Two data points. Both public. Both largely ignored by cloud engineering teams — until the bill arrives.
committed across two new fab facilities. Both operational 2028. Both 100% dedicated to HBM for AI accelerators.
↓ Why this matters — Chapter 3A 1TB SSD hit an all-time low of $69 in July 2023. The same drive costs $125+ today. The cheapest storage hardware on earth just got significantly more expensive.
Cloud storage prices follow hardware costs — with a lag. AWS cut EBS prices five times between 2008 and 2020 as NAND got cheaper. Now NAND is getting more expensive, and the fabs that made it are being converted to AI chips. The free lunch is over.
Here's the full picture — and what it costs you if you're still provisioning the old way.
When EBS launched in 2008, it was expensive. But AWS kept cutting prices, roughly every two years. Engineers who over-provisioned storage could rely on the next price drop to do the right-sizing for them. It was a rational bet — until it wasn't.
The SSD-backed EBS volumes most teams use — gp2, gp3, io1, io2 — are all built on NAND flash. It's the same memory technology in your phone, your laptop SSD, and every enterprise storage array. AWS buys it at wholesale, marks it up, and bills you per GB-month. (EBS also offers HDD-backed types like st1 and sc1 — but those aren't what's driving most enterprise storage spend.)
For 20 years, NAND followed a simple pattern: denser chips every 18–24 months → more gigabytes per wafer → lower cost per GB. AWS passed those savings on. The EBS price curve tracked NAND deflation almost exactly.
Here's the problem: just four companies produce almost all the NAND in the world. If something disrupts them, there's no spare capacity anywhere else in the system.
Training large AI models requires a very different type of memory: HBM (High Bandwidth Memory). It's stacked directly on top of Nvidia H100/H200 GPUs. And it's manufactured on the same advanced fabrication lines as NAND.
The business case for HBM is dramatically better than NAND. So what happened? The fabs shifted capacity. NAND production was cut. Supply tightened. The deflation stopped — and reversed.
And here's where it gets more deliberate: even as consumer demand softened in 2025, Samsung, SK Hynix, Micron, and Kioxia all actively cut NAND output further to drain inventory and push prices up. This isn't just a supply shortage — the manufacturers now have the incentive and the coordination to manage prices upward. → Blocks & Files, Jan 2026
SK Hynix HBM revenue grew 4.5× between 2023 and 2024, reaching 40% of its total DRAM revenue by Q4 2024. → SK Hynix earnings
Estimated NAND vs HBM/DRAM capacity split at SK Hynix — NAND is losing ground every year
SK Hynix originally planned its M15X facility as a NAND flash fab. The board voted to convert it to DRAM and HBM production instead. Capacity that was explicitly earmarked for SSDs got redirected to AI accelerators — before a single wafer was made.
The fab is targeting 50,000 wafers/month of HBM by Q4 2026. The NAND those wafers were meant to produce simply won't exist. → KED Global · → Tweaktown
The AI demand shock didn't hit a resilient supply chain. It hit one that was already stressed. Three separate events reduced the system's ability to absorb any shock at all.
Ukraine supplied ~90% of US semiconductor-grade neon — used in the lithography lasers that pattern chips. When Russia invaded in February 2022, the two main Ukrainian suppliers (Ingas & Cryoin) shut down. The supply buffer for fabs thinned overnight. → VentureBeat
October 2022: US banned exports of Nvidia A100/H100 to China. December 2022: YMTC — China's largest NAND maker — was added to the Entity List, cutting it off from US equipment. → Federal Register. Chinese NAND supply dropped; China accelerated its own fab build-out — a wildcard for future supply.
Taiwan holds ~92% of advanced (<10nm) chip manufacturing capacity globally. TSMC has no short-term substitute. Any disruption there — political, military, or natural — would be a supply catastrophe. That tail risk keeps capacity expansion conservative everywhere.
"None of these factors individually is decisive. Together, they put a floor under NAND costs that simply did not exist five years ago."
This isn't an AWS-specific story. The underlying NAND economics affect every cloud provider. Azure and GCP have followed an almost identical playbook — launching new, cheaper volume types rather than cutting prices on existing ones. Then going quiet.
Launched gp3 at $0.08/GB-month, 20% cheaper than gp2. gp2 stayed at $0.10. Neither has moved since. Last real cut: December 2020. → AWS Blog
Launched Premium SSD v2 as a more flexible, lower-cost alternative to Premium SSD v1 — decoupled IOPS/throughput provisioning, analogous to gp3. Premium SSD v1 pricing: unchanged. → Azure Blog
Launched Hyperdisk Balanced and Hyperdisk Throughput in 2023, with Hyperdisk Storage Pools claiming up to 30–50% TCO reduction. Persistent Disk pricing: unchanged. → Google Cloud Blog
The old playbook was rational: overprovision, the next price cut will wash the waste away. That worked for a decade. But industry data shows the average EBS volume is running at only 30–40% utilisation. The majority of what you're paying for is sitting idle.
In a deflationary environment, that waste self-corrected every 18–24 months. At flat pricing, it compounds forever.
Drag the sliders — assume 30–40% utilisation as a realistic baseline
Lucidity isn't a discount tool. It's the operational response to a structural market shift. It starts with a free Assessment — a complete picture of your storage utilisation in 25 minutes. Then the AutoScaler keeps it right-sized continuously. And Lumen gives you ongoing visibility and intelligence across your whole storage environment.
No agents, no downtime, no config changes. Get a full picture of utilisation across every EBS volume. Most teams are surprised by what they find.
Continuously monitors actual disk usage and adjusts provisioned capacity in real time. Scales up before you run out. Scales down when utilisation drops. No manual work ever.
Manual EBS resizing requires stopping instances or risky live migrations. Lucidity's approach avoids both. Right-sizing happens transparently, without any operational risk.
What customers typically see:
"The teams that treat storage right-sizing as a continuous discipline — not a one-time cleanup — will carry a permanent cost advantage over those still waiting for the market to bail them out."