AI memory costs squeeze cloud budgets

The gist

AI’s memory bill is exploding—DRAM and HBM are set to devour up to 70% of cloud capital spending by 2027, forcing a total rethink of infrastructure strategy.

What to know

  • DRAM prices are expected to peak at nearly $10/GB in 2026, with HBM costs staying high as suppliers like Nvidia and Samsung shift focus to high-margin server memory.
  • By 2027, memory will eat up almost 70% of cloud providers’ CapEx—up from just 47%—triggering up to 98% year-over-year budget jumps and a scramble to redesign AI architectures.
  • Even as supply ramps post-2027, wafer-hungry HBM and relentless AI demand mean cost pressures will linger, making efficiency upgrades and new memory tech mission-critical.

Memory Price Surge Reshapes AI

Skyrocketing DRAM and HBM costs are forcing AI system designers to confront a new era of economic bottlenecks and long-term spending spikes, fundamentally altering infrastructure planning.

Since mid-2025, the cost of memory components critical to AI infrastructure, particularly DRAM and HBM, has surged sharply due to robust demand and constrained supply, with DRAM average selling prices projected to peak near $10 per gigabyte in 2026 before easing to about $5 per gigabyte by 2030. Meanwhile, HBM’s cost trajectory diverges from DRAM’s; as accelerators like AMD’s MI455X push HBM capacity beyond 400 GB and stack heights and packaging complexity increase, the cost per bit for HBM is expected to remain elevated despite technological advances, posing a growing economic challenge for AI system designers.

The elevated DRAM pricing is not just a short-term spike but a significant economic factor that could inflate cumulative server infrastructure spending by nearly 10% through 2030, translating into hundreds of billions of dollars in additional costs. This impact is most acute in 2026 and 2027 when prices peak, underscoring how memory costs have become a substantial component of the AI infrastructure budget, far beyond what would be expected if DRAM prices had stabilized at $3.50 per gigabyte.

Underlying these price pressures are persistent supply constraints, although the landscape is expected to improve gradually as memory suppliers expand production capacity—including new facilities in China—and as server and processor vendors innovate to enhance memory utilization efficiency. These developments should help moderate pricing by 2028, but until then, the rising cost and complexity of memory, especially HBM, remain a critical economic bottleneck that influences not only AI infrastructure design choices but also supplier capacity allocation decisions across DRAM, HBM, and NAND flash markets.

Sources

Cloud Budgets Overrun by Memory

Explosive memory price hikes are driving cloud providers to radically overhaul AI architectures and pass unprecedented cost increases down the supply chain.

By 2027, memory costs are set to dominate cloud service providers' capital expenditures, surging from 47% in 2026 to nearly 70%, fundamentally reshaping AI infrastructure economics. This dramatic shift is driven by skyrocketing prices for server DRAM and High Bandwidth Memory (HBM), with DRAM contract prices expected to climb approximately 270% in 2026 and HBM prices potentially rising between 70% and 140% in 2027. Memory manufacturers are prioritizing production capacity for server and AI applications, allocating over half of DRAM bit supply to HBM and RDIMM, which intensifies supply constraints and pushes memory's share of CapEx even higher.

This surge in memory costs is prompting cloud hyperscalers to dramatically increase overall capital expenditure—projected to grow 98% year-over-year in 2026 and an additional 50% in 2027—while simultaneously reconsidering AI system architectures to manage cost pressures. Providers are exploring strategies such as reducing per-system memory capacity, adjusting RDIMM and HBM configurations, and even hardwiring model architectures into AI ASICs to optimize memory usage. These architectural shifts reflect memory's transition from a mere procurement line item to a core design constraint in AI infrastructure.

The memory price inflation ripple extends beyond cloud providers to chip suppliers and end customers. Nvidia, citing soaring memory costs, announced over a 15% price increase for AI servers containing its Vera Rubin and Grace Blackwell chips in early 2027, illustrating how memory-driven cost pressures cascade through the supply chain. Meanwhile, cloud operators like OVHcloud are passing on these increases to customers, with planned price hikes up to 87%, signaling a broader market-wide inflation fueled by constrained memory supply and soaring demand.

Underlying this cost upheaval is a structural reallocation of DRAM production capacity, heavily favoring server and AI accelerator applications over consumer electronics. Producing one gigabyte of HBM demands three to four times the semiconductor fabrication capacity compared to conventional DRAM, contributing to supply shortages and price inflation. Leading memory suppliers such as Samsung, SK Hynix, and Micron have leveraged this AI-driven demand surge to gain pricing power, with NAND Flash revenues soaring 77% quarter-over-quarter in Q2 2026, underscoring how memory-intensive AI workloads are permanently reshaping the semiconductor landscape.

Sources

AI System Design Gets a Rethink

Memory has become the central constraint in cloud infrastructure, pushing hyperscalers to innovate with leaner architectures, in-house chips, and aggressive procurement strategies.

Cloud service providers are undergoing a fundamental shift in infrastructure economics as memory costs surge to dominate capital expenditure, projected to account for 68% of hardware CapEx by 2027, up from 47% in 2024. This dramatic rise compels providers like NVIDIA’s customers to absorb significant price hikes, exemplified by a 15% increase in AI server prices primarily driven by memory expenses. To manage this, hyperscalers are aggressively optimizing AI system architectures by reducing per-system memory capacity, adjusting RDIMM configurations, and exploring AI ASICs with model architectures hardwired into silicon, balancing cost pressures while maintaining performance and shipment targets.

The soaring memory prices have accelerated a technology race among cloud providers centered on memory efficiency, prompting increased investment in in-house chip development and architectural innovation. This strategic pivot is driven by the permanent reallocation of DRAM production toward server and AI accelerator applications, with HBM and Registered DIMMs accounting for over half of total DRAM bit supply in 2026. Given that producing one gigabyte of HBM consumes three to four times the semiconductor fabrication capacity of conventional DRAM, cloud providers must carefully balance capital expenditure growth with memory efficiency to sustain AI infrastructure expansion amid constrained supply and elevated costs.

To mitigate the impact of volatile and rising memory prices, cloud providers are adopting multifaceted procurement strategies including long-term agreements with price caps, absorbing higher costs, and reducing per-server memory configurations. However, these measures offer only partial insulation, as contract prices for Server DRAM are forecasted to rise 13–18% quarter-over-quarter in 2026, compelling providers to further increase capital expenditures or innovate in architectural efficiency. This evolving landscape underscores memory’s transformation from a mere procurement line item to a core design constraint, reshaping cloud infrastructure planning and supplier relationships.

Suppliers Pivot to High-Margin Memory

Nvidia, Samsung, and peers are diverting wafer capacity to HBM and server DRAM, squeezing consumer markets and locking AI infrastructure into a high-cost supply dynamic.

Major memory and AI chip suppliers including Nvidia, Samsung, SK Hynix, and Micron are strategically reallocating their limited semiconductor fabrication capacity towards high-margin High Bandwidth Memory (HBM) and server DRAM production to meet the surging demands of AI workloads. This shift is significant as HBM and Registered DIMMs are projected to constitute 51% of total DRAM bit supply in 2026, underscoring a permanent reorientation of memory resources away from consumer markets toward AI and server applications.

The economics of memory production are undergoing a profound transformation, with server DRAM contract prices having surged 64% in the latter half of 2025 and forecasted to climb an additional 270% in 2026, while HBM prices are expected to rise between 70% and 140% in 2027. This dramatic price escalation empowers suppliers like Nvidia to justify substantial product price increases, compelling cloud service providers to either boost capital expenditures or innovate memory architectures to manage these soaring costs.

HBM production demands three to four times more semiconductor wafer area per gigabyte than conventional DRAM, creating a structural supply constraint that intensifies the reallocation of capacity toward AI infrastructure memory. Samsung, SK Hynix, and Micron, which collectively control about 95% of global DRAM production, have all pivoted their manufacturing focus accordingly, as HBM yields margins unattainable by traditional DRAM products.

SK Hynix has cemented its dominance in the HBM market, capturing over 56% of global HBM revenue by early 2026, reflecting its aggressive strategic positioning in AI memory solutions. Meanwhile, NAND Flash suppliers including Samsung, SK Hynix, and Micron have leveraged AI-driven demand to significantly boost pricing power and revenues, with Micron leading the pack by nearly doubling its NAND revenue quarter-over-quarter to $11.85 billion in Q2 2026, elevating its global market share ranking from fourth to third.

Sources

End-User Prices Feel the Squeeze

Cloud and device users alike are hit with rising bills as memory-driven inflation ripples from data centers to PCs and smartphones, tightening margins across the tech ecosystem.

By 2027, soaring memory prices are set to dominate cloud infrastructure spending, with DRAM and NAND flash projected to consume 68% of cloud providers' capital expenditures, up from 47% in 2026. This dramatic shift, driven by a 270% year-on-year increase in server DRAM costs and a 235% rise in enterprise SSD prices, is compelling operators like OVHcloud to pass these surging expenses onto customers, with planned price hikes reaching as high as 87%. Such aggressive cost pass-throughs signal a broader inflationary trend in cloud service pricing that end-users will increasingly feel.

The ripple effects of memory price inflation extend beyond cloud data centers, significantly impacting consumer device markets. This year alone, PC prices have surged by double-digit percentages amid shortages of standard memory components, contributing to a 5% drop in PC shipments, while smartphone shipments are expected to decline by 15%. These trends underscore how elevated memory costs are constraining consumer hardware affordability and availability, tightening the squeeze on end-users across multiple technology segments.

Elevated memory costs are also empowering AI chip manufacturers like Nvidia to increase their prices, further intensifying cost pressures on cloud providers. Nvidia's planned 15% price hike for products featuring its Vera Rubin and Grace Blackwell chips exemplifies how upstream hardware inflation cascades down the supply chain, compelling cloud operators to allocate more capital expenditure toward AI hardware and potentially pass these compounded costs onto customers. This dynamic highlights the interconnected nature of memory pricing and AI infrastructure economics.

Sources

Efficiency Becomes Mission Critical

Even as new fabs come online, persistent cost pressures are making memory efficiency, innovative architectures, and alternative technologies essential for sustainable AI growth.

Post-2027, the memory supply landscape is poised for notable expansion, with forecasts predicting a 27% increase driven by process migrations and new fabrication facilities. However, this supply growth may only partially temper cost pressures, as the wafer-intensive nature of high-bandwidth memory (HBM) production limits proportional increases in overall DRAM bit supply. Consequently, while pricing is expected to moderate from 2028 onward, escalating demand and complex manufacturing dynamics suggest that cost challenges will persist for cloud AI infrastructure.

Technological advancements in memory manufacturing, including hybrid bonding, higher-density HBM generations, and improved yields, are critical levers to offset rising costs associated with expanding HBM capacity in AI accelerators. These innovations, coupled with emerging architectural strategies such as enhanced model efficiency, refined memory management, and the integration of high-bandwidth flash (HBF) as a complementary, lower-cost memory tier, collectively aim to reduce reliance on expensive HBM and alleviate cost pressures on AI systems.

The industry’s response to mounting memory cost pressures is multifaceted, combining supply expansion and manufacturing improvements with architectural innovation. Cloud service providers and server OEMs are already strategically reducing per-unit memory capacities—shifting from 96GB or 128GB RDIMMs to smaller configurations—to optimize memory footprints. Additionally, the deployment of ASICs with optimized memory usage exemplifies how hardware-level efficiency gains are essential alongside supply-side measures to sustainably manage the long-term economics of AI infrastructure.

Despite the emergence of long-term agreements with price caps aimed at curbing memory price volatility, cloud providers remain vulnerable to significant cost hikes, especially for volumes exceeding contracted limits. TrendForce’s projection of a 70–140% rise in HBM contract prices in 2027 underscores the persistent market pressures. This dynamic reinforces the urgency for both supply-side expansion and architectural innovation to collaboratively mitigate the financial strain memory costs impose on AI cloud infrastructure.

Sources

Part of these trends

Get the stories behind the trends

Deep-dive reporting and the weekly brief, in your inbox.