Why Intel's Optane Ended: A Memory Caught Between DRAM and NAND | Asianometry Listening Notes
Notes after listening to Asianometry's 2026-09-20 episode on Intel and Micron's 3D XPoint / Optane, from phase change memory's origins to its 2022 wind-down. Educational content only, not investment advice and not a recommendation of any stock; technology products and the companies behind them carry high uncertainty.
Contents
The flying squirrel has five skills, yet masters none of them. Better to do one thing well than many things poorly.
— Yan Zhitui, Family Instructions for the Yan Clan, “On Economy of Effort” (Northern Qi, c. 6th century; my translation)
What this episode covers
The 2026-09-20 episode of Asianometry is about 3D XPoint, the memory Intel and Micron announced together in 2015, which Intel later sold under its own brand, Optane. The launch claims were big: reads and writes 1,000x faster than NAND flash, and 1,000x the endurance. But by 2022, Intel was announcing it would wind the business down.
The episode doesn’t stop at those seven years, though. It goes back to the 1960s and the physics behind this kind of memory, then walks through three market entries and three exits. Afterward I kept coming back to one thought: when a product is wedged between two mature products, its fate is set by how fast those two neighbors improve.
Key points
1. It stores data by changing a material’s state. DRAM (a computer’s working memory, which forgets everything when the power goes off) and flash both store electrons. 3D XPoint takes a different route: it’s a phase change memory. A certain material conducts easily when its atoms are lined up in an orderly crystal, but melt it and quench it into a glass-like jumble and its resistance jumps by three to four orders of magnitude, so measuring resistance is enough to read the bit. The catch is that getting the material back into its crystalline state is the slow step, and it often sets the memory’s speed limit.
2. It first reached the market in 1970. In the late 1960s, materials scientist Stanford Ovshinsky discovered this phase change behavior in a class of materials called chalcogenides. In 1970, his company and Intel’s Gordon Moore co-authored an article in Electronics announcing a “Read-Mostly Memory.” Features back then were about 10 micrometers wide, so it drew a lot of power, took milliseconds to write, and wore out. Once DRAM scaled up and EPROMs (erasable read-only memory chips) arrived, it faded. My favorite detail comes later: years afterward, someone asked a 79-year-old Moore about the article, and he said he didn’t recall it at all.
3. Shrinking transistors revived it around 2000. Smaller cells need smaller programming currents, so the old power problem suddenly looked solvable. A Micron co-founder and its longtime CTO formed Ovonyx with Ovshinsky in 1999, and Intel Capital invested the next year, because Intel’s NOR flash business was hitting scaling limits. Then NAND, denser and cheaper per bit, won the iPod and the phone, and worldwide NOR sales fell 9% in 2007. Intel spun NOR out in 2008, merging it with STMicroelectronics’ flash business to form Numonyx, which pinned its hopes of a breakout on phase change memory.
4. The mobile path was cut off by 3D NAND. In 2009, Samsung announced a 512-megabit phase change memory and claimed over 20% power savings when paired with DRAM. Micron bought Numonyx in 2010 for about $1.2 billion in stock, and in 2012 it shipped a 1-gigabit part that ended up in Nokia’s Asha feature phones. But in August 2013, Samsung began mass-producing 24-layer 3D NAND, kicking off the steepest cost-per-bit curve in semiconductor history, and by January 2014 Micron had pulled its mobile phase change memory.
5. The hype and the shipping dates never lined up. 3D XPoint was pitched as sitting between DRAM and NAND: as fast as DRAM, yet it doesn’t forget when the power goes off. Intel and Micron wouldn’t say what it was made of, and it took internet sleuths digging through the patents to find phase change memory. The episode guesses the secrecy came from the technology’s bad reputation. In 2016, a Micron executive said demand was through the roof and quoted a customer: “Do you know how much DRAM I buy? You’d better get me some 3D XPoint.” Two years later, Micron still hadn’t shipped a single part. In 2018 the partners split, and Micron paid $1.5 billion for Intel’s half of their Utah fab, a plant the two had put $6 billion into.
6. Optane worked, but the gains were far from the headline. The first product, in 2017, was a 16 or 32 GB PC accelerator module, and Intel claimed boot times cut in half and games loading 65% faster. Real-world tests were more modest, with gains that depended on the game or the app. Analysts suspected the standard NVMe software interface (the usual way a computer talks to its storage) and the PCIe bus (the data lanes on the motherboard) boxed the benefits in behind outside bottlenecks. The reviewers’ verdict: a cool technology that actually worked, which is more than you can say for a lot of things.
7. The data center battle stalled on customers rewriting software. DIMMs that slot into server memory sockets didn’t arrive until April 2019, four years after the announcement, and they required a specific Xeon, chipset, and motherboard. There were two ways to use them. Memory Mode worked out of the box, but whether it paid off depended on DRAM prices at the time. The real value was in App Direct Mode: a large database restarting normally spends an hour or more reloading data into DRAM, but with Optane the data stays put, and one directional test showed a 12.5x improvement. The catch was that programmers had to rethink data consistency. The number I cared about most in this stretch: in 2020, Intel cited 582 proofs of concept (small-scale customer trials), and only 40% used App Direct. When Intel sold its NAND business to SK hynix for $9 billion that same year, outsiders backed out an implied Optane loss of over $500 million. In March 2021, Micron stopped development, saying its Utah fab was chronically under-utilized and costing $400 million a year, and moved its bet to CXL, an open standard for pooling compute, memory, and storage. Then in July 2022, on an earnings call where revenue fell 22% year over year, Intel announced the wind-down.
Further thoughts
If the specs were that good, why didn’t it sell?
The first time I heard “1,000x faster,” my gut said a product like that can’t lose, but after the full episode I think the problem was the choice of comparison.
Optane’s rivals were two moving targets. Below it, NAND’s cost per bit fell year after year from 2013 on, while Optane sold at four to five times NAND’s price, a gap that got harder to justify each year. Above it, Memory Mode’s value depended on DRAM prices, so cheaper DRAM shrank the case. A spec lead is just a snapshot taken on launch day, while the neighbors’ cost curves move every quarter.
There was another layer: the bottleneck itself moved. Intel built Optane to fix the problem of getting data into the CPU fast enough. When Micron exited, it said machine learning meant GPUs and AI accelerators also needed more memory, and products like High Bandwidth Memory were competing for that job. A product built around one bottleneck loses its reason to exist when the bottleneck moves.
When I look at a new technology now, I ask two more questions: how fast are the incumbents on each side improving, and how many years until today’s lead gets eaten? Then a third: will the bottleneck it targets still be a bottleneck in five years? I won’t always answer these well, but they push me toward the cost curves and away from the launch-day multiplier.
Management says demand is through the roof. Should I believe it?
I’ve been swayed by “strong demand” on earnings calls more than once, so this episode made a clean case study for me.
In 2016, a Micron executive said capacity for the first couple of years was fully spoken for and major customers were lining up. Two years later there was no product. In 2020, Intel pointed to 582 proofs of concept, which sounds like a lot, until you see that only four in ten used the mode that delivered the real value. The distance between customers saying they want something and customers changing their own systems has a name: switching cost. App Direct required customers to rewrite software and rethink how data persists. That cost landed on the customer, and Intel couldn’t pay it for them.
The reading I’ve kept for myself is to look for numbers that can be graded later: when the product ships, which mode customers use, how full the fab runs. When Micron said in 2021 that its Utah fab was consistently under-utilized, that was the answer key. These numbers arrive later than any quote, but they don’t flatter anyone.
One more detail: after Micron left, customers could buy only from Intel, and the technology was proprietary. Buyers dread single sourcing. An ecosystem takes years of quiet work to build, and one public exit can set it back. When a company pushes a new platform, I now watch whether its partners are staying.
References
- Asianometry, “Intel’s Optane Memory: A Promise Unfulfilled,” 2026-09-20
- R. G. Neale, D. L. Nelson, and Gordon Moore, “Read-Mostly Memory” article, Electronics, 1970
- Benjamin Lee et al., “Architecting Phase Change Memory as a Scalable DRAM Alternative,” ISCA 2009 (Microsoft Research and Carnegie Mellon University)
- Micron’s March 2021 announcement ending 3D XPoint development
- Intel Q2 2022 earnings call
One thing to take with you
Judge an option against what the competition will look like next year, not what it looks like today. On launch day, Optane beat NAND by 1,000x. What it lost to was every year of NAND price cuts that followed.
Here’s something I’ve tried: pick one thing you’re weighing whether to switch to, maybe a new app, a new workflow, or a new phone, and write two lines on paper. Line one: where it beats the old option today. Line two: what the old option will look like a year from now. If you can’t fill in line two, spend ten minutes finding out what the old option changed over the past year, then decide.
This article is an educational discussion of investment method. It is not advice to buy or sell any individual security, offers no target prices, and does not analyze any current holding. Investing carries risk; make your own decisions or consult a qualified professional.