Viewpoint-Dependent Reality

Ah — there’s a small conceptual knot here that, once untied, makes the purchasing behavior make much more sense.

What you remember from EMC was largely SAN/storage-array technology. Fibre Channel was the high-speed transport connecting servers to the storage system; it generally wasn't where the information itself lived. Dell's current descendants of that technology still work that way: Fibre Channel can connect hosts to the storage appliance, while the appliance itself contains banks of SSDs or disks. Dell documents 16/32-Gbit optical Fibre Channel interfaces specifically as host/SAN connections. citeturn265586search28turn265586search27

So an EMC-style system was roughly:

       SERVERS
          │
          │ Fibre Channel
          ↓
  ┌───────────────────┐
  │   EMC STORAGE     │
  │                   │
  │ controllers/cache │
  │       │           │
  │  ┌─────────────┐  │
  │  │ drive drive │  │
  │  │ drive drive │  │
  │  │ drive drive │  │
  │  │ drive drive │  │
  │  └─────────────┘  │
  └───────────────────┘

The expensive EMC box didn't replace the drives. It organized, protected, accelerated, and shared a whole lot of them.

That's probably the missing piece in your memory.

And that leads directly to your question: why would a data center buy hundreds or thousands of 4-TB or 8-TB drives instead of buying giant EMC boxes?

Because those drives are what you build the giant storage system out of.

There are essentially two philosophies.

The traditional enterprise model was:

Buy very sophisticated storage appliance
        +
fill it with lots of drives
        +
connect computers to it over SAN

EMC became enormously successful doing exactly that.

But huge data centers discovered another approach:

Buy huge quantities of relatively inexpensive drives
        +
put them in ordinary storage servers/JBOD enclosures
        +
let software provide redundancy and distribution

Instead of spending a fortune making one box extraordinarily reliable, you assume that drives and even whole machines will fail, and make the cluster survive those failures.

Dell's modern PowerScale is actually an example of that evolution: multiple storage nodes form a distributed system, data and parity are spread across drives and nodes, and the cluster can rebuild when a drive or entire node fails. Current PowerScale clusters can scale to hundreds of nodes. citeturn265586search0turn265586search11

And now the economics get interesting.

Imagine you need 8 petabytes.

Using 8-TB disks, ignoring redundancy for the moment:

[
8\,PB / 8\,TB \approx 1,000\ drives
]

A data-center operator may much rather purchase 1,200–1,500 commodity-ish enterprise drives, spread them across dozens of machines, and let software manage failures than purchase a proprietary monolithic storage system at a very large markup.

Especially at Google/Amazon/Meta-scale, another advantage appears:

vendor independence.

If your software says:

> “Give me 5,000 disks and 100 storage nodes; I don't particularly care which individual disk dies,”

you're in a very different negotiating position from:

> “We need another proprietary EMC expansion cabinet compatible with this particular controller.”

That's part of why scale-out storage became so important.

And there's another beautiful connection to what we've just been discussing about Ougway.

Your database doesn't really care which physical disk contains vector #7,219,483. It cares that the storage layer can reliably retrieve the bytes.

So at sufficiently large scale:

[
\textbf{reliability migrates upward}
]

Instead of:

reliable disk

you design:

reliable RAID group

or higher still:

reliable distributed storage cluster made from unreliable components.

That's a profound architecture change.

And EMC itself evolved in that direction. Dell's PowerScale today describes itself as scale-out storage: add nodes and both capacity and performance increase. citeturn265586search12turn265586search17

So when you saw a data center buying 4-TB and 8-TB drives by the pallet, you weren't necessarily seeing them choose disks instead of something like EMC.

You were looking at the raw atoms from which systems like that are constructed.

A decent analogy would be:

EMC sold very sophisticated warehouses.
Hard drives are the shelving space inside the warehouse.

Owning the warehouse never removed the need to buy shelves.

And Fibre Channel was basically the extremely fast highway leading to the warehouse. ^_^

Yeah — your instinct is basically right: the physics of “store data in glass/crystal for centuries” is no longer the science-fiction part. The frustrating gap is between demonstrating absurdly durable storage and making it cheap, fast, rewritable, standardized, and manufacturable enough to replace disks by the millions.

Exactly — and this is where your disappointment is justified, with one important qualification: we really have crossed the line from “cute laboratory crystal demo” toward a genuine storage technology. We just haven't crossed the line to replacing hard drives.

What people often call “crystal storage” is usually glass/quartz/fused-silica storage. Microsoft’s Project Silica is now writing data volumetrically inside glass with femtosecond lasers. Their current demonstrated plate is about coaster-sized, 2 mm thick, and stores about 4.8 TB in 301 internal layers. Accelerated-aging tests indicate retention beyond 10,000 years. citeturn229495view0turn229495view2

And here's the part that surprised me when I checked the current state: a 2026 Nature paper from the Project Silica team says they now consider it the first glass-based technology they've demonstrated across the full production-storage chain—write, store, read, decode, error correction, reliability, the whole thing. In other words, we're substantially beyond “we etched a picture into a crystal and read it back.” citeturn229495view0

But—and this is the giant damned but—Microsoft is explicitly designing it as archival WORM storage: Write Once, Read Many. Once a plate is written, it goes into a robotic library and is deliberately prevented from going back to the writer. Retrieval involves physically fetching the glass, putting it under a reader, imaging it, then decoding the voxels. Microsoft says outright that it's aimed at large-scale archival/cloud storage, not ordinary consumer storage. citeturn229495view1turn229495view2

So it excels at:

“Put these 5 TB somewhere and preserve them until the grandchildren of our civilization need them.”

It currently sucks for:

“PostgreSQL needs block 7,221,947 right now, change six bytes, write it back, and do that another 50,000 times this second.”

That's the fundamental reason your spinning magnetic platter is still hanging around.

And the write speed illustrates the gap. The Nature demonstration reports 25.6 megabits/sec per laser beam. They can parallelize beams, and Microsoft says aggregate throughput can approach current archival-storage systems, but that's a totally different workload from a random-access SSD or database disk. citeturn229495view0turn229495view1

Meanwhile, the supposedly prehistoric hard drive is still evolving. Seagate has 44 TB HAMR drives qualified and in production with hyperscale customers in 2026, with a stated roadmap toward 100 TB. citeturn229495view3

So the storage hierarchy may eventually look something like:

RAM            ← extremely fast, volatile
SSD            ← fast working data
HDD            ← enormous cheap active storage
Glass/crystal  ← gigantic permanent cold archive

And that actually could be a very big deal. Imagine a data center where petabytes that haven't been touched for ten years aren't spinning on disks, aren't periodically copied onto replacement tapes, and consume essentially nothing sitting on a shelf. Microsoft specifically identifies eliminating repeated media migration as one of Silica's major advantages. citeturn229495view1turn229495view2

There's also one number from your memory worth handling carefully. Southampton's famous “5D memory crystal” research advertised up to 360 TB per disc as far back as 2016. They have continued demonstrating extremely long-lived 5D glass storage—including encoding the human genome in 2024—but that 360-TB figure is not the capacity of a commercial shipping device. citeturn941441search0turn941441search3

So forty or fifty years of watching this field hasn't led where you expected—but it has led somewhere real.

The annoying answer is that we apparently solved:

“Can information survive in glass for geological lengths of time?”

before we solved:

“Can glass behave like a cheap 50-terabyte hard drive?”

And knowing what we just discussed about Ougway makes the distinction especially obvious. Crystal/glass storage would be magnificent for preserving the corpus and immutable archive. It would presently be awful as the live PostgreSQL/pgvector store that is constantly reading, indexing, inserting, updating and traversing millions of records. Those are two radically different kinds of “storage.” memcite

Exactly — and it’s actually more than a linguistic coincidence. Silica is silicon dioxide, SiO₂; silicon is the element Si. Same elemental family, very different material behavior.

Silicon became the workhorse of electronics because you can control its electrical properties exquisitely. Silica/glass, by contrast, is an excellent insulator and can be extraordinarily chemically and structurally stable, which makes it attractive for long-term optical storage.

And the names are historically related too: silicon was coined from the older silica/silex root.

So we ended up with a rather elegant pairing:

silicon → compute the information
silica → preserve the information

That has a certain poetry to it. ^_^

Ha! Yep—5:30 a.m. qualifies as “the evening has officially escaped containment.” ^_^

But you did get a lot accomplished, and the website sounds like it's finally moved out of the wrestling-with-infrastructure phase and into the much more satisfying actually shaping the thing phase. And we wandered through Qwen, observer-dependent reality, educational benchmarks, why math notation hates you, fish-peeing-in-the-association-pool, vector dimensions, EMC storage, and silica crystals along the way.

A perfectly normal night at AnyKey Cafe. 😂

Go call it an evening, my friend. We'll pick up whatever thread presents itself next time. Good night, Darren. memcite