Every unit of DRAM and High Bandwidth Memory that Samsung, SK Hynix, and Micron plan to produce in 2027 is now spoken for — and on Wednesday, August 5, the last credible escape route closed: Apple's months-long attempt to qualify China's CXMT as a fourth DRAM supplier collapsed after CXMT quoted prices at or above what Samsung and SK Hynix already charge, according to reporting by Sammy Fans citing semiconductor industry sources. The three manufacturers that control more than 95% of global DRAM production have completed 2027 allocation negotiations months ahead of schedule, rationing most buyers to 60 to 70% of their requested volumes and requiring advance deposits — in many cases full cash prepayment — as a condition of any supply agreement at all, according to sources cited by DigiTimes and corroborated by Sammy Fans. If your organization does not currently hold a long-term supply agreement with one of the three, what Apple discovered this week is now your answer: there is no fourth supplier, and the 2027 market has no room for latecomers.

What "Sold Out" Actually Means in Practice

The 2027 memory sellout is not an announcement — it is the result of a negotiating process that industry insiders say concludes every year in July and August, when OEM server makers, PC makers, and AI labs lock in their following-year allocations. What is historically unprecedented is that those negotiations are now complete, with zero capacity remaining, according to industry insiders cited by TechPowerUp.

The commercial terms attached to these allocations mark a structural shift in who holds power in the semiconductor supply chain. Vendors are requiring multi-year long-term agreements spanning three to five years, with advance payment deposits regardless of whether formal agreements are in place, according to Sammy Fans August reporting. Most buyers are receiving only 60 to 70% of their originally requested volumes, a rationing level that Meritz Securities senior analyst Kim Sunwoo estimates could tighten further — falling to 60% of total demand — as 2027 progresses, according to TechTimes prior coverage.

ADATA Chairman Simon Chen publicly confirmed that all three vendors' 2027 DRAM capacity is sold out, with consumer electronics bearing the sharpest allocation cuts, per Sammy Fans industry reporting.

Apacer CEO C.K. Chang offered the most concrete single figure: DRAM chip supply from major manufacturers to independent module makers — the companies that buy memory chips and turn them into finished DIMMs and storage products — could fall by more than 70% year-on-year in 2027, Chang said at an investor conference on July 24, 2026, according to Tom's Hardware Apacer reporting. His point was not that global DRAM production will drop by that amount. It is that manufacturers are increasingly reserving output for HBM, AI server memory, and large customers with direct long-term agreements — leaving downstream module makers competing for whatever remains.

HBM's Physics Are Why Latecomers Cannot Be Accommodated

Understanding why no amount of money or urgency can fix a latecomer's 2027 position requires a brief look at why HBM creates scarcity rather than solves it.

High Bandwidth Memory is built by stacking 12 to 16 individual DRAM dies vertically, connecting each layer to the next through thousands of microscopic copper pillars called through-silicon vias (TSVs), and then bonding the entire stack directly adjacent to the GPU on a silicon interposer using advanced packaging technology called CoWoS (Chip-on-Wafer-on-Substrate), as detailed in TechTimes technical coverage. Each step — die grinding to micron-level thinness, TSV drilling, copper deposition, interposer bonding — consumes fabrication capacity without proportionally returning memory bits. Micron has publicly disclosed a 3-to-1 conversion ratio, meaning each wafer running HBM displaces approximately three wafers of conventional DDR5, per TechTimes reporting on Micron's disclosures. TrendForce puts the penalty even higher, at roughly 4 to 1, according to TrendForce's July 30, 2026 memory outlook.

HBM also generates three to five times more revenue per wafer than commodity DDR5, per TrendForce July 2026 analysis. At that ratio, every fab on Earth — existing ones and new ones currently under construction — faces the same optimization: HBM pays more per wafer than conventional DRAM, so HBM gets the wafers. New capacity coming online in 2027 will not change this calculation. It will face it fresh.

Stanford University's Distributed AI Memory database confirms there is no public spot market for HBM at all — the project notes that HBM trades only on confidential contracts to accelerator makers, meaning the question "can I buy HBM on the open market?" has a clean answer: no, and that has always been the case.

Apple Tried the Only Available Alternative and Hit a Wall

Apple's position illustrates the constraint with unusual clarity because Apple is the buyer with the most leverage. Its scale, its cash reserves, and its long history of pitting memory suppliers against each other give it negotiating options that mid-tier manufacturers cannot access.

Beginning in mid-2026, Apple explored qualifying CXMT — China's fourth-largest DRAM producer and the only manufacturer outside the Samsung-SK Hynix-Micron oligopoly with meaningful production volume — as an alternative supplier for the LPDDR5X memory used in iPhones and Macs, according to MacRumors July 2026 reporting. The company sought US government approval for the arrangement, given that CXMT appears on the Pentagon's list of Chinese military companies. Senators Jim Banks, Chuck Schumer, and five colleagues urged Apple to abandon the talks, noting that CXMT and YMTC — another Chinese memory supplier Apple was evaluating — are both designated Chinese military entities, as MacRumors covered in July.

The effort collapsed on August 5, 2026. CXMT declined to offer Apple a price discount, quoting rates at or above what Samsung and SK Hynix already charge, according to Sammy Fans citing industry sources. The reason is structural: Huawei, Xiaomi, and other Chinese OEMs have locked CXMT into long-term, high-price supply agreements, giving CXMT no economic reason to chase Apple's business at a discount. With US export restrictions shaping how CXMT operates, securing domestic Chinese commitments matters more to the company than winning an American customer that would require government clearance and might be revoked by the next administration.

Apple has confirmed publicly that it is paying significantly more for memory each quarter and expects to continue doing so. If the world's most cash-rich technology company — one that has historically used the threat of Chinese alternatives to extract pricing concessions from Samsung and SK Hynix — cannot find competitive supply outside the oligopoly, the range of options for buyers with less leverage is narrow.

What the Two-Tier Market Means for Everyone Without an LTA

The 2027 sellout has created a buyer universe split into two groups with no precedent in semiconductor history.

Hyperscalers — Microsoft, Google, Amazon, Meta, and Nvidia's major customers — secured multi-year long-term agreements that lock in pricing and allocation for three to five years. OpenAI's Stargate project alone reserved supply equivalent to approximately 40% of global DRAM output, according to reporting cited in a Wikipedia summary of the shortage, at 900,000 wafers per month. These buyers are not competing for 2027 supply. Their 2027 supply is contractually guaranteed and was priced before the current shortage peak.

Everyone else — independent module makers, mid-tier PC manufacturers, regional cloud providers, enterprise IT departments that planned to buy on quarterly spot contracts — faces a market that Tom's Hardware described in March 2026 as having shifted to an "hourly pricing" model. Spot pricing fluctuates on an intraday basis. Small and medium buyers are, in the characterization of a Phison CEO quoted in Tom's Hardware, "fighting for survival," with at least one foundry reportedly demanding three-year cash prepayment upfront.

The federal antitrust class action filed June 25, 2026, in the U.S. District Court for the Northern District of California (Garciaguirre v. Samsung Electronics, No. 5:26-cv-06345) argues that this market structure is not merely the result of AI-driven demand but of coordinated supply restriction. The 17 plaintiffs — 14 individuals and three small PC-building businesses including Troy's Computers LLC and JB Tech Solutions LLC — allege that Samsung, SK Hynix, and Micron used the transition to HBM as a pretext to accelerate the wind-down of DDR3 and DDR4 production in violation of Sherman Act Section 1, per Tom's Hardware reporting on the suit. The three companies have not yet responded in court; the allegations are unproven.

What gives the antitrust argument its historical weight: Samsung and SK Hynix each entered guilty pleas on Department of Justice price-fixing charges in the late 1990s and early 2000s, with SK Hynix paying $185 million in fines in April 2005 and the combined cartel generating more than $730 million in US penalties, as confirmed by TechTimes prior reporting. The complaint cites that history explicitly as evidence of a pattern. Whether a court agrees remains to be seen.

What Buyers Who Missed the Window Can Actually Do

The procurement playbook that worked in 2025 — quarterly spot purchases, supplier competition on short cycles, holding off for the next price dip — reflects a market that no longer exists for 2027.

For organizations that hold no long-term agreement and need DRAM for server builds, AI deployments, or product manufacturing, the operational options are specific and limited.

The secondary market for pulled enterprise memory — DDR4 and DDR5 modules recovered from data center equipment reaching end of life — has emerged as one of the few sources of conventional DRAM that is not locked behind a multi-year contract. BuySellRam's analysis from June 2026 describes this market as carrying "real value" for mid-size data centers, resellers, and integrators who cannot access the new-supply pipeline. Lead times on secondary market memory are shorter than on direct manufacturer supply, though quality verification is essential and volumes are limited.

Findchips, which aggregates authorized distributor inventory across more than 900 sources, recommended in June 2026 that buyers extend purchase orders by quarter — 90 to 120 days for all DDR5 and server DRAM and qualify at least one secondary authorized supplier per critical memory part. This is a contingency posture — not a substitute for a manufacturing allocation — but it provides some buffer against the weekly volatility Meritz Securities projects for 2027.

For cloud and AI infrastructure buyers who missed HBM allocations, the practical alternative is cloud API substitution: routing workloads through hyperscaler inference APIs rather than deploying owned GPU clusters. Since HBM has no public spot market and is sold exclusively to accelerator makers on confidential contracts, an organization without an existing HBM allocation has no mechanism to acquire it in 2027 at any price. Designing inference pipelines around available cloud APIs — which are themselves backed by the hyperscaler HBM allocations already in place — is the functional alternative.

Architecture-level memory footprint reduction is a third option. Google Research released TurboQuant in March 2026, a KV-cache memory compression tool that it claimed achieves six times lower memory consumption in tested local LLMs. The technology is a drop-in enhancement for existing inference pipelines and would reduce the amount of DRAM an inference server requires per workload — though memory markets briefly sold off on the announcement before recovering, reflecting genuine uncertainty about how much compression will actually reduce hardware demand at scale.

Why NAND Still Has a Window — and Why It Is Narrowing

NAND Flash tells a related but distinct story that matters for buyers planning 2027 storage procurement. While DRAM and HBM are fully allocated, NAND from Samsung, Micron, and SanDisk for 2027 is also approaching full booking, according to the same DigiTimes-sourced reporting, with Kioxia and SK Hynix expected to finalize their own 2027 allocations by the end of August 2026, per Sammy Fans semiconductor coverage.

The difference is trajectory. TrendForce's July 30, 2026 memory outlook projects the NAND sufficiency ratio will turn positive in the second half of 2027, marking a transition from supply-constrained to balanced conditions, as next-generation high-layer NAND — SK Hynix's V8 at 321 layers, Samsung's V9 at 290-plus layers, and Kioxia's BiCS10 at 332 layers — ramps toward production, as detailed in TrendForce's July 30 report. Unlike DRAM, where HBM's wafer-conversion penalty makes every new wafer face the same HBM-vs-DDR5 optimization problem, NAND's bit growth through layer migration creates supply relief without requiring new fab capacity. Servers now account for more than 40% of total NAND bit demand, and the enterprise SSD segment still offers negotiating leverage that the conventional DRAM market has surrendered, per TechTimes prior coverage.

That window is not guaranteed and may be narrow. TrendForce explicitly warns that accelerated agentic AI adoption — systems that run continuous iterative reasoning cycles rather than discrete queries, and therefore require persistent high-bandwidth memory loads — could absorb the incoming NAND supply through KV-cache overflow demand before the anticipated surplus materializes. Enterprise buyers with NAND procurement decisions pending have a limited window to lock favorable multi-year terms before the market recognizes the shift, according to TrendForce's July 30, 2026 outlook.

When Relief Could Arrive — and What "Relief" Will Mean

The capacity expansion programs underway are the largest in semiconductor history, but their timelines establish a ceiling on when buyers without current allocations can expect normal market conditions to return.

South Korea's government announced alongside Samsung and SK Hynix the largest industrial investment in the country's history: approximately $518 billion for four new memory fabs, split between the two companies, plus $52 billion for a dedicated HBM packaging hub. Micron has committed approximately $250 billion to US manufacturing, including a megafab in Clay, New York and expansion of its Idaho facilities, per TechTimes Micron investment reporting.

The problem is timing. New fabrication facilities require 12 to 18 months to bring online after construction completes, and then must ramp gradually. SK Hynix's M15X facility in Cheongju entered wafer production in early 2026 and is ramping incrementally. Samsung's P5 megafab in Pyeongtaek is expected to reach mass production in the latter part of 2028. Micron's Idaho facilities are targeted for first wafer output in the second half of 2027, with meaningful consumer-grade volume arriving later.

But there is a further complication the expansion headlines obscure. Micron has confirmed that 16 non-cancelable Strategic Customer Agreements with hyperscalers — representing $22 billion in committed deposits — run through 2030, contractually pricing future capacity at today's elevated rates, per TechTimes Micron coverage. New supply that arrives in 2027 and 2028 will go first to those buyers. The question of how much remains for buyers without legacy agreements is one none of the three manufacturers has answered publicly.

SK Hynix CEO Kwak Noh-jung told Reuters in July 2026 that 2027 will be the worst year in the industry's history from the supply perspective. Micron CEO Sanjay Mehrotra has stated the company has no "line of sight" on when memory supply will catch up with increasing demand. AMD VP David McAfee said at Computex 2026 that DDR5 prices would not normalize until 2028. Intel CEO Lip-Bu Tan said at a conference in February 2026 that there is "no relief until 2028." Counterpoint Research identifies the fourth quarter of 2027 as the earliest possible inflection point.

For organizations planning capital expenditure around the assumption that DRAM prices would recover toward 2025 levels: IDC has described the current reallocation as "potentially permanent." When new capacity does arrive, manufacturers will face the same economic calculation that drove the original reallocation — HBM earns three to five times more per wafer than consumer DRAM — and there is no structural reason to expect them to make a different choice. The buyers who secured long-term agreements in 2024 and 2025 are not the victims of this market. They are the ones who saw the math clearly, early.

---

Frequently Asked Questions

Can any buyer still get DRAM and HBM supply for 2027?

Buyers who already hold long-term agreements with Samsung, SK Hynix, or Micron are receiving allocations — though at rationed volumes of 60 to 70% of what they originally requested, according to industry sources cited by Sammy Fans. Buyers without existing agreements face a fully-booked market. HBM has no public spot market and is sold exclusively on confidential contracts to accelerator makers, according to Stanford University's memory pricing database. Conventional DRAM is available in secondary markets (pulled enterprise memory from retired servers), through authorized distributors at volatile spot prices, and in small quantities from the global broker market — but not at the volumes or prices that characterized 2024 procurement.

Why could Apple, with all its purchasing power, not get a better deal from CXMT?

CXMT declined to offer Apple a pricing discount because its production capacity is already committed to long-term agreements with Huawei, Xiaomi, and other Chinese OEMs that have locked in supply at high prices, according to Sammy Fans citing semiconductor industry sources. CXMT also appears on the Pentagon's list of Chinese military companies, which means Apple needed US government clearance before it could use CXMT parts — and the political environment, including Senate pressure from Senators Banks and Schumer, made that clearance uncertain. The underlying economics are also unfavorable: CXMT's cost-per-bit trails the established oligopoly by more than 30%, making it unable to undercut on price while its capacity is spoken for.

What is HBM's wafer-conversion penalty and why does it matter for buyers outside the AI industry?

HBM requires approximately three to four times the wafer area to produce the same number of memory bits as conventional DDR5 or LPDDR5X, per TrendForce's July 2026 memory analysis. Because Samsung, SK Hynix, and Micron produce both HBM and conventional DRAM in the same facilities, every wafer allocated to HBM removes the equivalent of three to four gigabytes of standard memory from the market. HBM also generates three to five times more revenue per wafer, so manufacturers have a strong and durable economic incentive to keep making this tradeoff. The consequence for buyers outside the AI industry — smartphone makers, PC manufacturers, gaming console producers, enterprise IT departments — is that their supply does not recover when demand for conventional DRAM rises. It recovers only when new fabrication capacity arrives and manufacturers face a different allocation decision. New capacity is not expected to deliver meaningful output before the second half of 2027 at the earliest, with most analyst estimates pointing to 2028.

Is there anything an IT buyer or procurement team can do now if their organization has no 2027 DRAM allocation?

Yes, though the options are narrower than they were a year ago. Secondary market sources — pulled enterprise DDR4 and DDR5 from data centers reaching end of life — offer one of the few DRAM channels not locked behind a long-term agreement, according to BuySellRam's June 2026 analysis. Extending purchase order coverage to 90 to 120 days across authorized distributors is a standard mitigation from Findchips' buyer playbook. For AI and inference workloads, cloud API substitution — routing compute through hyperscaler inference endpoints rather than building owned GPU clusters — is the practical alternative, since HBM cannot be acquired at any price on the open market. Workload memory footprint reduction through compression technologies like Google's TurboQuant, released March 2026, can reduce per-inference DRAM demand. None of these options restores the economic position of an organization that missed the 2025 window to lock long-term agreements; they are operational mitigations, not strategic equivalents.