My grievances with memory manufacturers: Samsung and others are suffocating the AI capital expenditure cycle
TechFlow Selected TechFlow Selected
My grievances with memory manufacturers: Samsung and others are suffocating the AI capital expenditure cycle
The author predicts the inflection point will arrive in mid-2027—much earlier than the market’s general expectation of 2030.
Navigating Web3 tides with focused insightsAuthor: P Equity Research
Compiled and translated by TechFlow
TechFlow Introduction: P Equity Research presents an underappreciated view: the “Big Three” memory manufacturers—Samsung, SK Hynix, and Micron—are pushing the AI capital expenditure cycle toward a breaking point through aggressive price hikes. DRAM contract prices have surged nearly 700% year-on-year, and memory is projected to account for 40% of cloud hyperscalers’ capex in 2027. The author forecasts a turning point in mid-2027—much earlier than the market’s consensus expectation of 2030. This is 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, collectively holding 89% share—with Samsung alone commanding 38%. This is an oligopolistic alliance.
Source: Counterpoint Research
This group of DRAM manufacturers has seized the supply-demand imbalance, raising prices quarter after quarter—to alarming levels.
The logic is simple: building advanced chips requires DRAM.
How DRAM Becomes HBM
Let’s pause briefly to explain how DRAM evolves into HBM.
HBM is created by stacking DRAM die vertically and interconnecting them via TSVs (through-silicon vias).
Source: SemiAnalysis
In conventional DRAM chips, data must travel to the edge of the silicon die to access wiring. HBM differs fundamentally: manufacturers use lasers and chemical etching to drill thousands of micron-scale holes directly through the center of the silicon wafer, then fill them with copper—these are the TSVs. They function like vertical shafts, penetrating the entire chip stack.
Between each DRAM layer, thousands of microscopic solder bumps—called microbumps—are placed. Upon heating the full stack, these bumps melt and electrically connect adjacent TSVs, forming a continuous, ultra-high-speed vertical data highway.
This is the complete transformation from DRAM to HBM.
Source: Bloomberg
As compute demand drives more advanced chips, HBM stack heights continue rising: HBM3 uses 12 layers; HBM4 will reach 16 layers. More layers mean higher bandwidth and greater capacity—that’s the direction.
Returning to DRAM demand: the more powerful the chips, the more memory they require—and the tighter the memory market becomes.
My Grievance with These Manufacturers: 60% Gross Margins Aren’t Enough?
These manufacturers could comfortably live like royalty on 60% gross margins—yet they keep squeezing further. I believe they’re deliberately sacrificing the AI capex cycle to extract even higher profits.
When gross margins will peak remains unclear—even now. That’s one reason I wrote this piece.
What’s certain is that prices will continue rising through the remainder of calendar year 2026 (CY26). DRAM contract prices have already surged nearly 700% year-on-year.
Source: Morgan Stanley
Micron, Samsung, and SK Hynix delayed large-scale capacity expansion until 2024–2025. All three have previously experienced boom-and-bust cycles: after price hikes, demand cools, supply overhangs emerge, and prices collapse.
Source: Morgan Stanley
I don’t blame them for waiting so long—two reasons explain it:
Past expansions depressed memory gross margins; waiting longer during the capex cycle improves visibility into future demand.
But here’s the problem: they now hold global pricing power—the ability to strangle the entire capex cycle. Yet this critical fact receives insufficient attention.
Memory to Account for 30% of Hyperscaler Capex in 2026—and 40% in 2027
Memory is expected to represent 30% of hyperscaler capex in CY26—and rise to 36.2% in CY27.
Source: SemiAnalysis
I believe even these estimates are conservative—memory prices have consistently outpaced forecasts. I project memory’s share of CY27 capex will reach 40%.
Take ALETHEIA CAPITAL as an example:
“We now expect average selling prices (ASPs) for server DRAM to jump another 30% in Q3 FY2026 (previously forecast at 10%–15%), with a further 10%–15% increase possible in Q4 FY2026 (in line with prior expectations). We anticipate HBM ASPs will double year-on-year in 2027.”
Source: ALETHEIA CAPITAL
They further project memory’s content value within AI hardware will climb from just over 40% in 2025 to over 70% in 2027—with some memory-intensive racks exceeding 90%.
Source: Company filings, P Equity Research
Samsung and Micron may reach gross margins above 70%; SK Hynix could hit mid-80% levels. Such profitability may persist through CY27 and into CY28.
Micron CEO Sanjay Mehrotra stated in a Bloomberg interview that meaningful new capacity won’t come online until 2028.
Video: https://x.com/MilkRoadAI/status/2066231053749006634/video/1
Not until 2028?
Memory costs may not peak until 2028—while hyperscalers, already squeezed on free cash flow (FCF), will have no choice but to adjust their spending to offset surging memory expenses.
Microsoft Spends $25 Billion Extra on Memory and Chips
Source: Tom's Hardware
To counter rising memory and chip prices, Microsoft increased its capex by $25 billion. $25 billion.
Other hyperscalers haven’t disclosed specific figures tied directly to memory costs—but their wording is consistent, or at least obliquely acknowledges the issue:
Meta said “component prices are higher this year, especially memory”; Microsoft cited “higher component prices”; Amazon noted “memory prices have surged due to constrained supply and strong industry-wide demand.”
Source: EPOCH AI
No matter whom you ask, memory has become a universal cost threat. It accounts for 64% of total component costs in Q4—and is likely to exceed 70% by end-2026.
So what can hyperscalers do? Nothing. Even long-term agreements (LTAs) offer little relief.
Put plainly, hyperscalers face soaring memory costs because they need both HBM and memory modules. HBM production consumes three times the fab capacity of standard server memory. Factories are rapidly retooling equipment for HBM—causing standard server memory supply to collapse and prices to surge.
LTAs impose hard caps on discounted purchase volumes. The AI boom arrived so suddenly that hyperscalers exhausted their LTA allocations almost instantly. Additional demand must be fulfilled at prevailing spot-market prices.
Source: TrendForce
Hyperscalers have no choice but to sign new LTAs with memory vendors—now spanning 3 to 5 years rather than one. Chipmakers seek to lock in supply quickly to hedge against DRAM’s rapid price increases. More troublingly, these LTAs commit to older memory generations unlikely to see broad adoption in the future. Transitioning from HBM3 to HBM4 will trigger yet another price jump.
Hyperscalers remain passive players—pricing power rests firmly in the oligopoly’s hands.
Free Cash Flow Hits Bottom: 98% of Operating Cash Flow Absorbed by Capex
Hyperscalers have no option but to continually issue equity and debt. Google and Meta (hinting at potential issuance?) are doing so—and Amazon may soon follow.
Free cash flow is rapidly evaporating: hyperscalers are allocating 98% of operating cash flow to capex—the highest level since the dot-com bubble.
Source: Goldman Sachs
Meanwhile, Morgan Stanley forecasts robust capex in 2027—approximately $1.1 trillion.
Source: Morgan Stanley
Breakdown: ~40% of that sum—roughly $440 billion—goes to memory. That amount nearly equals hyperscalers’ total capex for all of 2025.
Two points trouble me:
First, equity and debt markets are already sending negative signals to participants—cash reserves are depleting, and price-to-sales and price-to-free-cash-flow multiples are spiking.
Second, mounting cost pressure may slow capex growth—or halt it earlier than expected. Per my estimate, signs of pausing will begin appearing on earnings calls around mid-2027.
I believe this second point will confront memory manufacturers by end-2026—much sooner than many anticipate.
Going forward, the top question on earnings calls will be component pricing—especially memory—and how it compresses spending budgets. I don’t believe hyperscalers will ignore this and continue ramping capex without reservation.
This is merely my view.
Chipmakers Are Already Seeking Ways to Reduce Memory Use
AMD, NVIDIA, and Google are already pursuing memory-optimization strategies.
NVIDIA’s next-generation Rubin NVL72 rack may cut CPU-side SOCAMM DRAM from ~55TB per rack to ~28TB—a near 50% reduction. This makes sense: VR200’s BOM shows memory costs rose 435% versus GB300.
Source: Morgan Stanley
SOCAMM isn’t HBM—but the same cost pressure drives cost-saving innovation, whether AMD pools memory via MEXT (making flash behave like DRAM) or simply slashes SOCAMM DRAM.
Chipmakers actually have even less flexibility: they’re already paying premium prices for HBM—and adding SOCAMM costs on top? That hurts. They’re getting squeezed from both sides.
Memory Remains Cyclical—The Turning Point Is Mid-2027
Finally, let’s discuss memory’s cyclical nature.
I reject the claim that “memory is no longer cyclical.”
Even if everything I’ve argued proves wrong—and capex truly remains strong for a decade—you’d still inevitably hit the boom-and-bust cycle. Anyone disputing this must assume memory demand grows annually *and* hyperscaler spending never enters a cycle—which is impossible.
Source: SEMI
These aggressively expanding manufacturers are betting on sustained capex growth (unlikely given current trends) and persistent memory demand (which itself depends on continued capex growth).
My projection: DRAM pricing peaks in 2027.
SK Hynix gross margin ~80%; Micron ~78%–80%; Samsung ~70%–75%.
Price curves plateau—even as capacity remains tight—around February or March. Then, around mid-2027, you’ll detect signals of slowing or halted capex growth.
I expect most memory stocks to begin reversing gains at this inflection point, as investors front-run the coming gross margin compression.
By 2028, more capacity comes online (supply remains tight), but demand expectations weaken—pushing gross margins down to just above 60%. From 2028 to 2030, additional capacity ramps up, easing supply constraints, while capex fails to meaningfully grow. I predict the true bust occurs in this phase—prompting significant stock price retracements beginning in late 2027.
Everyone believes memory strength will last through end-2030. My forecast: gross margin compression begins mid-2027—and many memory stocks’ rallies will reverse.
That said, if hyperscalers announce significantly stronger capex for 2028 on their 2027 earnings calls, this article will have been entirely misguided—and I’ll look foolish. Time will tell. But I believe this is the path memory will take.
Why I’m Not So Optimistic
I’m less optimistic about memory than others—for several reasons:
Memory vendors are excessively greedy on margins; I maintain memory remains cyclical—the “no-cycle” thesis rests entirely on the assumption that capex never enters a cycle; chipmakers’ pursuit of memory-saving solutions proves they’re already fatigued by high costs; CFOs’ cash is nearly 100% consumed by capex, and memory will account for 40% of 2027 costs—further debt issuance and equity dilution are no longer sustainable.
The only positive outcome would be a sudden wave of aggressive supply dumping that collapses the Big Three’s memory prices—enabling the same capex to yield far greater output.
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
@pequityresearch
