Back to list
Industry NewsMemory PricesAI HardwareSemiconductor

Historical Memory Price Trends 1960-2026: Analyzing the Evolution of DRAM, NAND, and HBM Costs

A comprehensive new dataset from DAM Stanford University tracks the historical trajectory of memory and storage prices from 1960 through projected 2026. The data highlights the massive deflationary trends in DRAM and NAND flash while shedding light on the emerging costs of High Bandwidth Memory (HBM) used in AI accelerators. By integrating the classic McCallum dataset with modern retail and analyst estimates, the report provides a detailed breakdown of DRAM generations from SDRAM to DDR5 and HBM iterations up to the projected HBM4 launch in Q3 2026. Furthermore, it offers modeled estimates of accelerator costs for industry leaders like Nvidia, AMD, Google, and Amazon, revealing how components like HBM and CoWoS packaging are reshaping the economics of high-performance computing.

Hacker News

Key Takeaways

  • Long-term Price Deflation: Memory prices across DRAM and NAND flash have seen a consistent decline on a log scale from 1960 to the present day.
  • HBM Market Dynamics: Unlike DRAM, HBM lacks a public spot market; pricing is based on confidential contracts and analyst estimates from firms like TrendForce and SemiAnalysis.
  • AI Accelerator Cost Structure: Modern AI hardware costs for Nvidia, AMD, Google, and Amazon are increasingly driven by HBM, logic dies, and advanced packaging (CoWoS).
  • Future Projections: HBM4 is projected to launch in Q3 2026, continuing the trend of increasing memory bandwidth and density.
  • Methodological Rigor: The dataset combines historical records from John C. McCallum with real-time Amazon retail pricing and Epoch AI modeled estimates.

In-Depth Analysis

The Evolution of DRAM and NAND Flash Pricing

The historical data provided by DAM Stanford University illustrates a profound shift in the affordability of digital storage. Starting from the early 1960s, the price per gigabyte for memory has plummeted, a trend captured through the integration of the John C. McCallum dataset and modern tracking via Keepa Amazon prices. The analysis categorizes DRAM evolution into distinct generations: Pre-DDR (including SDRAM and core memory), followed by the sequential releases of DDR, DDR2, DDR3, DDR4, and the current DDR5 standard.

While DRAM has historically been the primary focus of memory cost analysis, NAND flash has emerged as a critical secondary series. The dataset tracks the cheapest consumer-NVMe prices starting from 2016, with approximate anchors for earlier periods. This log-scale visualization reveals that while both technologies follow a downward price trajectory, their roles in the computing ecosystem have diverged, with NAND providing high-capacity storage and DRAM focusing on volatile, high-speed access. The methodology emphasizes that these figures represent the cheapest listed retail prices in nominal USD, rather than inflation-adjusted or contract-specific prices.

The Economics of AI Accelerators: HBM and Logic Costs

A significant portion of the analysis is dedicated to the cost breakdown of AI accelerators, utilizing modeled estimates from Epoch AI. This data covers the four largest designers in the space: Nvidia, AMD, Google (TPU), and Amazon (Trainium). The cost of these high-performance chips is no longer just a matter of silicon logic. Instead, it is a complex stack of expenses including the logic die, HBM (High Bandwidth Memory), auxiliary components, and sophisticated packaging technologies like CoWoS (Chip on Wafer on Substrate).

The shift toward HBM is particularly noteworthy. Because HBM is sold through confidential contracts directly to accelerator makers, there is no public spot market. Consequently, the data relies on sparse industry-analyst estimates. The transition from HBM2e to HBM3 and HBM3e shows a clear progression in bandwidth and capacity, with HBM4 expected to enter the market in the third quarter of 2026. The metric of $/TBps (cost per unit of memory bandwidth) is introduced as a vital KPI for evaluating the efficiency of these memory stacks relative to their price.

Methodology and Data Integrity

The dataset serves as a modern extension of the spirit of John C. McCallum’s classic work. For DRAM, the history is maintained through the McCallum dataset and extended from mid-2024 using Keepa Amazon prices. For NAND, the focus is on the cheapest consumer-grade NVMe drives. The HBM figures represent a different challenge, as they are modeled estimates rather than transaction prices. This distinction is crucial for researchers and industry analysts who must account for the lack of transparency in the enterprise memory market compared to the consumer retail market. The use of production-volume-weighted averages for accelerator costs ensures that the data reflects the actual market impact of dominant players like Nvidia.

Industry Impact

The data presented has significant implications for the future of the semiconductor and AI industries. As AI models grow in complexity, the reliance on HBM suggests that memory bandwidth, rather than just raw compute power, is becoming the primary bottleneck and cost driver. The projected launch of HBM4 in 2026 indicates that the hardware roadmap is already being laid out to support the next generation of large language models and generative AI tools.

For cloud providers like Google and Amazon, the ability to design in-house accelerators (TPU and Trainium) allows for potential cost optimizations across the HBM and logic die stack. However, the rising costs of advanced packaging and the specialized nature of HBM mean that the entry barrier for high-end AI hardware remains exceptionally high. This dataset provides the transparency needed for stakeholders to understand the underlying economic pressures of the AI infrastructure boom.

Frequently Asked Questions

Question: How does HBM pricing differ from standard DRAM pricing in this dataset?

Unlike DRAM, which has a public spot market and retail pricing tracked via platforms like Amazon, HBM pricing is based on confidential contracts. The dataset uses modeled estimates from analysts like TrendForce and SemiAnalysis to approximate HBM costs, as there is no public transaction data available.

Question: What components are included in the AI accelerator cost breakdown?

The cost breakdown for accelerators from companies like Nvidia and AMD includes the logic die, HBM stacks, packaging (specifically CoWoS), and auxiliary components. These are presented as production-volume-weighted averages.

Question: When is the next generation of High Bandwidth Memory expected to launch?

According to the projected data and industry estimates, HBM4 is expected to launch in the third quarter (Q3) of 2026.

Related News

How Google Duplex Evolved to Call Businesses on Your Behalf in Gemini
Industry News

How Google Duplex Evolved to Call Businesses on Your Behalf in Gemini

A Google Research Blog entry titled “How Google Duplex evolved to also call businesses for you in Gemini” describes its subject as the evolution of Google Duplex to support calls to businesses on a user’s behalf in Gemini. The entry is dated November 20, 2026, and the supplied article text consists only of the label “Generative AI.” The headline identifies the capability and connects it to Duplex, but the available text does not explain how the system evolved, how users request a call, or where the feature is available. The headline establishes the topic, but the supplied material does not provide a technical account or rollout details.

How AI and Lean Are Helping Automate Formal Verification to Fight Software Slop
Industry News

How AI and Lean Are Helping Automate Formal Verification to Fight Software Slop

In Latent Space, Richard MacManus examines formal verification as an emerging way to fight what the headline calls “slop.” The central claim is that proving software correct with mathematics has traditionally required specialist work, and that AI and Lean are now helping automate that process. The distinction is important: the reported change concerns how the work of proving correctness is done, not what counts as a proof. The supplied excerpt presents automation as assistance, without saying that it replaces specialists or completes proofs independently. It also does not specify which parts of the process AI and Lean help automate, how much work they save, or which software has been verified this way.

AI and Human Claim Solutions to the 167-Year-Old Riemann Hypothesis
Industry News

AI and Human Claim Solutions to the 167-Year-Old Riemann Hypothesis

TechCrunch AI reports that AI claimed a solution to the Riemann hypothesis, a mathematical problem the headline describes as 167 years old. A human also put forward a competing solution. The report, written by Tim Fernholz and published on November 20, 2026, describes mathematicians as astonished by the AI-driven result. The supplied excerpt does not explain either proposed solution, identify the AI system or the human claimant, or say whether either claim has been verified. The distinction between a claimed solution and a confirmed one is central to reading the report: the excerpt establishes that two claims are being reported, but it does not establish that either has resolved the hypothesis.