A market increasingly sold before it is made

The memory industry is entering a phase in which availability, not merely price, is becoming the central issue for large buyers. A DigiTimes analysis published on August 8 reported that major suppliers had locked in supply commitments and price floors through 2030, backed by roughly US$38 billion in deposits, collateral and related financial guarantees. The reported arrangements cover a market led by Samsung Electronics, SK hynix and Micron, with SanDisk also cited in the broader tally.

The important qualification is that this money should not automatically be treated as revenue or profit for the manufacturers. It is better understood as security supporting long-term procurement contracts: customers put capital at risk in exchange for priority access to scarce capacity, while suppliers gain confidence to plan production and investment. The detailed contract terms, timing and accounting treatment have not been publicly disclosed, so the US$38 billion figure should be regarded as an industry estimate rather than a fully independently auditable market total.

Even so, the report captures a significant change in the balance of power. Memory has historically been one of the semiconductor industry’s most cyclical businesses. Manufacturers added capacity, oversupply emerged and prices fell sharply; then investment was curtailed and a new shortage followed. The present cycle is unusual because cloud operators and AI-system builders are seeking multi-year visibility on both volume and price, rather than relying mainly on quarterly negotiations.

AI demand is changing the mix of memory production

The shortage is not simply a matter of more PCs or smartphones being sold. AI infrastructure requires several forms of memory at once: high-bandwidth memory, or HBM, placed close to powerful accelerators; server DRAM for host systems; and NAND flash storage for the expanding volumes of data used in training and inference.

HBM is particularly influential because it is technically demanding and consumes manufacturing resources that could otherwise support more conventional DRAM products. Suppliers are therefore prioritising advanced-node DRAM, HBM and higher-value server products. The effect can be felt further down the supply chain: capacity that is technically or economically redirected towards AI systems is no longer available to the same extent for mainstream PC, mobile or embedded-memory products.

Industry researchers have described the result as a structural constraint rather than a brief disruption. IDC has noted that leading producers have shifted attention from conventional DRAM and NAND used in consumer devices towards AI data-centre memory, while TrendForce has forecast sharply higher memory-market spending as agentic AI expands server requirements. Samsung has said it expects its HBM sales to more than triple in 2026 from the previous year, illustrating the commercial incentive to concentrate on this segment.

That does not mean every memory chip is interchangeable or uniformly scarce. Product qualification, process technology, performance requirements and packaging capacity all matter. A buyer that needs leading-edge HBM cannot necessarily substitute standard DRAM, while a laptop maker may be able to qualify an additional supplier over time. Nevertheless, the broad direction is clear: AI investment is raising demand for memory faster than readily usable supply can expand.

Why long-term contracts matter

Multi-year agreements can make sense for both sides. Hyperscalers and system manufacturers avoid the risk that a shortage prevents them from deploying servers or shipping products. Memory producers receive clearer demand signals, reduce exposure to sudden order cancellations and can support large capital-investment programmes with greater confidence.

The cost is reduced flexibility. Buyers that commit to volumes, price floors or collateral may have less ability to benefit if the market weakens. Suppliers, meanwhile, must still execute their technology roadmaps and bring promised output online. The arrangements may dampen the volatility of spot pricing, but they do not abolish the cycle; they transfer some risk from the supplier to the customer and make demand forecasts more consequential.

For smaller device makers, the change could be uncomfortable. Companies without the purchasing scale to negotiate direct multi-year allocation may face higher contract prices, longer lead times or less attractive product mixes. Their alternatives include redesigning products around available components, reducing memory configurations, using qualified second sources where possible and accepting lower margins. Consumer prices for RAM modules and solid-state drives may reflect these pressures, but retail pricing will also depend on channel inventories, distributors and the mix of products sold.

New fabs will help, but not immediately

The prospect of sustained high prices is prompting a new investment cycle. SEMI expects global 300mm memory-equipment investment to exceed US$50 billion in 2026 and approach US$80 billion by 2029. Micron, for its part, said in July that it was increasing planned US investment and technology spending to more than US$250 billion through 2035, including a long-term aim to produce 40% of its DRAM in the United States.

These headline figures should not be confused with an immediate supply solution. Building a fabrication plant is only the first stage. Equipment installation, process qualification, yield improvement and the transition to commercially useful output take time. Advanced DRAM and HBM add further complexity because their effective capacity depends on technology migrations, manufacturing yields and sophisticated packaging. A new facility can therefore take years to materially alter the volume available to customers.

This is why 2029 has emerged as a plausible turning point in some market commentary. It is not a firm expiry date for the shortage. Rather, it is a horizon by which investments now being announced could begin to have a meaningful cumulative effect, provided construction, equipment supply and technical ramp-up proceed as planned.

The route to normalisation remains uncertain

Several factors could shorten or extend the current tightness. Demand may cool if AI infrastructure spending becomes more disciplined, if system designers improve memory efficiency, or if customers defer deployments. Conversely, shortages could persist if HBM adoption accelerates, new fabs ramp more slowly than expected or geopolitical and supply-chain risks interrupt production.

Additional competitors could also reshape the market. Chinese suppliers are expanding their presence in conventional DRAM, while established manufacturers are investing heavily in leading-edge products. Greater competition could improve supply choices over time, although qualifying alternative memory for high-performance systems is a long process and does not instantly resolve shortages in the most advanced categories.

The reported collateral arrangements are therefore a signal of current confidence and scarcity, not proof that suppliers can dictate conditions indefinitely. They show that major customers consider secured access to memory important enough to commit capital years ahead. Whether that leverage lasts until 2029 will depend less on any single contract than on the race between AI-driven consumption, manufacturing execution and the industry’s willingness to add capacity without creating the next oversupply downturn.

Sources