Citi: AI Continual Learning Will Extend the Memory Shortage Through 2031

Citi: AI Continual Learning Will Extend the Memory Shortage Through 2031

Custom Image

Citi Warns of a Multi-Year Memory Supply Crunch as AI Enters the Continual Learning Era

Citi Research released a detailed analysis on September 14, 2026, projecting that the shift of artificial intelligence systems into a continual learning phase will fundamentally reshape demand patterns for high-bandwidth memory, server DDR5, and enterprise solid-state drives. This transition moves AI beyond static training and inference models toward systems that continuously update knowledge and retain prior data, creating sustained pressure on memory supply. The bank’s forecasts show DRAM supply-demand ratios turning sharply negative in 2027 and 2028, while NAND follows a similar path, with the imbalance expected to persist through 2031.
 
These projections arrive amid recent market concerns over AI capital expenditure returns and high-bandwidth memory specification adjustments, yet Citi interprets the latter as evidence of supply constraints rather than softening demand. Supporting data from industry trackers indicate memory already accounts for more than half of semiconductor revenue in recent quarters, underscoring the structural nature of the tightness. Continual learning will simultaneously expand demand across HBM, server DRAM, and enterprise storage starting in 2027, outpacing feasible capacity additions and locking in multi-year shortages that reshape pricing power and investment priorities for leading memory suppliers.

HBM Bit Demand Projected to Climb 62 Percent in 2027 Under Continual Learning

Citi’s September 14 report estimates high-bandwidth memory bit demand will rise 62 percent year-over-year to 75.2 billion gigabits in 2027 before accelerating a further 69 percent to approximately 127 billion gigabits in 2028. These figures roughly double earlier projections and reflect the dual requirements of model updating and persistent access to historical data that continual learning imposes. Demand sources extend beyond a single accelerator vendor; Broadcom’s HBM needs are expected to expand from about 9 billion gigabits in 2026 to more than 41 billion by 2028, while Google’s requirements grow from roughly 4.7 billion to 14 billion gigabits over the same window.
 
On the supply side, monthly wafer capacity is forecast to increase from 420,000 wafers in 2027 toward 700,000, yet accelerating custom ASIC shipments are projected to push the HBM supply-demand ratio to a deficit of 36 percent by 2028. Long lead times for advanced packaging and new fab construction limit the speed of response, leaving the market structurally short even as capital spending rises. Industry checks cited by the bank confirm that customer inquiries for 2027-2028 allocations continue to intensify, reinforcing the view that specification adjustments represent rationing rather than demand weakness. This dynamic positions HBM as the primary bottleneck for next-generation AI systems that must retain and refine knowledge over extended operational periods.

Server DDR5 Demand Set to Surge 51 Percent as Continual Learning Scales

Global DRAM demand is forecast to expand 30.2 percent in 2027 while supply grows only 18.8 percent, producing an 8.7 percent supply-demand deficit that widens to 9.7 percent in 2028 as demand rises another 35 percent against 22 percent supply growth. Server DRAM specifically is projected to jump 51 percent from 226.3 billion units (1Gb equivalent) in 2026 to 341.7 billion units in 2027, representing roughly 67 percent of total DRAM consumption. Continual learning drives this acceleration because models require frequent parameter updates alongside rapid access to previously acquired knowledge, elevating the role of high-performance server DDR5 modules.
 
HBM production already cannibalizes wafer starts that would otherwise support commodity DRAM, while slower technology-node transitions further constrain output. Citi notes that even aggressive capital expenditure increases will not close the 2027 gap given multi-year timelines for greenfield capacity. The resulting tightness is expected to support continued average selling price gains, with blended DRAM prices projected to rise 23.1 percent in 2027 after the extraordinary advances recorded in 2026. These conditions create sustained pricing power for suppliers that maintain leading positions in both HBM and high-performance server memory.

Enterprise SSD Demand Growth of 53 Percent Tied to Data Retention Needs

Enterprise solid-state drive demand is projected to expand 52.9 percent in 2027 as continual learning systems shift key-value cache workloads to external storage and adopt high-capacity QLC configurations for AI servers. This growth stems directly from the requirement to retain previously learned data while models continue training on new tasks, a process that multiplies storage intensity compared with static inference workloads. Citi anticipates overall NAND demand rising 29 percent in 2027 and 33 percent in 2028, exceeding supply growth of 21 percent and 25 percent and producing supply-demand ratios of minus 6.1 percent and minus 5.5 percent, respectively.
 
Samsung’s NAND capacity is expected to decline 4.7 percent in the near term as resources prioritize DRAM and HBM, limiting industry-wide expansion. The combination of reduced HBM capacity per accelerator, Nvidia’s cache offload trends, and rising personal and physical AI deployments further amplifies eSSD requirements. Even with NAND capital expenditure rising 34.2 percent to 21.8 billion dollars in 2027, long construction and qualification cycles prevent rapid closure of the shortfall. These factors convert storage into a structural rather than cyclical constraint for AI infrastructure.

DRAM Supply Growth Limited to 19 Percent in 2027 Despite Capex Ramp

Citi forecasts total memory capital expenditure to surge 46.5 percent to 80.4 billion dollars in 2027, with DRAM spending alone climbing 51.6 percent to 58.6 billion dollars. Samsung is expected to allocate 20.6 billion dollars, SK Hynix 17.5 billion dollars, and Micron 15.8 billion dollars. Despite this scale of investment, greenfield capacity additions face multi-year lead times for cleanroom construction, tool installation, and process qualification, rendering them insufficient to balance 2027 demand. Technology migration remains slower than historical norms, and HBM packaging consumes disproportionate wafer and assembly resources.
 
The bank therefore concludes that supply will lag demand growth of 30 percent in 2027 and 35 percent in 2028, locking in deficits of 8.7 percent and 9.7 percent. Historical semiconductor cycles show that once inventories fall below normal levels and demand remains structurally elevated, price and allocation tightness can persist for several years. Current inventory metrics already sit well below traditional thresholds across suppliers, cloud operators, and channels, providing additional support for the prolonged shortage thesis.

NAND Capacity Constraints Intensify as Resources Shift to HBM

Industry-wide NAND expansion appetite remains muted as leading suppliers redirect capital and engineering resources toward higher-margin HBM and server DRAM. Samsung’s NAND output is projected to contract 4.7 percent in the near term, while overall supply growth of 21 percent in 2027 and 25 percent in 2028 trails demand increases of 29 percent and 33 percent. Continual learning amplifies the imbalance by elevating enterprise SSD requirements for long-term data retention and cache offloading.
 
Citi projects NAND average selling prices to advance 45.3 percent in 2027, extending the strong upward trajectory observed in 2026. Low inventory levels and supplier stocks at approximately 2.6 weeks against a normal five-week baseline further limit the industry’s ability to absorb demand spikes. These conditions create a self-reinforcing cycle in which capacity additions remain cautious until sustained high pricing justifies broader investment, extending the shortage window well beyond traditional cycle lengths.

Personal AI and Physical AI Adoption Underpin Multi-Year Shortage Extension

Citi argues that the full-scale proliferation of continual learning, personal AI agents, and physical AI systems will provide the fundamental demand floor supporting undersupply through 2031. Personal AI requires persistent memory of user interactions and context, while physical AI applications in robotics and autonomous systems generate continuous streams of sensor data that must be stored and refined. These use cases expand the total addressable market for both high-bandwidth and high-capacity storage beyond current hyperscaler training clusters.
 
The bank’s analysis indicates that once these applications reach critical mass, demand growth rates will remain elevated even if pure training workloads moderate. Supply-side responses, constrained by capital intensity and technical complexity, cannot match this trajectory within the forecast horizon. Consequently, the memory market transitions from a cyclical recovery into a structural supercycle characterized by sustained deficits and elevated pricing.

Top Memory Suppliers Positioned for Structural Pricing Power

Citi identifies Samsung Electronics, SK Hynix, Micron, SanDisk, and Kioxia as the primary beneficiaries of the projected multi-year tightness. Samsung and SK Hynix hold leadership positions across both HBM and high-performance server DDR5, placing them at the center of the demand surge. Micron benefits from its expanding HBM footprint and DRAM scale, while SanDisk and Kioxia capture the enterprise SSD and high-density NAND opportunity.
 
Equipment and materials suppliers, including Applied Materials, Lam Research, Montage, TES, Eugene Technology, and TechWing, are also highlighted as indirect beneficiaries of the elevated capital spending cycle. The combination of constrained supply and rising average selling prices is expected to translate into expanded margins and cash-flow generation for these companies through the remainder of the decade. Market data already shows memory contributing more than 50 percent of semiconductor revenue in recent quarters, confirming the sector’s outsized role in the AI infrastructure build-out.

Q2 Semiconductor Revenue Records Underscore Memory Strength

According to Omdia data cited in conjunction with the Citi analysis, global semiconductor revenue reached an all-time high above 425 billion dollars in the second quarter of 2026, with memory accounting for 54 percent of the total, the first time the category has exceeded half of industry sales. The sector has recorded four consecutive quarters of double-digit sequential growth, a rarity in historical data spanning nearly 100 quarters.
 
AI infrastructure investment is lifting demand not only for HBM and DRAM but also for associated non-memory components. These figures provide empirical support for Citi’s contention that the current cycle differs from previous memory recoveries driven primarily by consumer electronics. The sustained contribution of memory to overall semiconductor growth reinforces the view that supply constraints will remain binding as continual learning workloads scale.

Capital Expenditure Surge of 46.5 Percent Still Insufficient for 2027 Balance

Even with global memory capital expenditure projected to rise 46.5 percent to 80.4 billion dollars in 2027, Citi concludes that the timing of capacity additions will leave the market short. DRAM spending is expected to reach 58.6 billion dollars, and NAND 21.8 billion dollars, yet greenfield projects require years from groundbreaking to qualified volume production.
 
Existing facilities face limitations from tool lead times, skilled labor availability, and the technical complexity of advanced packaging required for HBM. The bank therefore maintains that supply growth of 19 percent for DRAM and 21 percent for NAND in 2027 cannot match the corresponding demand increases. This mismatch sets the stage for continued allocation tightness and pricing strength into subsequent years, independent of near-term fluctuations in AI capital expenditure sentiment.

Despec Adjustments Interpreted as Supply Rationing Rather Than Demand Softness

Recent debate surrounding potential reductions in high-bandwidth memory specifications per accelerator has been interpreted by some market participants as a signal of cooling demand. Citi takes the opposite view, characterizing the adjustments as a rational response to limited HBM availability that allows chipmakers to maximize the number of accelerators shipped within constrained supply.
 
The bank’s upward revision of HBM bit demand forecasts, roughly doubling prior estimates, supports this interpretation. Customer inquiries for 2027 and 2028 allocations continue to rise, and major custom silicon programs at Broadcom and Google show expanding rather than contracting memory footprints. These observations indicate that the underlying demand trajectory remains intact and that specification changes represent efficient resource allocation under scarcity conditions.

Inventory Levels Remain Structurally Low Across the Supply Chain

Supplier, cloud-provider, and channel inventories for both DRAM and NAND sit well below historical norms, according to industry data referenced alongside the Citi report. NAND supplier stocks are estimated near 2.6 weeks versus a normal five-week level, while cloud inventories hover around three weeks against a seven-week baseline. Channel inventories similarly remain compressed. These low levels leave little buffer against demand acceleration from continual learning workloads.
 
In previous cycles, inventory rebuilding has required multiple quarters of elevated production; under the projected demand growth rates, rebuilding becomes even more difficult. The absence of excess stock therefore amplifies the impact of any incremental demand, reinforcing the multi-year shortage outlook through 2031.

Future Structural Shortage Thesis Extends Through 2031

Citi’s core conclusion is that the combination of continual learning, personal AI, and physical AI will generate demand growth that consistently outstrips the industry’s ability to add qualified capacity on relevant timelines. Supply-demand deficits of 8.7 percent for DRAM and 6.1 percent for NAND in 2027 are expected to persist in modified form for several additional years. Capital intensity, technical complexity, and the prioritization of HBM over other memory types create structural barriers to rapid balancing.
 
As a result, the memory market is positioned for an extended period of elevated pricing, allocation discipline, and strong cash generation for leading suppliers. The analysis provides a clear framework for understanding why recent concerns over AI investment returns are unlikely to reverse the underlying tightness in memory semiconductors through the end of the decade.

🔥 Beyond the Headlines: What KuCoin 5.0 Means for You

Market news moves fast — but where you act on it matters just as much. This October, KuCoin launches KuCoin 5.0, transforming KuCoin into a rebuilt platform. Here's what actually changes for you:
  • One account for everything. Older platforms split your money across separate "spot," "margin," and "futures" accounts and expected you to understand why. KuCoin 5.0's unified account removes that entirely — deposit once, and everything is simply there.
  • Stocks, indices, and commodities. KuCoin 5.0 expands beyond crypto into global markets. When crypto chops sideways and equities rally (or the reverse), you rotate in minutes instead of opening a brokerage account and waiting days for fiat rails.
  • Real-world assets (RWA). Tokenized exposure to traditional assets like commodities, right inside your crypto account. One of the fastest-growing segments in global finance is no longer reserved for institutions — you access it from the same balance you trade with.
  • Earn while you learn. Not ready to trade? KCUSD lets your stablecoins earn daily, auto-compounding interest. The lowest-stress way to put your idle deposit to work for 4% yield.
  • An AI assistant in plain language. Ask questions, get market context, understand what you're looking at — built into the platform, no jargon required.
  • An app that doesn't overwhelm. Faster, cleaner, and consistent — intuitive from the first tap, not after a tutorial.
  • Safety you can check, not just trust. A MiCAR-licensed EU entity, Proof of Reserves you can verify yourself, and internationally certified security (SOC 2 Type II, ISO 27001:2022).
 
Create your account in minutes — and start on the platform built for where crypto is going, not where it's been.

FAQs

What exactly does continual learning mean in the context of AI memory demand?

Continual learning refers to AI systems that continuously update their models with new tasks and knowledge while retaining access to previously acquired data. This process requires simultaneous high-bandwidth memory for active training updates and substantial storage capacity for historical data, driving concurrent demand growth across HBM, server DDR5, and enterprise SSDs starting in 2027, according to Citi’s analysis.
 

How large are the projected supply-demand deficits for DRAM and NAND?

Citi forecasts DRAM supply-demand ratios of minus 8.7 percent in 2027 and minus 9.7 percent in 2028, with demand growing 30 percent and 35 percent against supply growth of 19 percent and 22 percent. For NAND, the ratios are projected at minus 6.1 percent and minus 5.5 percent as demand rises 29 percent and 33 percent while supply expands 21 percent and 25 percent.
 

Which companies does Citi highlight as primary beneficiaries?

The bank names Samsung Electronics, SK Hynix, Micron, SanDisk, and Kioxia as its top memory picks, citing their positions in HBM, server DDR5, and high-density enterprise storage. Equipment suppliers such as Applied Materials and Lam Research are also expected to benefit from elevated capital spending.
 

Why will increased capital expenditure not close the shortage quickly?

Even with memory capital expenditure rising 46.5 percent to 80.4 billion dollars in 2027, greenfield capacity requires multi-year lead times for construction, tool installation, and process qualification. Existing facilities face constraints from packaging complexity and slower technology transitions, leaving 2027 demand unmatched by available supply.
 

How does Citi interpret recent HBM specification adjustments?

The bank views potential reductions in memory content per accelerator as a supply-constrained efficiency measure that allows more accelerators to be shipped within limited HBM availability, rather than evidence of weakening demand. Upward revisions to HBM bit demand forecasts support this assessment.
 

What role do personal AI and physical AI play in the longer-term outlook?

Citi expects these applications to create a durable demand floor by requiring persistent memory of user context and continuous storage of sensor data. Once widely adopted, they are projected to sustain elevated memory consumption even if pure training workloads moderate, supporting shortages through 2031.
 
Disclaimer: This content is for informational purposes only and does not constitute investment advice. Investments carry risk. Please do your own research (DYOR).