rare earth elements

Three Corporations Ration the Physical Substrate of Global Computation, and No Government Authorized the Triage

The Death of the Commodity

For decades, DRAM was the commodity nobody watched. A gigabyte was a gigabyte. Price followed volume, volume followed demand, and the market behaved like grain futures—cyclical, predictable, occasionally volatile, ultimately boring. That world ended in 2025. TrendForce data showed DRAM contract prices surging 171.8 percent year-over-year by the third quarter, consumer DDR5 kits doubled in retail price within four months, and total contract prices including HBM were projected to rise 50 to 55 percent in a single quarter. The industry calls this a “memory supercycle.” The term flatters what is actually happening. A supercycle implies natural market dynamics—supply tightening, prices rising, capacity expanding, equilibrium restoring. This is not a cycle. It is a structural reallocation of the physical substrate of computation from the many to the few.

The commodity model assumed fungibility. A gigabyte of DRAM going into a desktop module was interchangeable with a gigabyte going into a server. That assumption is dead. The gigabyte being stacked into a High Bandwidth Memory chip for an AI accelerator competes for the same silicon wafer starts as the gigabyte destined for a laptop, but the AI customer pays five to ten times more per unit. EE Times reported that advanced server-grade memory modules now carry profit margins as high as 75 percent, far exceeding the thin margins on consumer PC modules. When wafer capacity is finite and one buyer outbids all others, the market does not self-correct. It triages.

The fallacy at the center of this crisis is what this paper calls the Free Market Memory Myth—the assumption that DRAM pricing follows open-market dynamics when it is governed by a structural oligopoly whose wafer-allocation decisions are driven by AI demand capture and geopolitical weaponization, not consumer economics. No antitrust framework, no trade policy, and no defense doctrine currently accounts for a world in which three corporations ration the physical substrate of computation. That absence is the convergence gap.

Three Boardrooms, One Chokepoint

The global DRAM market is controlled by three manufacturers. As of the third quarter of 2025, Counterpoint Research reported SK Hynix at 34 percent, Samsung at 33 percent, and Micron at 26 percent of DRAM revenue—a combined 93 percent. China’s CXMT holds roughly 5 percent. Everyone else is rounding error. In High Bandwidth Memory specifically, the concentration is absolute: SK Hynix held 57 percent, Samsung 22 percent, and Micron 21 percent of HBM sales in Q3 2025. There is no fourth supplier in HBM. There is no alternative.

These three companies are not a cartel in the OPEC sense. They do not coordinate pricing in a smoke-filled room. They are a structural oligopoly in which each actor’s rational self-interest—maximize HBM margin—produces a collective outcome—consumer and sovereign scarcity—that no single actor chose but none will reverse. The financial incentive is overwhelming. When the choice is between a product that earns pennies and one that earns dollars from the same wafer, the boardroom math is not ambiguous. Memory manufacturers have effectively sold out their HBM capacity for the year, with the top three prioritizing value over volume.

Samsung, the undisputed volume king for more than three decades, lost its throne in the first quarter of 2025 when SK Hynix overtook it in DRAM revenue for the first time since the company’s founding in 1983. The displacement was driven entirely by HBM. SK Hynix bet early on NVIDIA’s accelerator architecture, became the primary HBM supplier for both the Hopper and Blackwell GPU platforms, and locked in multi-year supply agreements that gave it pricing power no defense planner anticipated. SK Hynix indicated it had already sold all of its 2026 production capacity for HBM, DRAM, and NAND. Samsung stumbled on HBM3E yield issues and quality qualification failures at NVIDIA, falling to third place in the very market segment driving the industry’s transformation. The wounded giant is now racing to regain ground with HBM4, but the structural advantage has shifted.

Then there is Micron—the only American manufacturer of advanced DRAM and the only domestic producer of HBM. The U.S. government treats Micron as critical infrastructure. The Commerce Department awarded Micron $6.4 billion in direct CHIPS Act funding, supporting a planned $200 billion total investment in domestic memory manufacturing and R&D. Micron is the only U.S.-based manufacturer of advanced memory chips, and currently 100 percent of leading-edge DRAM production occurs overseas, primarily in East Asia. When the federal government subsidizes your fabs at this scale, your incentive to produce cheap consumer RAM does not merely diminish. It evaporates. In December 2025, Micron announced it would exit the Crucial consumer business entirely to redirect capacity toward enterprise and AI customers. The American Fortress is real. It is also not building for you.

The architecture here mirrors the critical minerals chokepoint identified in GAP 1. Replace “rare earths” with “wafer starts” and the geometry is identical: a small number of suppliers controlling an irreplaceable input to national power, with no mechanism for sovereign nations to ensure allocation during crisis.

The Silicon Triage

The center of gravity in this crisis is not demand. Demand is infinite and irrelevant to the chokepoint. The center of gravity is wafer-start allocation—the quarterly decision, made inside three boardrooms, that determines whether finite silicon goes to HBM stacks for AI accelerators or DDR5 modules for everything else. That decision is the triage.

The physics are unforgiving. HBM3E consumes roughly three times the silicon wafer area of standard DDR5 per gigabyte. The ratio is driven by two factors: HBM dies are physically larger, and the vertical stacking process—through-silicon vias connecting multiple DRAM layers—introduces yield losses that compound at every layer. An eight-layer stack must produce eight good dies; a twelve-layer stack, twelve. Industry sources confirm that HBM wafer sizes increase 35 to 45 percent versus equivalent DDR5, while yields run 20 to 30 percent lower. The advanced packaging lines required for HBM—SK Hynix’s mass reflow molded underfill process, TSMC’s CoWoS interposers—are not interchangeable with conventional DRAM production equipment. SK Hynix has told investors that its advanced packaging lines are at full capacity through 2026. Samsung and Micron face identical constraints. The tools, masks, and equipment for HBM occupy space that would otherwise produce DDR5 or LPDDR5. Every HBM chip that ships to an NVIDIA datacenter is silicon that did not become consumer memory.

This is not waste. This is triage—the medical term is precise. The term this paper coins for the phenomenon is the Silicon Triage: the deliberate reallocation of finite semiconductor wafer capacity from consumer and sovereign computing to AI datacenter infrastructure, creating a de facto global rationing system administered by three corporations. No government voted on it. No treaty authorized it. No regulatory body oversees it. And yet it determines which nations can compute and which cannot.

The inventory data confirms the triage is real and accelerating. DRAM supplier inventory fell from 17 weeks in late 2024 to just two to four weeks by October 2025. Two to four weeks of inventory is not a market operating under pressure. It is a market operating without a buffer. Any disruption—a fab shutdown, an earthquake, a single procurement decision by a hyperscaler—triggers immediate price explosions. And a single procurement decision did exactly that. In October 2025, OpenAI signed deals to secure approximately 900,000 DRAM wafers per month for its Stargate Project—roughly 40 percent of global DRAM output. The simultaneous, secretive nature of these agreements triggered market panic and cascading stockpiling across the industry. Major OEMs began stockpiling memory chips in anticipation of further supply constraints. The hoarding compounded the shortage, as it always does.

IDC analysts stated the dynamic plainly: every wafer allocated to an HBM stack for an NVIDIA GPU is a wafer denied to the LPDDR5X module of a mid-range smartphone or the SSD of a consumer laptop. The consequences are cascading. IDC projects the global PC market and smartphone sales could decline significantly in 2026 under downside scenarios as memory costs reshape product roadmaps across the industry. TrendForce has downgraded its 2026 notebook shipment forecast from growth to decline as rising memory costs compress margins across consumer electronics. The automotive industry, where DRAM powers advanced driver assistance systems and digital cockpits, faces growing operational disruption as the sector accounts for less than 10 percent of global DRAM demand and lacks the bargaining power to compete with hyperscalers for allocation.

The triage is not abstract. It is priced into the hardware ordinary citizens buy. Samsung raised prices for thirty-two-gigabyte DDR5 modules from one hundred forty-nine dollars to two hundred thirty-nine dollars—a sixty percent increase in a single quarter. Asus raised PC product prices in January 2026, citing memory costs directly. A typical server requires thirty-two to one hundred twenty-eight gigabytes of memory. An AI server can require a terabyte. When three companies control the global supply and one class of customer can outbid every other, the triage is not a metaphor. It is a procurement reality that no elected official voted to impose.

Samsung’s co-chief executive told Reuters the shortage was “unprecedented” and warned that constraints could persist for months or years as AI infrastructure competes for wafers. The word was precise. There is no historical precedent for a shortage driven not by supply failure but by deliberate supply reallocation toward a single customer class. What makes this crisis different from the 2020–2023 chip shortage is the cause. That shortage was driven by pandemic disruption—factory closures, logistics failures, demand whiplash. It was painful and temporary. The Silicon Triage is driven by structural reallocation of manufacturing capacity toward higher-margin products. It is not a disruption. It is a business model. And it will not self-correct because the margin differential that drives it only widens as AI demand grows.

The Geopolitical Vice

The Silicon Triage operates inside a geopolitical vise that tightens from both directions simultaneously. On one jaw: American export controls designed to deny China the memory architecture required for advanced AI. On the other: Chinese retaliation targeting the critical minerals required to manufacture that memory. The vise guarantees that prices will not return to pre-crisis levels, because the crisis is now structural rather than cyclical.

On December 2, 2024, the Bureau of Industry and Security imposed the first country-wide export controls on High Bandwidth Memory, restricting the sale of HBM from HBM2E and above to China and adding 140 Chinese entities to the Entity List. The controls treated HBM as equivalent to weapons-grade technology—which, in the context of training frontier AI models, it functionally is. Memory bandwidth is the binding constraint on AI accelerator performance. Without HBM, you cannot train large language models at scale. Without large language models, you cannot build the AI systems that will determine military, economic, and intelligence dominance for the next generation. The CSIS analysis was direct: the 2024 controls targeted a key vulnerability in China’s ability to produce advanced AI chips by banning HBM sales from HBM2E and above. In September 2025, BIS removed the named Chinese facilities of Samsung and SK Hynix from the Validated End-User program, effective December 31, 2025—further constricting the pathways through which memory technology reaches Chinese manufacturers.

China’s response was instantaneous and symmetrical. On December 3, 2024—one day after the HBM controls—China’s Ministry of Commerce banned exports of gallium, germanium, antimony, and superhard materials to the United States. These are not obscure elements. Gallium and germanium are foundational to semiconductor manufacturing. China dominates global production and processing of all four materials. A U.S. Geological Survey report estimated that a simultaneous gallium and germanium export ban could cost the American economy $3.4 billion in GDP. The retaliation escalated throughout 2025. Beijing imposed export controls on tungsten and tellurium in February, seven rare earth elements in April, and by October 2025 asserted jurisdiction—for the first time—over foreign-made products containing Chinese-origin rare earth materials. The architecture was no longer tit-for-tat. It was systemic.

Following the Trump-Xi meeting in late October 2025, China suspended the most aggressive rare earth controls for one year. But the underlying export control architecture remains intact—the suspension is a pause in escalation, not a strategic reversal, and China’s April 2025 licensing requirements for seven rare earth elements continue without interruption. Beijing demonstrated that it possesses—and is willing to deploy—a mirror-image chokepoint to match Washington’s semiconductor controls. Memory chips versus critical minerals. Each side holds a knife to the other’s supply chain. Neither can cut without being cut.

Meanwhile, China is building its own alternative. CXMT, the state-funded DRAM manufacturer based in Hefei, is the world’s fourth-largest DRAM producer, preparing a $4.2 billion IPO on Shanghai’s Star Market after revenue surged nearly 98 percent in the first nine months of 2025. CXMT is producing DDR5 and LPDDR5X, demonstrating chipmaking capabilities that surprised Western analysts despite U.S. export restrictions—including DDR5-8000 and LPDDR5X-10667 speeds achieved without access to leading-edge fabrication tools. By early 2025, CXMT had doubled its monthly wafer output to 200,000, with forecasts pointing to 300,000 by 2026. But CXMT cannot produce HBM2E or above. It lags the triopoly by one-and-a-half to five years in process technology. And its expansion—while impressive in commodity DRAM—will not relieve the HBM bottleneck driving the global shortage. China can build its own commodity memory. It cannot yet build the memory that powers frontier AI. The implications for sovereign AI capability are stark: any nation dependent on the triopoly for HBM allocation is dependent on three boardrooms for its ability to train advanced AI models. No treaty governs that dependency. No alliance manages it.

But that gap is closing faster than Western analysts projected. ChangXin Memory Technologies has grown its global DRAM market share from near zero in 2020 to approximately five percent by 2024, and is targeting HBM3 production by 2026–2027. Yangtze Memory Technologies—China’s NAND champion—is entering DRAM fabrication and exploring a partnership with CXMT to leverage its Xtacking hybrid bonding technology for HBM assembly. The collaboration matters because HBM is fundamentally a packaging challenge as much as a DRAM challenge, and YMTC’s wafer-to-wafer bonding expertise is among the most advanced in Asia.

The strategic intent is undisguised. Huawei’s three-year Ascend AI chip roadmap includes the Ascend 950PR in the first quarter of 2026, notable for its planned use of domestically produced HBM. China’s forthcoming Fifteenth Five-Year Plan explicitly targets memory industry expansion and HBM development as national priorities, backed by Big Fund III, launched in 2024. The Bureau of Industry and Security added HBM-specific export controls in late 2024, but CXMT—one of China’s four largest chip fabrication companies—remains absent from the Entity List. The export controls are chasing a target that is building its own supply chain underneath them.

The convergence this paper identifies is the intersection of three vectors that separate institutions manage in isolation: semiconductor export controls administered by BIS, critical mineral policy managed by the State Department and USGS, and AI infrastructure procurement negotiated between private hyperscalers and private memory manufacturers. No single institution sees the unified chokepoint. The Silicon Triage operates at that intersection, invisible to the bureaucratic architecture designed to govern each vector independently.

The Response Gap

The United States currently holds less than two percent of the world’s advanced memory manufacturing capacity. The CHIPS and Science Act of 2022 was designed to change that. Micron received up to 6.165 billion dollars in direct funding to support a twenty-year vision that would grow America’s share to approximately ten percent by 2035. SK Hynix received an award to build a memory packaging plant in West Lafayette, Indiana. Samsung received 6.4 billion dollars for facilities in Texas. These are serious commitments. They are also structurally late.

The majority of CHIPS funding has been finalized but not disbursed, leaving billions in possible limbo if contracts are not carried out. The Trump administration’s federal workforce reductions have targeted the Department of Commerce and NIST—the agencies responsible for disbursement. The Semiconductor Industry Association warns that the Section 48D advanced manufacturing investment tax credit—the twenty-five percent incentive that catalyzed over five hundred forty billion dollars in announced private investment—is set to expire on December 31, 2026. Nine months from this writing. The bipartisan BASIC Act to extend it has not passed.

Meanwhile, new fabrication plants take three to five years to reach volume production. TSMC’s Arizona facility has been delayed repeatedly, with the company citing construction costs four to five times higher than in Taiwan. Intel’s Ohio fab has slipped into 2026. SK Hynix’s Indiana plant is not expected to produce at scale until 2027. The gap between the threat timeline and the response timeline is measured not in months but in years—and the threat is not waiting.

The Doctrine: Five Pillars of Compute Sovereignty

The convergence gap demands doctrine, not commentary. The following five pillars define a framework for treating memory allocation as what it has become—a matter of national sovereignty and strategic resilience.

Sovereign Memory Reserves. Nations maintain strategic petroleum reserves against energy supply disruption. No equivalent exists for semiconductor memory. The United States should establish a Strategic Compute Reserve—a national stockpile of DRAM and HBM sufficient to sustain critical AI, defense, and infrastructure computing through a supply disruption of defined duration. The model is not speculative. The Strategic Petroleum Reserve was created in 1975 after the Arab oil embargo demonstrated that energy dependence was a national security vulnerability. The memory market in 2025 demonstrated the identical lesson. The precedent exists. The mechanism exists. The political will does not, because policymakers have not yet understood that memory is infrastructure, not product.

Wafer Allocation Transparency. The triopoly’s quarterly wafer-start allocation between HBM and conventional DRAM is currently proprietary. This is the single most consequential resource-allocation decision in the global technology economy, and it is made behind closed doors with no public accountability. Any memory manufacturer receiving government subsidy—including CHIPS Act funding—should be required to disclose wafer-start allocation ratios between product categories on a quarterly basis. If taxpayers fund the fabs, the public sees the triage math. This is not regulation of private enterprise. It is a condition of public subsidy. The principle is already established in defense contracting, where cost-plus structures require financial transparency. The same principle applies when the subsidy is $6.4 billion.

Allied Memory Compact. NATO maintains fuel-sharing agreements for wartime operations. It has no silicon-sharing agreements. An Allied Memory Compact would establish a framework for memory allocation during supply crisis—who gets priority, how shortfalls are distributed, what triggers emergency reallocation. The 2025 shortage demonstrated that allied nations competing against each other for the same constrained memory supply weakens all of them simultaneously. Japan, South Korea, and the EU are all dependent on the same three manufacturers for defense-relevant compute memory. A compact does not solve scarcity. It prevents scarcity from becoming a mechanism for allied fragmentation—which is precisely what adversarial actors would exploit.

Domestic Fabrication Floor. Micron’s $200 billion investment commitment is a beginning, not an endpoint. A statutory Domestic Fabrication Floor should define a minimum percentage of national memory consumption that must be produced on domestic soil—not as aspiration but as enforceable threshold, with consequences for falling below it. The current reality—100 percent of leading-edge DRAM production overseas—is a vulnerability that no amount of subsidy addresses until the fab lines are operational and producing at scale. The CHIPS Act funds construction. Doctrine must define the floor. Without it, the subsidy is a one-time investment with no structural guarantee, and the next administration can redirect priorities without constraint.

Compute Access as Critical Infrastructure. Access to sufficient computing memory should be reclassified as critical infrastructure, equivalent to the power grid, water supply, and telecommunications networks. This is not metaphor. When memory scarcity prevents a hospital from upgrading its diagnostic AI, when a defense contractor cannot source the DRAM for an avionics system, when a national laboratory cannot build the compute cluster required for climate modeling—the failure mode is identical to a power outage or a water main break. The difference is that power and water are regulated as public utilities. Memory is still treated as a market commodity subject to private allocation. The Silicon Triage has demonstrated that this classification is obsolete. Reclassification would trigger regulatory frameworks—allocation priority during shortage, price stabilization mechanisms, mandatory reserves—that currently do not exist because the commodity assumption has never been challenged. It is being challenged now.

The question this paper leaves with its reader is not whether memory scarcity is real. The inventory numbers confirm it. The price data screams it. The question is whether the institutions responsible for national security and economic sovereignty will recognize that three boardrooms now control the physical capacity to think—and whether that recognition will arrive before the next triage decision is made. The triage will not end. It will bifurcate. And the governments that failed to see the first one forming are unlikely to see the second one until it is already operational.

RESONANCE

References and Source Attribution

Astute Group. (2026). “Memory makers divert capacity to AI as HBM shortages push costs through electronics supply chains.” Summary: Reports Samsung co-CEO calling the shortage unprecedented and confirms the three-to-one HBM-to-DDR5 wafer consumption ratio.

Astute Group. (2025). “SK Hynix Holds 62% of HBM, Micron Overtakes Samsung, 2026 Battle Pivots to HBM4.” Summary: Tracks HBM market share shifts among the three dominant suppliers and documents Asus price increases tied to memory costs.

Bureau of Industry and Security. (2024). Press release: Commerce strengthens export controls to restrict China’s capability to produce advanced semiconductors. Summary: Announces new HBM export controls, 140 Entity List additions, and expanded semiconductor manufacturing equipment restrictions.

Center for Strategic and International Studies. (2024). “Where the Chips Fall: U.S. Export Controls Under the Biden Administration from 2022 to 2024.” Summary: Analyzes the evolving export control regime including HBM restrictions targeting China’s AI capabilities.

CNBC. (2025). “China suspends some critical mineral export curbs to the U.S. as trade truce takes hold.” Summary: Reports China’s one-year suspension of rare earth and critical mineral export controls following the Trump-Xi meeting.

Congressional Research Service. (2025). “U.S. Export Controls and China: Advanced Semiconductors.” R48642. Summary: Documents BIS removal of Samsung and SK Hynix Chinese facilities from the Validated End-User program effective December 31, 2025.

Council on Foreign Relations. (2025). McGuire testimony before House Foreign Affairs Committee: “Protecting the Foundation: Strengthening Export Controls.” Summary: Documents that CXMT remains absent from the Entity List despite being one of China’s four largest chip fabrication companies.

Counterpoint Research via Semiecosystem. (2025). “SK Hynix’ Lead Shrinks in DRAM, HBM.” Summary: Reports Q3 2025 DRAM revenue and HBM market share data for all major manufacturers.

Digitimes. (2025). “China’s CXMT muscles into DRAM’s top tier.” Summary: Reports CXMT’s doubling of monthly wafer output to 200,000 with forecasts to 300,000 by 2026.

EE Times. (2026). “The Great Memory Stockpile.” Summary: Documents the zero-sum wafer allocation dynamic, HBM margin superiority, and the structural nature of the memory shortage.

Everstream Analytics. (2026). “Global Memory Chip Shortage Worsens.” Summary: Documents DRAM inventory decline from 17 weeks to two-to-four weeks and SK Hynix pre-selling all 2026 production capacity.

Financial Content / TokenRing. (2025). “AI-Driven DRAM Shortage Intensifies as SK Hynix and Samsung Pivot to HBM4 Production.” Summary: Reports HBM yields between fifty and sixty percent and the three-to-four standard chip cannibalization ratio per HBM unit produced.

Foundation for Defense of Democracies. (2025). “China Pauses Some Rare Earth Export Curbs While Retaining Levers of Control.” Summary: Analyzes the November 2025 suspension as a pause in escalation with underlying control architecture intact.

Global Trade Alert. (2025). “A Widening Net: A Short History of Chinese Export Controls on Critical Raw Materials.” Summary: Tracks China’s escalating export control regime from 2023 through October 2025 including expansion to rare earth technologies.

IDC. (2026). “Global Memory Shortage Crisis: Market Analysis and the Potential Impact on the Smartphone and PC Markets in 2026.” Summary: Analyzes the zero-sum wafer allocation dynamic and projects significant declines in smartphone and PC markets under downside scenarios.

IEEE Spectrum. (2024). “Chips Act Funding: Where the Money’s Going.” Summary: Reports SIA finding that more than half of newly created U.S. semiconductor jobs by 2030 are on course to go unfilled.

Information Technology and Innovation Foundation. (2025). “U.S. Semiconductor Manufacturing Tax Credits Need to Be Extended and Broadened.” Summary: Documents the Section 48D tax credit expiration date and its role in catalyzing over five hundred forty billion dollars in private investment.

KED Global. (2025). “SK Hynix beats Samsung to become global No. 1 DRAM maker.” Summary: Reports SK Hynix overtaking Samsung in DRAM revenue for the first time since 1983, driven by HBM leadership.

Manufacturing Dive. (2025). “US Chip Production Targets Edge Further Out of Reach Under Trump Administration.” Summary: Reports that CHIPS funding has been finalized but not disbursed, with federal workforce reductions threatening disbursement capacity.

Micron Technology. (2025). Press release: “Micron and Trump Administration Announce Expanded U.S. Investments.” Summary: Announces $200 billion domestic manufacturing commitment, $6.4 billion in CHIPS Act funding, and plans to bring HBM packaging to the United States.

Micron Technology. (2025). Press release: “Micron Announces Exit from Crucial Consumer Business.” Summary: Announces decision to exit the 29-year-old Crucial consumer brand and redirect all capacity toward enterprise and AI customers.

National Governors Association. (2025). “CHIPS and Science Act: Implementation Resources.” Summary: Documents Micron’s 6.165 billion dollar CHIPS Act award and the target of growing U.S. advanced memory share from less than two percent to ten percent by 2035.

National Institute of Standards and Technology. (2025). Fact sheet: President Trump secures $200 billion investment from Micron Technology. Summary: Confirms Micron as the only U.S.-based manufacturer of advanced memory chips and details CHIPS Act funding for domestic fabrication.

Network World. (2026). “Samsung Warns of Memory Shortages Driving Industry-Wide Price Surge in 2026.” Summary: Reports Samsung DDR5 price increases of sixty percent in a single quarter and SK Hynix confirmation that all capacity is sold out for 2026.

Optilogic. (2025). “How China’s Rare Earth Metals Export Ban Will Impact Supply Chains.” Summary: Documents China’s December 2024 retaliatory export ban on gallium, germanium, antimony, and superhard materials.

ORF America. (2025). “China’s Critical Mineral Export Controls: Background and Chokepoints.” Summary: Estimates $3.4 billion U.S. GDP loss from simultaneous gallium and germanium ban and maps China’s critical mineral leverage.

Semiconductor Industry Association. (2025). “Chip Incentives and Investments.” Summary: Reports that the Section 48D advanced manufacturing investment tax credit is set to expire in 2026 and warns the investment trajectory is at risk.

SoftwareSeni. (2026). “Understanding the 2025 DRAM Shortage and Its Impact on Cloud Infrastructure Costs.” Summary: Reports OpenAI’s Stargate Project securing approximately 900,000 wafers per month, roughly 40 percent of global DRAM output.

South China Morning Post via Yahoo Finance. (2025). “China’s DRAM giant CXMT plans $4.2 billion IPO.” Summary: Details CXMT’s IPO plans, 97.8 percent revenue growth, and position as the world’s fourth-largest DRAM manufacturer.

TechSpot. (2025). “AI boom drives record 172% surge in DRAM prices as shortages hit memory market.” Summary: Reports TrendForce data showing 171.8 percent year-over-year DRAM contract price increases driven by AI server demand.

Tom’s Hardware. (2026). “Chinese Semiconductor Industry Gears Up for Domestic HBM3 Production by the End of 2026.” Summary: Reports CXMT targeting HBM3 production and YMTC/XMC developing HBM packaging technologies using hybrid bonding.

Tom’s Hardware. (2025). “Here’s why HBM is coming for your PC’s RAM.” Summary: Explains HBM’s three-times wafer consumption ratio versus DDR5, advanced packaging constraints, and cascading consumer price effects.

Tom’s Hardware. (2025). “China’s banned memory-maker CXMT unveils surprising new chipmaking capabilities.” Summary: Documents CXMT DDR5-8000 and LPDDR5X-10667 products achieved without access to leading-edge fabrication tools.

Tom’s Hardware. (2025). “YMTC and CXMT Team Up to Accelerate Chinese Domestic HBM Production.” Summary: Documents the YMTC-CXMT partnership leveraging Xtacking hybrid bonding technology for domestic HBM assembly.

TrendForce. (2025). “China’s NAND Giant YMTC Reportedly Moves into HBM Using TSV, Following CXMT and Huawei.” Summary: Reports Huawei’s Ascend 950PR roadmap with domestically produced HBM planned for Q1 2026.

TrendForce. (2025). “Global DRAM Revenue Jumps 30.9% in 3Q25.” Summary: Reports Q3 2025 DRAM revenue data and projects contract price increases of 45 to 55 percent quarter-over-quarter in Q4 2025.

TrendForce. (2024). “HBM and Advanced Packaging Expected to Benefit Silicon Wafer.” Summary: Reports HBM wafer size increases of 35 to 45 percent versus DDR5 and yield rates 20 to 30 percent lower.

TrendForce. (2025). “Memory Price Surge to Persist in 1Q26.” Summary: Reports downgraded notebook shipment forecasts and rising BOM costs forcing brands to raise prices or cut specifications.

Yole Group. (2025). “China’s Next Move: The Five-Year Plan That Could Reshape Semiconductors.” Summary: Documents China’s Fifteenth Five-Year Plan priorities including memory industry expansion, HBM development, and equipment localization.