Why Memory Prices Are Rising Fast, and What It Means for SSDs, SD Cards, and USB Drives

On April 29, 2026, Meta, Microsoft, and Alphabet all held earnings calls on the same afternoon. When the three companies gave their 2026 capital expenditure guidance, they used surprisingly similar language.

Meta wrote directly in its 8-K that the guidance reflected the company’s expectation of higher component prices this year. An 8-K is a major-event disclosure filed by public companies with the U.S. SEC, and the wording is usually extremely conservative. It is not common to cite a specific cost item as justification for raising capex guidance.

Microsoft CFO Amy Hood was even more explicit on the earnings call. She said that of the company’s roughly $190 billion in 2026 capital expenditure, about $25 billion was related to the impact of rising component prices. Alphabet raised its guidance from $175-185 billion to $180-190 billion, and also signaled that 2027 capex would increase significantly.

A week earlier, Amazon had already confirmed that its 2026 spending level would remain around $200 billion. Adding the four together, using the top end of their guidance, hyperscaler capex in 2026 comes to roughly $725 billion. Using the midpoint, it is closer to $712.5 billion.

For comparison, the four companies spent about $250 billion in total in 2024, and $400-450 billion in 2025. In two years, the annualized pace of capex has nearly tripled, which is already shocking. But the bigger issue the market should be asking about is that component price increases are now being named directly. When CFOs from two companies point to the same cost item in the same week, it usually means the issue is large enough that they feel they have to disclose it.

Put this signal together with Samsung and SK Hynix’s Q1 earnings from the same week, and the picture becomes clearer. On the demand side, cloud companies are talking about component inflation. On the supply side, memory manufacturers are reporting rapidly rising prices and profits. Both sides are pointing to the same thing: the 2026/2027 memory market may be tighter than the market originally expected.

The Real Meaning Behind More Than $700 Billion

Here are the four companies’ numbers:

There are a few details worth unpacking.

Meta’s $10 billion increase was mostly attributed by Susan Li to component costs, especially memory. Li again said the company continues to underestimate compute demand, a theme she has repeated since 2024. In previous instances, that statement was also accompanied by upward capex revisions. Meta’s 8-K clearly stated that the $10 billion increase mainly reflected component price increases, while additional data center capacity was treated as a secondary factor.

Microsoft’s $190 billion figure was the first time Amy Hood gave a concrete calendar-year 2026 number. She specifically pointed out that about $25 billion was related to higher component prices and finance leases. Short-lived assets, meaning fast-depreciating hardware such as GPUs and CPUs, account for roughly two-thirds of capex. Read the other way: within short-lived assets, the cost movement in memory has become large enough that it needed to be disclosed on the earnings call.

At Alphabet, Sundar Pichai said the company remains compute-constrained in the short term. Combined with cloud backlog reaching $460 billion, nearly doubling quarter over quarter, the signal is clear: demand remains meaningfully stronger than current supply. CFO Anat Ashkenazi described 2027 capex as increasing significantly, which Wall Street broadly interpreted as preparation for another upward revision.

Amazon was the only one of the four that did not raise guidance that week. Its $200 billion figure had already been given in February. AWS grew 28% quarter over quarter, reaching a 15-quarter high, but Andy Jassy also said compute supply has not kept up with demand.

Looking at the four together, there is a structural shift. Over the past year, cloud providers usually explained capex increases as “demand exceeded expectations, so we need to build more data centers.” This time, “the same amount of compute now costs more to buy” has also become a major explanation. Meta has officially said its $10 billion increase mainly reflects component price inflation, with new data center costs being secondary. Microsoft’s $25 billion attribution is even more direct.

This is not a story about a strong business cycle. It is a story about cost structures getting out of control. CFOs are putting this into 8-K filings because a single supply-chain bottleneck has become large enough to affect the credibility of financial guidance.

Supply Is Confirming Demand

If the demand side is showing stress, does the supply side confirm it? Samsung and SK Hynix’s late-April Q1 2026 earnings give the answer.

SK Hynix Q1 2026: revenue of KRW 52.5763 trillion, operating profit of KRW 37.6103 trillion, and operating margin of 72%. For comparison, TSMC’s operating margin during the same period was around 49%. A memory company having a higher operating margin than the world’s top foundry is extremely rare. SK Hynix said on its earnings call that its full-year 2026 DRAM, NAND, and HBM capacity had already sold out, and HBM orders had already been booked into the second half of 2027. Customer demand for HBM4 is higher than supply capacity for the next three years.

Samsung Q1 2026: consolidated revenue of KRW 133.9 trillion and operating profit of KRW 57.23 trillion, up 756% year over year, with the memory division up 250%. The company said memory ASPs continued rising, HBM4 had already started shipping for NVIDIA’s Vera Rubin platform, and HBM4E samples were expected to begin in May. In other words, even without relying on any single third-party valuation, Samsung’s own wording is enough to show that the memory pricing environment is strengthening quickly.

Third-party price trackers support this. TrendForce data shows DRAM contract prices rose 90-95% quarter over quarter in Q1, while NAND rose 55-60%. Counterpoint estimates DDR5 will rise another 40% in Q2 and 20% in Q3.

When supplier operating margins are breaking historical records, prices are rising quickly, and capacity is sold out into 2027, this no longer looks like a normal cycle. Since the 1990s, the standard memory cycle has usually lasted three to four years, moving from oversupply to tight supply and then to price correction. This time looks more like demand has pushed supply all the way to its limit. SK Hynix’s investor relations language has moved from “favorable pricing environment” to “customer demand exceeds supply capacity,” which is much more aggressive than the conservative wording memory companies usually use.

The component price increases named by the demand side and the price increases and tight capacity seen by the supply side are two sides of the same issue. The next question is: why is this price increase so extreme?

How HBM Is Stretching the Memory Market

To understand the 2026 memory market, start with NVIDIA’s Vera Rubin rack.

A single Rubin GPU package uses eight HBM4 stacks. A full NVL144 rack has 72 GPU packages, or 576 HBM4 stacks in total. By comparison, in the H100 era, a DGX H100 had eight GPUs with five HBM3 stacks each, or 40 stacks for the full system. From H100 to a Rubin rack, HBM usage increases by 14 times.

That multiple directly affects system cost. Industry estimates put the price of one HBM4 stack at around $500, roughly 40% higher than HBM3E. The HBM bill of materials for one Rubin rack is close to $300,000. A GB300 rack using 12-layer HBM3E and 576 stacks also has HBM material cost close to $200,000. SemiAnalysis’s key conclusion in its GB300 bill-of-materials analysis was simple: HBM has become the largest single cost item inside the GPU package.

From a system perspective, HBM’s share of AI accelerator material cost has moved from below 20% in the Hopper era to 30-40% in the Blackwell/Rubin era. This structural change is the root reason cloud providers are naming component price increases in 8-K filings. They are not just buying more expensive GPUs; they are buying the increasingly expensive HBM surrounding the GPU.

But that is only half the story. HBM uses advanced-process wafers and shares capacity with regular server DDR5. The three major manufacturers prioritize wafers for HBM because HBM gross margin has previously exceeded 80%. Whatever wafers remain can be used for DDR5. The problem is that hyperscalers are not only buying GPUs with HBM; their server motherboards also need server DDR5. As HBM consumes wafer capacity, DDR5 becomes severely short because its capacity is squeezed.

Then Q1 2026 brought a turning point: the market began discussing whether DDR5 gross margin could surpass HBM. SK Group chairman Chey Tae-won and Micron’s CEO both implied in different settings that server DDR5 gross margin in 2026 could reach 80%, while HBM may fall back toward 60% due to depreciation pressure from new-generation capex. The shortage in general DRAM appears to have reached the point where, beyond the original “HBM is the golden goose” logic, manufacturers have found another golden egg in server DDR5.

This shift matters a lot for market structure. If only HBM were rising, cloud companies could still absorb the cost through product mix adjustments. But if both HBM and server DDR5 are rising, and DDR5 is rising even faster because of wafer crowding, hyperscalers have very little room to avoid it. Meta putting component inflation directly into its 8-K likely reflects more than one product becoming expensive. Both GPU racks and general servers are under pressure.

There is also a second-order beneficiary that most people do not notice. When original manufacturers allocate almost all wafer capacity to HBM and advanced server DDR5, consumer DDR4 and specialty DRAM face structural shortages. That extends order visibility for Taiwanese module makers. The revenue momentum at Nanya, Winbond, and ADATA in the first half of 2026 is essentially a byproduct of hyperscaler capex pressure.

This Price Increase Is Not Only About HBM

Zoom out further, and SanDisk’s earnings, released on April 30, add another important piece.

SanDisk Q3 FY2026 revenue reached $5.95 billion, up 97% quarter over quarter. The company’s press release was clear about the reason: the outperformance came from two things, strong data center business growth and higher prices. Data center revenue reached $1.467 billion in the quarter, up 233% quarter over quarter and 645% year over year. Non-GAAP gross margin jumped from 51.1% in the prior quarter to 78.4%.

This is not an HBM earnings report, so it cannot directly prove that HBM is even more short. But it supports another point very well: AI data center purchasing pressure is pushing up HBM, DRAM, data center NAND, and enterprise SSDs at the same time. Upstream buyers say components are getting more expensive, and downstream suppliers say revenue beat expectations because prices were higher. The two sides line up.

More importantly, SanDisk talked not only about pricing but also business model. The company said that by the end of Q3 it had signed three new business-model agreements, and in Q4 it signed two more, all with clear financial commitments in multi-year customer partnerships. This means large customers are no longer just fighting for spot supply. They are moving upstream into long-term agreements to lock supply, pricing, and capacity.

Put that together with the capex guidance from the four major cloud providers, and the meaning becomes more complete. This upward cycle is not just a single-point HBM story, and not just a DRAM cycle story. The entire data center memory and storage chain is moving toward higher value, higher constraints, and stronger supplier bargaining power.

Why This Is Different From 2018

Every time memory prices rise this aggressively, older investors ask the same question: didn’t the last supercycle look like this too, before prices collapsed by half?

Go back to 2017-2019. DRAM average selling prices rose 165% from the 2016 low to the August 2018 high. Server DDR4 reached a historical high in Q3 2018. Then the collapse started in Q4 2018. According to TrendForce, DRAM ASPs fell about 20% in 2019 overall, while server-grade DDR4 is commonly cited as falling about 50% from the peak. SK Hynix’s 2019 operating profit fell 87% year over year, Samsung’s semiconductor operating profit fell to less than half of the 2018 peak, and Micron’s full-year revenue fell 23%. None of the three posted an annual loss, but the industry term “memory winter” was born at that time.

The core reason for that collapse was that the three major manufacturers over-expanded at the same time in 2017-2018. Samsung started its Pyeongtaek P1 fab, a $30 billion investment. SK Hynix built M15 and expanded its Wuxi DRAM fab, planning KRW 46 trillion over five years for three new fabs. Micron expanded its Hiroshima DRAM fab. That capacity was set to come online from the second half of 2018 to the first half of 2019, just as smartphone demand saturated, cloud providers entered a brief capex digestion period, and the PC market weakened. Those three demand headwinds hit together, and oversupply immediately caused the market to roll over. That cycle followed the textbook pattern: price increase, full expansion, oversupply 18 months later.

The 2026 cycle has three structural differences that make “over-expansion” harder to repeat.

First, orders are coming before capacity expansion. SK Hynix’s Q1 earnings call announced a significant increase in 2026 capex, including an additional $15 billion for the Yongin cluster and M15X. But that is built on the premise that three years of HBM4 orders have already sold out. In other words, this capex is for fulfilling signed contracts, not betting on the future. In the 2018 cycle, a meaningful portion of expansion was driven by price increases and a bet that the cycle would continue. This time, customer long-term agreements mean that even if demand weakens, contracts will support the manufacturers’ baseline.

Second, HBM is customized capacity. HBM is not like ordinary DRAM that can be switched to other uses at any time. Each HBM stack is tied from design to validation to a specific customer’s accelerator, whether NVIDIA, AMD, Google TPU, or AWS Trainium. SK Hynix’s HBM4 designed for Vera Rubin cannot suddenly be sold into phones or PCs. Capacity is physically bound by contracts. It does not have the same flexibility as 2018, when a supplier could add 20% more wafers and redirect output if prices changed.

Third, the three major manufacturers are speaking with one voice. In 2018, Samsung was famously aggressive on expansion and deliberately increased capacity near the price peak to grab share. This time, Samsung’s language is more conservative, saying it will “proactively capture demand opportunities.” SK Hynix emphasizes caution. Micron emphasizes disciplined capex. All three are stressing discipline. No one is currently stepping out to expand aggressively for share.

Taken together, this time’s supply discipline is not because manufacturers have suddenly become smarter. It is because the physical nature of HBM orders prevents them from building recklessly. In 2018, producers lost control because DRAM was a commodity product and extra wafers could still find a market. HBM in 2026 does not have that freedom. This is not the industry becoming more rational by choice; it is the market structure forcing rational behavior.

The Hole Inside EPS Estimates

Finally, the article brings the focus back to the cloud providers themselves.

Meta’s 8-K raised capex by $10 billion, almost entirely because of component price increases. Microsoft carved out $25 billion of its $190 billion capex for component inflation. Together, this means a meaningful portion of these two companies’ 2026 capex increase is not buying new compute. It is paying more money for the same amount of compute.

At the valuation level, this is an easy flaw to overlook. In the past, when the market saw hyperscaler capex guidance raised, the logic was usually: “higher capex = stronger-than-expected demand = future revenue growth.” But this time, part of the capex increase is “same compute volume, higher unit cost.” In other words, capex is still rising, but the compute output bought by each dollar is being discounted.

On April 30, JPMorgan downgraded Meta from overweight to neutral. Its model estimated Meta’s 2027 capex at $202 billion, up 42% year over year, with 2026 free cash flow turning negative at minus $4 billion and 2027 free cash flow at minus $24 billion. The downgrade was not because Meta’s advertising business had a problem. The core logic was that Meta’s capex discipline is becoming harder to predict. When component inflation turns capex movement into an uncontrollable variable, traditional discounted cash flow assumptions become much more fragile.

Looking into 2027, this becomes a serious question. If hyperscaler capex budgets keep being eaten by component price increases, they face two awkward choices: keep raising capex to maintain the same compute expansion pace, putting more pressure on free cash flow; or slow compute expansion and accept forced demand cooling, which means lower cloud growth expectations.

Who Breaks First?

The sharpest question is not whether component prices will keep rising. It is who can no longer absorb it first.

The first breaking point is usually financial, not technical. The hyperscalers with the deepest cash flow may not be the first to fall. More likely, the first group to suffer will be those with shallower cash flow who still want to compete for high-end components. This includes second-tier cloud companies, AI startups relying on external funding to build infrastructure, and enterprise AI projects whose monetization speed cannot keep up with compute depreciation.

The second breaking point is unit economics. As long as adding another GPU rack, more HBM, and more server DDR5 can generate enough revenue and gross profit, CFOs will keep buying despite the pain. But if each additional unit of memory no longer produces corresponding cash flow, spending starts to slow. The real breaking point is when the same amount of money buys compute that is no longer enough to justify continued investment.

The third breaking point is organizational tolerance. When the market realizes that a growing portion of capex is merely paying for more expensive components, without bringing proportional gains in compute, customers, or free cash flow, management’s tone will change first. At that point, it is not necessarily that the company has no money. More often, the board, shareholders, and analysts are no longer willing to give the same level of patience.

So the first to break may not be Meta, Microsoft, or Alphabet, which sit at the top of the supply chain and have the most cash. It is more likely to be two groups: outsiders trying to compete for supply without long-term contracts or bargaining power, and internal departments whose monetization is slow but whose compute burn is high. The first group may simply be priced out. The second group is more likely to be cut by internal capital allocation.

If one had to pick a giant that the market will question first, Meta is likely near the front. But that means it may be the first to face questions about capital allocation and forced public trade-offs, not necessarily the first that cannot afford it.

The deeper issue is that the core bottleneck in this supply chain is concentrated among Samsung, SK Hynix, and Micron. The first two are in South Korea, and Micron is in the U.S. When 10-15% of incremental capex from U.S. hyperscalers flows directly into the operating profits of these memory manufacturers, a supply chain once seen as a pure commodity business is gradually becoming one of the most important bargaining points in the AI compute race.

The full version of this story will not be visible until 2027. But the earnings reports from the last week of April 2026 have already written the main plotlines: HBM and server DDR5 pricing pressure does not look likely to disappear in the short term; cloud provider capex will likely continue to be eroded by component costs; and the high profitability of the three major memory makers has not yet shown clear signs of easing.

When Meta writes component price increases into its 8-K, and Microsoft and Alphabet both acknowledge compute and component pressure in different ways during the same week, this is still only the beginning of the story for the next two years.