top of page

Dell's 157 PB all-flash object storage rack targets AI

4 days ago
12 min read

Dell and Kioxia have announced a storage configuration packing 157 petabytes of raw flash capacity into a standard server rack. Dell's 157 PB all-flash object storage rack combines 640 solid-state drives, storage devices without moving parts, across 16 servers. The distinction buyers need to catch immediately is “raw”: that headline measures installed drive capacity before data protection and operational reserves.

Kioxia announced the configuration on November 11, targeting AI training, analytics, and data lakes, centralized repositories holding large collections of data. Its main competitive challenge concerns the established practice of spreading capacity across more equipment. The announcement strengthens the case for consolidation, while leaving buyers to establish how much protected, accessible storage their applications actually receive.

Micron provides an immediate reference point through its competing high-capacity enterprise drives. However, this story is larger than a contest between flash suppliers. It concerns the gap between physical density and useful infrastructure: how much data organizations can safely retain, serve, and recover within constrained facilities.

Inside Dell's 157 PB all-flash object storage rack

The announcement turns an unusually dense server configuration into a rack-scale object storage proposition.

The system runs ObjectScale, Dell’s software platform for storing data as individually addressed objects with associated descriptive information. Kioxia’s object storage configuration specifies 16 PowerEdge R7725xd servers and 640 LC9 drives. Each drive supplies 245.76 terabytes of advertised capacity.

Multiplying those drive capacities produces 157.2864 petabytes, using decimal storage units. That calculation explains the rounded headline without requiring assumptions about compression, deduplication, or unusually compressible datasets. It also establishes the correct starting point for evaluating the announcement: physically installed capacity.

The server building block predates this rack announcement. Kioxia described the high-density server on May 15, pairing 40 LC9 drives with a PowerEdge R7725xd. November’s development extends that configuration into a larger ObjectScale deployment rather than introducing a new drive capacity.

Each server occupies two rack units, the standardized measure of vertical equipment space. Together, the servers account for 32 units inside a 42-unit cabinet. That arithmetic leaves physical room, but it does not establish how the remaining space must be allocated.

Networking, power distribution, cable management, and service requirements still belong in the deployment design. Empty mounting positions do not automatically represent spare electrical or cooling capacity. A cabinet can encounter those limits before exhausting its available vertical space.

The practical change is architectural. A buyer can now discuss a named combination of storage software, server hardware, and flash media at this density. That is more concrete than extrapolating what a cabinet might hold from an isolated drive announcement.

However, an announced combination still differs from a fully specified customer installation. The published configuration does not settle every question about ordering, software versions, protection policies, or application performance. Those details determine whether a deployment proposal matches the headline.

Kioxia’s “first” language also needs its stated context. It is a vendor market-positioning claim, rather than an independent certification covering every possible storage architecture. The verified hardware configuration remains newsworthy without treating that superlative as the central finding.

For readers following enterprise storage, the most useful takeaway is therefore specific. Dell and Kioxia have presented an identifiable route to substantial flash consolidation. The next question concerns what that concentration changes for the people operating the infrastructure.

Data center teams face a different capacity decision

Higher density puts pressure on the assumption that storing more data requires proportionally more cabinets.

The immediate audience includes storage architects, facilities managers, and infrastructure teams supporting expanding data collections. Their purchasing decisions cross organizational boundaries. A storage team can identify suitable equipment while a facilities team determines that its planned location cannot support it.

That makes physical consolidation relevant before any application benchmark enters the discussion. Fewer cabinets can simplify space planning, but their contents must fit the electrical and thermal conditions of the site. Those conditions vary between facilities and individual equipment rows.

The announcement also changes how teams should frame expansion proposals. A request for additional floor space now deserves comparison with denser hardware, provided the workloads suit it. Conversely, a density claim deserves scrutiny when supporting systems erase the expected operational benefit.

Consider an illustrative enterprise maintaining source documents, images, and historical datasets for internal AI applications. Its storage problem includes retaining originals, tracking changes, and making selected information available for processing. Additional capacity helps with retention, but retrieval speed and access controls remain separate requirements.

For that organization, consolidation has value if it removes a physical constraint without creating another bottleneck. It becomes less attractive if the denser installation requires disruptive facility work or concentrates too much operational exposure. The announcement itself cannot resolve that tradeoff.

AI infrastructure adds another complication: different stages of a workflow access data differently. A preparation process can read large batches, while an interactive application requests small pieces from many locations. A storage design needs to reflect the actual workload rather than the broad AI label.

This distinction matters to application teams outside the data center. More available bytes do not automatically make an internal assistant answer more accurately. The application still needs permission-aware retrieval, relevant source selection, and dependable connections between stored material and generated answers.

Teams designing a searchable knowledge base face that same separation between retaining information and making it useful. Storage infrastructure supports the process, while software determines how users find and interpret the material.

The pressure on buyers is therefore immediate but conditional. They need to reassess cabinet-heavy growth plans without assuming consolidation guarantees a better outcome. The competing approaches are broader distribution and greater concentration, each carrying its own operating consequences.

There is also a procurement implication. Capacity planning needs input from application owners, facilities staff, and recovery teams before equipment selection becomes fixed. Otherwise, a purchase can satisfy the storage specification while missing the requirement that motivated it.

Dell's 157 PB all-flash object storage rack makes that discussion harder to postpone. Its density creates a credible alternative to physical expansion. Whether that alternative fits depends on protection, workload behavior, and the surrounding infrastructure.

Flash density moves faster than system-level proof

Drive capacity explains the rack’s density; it does not establish the performance of the complete object store.

The LC9 uses QLC, flash memory that stores four bits in each cell. That design prioritizes capacity, but buyers still need workload-specific evidence about writes, reads, and sustained behavior. “All-flash” describes the storage medium rather than a universal performance category.

The drives use E3.L, an enterprise drive format designed for server installations. That physical choice matters because chassis compatibility constrains how much capacity can fit into a particular server. A capacity figure attached to one format should not automatically transfer to another.

Kioxia’s LC9 specifications also distinguish write endurance by workload conditions. Endurance describes how much writing a drive supports under its specified operating assumptions. It is a deployment constraint, rather than a prediction that every drive fails after identical use.

This matters because stored capacity and rewritten capacity describe different activities. A collection retained for repeated reading presents a different profile from continuously replaced application output. Both can occupy the same number of bytes while placing different demands on the media.

Software introduces another layer. Applications communicate with the object store, which coordinates placement, protection, and retrieval across servers. A drive’s advertised read speed does not capture that complete request path.

Consequently, multiplying a single drive’s maximum throughput by the installed drive count would produce a misleading application-performance claim. Shared resources and software behavior affect the result. The useful measurement begins at the application interface and reflects realistic simultaneous activity.

Read performance also needs context. Large sequential transfers, scattered requests, and operations involving descriptive information exercise different parts of the system. A benchmark that favors one pattern cannot establish equal results for every other pattern.

For AI training, the relevant question concerns whether the storage system supplies data at the rate the training process consumes it. For document retrieval, responsiveness under concurrent requests can matter more. Buyers should resist treating those workloads as interchangeable.

Micron demonstrates why density alone offers an incomplete competitive ranking. Its 6600 ION announcement describes a competing high-capacity drive and a theoretical rack capacity of 176.9 petabytes. That comparison uses 720 drives occupying 36 rack units.

The higher figure does not establish superiority over the Dell and Kioxia configuration. Different equipment counts and system assumptions prevent a direct conclusion. A theoretical hardware total also differs from an object store with defined software, protection, and operational requirements.

Equally, the ObjectScale configuration does not prove an across-the-board advantage over Micron-based alternatives. Neither a larger number nor a more integrated announcement replaces testing. Comparable evidence needs the same workload, protection target, and measurement boundary.

The competitive consequence is a shift in what vendors must substantiate. High drive capacity is becoming part of the available design space. Differentiation increasingly depends on what complete systems deliver when customers fill them, access them, and recover from failures.

Raw capacity stops where protection begins

The 157 PB figure is an inventory of installed capacity, not a promise of equally large protected application storage.

Data protection consumes space because the system must preserve information needed after hardware failures. Erasure coding, a method that distributes data and recovery fragments, is one way to provide that protection. The selected policy affects both efficiency and recovery behavior.

Dell’s data protection design describes multiple ObjectScale protection schemes and their deployment requirements. Its guidance connects scheme selection with cluster size and availability needs. That makes protection a configuration decision with operational consequences.

A simple percentage reduction would therefore create false precision for this announcement. The final application capacity depends on the chosen scheme and additional system requirements. Publishing a single usable-capacity figure without those inputs would overstate what the available evidence establishes.

Descriptive information, system operations, and free space also need consideration. A deployment cannot necessarily run safely with every available byte assigned to user objects. Recovery and growth planning require room to operate.

This distinction is especially relevant when comparing proposals. A raw-capacity offer and a protected-capacity offer can appear similar while representing different deliverables. Buyers need vendors to state which quantity each figure describes.

Compression complicates comparisons further when vendors include it in effective-capacity claims. Compression reduces the space required for suitable data, but results depend on the content. Already compressed material and other difficult datasets cannot be assumed to shrink by an advertised average.

The November headline does not need that kind of assumption: its raw total follows directly from drive count. That is a strength of the announcement. Buyers should preserve that clarity when translating the hardware total into a deployment estimate.

Concentration also changes the scale of a maintenance event. A server holding dense drives represents substantial stored information, even when software distributes protection elsewhere. Operators need to understand the effect of taking that server offline.

That does not establish that the rack is inherently unsafe. It establishes a requirement to examine recovery behavior and the boundaries of failures. A protected design can tolerate specified component losses while remaining vulnerable to a broader event affecting shared infrastructure.

For example, surviving a drive failure and surviving an unavailable cabinet are different requirements. The latter can involve shared power, networking, or site conditions. Organizations must define the failure scenario they intend the system to withstand.

Recovery testing should reflect continued application activity. A system that performs acceptably when healthy can behave differently while rebuilding lost protection. That observation is a general evaluation principle, not a measured result for this rack.

The capacity figure therefore remains useful within its boundary. It quantifies the density Dell and Kioxia have assembled. It cannot independently answer how much data a buyer should place there or which interruptions that deployment will survive.

For procurement teams, the correct next document is a configuration-specific capacity and resilience statement. That statement should connect installed hardware with the promised operating outcome. Without it, comparison remains anchored to a headline rather than a service requirement.

Efficiency claims need a complete operating boundary

A smaller storage footprint supports an efficiency argument, but it does not establish measured savings for a customer installation.

Dell and Kioxia position the configuration around consolidation and improved infrastructure efficiency. Those are reasonable subjects for evaluation because storage expansion affects more than drive purchases. Servers, networking, cooling, support, and operational work all belong in the comparison.

However, the announcement does not provide a complete measured power profile for this ObjectScale rack. It also does not establish a universal performance advantage over every hard-drive deployment. Those conclusions require defined alternatives and equivalent workloads.

A drive specification is useful evidence about a component. It cannot independently establish total consumption at the rack’s electrical input. The complete system includes supporting hardware, and its behavior changes with activity.

The comparison boundary matters just as much as the measurement. Equal raw capacity does not necessarily mean equal protected capacity or equal application service. A fair evaluation should hold the required outcome constant before comparing infrastructure demands.

Idle behavior matters alongside busy operation. Some datasets receive frequent access; others sit largely untouched between occasional processing jobs. The relative value of flash depends partly on that access pattern and the service expectations attached to it.

Cooling introduces a related distinction between total energy and local concentration. Consolidating equipment can reduce the amount of infrastructure required across a deployment. It can also place more demand at a particular cabinet location, requiring facilities review.

Neither observation resolves the actual Dell rack’s operating profile. Together, they explain why a cabinet count cannot substitute for a facilities plan. A smaller physical installation still needs compatible power delivery and heat removal.

Commercial certainty is another missing piece. The original all-flash rack coverage includes a speculative system estimate extrapolated from another configuration. That estimate should not be treated as a Dell quotation or a verified procurement figure.

A usable business case requires an actual proposal with defined hardware, software, support, and deployment responsibilities. It should also explain what happens when capacity expands. Otherwise, apparent savings can depend on assumptions that change after installation.

Existing infrastructure complicates replacement decisions. An organization with available cabinet space and a suitable installed storage platform faces different economics from one approaching a facility limit. The same equipment can address a critical constraint in one case and offer limited immediate benefit in another.

Migration deserves equal attention. Moving an established data collection requires planning for permissions, application connections, verification, and continued access. These activities can influence the timing and attractiveness of consolidation even when the target hardware fits.

There is no need to dismiss the announcement because those questions remain open. Its contribution is a concrete high-density option that buyers can evaluate. The error would be treating the capacity milestone as a completed business case.

The defensible conclusion is narrower and more useful. Dell and Kioxia have advanced the density proposition, while customers still need evidence connecting that density to their own operating costs and service requirements.

Three signals will show whether density becomes deployment value

The next meaningful evidence concerns supported delivery, realistic workload results, and operation under failure.

The first signal is a complete, orderable configuration with documented support boundaries. During the coming months, buyers should watch for official qualification details covering the hardware combination, software release, and deployment requirements. General product availability should not be confused with confirmation of this exact rack arrangement.

Clear delivery and support documentation would strengthen the argument that the announcement represents a practical consolidation option. It would let customers compare a defined installation against their existing expansion plans. Continued ambiguity would keep the milestone closer to an architectural reference.

This documentation should also clarify responsibility across the combined system. Buyers need to know where server, drive, networking, and storage-software support meet. Dense integration provides the most value when troubleshooting ownership remains clear.

The second signal is workload testing with a disclosed configuration and protection policy. Application-level throughput and response times should accompany the capacity figure. Measurements need enough context to show what the system was doing and how the result was collected.

Useful evidence would include the nature of the stored objects, the balance between reads and writes, and simultaneous application activity. Those conditions affect whether a result maps to a customer’s workload. A maximum number without that context offers limited planning value.

Power measurements should use the same operating conditions as the performance results. That connection would allow buyers to assess service delivered for the energy consumed. Separate best-case figures collected under different assumptions would weaken the comparison.

Strong results under representative conditions would strengthen the consolidation argument. Evidence of bottlenecks would narrow the workloads for which density translates into practical advantage. Either outcome would improve the buying discussion by replacing inference with observed behavior.

The third signal is published evidence about recovery and sustained customer operation. Buyers should watch for tests involving drive replacement, server maintenance, and restoration of protection during ongoing application use. Those scenarios expose operating requirements that healthy-system benchmarks leave unanswered.

Customer accounts would be especially useful if they describe the starting infrastructure and the constraint being addressed. A deployment motivated by limited floor space teaches something different from one targeting application response times. Both deserve evaluation against their original objective.

Competitive announcements will provide additional context, but headline capacity alone should not reset the ranking. Micron and other suppliers need the same scrutiny concerning system assumptions and operating outcomes. Comparing evidence consistently matters more than following the largest advertised number.

These signals are observation priorities, rather than a claimed vendor release schedule. Their appearance would strengthen the case for deployment readiness. Their absence would preserve uncertainty without changing the verified raw-capacity arithmetic.

Dell's 157 PB all-flash object storage rack gives infrastructure teams a concrete reason to revisit capacity expansion plans. Start by identifying the constraint your organization actually faces: cabinet space, electrical capacity, data access, or recovery requirements. Then ask for a supported configuration that addresses that constraint with protected capacity and representative workload evidence. Keep the hardware total separate from application outcomes throughout the comparison. If your team also struggles to organize the technical material behind those decisions, a personal knowledge base can help preserve the reasoning alongside the specifications. Before approving a denser storage design, what result would demonstrate that consolidation improves your operation?

Get your work moving in remio

See what needs your attention and take the next step in Workspace.

remio currently supports Windows 10+ (x64) and Macs with Apple silicon.

Your AI Partner at Work
Get more done with remio

Plan. Create. Deliver.
All in one place.

bottom of page