A Milpitas company that decided the SSD should stop being dumb. Ten years in, the drives quietly run a chunk of the workloads everyone else is busy talking about.
Above: the company name, set in italic because the product is fast. Not a logo. We tried.
In a rack somewhere in Northern Virginia, a database is running faster than its specs say it should. The CPU isn't pinned. The cache isn't doing anything clever. Look closer and the trick is sitting one layer down: a ScaleFlux SSD, quietly compressing data on its way to flash, decompressing it on the way back, and handing the host server a disk that behaves like it's three times its size.
This is what ScaleFlux sells - drives that look like ordinary NVMe SSDs to the operating system and behave like a small computer to the data. The company calls the category "computational storage." The industry has talked about the idea for a decade. ScaleFlux is one of the few that actually ships it.
It is, on paper, a deeply unfashionable business. Hardware. Long sales cycles. Customers who measure success in milliwatts and read amplification ratios. The kind of company nobody writes a profile about - which is exactly why we are.
For thirty years, the deal was simple. The CPU did the thinking. The disk did the storing. Memory sat between them and tried to keep the peace.
Then the workloads changed. Modern databases spend more time moving bytes than computing on them. AI training pipelines are bottlenecked not by GPU math but by how fast data can be staged from storage. Every byte of data on a modern SSD has been compressed, decompressed, encrypted, decrypted, and shuffled across a PCIe bus that was never designed for the trip.
The orthodox answer is to buy a bigger server. The ScaleFlux answer is to buy a smarter drive.
Compression, encryption, deduplication, atomic writes - all done by the CPU. Burns cycles you paid for. Heats up a rack you have to cool.
Put a programmable ASIC inside the SSD controller. Let it handle the boring, repetitive byte-shuffling. Free the CPU for things only it can do.
Effective capacity goes up. Endurance goes up because you write fewer raw bytes. Power per useful operation goes down. Your finance team notices.
ScaleFlux was founded in October 2014, in the part of Silicon Valley that does not get profiled. The three co-founders had between them the bulk of a generation's institutional memory about how flash memory actually works - which, it turns out, is exactly what you need if you intend to put a computer inside one.
Veteran of Fusion-io and SandForce. Helped ship one of the industry's first 40nm LDPC read-channel chips back in 2008. Has been quietly building flash controllers longer than most VCs have been writing checks.
Tenured professor at Rensselaer Polytechnic Institute. Provides the academic horsepower behind the company's controller algorithms - and the unusual sight of a research lab whose papers ship as commercial products.
Runs the systems-and-firmware side. The reason a card meant to do exotic things looks, to the operating system, like a perfectly ordinary NVMe drive.
Their bet was simple to describe and unfashionable to make: that the next decade of storage innovation would not come from faster flash but from smarter controllers. So far the bet appears to be holding.
The product catalog reads like ordinary enterprise storage - U.2, E1.S, HHHL form factors, capacities in the multi-terabyte range, NVMe over PCIe Gen4. Plug one in and Linux sees an SSD. What it does not see is what is happening on the other side of the cable.
Gen4 computational SSD with transparent in-line compression, encryption and programmable cores. Capacities to 16TB. The drive that introduced most of the world to ScaleFlux.
Newer generation aimed squarely at AI/ML and cloud workloads. Higher capacity, better performance-per-watt, same drop-in NVMe profile. Launched on the company's tenth anniversary.
A PCIe accelerator card that retrofits computational features onto existing drives. For customers who can't rip and replace.
The tooling layer that makes the drive's compute features programmable - compression ratios, atomic writes, multi-tenant virtualization. The bit that turns hardware into a platform.
Memory expansion devices over Compute Express Link. The other half of the bet: if storage can be smart, so can memory.
Underneath all of it: a portfolio covering LDPC, on-drive compute, self-managed DRAM, and storage security. The boring stuff that compounds.
Hardware claims are easy. Hardware that ships is harder. The figures below come from ScaleFlux's public material and third-party reviews - they describe the drives as they appear under normal database and AI pipeline loads, not the synthetic best case.
Bars approximate. Reality varies by workload. The endurance figure assumes compressible data, which is most data.
The customer list is the part ScaleFlux talks about least and the part that matters most: hyperscale operators, enterprise database shops running Postgres and MySQL and ClickHouse, financial-services firms with latency budgets measured in microseconds, and a growing roster of AI infrastructure customers who would rather not say so out loud. The pattern is consistent. The drives go in for one workload and end up in three more.
The company line is that ScaleFlux helps customers "harness data growth as a competitive advantage." The honest line is shorter: do more with the racks you already own.
It is a mission with awkward timing - and excellent timing. Awkward because the AI buildout is, at the moment, a story about adding capacity at any cost. Excellent because the bill for that buildout is coming, and someone is going to have to find efficiency that does not require ripping out the data center floor.
That, in the end, is what computational storage actually sells: efficiency that does not require a forklift.
Two curves are about to collide. AI training and inference are pushing data-pipeline demands to a place CPUs weren't designed to go. Power and cooling budgets are constraining new builds before the racks are even ordered. Computational storage sits exactly in the middle of those curves.
ScaleFlux's CXL memory work hints at the next move: pushing the same logic - smart hardware that does work the CPU shouldn't have to - into the memory tier. If they pull it off, the line between storage, memory, and compute starts to blur in a way that is genuinely useful, not just slide-worthy.
Whether ScaleFlux is the company that owns the category long-term is a question for a longer feature. That the category itself is real, and growing, is no longer in serious dispute.
The database in Northern Virginia is still faster than its spec sheet. The CPU is still un-pinned. The cache is still doing nothing clever. Down at the bottom of the stack, the ScaleFlux SSD is still quietly turning compressible data into less of it, and turning that arithmetic into capacity, endurance, and watts that didn't have to be spent.
Ten years ago, three engineers in Milpitas bet that the SSD would stop pretending. The bet is paying. The rack is cooler. The bill is smaller. The data still arrives.