CXL Specification Advances to 4.0, QLC Orders Already Booked Through 2028: The Two Storage Roadmaps Diverge in Pace

Abstract

CXL does not lack standards but lacks platforms: the 4.0 specification and first deployments are in place, the CPU platform is delayed to 2027, and volume ramp is still pending;

QLC has passed the cost inflection point: capacity has grown about fourfold in three years, orders and prices are resonating, and the HDD supply gap has opened a substitution window;

Commercialization hinges on three factors: platform, ecosystem, and economics; the inflection point is determined by a reversal in cost structure, not by technological maturity.

The storage industry in 2026 is sending two opposite signals: CXL has advanced its standard to version 4.0 and has its first cloud provider deployments, yet it is still waiting for a CPU platform that will not arrive until 2027; QLC, meanwhile, has used a roughly fourfold increase in drive capacity over three years and steadily rising orders to fully open the window for replacing hard disk drives.

CXL standard in place, only the platform missing

On November 18, 2025, the CXL Consortium released version 4.0 of the specification, doubling the link rate from 64 GT/s to 128 GT/s and aligning directly with the not-yet-commercial PCIe 7.0. At almost the same time, Microsoft and Astera Labs announced at the same event the industry's first publicly named CXL memory deployment: Azure M-series virtual machines attach pooled memory through the Leo controller, underpinned by Intel Xeon 6, and are currently in private preview.

On the QLC side, there is no such thing as a "preview." On May 5, 2026, Micron announced that its 245TB 6600 ION enterprise SSD had officially shipped, making it the largest-capacity SSD currently on sale.

Within the same storage supply chain, the rhythms of these two developments are completely different. The CXL story has been told for seven years and has finally won its first customer willing to be publicly named; QLC took only three years to raise single-drive capacity from 61.44TB to 245TB, with product, price, and order curves all rising at once. One technology is enduring the eve before dawn, while the other is stepping on the accelerator; the point of divergence is worth examining separately.

The pace of CXL standard iteration is rare among interconnect protocols: 1.0 was released in March 2019; 2.0 in 2020 introduced memory pooling and switches; 3.0 in August 2022 raised the rate to 64 GT/s; 3.1 and 3.2 followed in November 2023 and December 2024; and 4.0 in November 2025 doubled the rate again. Six versions of the specification have been written, yet the problem is stuck at the next layer—CPUs capable of running CXL 3.x have not yet entered mass production.

The carrier platform for CXL 3.x is Intel's next-generation server chip, Diamond Rapids. Details disclosed at Hot Chips 2026 show that the chip has up to 256 cores and provides 128 PCIe 6.0/CXL 3.x lanes, with its launch delayed to 2027. The Xeon 6 available today remains at PCIe 5.0 and CXL 2.0. With the platform arriving late, memory modules can only wait as well: Samsung delayed mass production of its CXL 3.1 memory modules by about nine months, and Digitimes attributed the reason directly to delays in Intel and AMD server platforms.

An even harsher blow came from the storage manufacturers themselves. In July 2026, South Korean media outlet ZDNet Korea reported that Samsung, SK Hynix, and Micron had successively halted or sharply scaled back their in-house CXL controller development.

On the demand side, however, real examples have emerged. Meta's in-house chip Vistara reuses DDR4 memory modules from retired servers through bridge controllers, reducing server counts by up to 25% in some machine learning inference workloads; Microsoft's M-series virtual machines take the route of Astera Labs' commercial controllers, targeting in-memory databases such as SAP HANA and using CXL memory to raise the memory ceiling of virtual machines in exchange for lower overall TCO.

Yole's forecast of more than $15 billion by 2028 comes with one precondition: that the CPU roadmap proceeds on schedule. Diamond Rapids slipping to 2027 means that large-scale volume ramp of CXL 3.x will at the earliest occur after 2027. The standard is in place, the controllers are in place, the first deployments are in place—only the platform is missing.

The steep curve: QLC's three-line resonance

QLC's product curve is much steeper. In 2023, Solidigm's D5-P5336 pushed enterprise SSDs to 61.44TB; in 2024, the same series doubled to 122.88TB; in July 2025, Kioxia released the LC9 series 245.76TB product, whose 32-die stacked BiCS QLC flash won an FMS exhibition award; in August of the same year, SanDisk showcased a 256TB UltraQLC platform sample drive at FMS 2025; in May 2026, Micron's 6600 ION officially shipped at 245TB, equipped with its own ninth-generation QLC NAND.

What is pushing this curve upward is the collapse on the HDD side. In February 2026, Western Digital announced that its full-year high-capacity capacity was essentially sold out; Seagate subsequently pushed allocation of nearline drive shipments out to 2028. In the Japanese retail market, the average price of HDDs above 16TB has more than doubled compared with November 2025. The speed at which AI data lakes are consuming HDD capacity has exceeded the expansion speed of the three HDD manufacturers, leaving procurement orders with no choice but to turn to QLC.

Manufacturers have long done this math. Solidigm compared a 122TB QLC SSD plus VAST system with a traditional HDD plus Ceph solution: ten-year total cost of ownership fell from $85.62 million to $35.19 million, saving about 58.9%. Micron's description of the 245TB drive is even more direct—a single drive approaches a quarter of a PB, whereas in the past this would have occupied an entire rack.

Behind such figures is a reversal in industry perception. QLC was once categorized as a low-cost consumer-grade product because of its short write endurance, while enterprise nearline storage was the hereditary domain of HDDs; now, with NAND stacking and packaging density continuously increasing, single-drive capacity has been pushed past the 200TB threshold, and what the main battlefields of cold data—data lakes, backup, and archiving—value is precisely the per-rack capacity that this density brings. The HDD moat remains only in write-intensive workloads, and most nearline storage scenarios do not stand inside that moat.

Distinguishing the eve from the acceleration

Placing the two lines side by side within the same framework, the criteria are actually the same set: platform, ecosystem, and economics. QLC is standing firm on all three legs—the PCIe host interface is readily available, the NVMe ecosystem is mature, and cost per TB in 2026 is lower than that of HDDs. CXL already has a foundation in standards, products, and first deployments, but the platform must wait for Diamond Rapids in 2027, the ecosystem still awaits validation at volume, and economics have just been dragged to a critical point by this round of DRAM market conditions.

DRAM price increases are a tailwind for QLC and a double-edged sword for CXL. TrendForce expects DRAM contract prices to rise 90% to 95% quarter over quarter in the first quarter of 2026; memory market sites show DDR5 16Gb die prices up about 480% year over year. With memory this expensive, pooling is showing its cost-saving value for the first time—distributing idle capacity from fully configured servers to machines that need it; in that SC25 panel discussion, what the four manufacturers repeatedly talked about was the word "utilization." Yet the same round of price increases is also pushing up the cost of CXL modules themselves, and the premium customers pay for pooling moves with the market. The eve's alarm clock is already ringing; no one can say exactly which mark it has reached.

Standards arriving before platforms is not unprecedented in storage history. The DDR5 specification was finalized in mid-2020, and the first consumer platform support did not arrive until the end of 2021—a year and a half in between—and the industry still proceeded step by step. CXL's wait this time is somewhat longer, but the path is the same: specification freeze, device availability, platform follow-through, application volume ramp—only after all four steps is the page turned. At present it stands between the second and third steps; misjudging the position of this step means either waiting a year for nothing or betting a year too early.

Commercialization actions can be scheduled along this timeline: the procurement and migration window for QLC is right now, and the HDD supply gap will extend to 2028, so acting late means buying only goods at higher prices; CXL is worth small-scale validation in 2026 to 2027, as Azure's M-series virtual machines have already opened private preview, and by the time Diamond Rapids launches there will be measured data on hand for comparison.

The inflection point never arrives at the moment when a technology is most mature; it arrives at the moment when the cost structure flips. QLC crossed that line in 2026, with orders and prices stamping its passport; CXL is still on the other side of the line—the specification has reached 4.0, deployments have entered cloud providers' data centers, and all that is missing is a reason for the CFO to sign off. With every round of DRAM price increases, that reason gets one step closer.