak only in 2028, and cloud providers, with already tight free cash flow (FCF), can only adjust their spending to hedge against rising memory costs.

Microsoft has invested an additional $25 billion in memory and chips

Chart source: Tom's Hardware

To cope with rising memory and chip prices, Microsoft increased capital expenditure by $25 billion. $25 billion.

Other cloud providers haven’t given specific figures directly linked to memory costs, but their wording is similar or they’ve indirectly admitted:

Meta said “component prices are higher this year, especially memory”; Microsoft said “component prices are higher”; Amazon said “memory prices have soared due to supply constraints and strong industry demand.”

Chart 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 Q4 and is likely to exceed 70% by the end of 2026.

What can cloud providers do? Nothing much. Even long-term agreements (LTAs) won’t save them.

In short, cloud providers face skyrocketing memory costs because they need to buy both HBM and memory modules. HBM capacity consumption is three times that of regular server memory. Factories are frantically shifting equipment to produce HBM, causing supply of regular server memory to collapse and prices to surge.

LTAs have a hard cap on the volume they can buy at discounted prices. The AI boom arrived too quickly, and cloud providers nearly exhausted their contract quotas instantly. The excess demand can only be purchased at current market prices.

Chart source: TrendForce

Cloud providers have no choice but to sign new LTAs with memory manufacturers. These contracts are now 3 to 5 years long, as chip makers want to lock in supply quickly to hedge against rapid DRAM price increases. More troubling, these LTAs lock in older memory that won’t see large-scale adoption in the future. Moving from HBM3 to HBM4 will push prices even higher.

Cloud providers remain in a passive position, with pricing power firmly in the hands of this alliance.

Free cash flow hits bottom: 98% of operating cash flow is consumed by capital expenditure

Cloud providers have no choice but to keep issuing equity and bonds. Google, Meta (hinting at possibly issuing?), and perhaps Amazon will follow soon.

Free cash flow is rapidly depleting, with 98% of operating cash flow poured into capital expenditure. This is the highest level since the internet bubble.

Chart source: Goldman Sachs

Meanwhile, Morgan Stanley forecasts that capital expenditure will remain strong in 2027, around $1.1 trillion.

Chart source: Morgan Stanley

Calculations show that about 40% of this money—roughly $440 billion—will go to memory. This is roughly equal to the total capital expenditure for the entire year of 2025.

Two points worry me:

First, equity and debt financing in the market are already sending negative signals—cash is bottoming out, and the price-to-sales ratio and free cash flow multiples are exploding.

Second, cost pressures may slow the growth of capital expenditure, or even cause an earlier pause than expected. In my estimation, around mid-2027, earnings calls will start hinting at a slowdown.

I believe point two will be realized by the end of 2026, much earlier than many expect, as memory manufacturers approach the peak of margins.

From then on, earnings calls will repeatedly emphasize component prices—especially memory—and how they squeeze expenditure budgets. I don’t think cloud providers will ignore this or continue to ramp up capital spending without concern.

This is just my view.

Chipmakers are already looking for ways to save memory

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

Nvidia’s next-generation Rubin NVL72 cabinet’s CPU-side SOCAMM DRAM might be cut from about 55TB per cabinet to around 28TB, nearly halving. This makes sense, as VR200’s BOM shows memory costs increased by 435% over GB300.

Chart source: Morgan Stanley

SOCAMM isn’t HBM, but the cost-cutting logic is the same—whether AMD uses MEXT to pool memory (making flash storage behave like DRAM) or directly cuts SOCAMM DRAM.

Chipmakers really have no choice: they’ve already spent on HBM, and adding SOCAMM costs? Painful. They’re caught between two cuts.

Memory remains cyclical, with a turning point in mid-2027

Finally, a word on the cyclicality of memory.

I disagree with the idea that “memory no longer has cycles.”

Even if I’m completely wrong, if capital expenditure remains strong for ten years, you’ll still inevitably hit the boom-bust cycle. Anyone arguing otherwise must assume: memory demand grows every year, cloud providers’ spending never enters a cycle—that’s impossible.

Chart source: SEMI

These aggressive capacity expansion companies are betting on continued growth in capital expenditure (which seems unlikely given current trends) and sustained high demand for memory (which again depends on ongoing capital spending growth).

My projection is that in 2027, DRAM prices will start to peak:

SK Hynix’s margins around 80%; Micron around 78-80%; Samsung around 70-75%.

Prices will flatten amid still tight capacity, roughly around February or March. Then, in mid-2027, signals of slowing or even pausing capital expenditure will emerge.

I believe most memory stocks will start to give back gains at this point, with investors pricing in the upcoming margin contraction early.

By 2028, more capacity will come online (supply still tight), but demand expectations will soften, margins will continue to decline to around 60%. From 2028 to 2030, capacity will keep increasing, supply will ease, and capital expenditure will stagnate. I predict a real crash will occur during this phase, with many stock prices starting to decline from late 2027.

Everyone believes memory will stay strong until the end of 2030; my forecast is that margin contraction begins in mid-2027, and many memory stocks’ gains will reverse.

Back to the point: as long as in 2027 cloud providers say their 2028 capital expenditure will be significantly stronger, my article is pointless, and I’d look like a fool. The truth is, time will tell, but I believe memory’s next move is along this path.

Why I’m less optimistic

I’m less optimistic about memory for a few reasons:

Memory manufacturers are too greedy for margins; I believe memory still has cycles—“no-cycle theory” relies entirely on perpetual growth in capital expenditure; chipmakers are already finding ways to cut memory costs, which proves they’re fed up with high costs; CFOs’ cash is almost 100% consumed by capital expenditure, and memory will still account for 40% of costs in 2027—raising more debt or issuing more shares is no longer sustainable.

The only positive scenario is a sudden surge of supply dumping, crashing the prices of memory from the three giants. That way, the same capital expenditure could produce more output.