A Requiem for Optane
Intel's Optane storage and memory products are among the many technologies that died because they arrived before the world needed them. Expensive, badly understood, and awkwardly priced, 3D XPoint survived just five years before Intel pulled the plug.
Things might have been different if it were launched today. The shift from AI training to inference has driven tremendous demand for DRAM and NAND flash. The same properties that made Optane a niche product a few years ago would have made it ideally suited to write-intensive AI workloads.
What 3D XPoint Actually Was
Intel and Micron co-developed Optane — or more precisely, 3D XPoint memory — and unveiled it in 2015. The goal was to bridge the gap between conventional NAND flash used in SSDs and DRAM used in DDR4 (and later DDR5).
Like NAND, 3D XPoint was non-volatile, meaning data persists when unpowered and making it appropriate for storage applications. But unlike NAND, 3D XPoint didn't store data using trapped electrons. Instead, it relied on a phase change material that was both extremely fast — achieving sub-10-microsecond latencies (even lower for later PMem modules) — and absurdly write-endurant.
Intel's penultimate Optane SSDs boasted endurance of 100 drive writes a day and a mean time between failures of two million hours. Those figures are extrapolated based on how 3D XPoint reads and writes data, but Intel was probably being conservative with its claims. These specs remain unrivaled today.
That one-two punch of endurance and latency, particularly for random reads and writes, meant Optane could function as a second tier of system memory. Optane SSDs are extremely low latency for storage-class memory, but they're still orders of magnitude less responsive than DRAM, which can hit 100 nanoseconds or lower. Intel's PMem DIMMs were capable of latencies of roughly 350 nanoseconds.
Intel's Optane persistent memory also had capacities up to 512 GB per DIMM, at a time when the most you could expect out of DDR4 was about 128 GB — and only if you had deep pockets. These persistent memory DIMMs could behave like main memory, with standard DDR4 functioning as a massive L4 cache when enabled, as a storage pool, or in an application-aware memory mode that exposed Optane memory directly to select applications.
Why Optane Failed: Pricing and Timing
Despite Optane's many strengths, it was awkwardly priced. For the same memory density, 3D XPoint was consistently more expensive than NAND and, while cheaper than DRAM, significantly less performant. There was rarely a scenario in which users were better off with Optane than they would be just buying more NAND or fewer, bigger DIMMs.
To make matters worse, as the boffins at TechInsights noted at the time, while 3D XPoint had its merits, it wasn't improving quickly enough to keep up with NAND flash on bit density. By 2021, Micron had had enough. The memory vendor announced it would end development of 3D XPoint products. With no source of new silicon for its SSD and persistent memory products, the writing was on the wall for Optane.
Intel's then-CEO Pat Gelsinger officially pulled the plug in 2022, ending development of new products under the Optane banner. The P5810X and P5811X, the final Optane products to roll off the assembly line, launched in late 2022.
The AI Inference Workload That Came Too Late
The same year Intel pulled the plug on Optane, a new workload was emerging that would completely turn the datacenter on its head — and might very well have been the killer app to justify 3D XPoint's continued development.
In November 2022, OpenAI, then a relatively obscure AI startup, lifted the curtain on ChatGPT, kicking off an AI arms race and sparking a massive influx of money into the tech sector.
Large language models are among the most resource-intensive workloads in the world. But until 2025, most of that compute was dedicated to training bigger, smarter, and less hallucination-prone models. The arrival of DeepSeek early in that year signaled the shift toward inference, a workload that needs plenty of compute, memory, and storage.
Modern inference engines are tuned to maximize efficiency. One way is by caching the key-value pairs used to track model state. Chatbots and agents are inherently iterative. For multi-turn sessions, every prompt contains not only new information, but also every request and response that came before it. Recomputing all of that every time a new request is made is wasteful, so instead that information is computed, cached, and reused.
The problem is at large context sizes and high concurrency, these KV caches can get rather large. A single 64,000-token sequence in a model like DeepSeek R1 can chew up four gigabytes of GPU memory. Multiply that across hundreds or thousands of users and it adds up quickly.
Modern GPUs don't have much memory, and what they do have is usually tied up hosting model weights. Many inference engines therefore support KV cache offloading, either natively or through plugins. Once GPU memory is exhausted or chat sessions stale, older chats are ejected to system memory.
But DRAM is expensive, in short supply, and may not even be enough. Nvidia is reportedly cutting the amount of LPDDR5X shipped as part of its Vera Rubin platform. Because of this, KV caches where a user has walked away and the session has been idle for an hour or more may be offloaded to flash storage arrays for longer-term storage.
The problem is that KV caching is a write-intensive workload and NAND has a finite number of writes it can make before it's shot. There's a lot of work being done to mitigate this, but the fact remains: if you write enough data to NVMe storage, eventually it wears out.
Optane's otherworldly write endurance and low latency, particularly when concerning small random writes which dominate KV-caching, would have made it a perfect choice for this workload.
What Came After: CXL and the SCM Void
One of the nails in Optane's coffin was that Compute Express Link (CXL) was already on the horizon. If the primary reason for adopting Optane was to get your hands on a large pool of reasonably fast DRAM-like memory, then CXL offered all the benefits and none of the compromises. Early CXL implementations were essentially a remote memory controller to which you could attach your choice of memory, DDR4 or DDR5.
Need more memory than your CPU supports? Just plug a CXL memory expander into a free PCIe slot and you were off to the races. Later revisions of the memory-coherent protocol added support for pooling, and later sharing.
However, CXL hasn't changed the fact that DDR5 is expensive at the best of times, and we are currently living in the worst of times for DRAM prices. (While your CPU may only support DDR5, the CXL controller might still be able to use DDR4. This is exactly what Meta is doing to boost the memory capacity of its systems on the cheap.)
In the absence of Optane, memory vendors have attempted to fill the void with write-optimized NAND of their own. Kioxia's XL-Flash is one such example of storage class memory (SCM) that promises many of the same benefits as Optane, including advertised write latencies under 10 microseconds and endurance up to 60 drive writes a day, through the use of SLC (1-bit per cell) memory. There's also a crossover with CXL here: XL-Flash can be exposed as a CXL device to simplify memory tiering.
Samsung also attempted to replicate many of Optane's qualities with its Z-NAND tech, but the tech still falls well short of 3D XPoint. Samsung is reportedly revamping the tech and aiming for a 15x performance increase over conventional NAND.
Alas, neither quite lives up to Optane's legacy. The technology was simply too far ahead of its time. Had Intel and Micron waited just a few years longer, it might have ridden the AI wave all the way to victory.
Related Reading
- LLM KV Cache Compression: Quantization, Eviction & Paging Strategies for Cost-Throughput Optimization in 2026
- The Memory Chip Squeeze: AI's Hidden Inflation Problem
- The Memory Tipping Point: Why AI's Appetite Changed Everything
