Liquid-Cooled Enterprise SSDs Have Arrived: Inside the CM9-DLC and the Shift to Direct Liquid Cooling

A liquid cooled enterprise SSD was, until this week, a custom engineering project rather than a product you could order. That changed on October 12–13, 2026, when Kioxia announced direct liquid cooling (DLC) models for its CM9 Series enterprise SSDs — the CM9-DLC lineup — timed to the OCP Global Summit 2026, running October 12–15 in San Jose. The new drives bring a cold-plate cooling option to the E3.S form factor, keep support for traditional air cooling, and are explicitly positioned for AI and other high-performance enterprise environments.

Why does an SSD need liquid cooling at all? Because the storage layer in an AI data center no longer sits quietly in the background. PCIe 5.0 controllers and their NAND packages can sustain power draws that would throttle a passively cooled drive, and the racks they live in are already being redesigned around liquid cooling for GPUs. Once the cold plates are in the rack, extending that cooling loop to the SSDs is the logical next step — and the CM9-DLC is the first mainstream signal that the storage industry agrees. For context on the summit where this news landed, see our OCP Global Summit 2026 storage preview.

The announcement: what the CM9-DLC actually is

The CM9-DLC adds direct liquid cooling support to the existing CM9 Series enterprise SSD lineup in the E3.S form factor, targeting cold-plate-enabled server designs. The drives remain compliant with PCIe 5.0, NVMe 2.0, NVMe-MI 1.2c, and the OCP Datacenter NVMe SSD v2.5 specification, with dual-port support — a checklist aimed squarely at hyperscale and OEM buyers who validate against open specifications.

The published specifications, per the October 12–13 announcement:

• Interface: PCIe 5.0 x4, NVMe 2.0, NVMe-MI 1.2c; OCP Datacenter NVMe SSD v2.5 compliant
• Form factor: E3.S with cold-plate (direct liquid cooling) support; also works in air-cooled environments
• Capacities: up to 30.72TB in the read-intensive (1 DWPD) endurance class, up to 25.6TB in mixed-use (3 DWPD)
• Sequential performance (128 KiB, QD32): up to 14.8 GB/s read, 11 GB/s write
• Availability: currently sampling to select customers; showcased at OCP Global Summit 2026

Two design details matter more than the headline specs. First, the dual-mode cooling support means buyers do not have to commit their whole fleet to liquid infrastructure on day one — a drive that can run air-cooled today and cold-plate-cooled tomorrow dramatically lowers the adoption risk. Second, the 1 DWPD and 3 DWPD endurance classes, explicitly tied to SSD context caching for AI workloads, tell you exactly where this drive is meant to live: in the tier that feeds GPUs and accelerators with cached model data, KV caches, and training checkpoints.

Background: why data center SSDs got too hot for air

Three forces converged to make liquid-cooled storage inevitable.

1. PCIe 5.0 doubled the heat, not just the speed. A PCIe 5.0 x4 NVMe SSD can sustain 14 GB/s of throughput, but its controller runs far hotter than a Gen4 equivalent at full tilt. In densely packed 1U and 2U servers, enterprise drives already approach thermal limits under sustained AI checkpointing or inference workloads, and the standard response — heavier heatsinks, more directed airflow — costs rack space and fan power.

2. Racks are already going liquid for GPUs. AI training racks now push tens of kilowatts, far beyond what air cooling can handle economically. Operators are installing cold plates, rear-door heat exchangers, and facility water loops. Once that plumbing exists, cooling the SSDs through the same loop costs almost nothing per drive — but it requires the drives to be built for it, with thermal interfaces and form factors that mate to cold plates. That is exactly the gap the CM9-DLC fills.

3. The E3.S form factor was designed for this transition. E3.S was created for dense, hot-aisle-free compute-and-storage servers: it is longer and thinner than U.2, allows front-serviceability, and has defined thermal envelopes that accommodate cold plates. If you are still deciding between enterprise form factors, our U.2 vs EDSFF explainer walks through the trade-offs.

Technical deep-dive: how direct liquid cooling works on an SSD

Direct vs. indirect: the cooling taxonomy

"Liquid cooling" covers several very different architectures, and the distinction matters for buyers:

• Direct liquid cooling (DLC): a cold plate — a machined plate with coolant channels — makes direct contact with the drive's thermal interface, carrying heat away through a liquid loop. This is what the CM9-DLC supports.
• Immersion cooling: the entire server or drive is submerged in dielectric fluid. Extremely efficient, but it changes every service procedure and is rare outside specialized deployments.
• Rear-door heat exchangers (RDHx): still fundamentally air cooling inside the server; liquid only carries heat out at the rack door. Your SSDs never see liquid.

DLC is the pragmatic middle ground: it targets the hottest components with liquid while leaving the rest of the server architecture — and, crucially, the SSD's own firmware and qualification — largely intact.

Why 14.8 GB/s matters thermally

A PCIe 5.0 SSD sustaining 14.8 GB/s sequential reads is moving the theoretical maximum of a Gen5 x4 link. At that throughput, controller power can exceed 15–20W on the drive alone, concentrated in a few square millimeters of silicon plus the NAND packages. Air cooling a 30.72TB drive at full throughput in a 45°C inlet environment is a genuine engineering problem; a cold plate converts it into a plumbing problem, which data centers are already solving at rack scale.

The KV cache angle

Kioxia's announcement specifically mentions AI architectures that rely on large-scale memory and KV (key-value) cache resources. This is not marketing garnish. As inference serving moves toward larger context windows and persistent KV caches, the working set often overflows GPU HBM and lands on fast NVMe — a tier sometimes called "SSD context caching." That tier must sustain extremely high read throughput with minimal tail latency, which means the controller runs hot for hours at a time. It is precisely the workload profile where air-cooled drives throttle and liquid-cooled drives earn their keep.

Dual-port and OCP compliance

Dual-port NVMe support means the drive can present itself to two hosts or controllers simultaneously — standard for high-availability storage arrays and for the disaggregated architectures hyperscalers favor. Compliance with OCP Datacenter NVMe SSD v2.5 matters because it is the specification hyperscale buyers write into RFPs; a drive that meets it can be second-sourced across vendors without re-qualification drama.

Benchmarks and scenarios: where liquid cooling pays off

Think of the CM9-DLC's value in three deployment scenarios:

Scenario 1: AI inference serving with SSD context caching

Payoff: highest. Sustained high-queue-depth reads from many concurrent inference requests keep the controller at maximum power indefinitely. A cold plate holds the NAND and controller at stable temperatures, which reduces thermal throttling events and keeps tail latency (p99/p99.9) consistent — the metric that actually determines user-perceived quality in LLM serving. Energy savings compound: liquid loops move heat at far higher efficiency than fans, and DLC infrastructure is documented to enable more servers per rack than air cooling for the same power envelope.

Scenario 2: Mixed enterprise virtualization on PCIe 5.0

Payoff: moderate. Typical VM and database workloads are bursty rather than sustained; the drive spends much of its life well below thermal limits. Here the DLC option is an insurance policy and a density play: if you are already deploying cold-plate racks, adding the SSD to the loop is cheap, and the drive's air-cooled fallback means you can mix racks without maintaining two drive SKUs.

Scenario 3: Traditional air-cooled data center

Payoff: near zero today — and that is fine. The CM9-DLC's dual-mode design means a buyer can standardize on one SKU while only a fraction of the fleet runs on cold plates. There is no forced infrastructure upgrade. That is the strategically smart part of this announcement: it lets early adopters pilot liquid-cooled storage inside an otherwise conventional environment.

Buying checklist: should your next enterprise SSD be liquid-cooled?

This drive is sampling to select customers — it is not a retail product and will not be. But its arrival sets the agenda for enterprise SSD procurement through 2027. Use this checklist:

1. Cooling roadmap: Is your facility already planning cold-plate or liquid-assisted racks for GPU nodes? If yes, SSDs are a near-zero marginal addition. If no, air-cooled enterprise SSDs remain the rational buy.
2. Workload profile: Do your drives sustain high throughput for hours (checkpointing, inference caches, analytics scans)? Sustained workloads benefit; bursty workloads rarely hit thermal limits.
3. Form factor standardization: Are you on E3.S or moving there? DLC options are emerging first in E3.S, not U.2.
4. Endurance class: Match DWPD to the workload — 1 DWPD read-intensive for caching and inference, 3 DWPD mixed-use for heavier write mixes. Overspending on endurance classes you will never use is the most common enterprise SSD procurement error; see our enterprise vs consumer SSD price split for how endurance and warranty shape pricing.
5. OCP compliance in RFPs: If you buy against OCP Datacenter NVMe SSD specifications, verify which revision your validation pipeline requires — the CM9-DLC targets v2.5.
6. Dual-sourcing: Sampling-stage products carry lead-time and allocation risk. Keep a qualified second source until volume supply is confirmed.

Common mistakes to avoid

• Assuming liquid cooling means better performance per drive. It does not raise the PCIe 5.0 ceiling; it lets the drive sustain its rated performance under conditions where an air-cooled drive would throttle. The spec sheet numbers are the same; the consistency is what you are buying.
• Treating DLC as an all-or-nothing migration. Dual-mode drives exist precisely so you do not have to. Pilot on the hot tier first.
• Forgetting the facility side. A cold-plate SSD in a rack with no coolant loop is just an SSD. The cooling decision is made at the facility level, not the drive level.
• Confusing announcement with availability. "Sampling to select customers" is the earliest commercial stage. Volume timelines, pricing, and full qualification data come later — plan procurement calendars accordingly.
• Overlooking NAND endurance math for AI writes. Training checkpoints can be enormous and frequent; size your DWPD class from measured write amplification, not from capacity alone.

What rivals are likely to do next

No vendor gets to own a thermal interface. Direct liquid cooling for SSDs is a mechanical and specification problem — cold-plate dimensions, inlet temperature envelopes, leak-detection wiring — not a silicon secret, which means the CM9-DLC announcement is best read as the starting gun rather than the finish line. The likely sequence: rival enterprise SSD makers qualify their own DLC-capable E3.S drives through 2027, server OEMs publish standardized cold-plate keep-out zones and connector placements so one plate design fits drives from multiple vendors, and the OCP ecosystem folds SSD liquid cooling into its open specifications the way it already standardized air-cooled drive thermals.

The interesting second-order effect is on the controller vendors. A PCIe 5.0 NVMe controller designed with cold-plate cooling as the primary thermal path can be tuned differently — higher sustained clock targets, tighter thermal margins, less silicon spent on throttling logic. Today every controller is designed for the worst-case air-cooled rack; once a critical mass of data-center racks assumes liquid, controller architects can optimize for that world. That is a multi-year shift, but it starts with exactly this kind of announcement.

There is also a cost angle that favors rapid adoption. Facility liquid loops are dominated by the fixed cost of pumps, coolant distribution units, and plumbing. Adding an SSD cold plate to an already-liquid rack is a small incremental cost, while the alternative — engineering ever-more-elaborate air cooling for a 25W+ drive in a dense chassis — gets more expensive every generation. When the marginal cost of cooling an SSD with liquid is lower than the marginal cost of cooling it with air, the transition becomes a procurement no-brainer rather than a technology bet.

2026 market context: storage follows compute into the liquid era

The CM9-DLC did not appear in a vacuum. Through 2026, every layer of the memory market has been reshaped by AI demand: TrendForce projects conventional DRAM contract prices up 10–15% and NAND flash contract prices up 15–20% in Q4 2026, driven by cloud providers expanding AI infrastructure procurement. Memory makers are prioritizing HBM, server DRAM, and enterprise storage — the categories with the strongest demand and margins — which is exactly why enterprise SSD innovation is accelerating while consumer supply tightens.

The pattern is consistent: where AI demand concentrates, engineering follows. We saw it in HBM, in server DRAM allocation, and now in storage thermals. The liquid-cooled SSD is less a product announcement than a market signal — the storage tier is being redesigned as a first-class citizen of the AI rack, not as an afterthought bolted to the motherboard.

Expect 2027 to bring three things: rival DLC-capable enterprise SSDs (no vendor holds a thermal interface exclusively), standardized cold-plate mechanical definitions for E3.S across more OEMs, and — the one to watch — the first SSDs whose spec sheets assume liquid cooling, with air-cooled operation as the fallback rather than the default. The CM9-DLC's dual-mode design is the transition vehicle; the destination is a rack where the SSD is simply another component on the coolant loop.

FAQ

What is a liquid-cooled enterprise SSD?

An enterprise SSD designed to transfer heat to a liquid cold plate (direct liquid cooling) instead of relying solely on airflow. The Kioxia CM9-DLC, announced October 2026, is among the first mainstream examples, supporting cold-plate cooling in the E3.S form factor while remaining air-cooling compatible.

Why would an SSD need liquid cooling?

PCIe 5.0 controllers sustaining up to ~14.8 GB/s generate substantial heat, and AI workloads like SSD context caching keep drives at full throughput for hours. In dense, already-liquid-cooled AI racks, extending the coolant loop to SSDs prevents thermal throttling and improves energy efficiency.

What are the CM9-DLC's key specifications?

PCIe 5.0, NVMe 2.0, NVMe-MI 1.2c, OCP Datacenter NVMe SSD v2.5 compliant, dual-port, E3.S form factor, up to 30.72TB (1 DWPD read-intensive) or 25.6TB (3 DWPD mixed-use), up to 14.8 GB/s sequential read and 11 GB/s write.

Is the CM9-DLC available to buy now?

No — it is currently sampling to select customers, the earliest commercial stage. It was showcased at OCP Global Summit 2026 (October 12–15). Volume availability, pricing, and full qualification data have not been announced.

Do I need a liquid-cooled rack to use it?

No. The CM9-DLC supports both cold-plate and traditional air-cooled environments, so it can be deployed in conventional racks today and moved to liquid-cooled racks later without changing the drive SKU.

What is SSD context caching in AI workloads?

A technique where fast NVMe SSDs hold cached model data — including KV (key-value) caches for large-context inference — that overflows GPU memory. It demands sustained high read throughput with low tail latency, which is exactly the workload profile that benefits from liquid cooling.

How does direct liquid cooling differ from immersion cooling?

Direct liquid cooling uses a cold plate in contact with the component, leaving the rest of the server architecture intact. Immersion cooling submerges entire systems in dielectric fluid — more efficient but far more disruptive to service procedures and infrastructure.

Will liquid cooling make my SSD faster?

Not above its rated specifications. Liquid cooling lets the drive sustain its peak performance under thermal conditions where an air-cooled drive would throttle. You are buying performance consistency and energy efficiency, not higher benchmark numbers.

What does OCP Datacenter NVMe SSD v2.5 compliance mean?

It means the drive meets the Open Compute Project's specification for data-center NVMe SSDs, which hyperscale buyers use in procurement. Compliance simplifies qualification and second-sourcing across vendors.

What should I watch for next in liquid-cooled storage?

Rival DLC-capable enterprise SSD announcements, standardized E3.S cold-plate mechanical definitions from more OEMs, and eventually spec sheets that assume liquid cooling as the default operating mode.