Pure Storage's FlashBlade//EXA '10 TB/sec' Claim: When Vague Numbers Replace Real Benchmarks
Pure Storage claims FlashBlade//EXA delivers 'more than 10 TB/sec' read performance. We analyze why this vague claim tells us almost nothing about real-world performance.
In March 2025, Pure Storage announced FlashBlade//EXA with a headline claim: “more than 10 terabytes per second read performance in a single namespace.” This sounds impressive until you start asking basic questions about what this number actually means.
The “More Than” Problem
“More than 10 TB/sec” is marketing speak for “we don’t want to tell you the actual number.” It could be 10.1 TB/sec. It could be 15 TB/sec. It could theoretically be 50 TB/sec. The phrase reveals nothing except that Pure Storage measured something above 10 TB/sec under some unspecified conditions.
Compare this to credible benchmark reporting:
- Bad: “More than 10 TB/sec”
- Good: “14.2 TB/sec peak, 11.8 TB/sec sustained over 1 hour”
The vague claim invites prospects to imagine the best-case scenario while Pure Storage avoids being held accountable to any specific number. This is Benchmarking 101: make the claim impressive but vague enough that you can’t be proven wrong.
Missing Context 1: What Configuration?
10 TB/sec sounds impressive, but how many nodes does it take to achieve this? Per-node throughput matters far more than aggregate cluster throughput.
Scenario A: 10 TB/sec from 10 nodes = 1 TB/sec per node Scenario B: 10 TB/sec from 100 nodes = 100 GB/sec per node
These scenarios have radically different cost, power, and operational profiles. Without knowing the cluster size, “10 TB/sec” is meaningless for capacity planning or cost comparison.
Pure Storage’s marketing materials emphasize FlashBlade//EXA is built for “demanding AI and HPC workloads” and represents “the world’s most powerful data storage platform.” These superlatives don’t substitute for actual configuration details.
Missing Context 2: What Workload?
“Read performance” could mean many different things:
Sequential reads: Reading large files sequentially (most favorable for throughput) Random reads: Reading small blocks from random locations (realistic for many workloads) Mixed workload: Combination of read/write, sequential/random Block size: 4KB? 1MB? 16MB? Queue depth: How many outstanding I/O operations? Number of clients: How many concurrent readers?
Peak sequential read performance with large block sizes and high queue depth is the easiest benchmark to optimize for—and the least representative of real-world AI training workloads, which often involve random access to checkpoints, model weights, and training data.
Without specifying the workload, “10 TB/sec read” could be achieved under highly artificial test conditions that bear little resemblance to actual customer deployments.
Missing Context 3: Burst vs. Sustained
Is “10 TB/sec” a burst rate measured over seconds, or sustained throughput measured over hours?
Flash storage systems can often achieve impressive burst performance by leveraging DRAM caches, but sustained performance is limited by the underlying flash media and network bandwidth. A system might burst to 15 TB/sec for 30 seconds but sustain only 8 TB/sec over an hour as caches fill and thermal throttling kicks in.
AI training jobs run for hours or days. Burst performance is irrelevant if the system can’t sustain it. Pure Storage’s omission of any time dimension suggests they’re reporting burst performance, not sustained throughput.
Missing Context 4: Write Performance
Pure Storage emphasizes read performance but says nothing about writes. AI training workloads involve:
Checkpointing: Periodic writes of model state (write-heavy) Data loading: Reading training data (read-heavy) Gradient updates: Small random writes (write and latency sensitive)
A system optimized for sequential reads might perform poorly for writes or small random I/O. Pure Storage’s silence on write performance suggests it’s not as impressive as the read numbers.
The claim that FlashBlade//EXA “targets AI checkpointing write performance” in separate coverage is telling—if write performance were stellar, Pure Storage would publish specific numbers rather than vague positioning statements.
The “50% Better” Claim
Pure Storage also claims FlashBlade//S R2 delivers “up to 50% more performance than the previous generation” and “up to double the performance in write-bound workloads compared to the original FlashBlade//S.”
“Up to” is another classic marketing qualifier. It means “in the best case we measured.” The median improvement might be 10%, but Pure Storage can truthfully claim “up to 50%” if they found one workload where it happened.
“50% more performance” also lacks specificity. Performance of what? IOPS? Throughput? Latency? All three metrics can’t improve by 50% simultaneously—they trade off against each other.
The Pattern
Pure Storage’s FlashBlade//EXA claims follow a predictable pattern:
- Impressive headline number: “10 TB/sec”
- Vague qualifiers: “More than,” “up to”
- Missing methodology: No workload, configuration, or test duration details
- Superlative positioning: “World’s most powerful” without comparative data
- Marketing without measurement: Claims designed to sound good, not to be verified
This isn’t engineering communication—it’s marketing.
What Would Credible Claims Look Like?
Pure Storage should publish:
Specific numbers: Not “more than 10 TB/sec” but “14.2 TB/sec peak, 11.1 TB/sec sustained” Configuration details: How many nodes? How much flash capacity? Network topology? Workload characteristics: Sequential or random? Block size? Read/write mix? Number of clients? Time dimension: Burst (seconds)? Sustained (minutes/hours)? Write performance: If targeting AI checkpointing, publish write throughput and latency numbers Reproducible methodology: Share enough details that customers can validate claims in their own environments Comparative data: If claiming “world’s most powerful,” provide apple-to-apple comparisons with competitors
The Economics Problem
FlashBlade//EXA is positioned as enterprise storage for AI workloads. These deployments cost millions of dollars. Customers making multi-million dollar purchasing decisions deserve more than “more than 10 TB/sec” and “up to 50% better.”
Without specific performance characteristics, customers can’t model whether FlashBlade//EXA meets their needs. A cluster that achieves 10 TB/sec sequential reads but only 2 TB/sec random reads might be perfect for one workload and terrible for another. Pure Storage’s vague claims make informed purchasing decisions impossible.
The Industry Problem
Pure Storage isn’t unique in publishing vague performance claims. The entire storage industry has normalized this:
- VAST Data: “10x faster than Kafka” (no methodology)
- NetApp: “2X performance” (compared to what?)
- Dell ECS: “World’s fastest” (by which metric?)
When every vendor publishes unverifiable superlatives, the numbers become meaningless. Customers are forced to conduct their own benchmarks because vendor-published performance claims can’t be trusted.
This is a failure of the industry. Storage vendors have a responsibility to publish credible, reproducible performance data. Tech journalists have a responsibility to demand methodology before amplifying vendor claims. Until both happen, the cycle of vague marketing claims continues.
Conclusion
Pure Storage’s “more than 10 TB/sec” claim for FlashBlade//EXA tells us almost nothing about real-world performance. Without configuration details, workload characteristics, time dimensions, or reproducible methodology, this number serves marketing purposes rather than technical evaluation.
FlashBlade//EXA may genuinely be a high-performance storage system well-suited for AI workloads. But we can’t evaluate that from Pure Storage’s published claims. The phrase “more than 10 TB/sec” is optimized for press releases, not for engineers making purchasing decisions.
Storage purchasing involves significant capital and multi-year operational commitments. Vendors claiming industry-leading performance should show their work. Until Pure Storage publishes detailed methodology and reproducible benchmarks, treat their performance claims as aspirational marketing rather than verified fact.
For more analysis of vague storage vendor claims, see our coverage of VAST Data’s unverifiable performance assertions, NetApp’s “2X performance” ambiguity, and the broader pattern of unverifiable benchmarks across the storage industry.
References:
- Pure Storage: “Pure Storage Introduces FlashBlade//EXA” (March 2025)
- Next Platform: “Pure Storage FlashBlade//EXA Boosts AI Performance, Scalability” (April 2025)
- Pure Storage: “Pure Storage Unveils Next-generation Storage Products” (June 2025)