How a $500 SSD Upgrade Undercuts Nvidia's $4,000 AI Box

There's a certain category of hardware where the price jumps feel less like engineering costs and more like gatekeeping. Alex Ziskind's latest video explores this phenomenon through a simple question: What if you could quadruple your storage and upgrade to faster drives for less than the premium manufacturers charge?

The GB10 family—ASUS GX10, Dell GB10, MSI Edge Expert, and Nvidia's DGX Spark—represents the current generation of compact AI workstations. They share identical core specs: the same chip, 128GB of memory, the same board. What varies is cooling design and, critically, storage configurations. Some ship with 1TB, some with 2TB, and the premium models with 4TB. That last category commands serious money.

Ziskind walks through the pricing landscape with the careful attention of someone who's actually trying to buy this stuff. The Nvidia DGX Spark with 4TB costs $4,000. Dell's 4TB version runs $4,700. HP's hits $5,100. Even the 1TB models aren't cheap—Lenovo's ThinkStation PGX starts at $3,500. The most affordable entry point? The ASUS GX10 with 1TB at $3,000.

"Still a big chunk of change, but it's not 4,000. It's not 5,000," Ziskind notes. Then he asks the obvious question: Why not upgrade it yourself?

The Storage Economics

The drives inside these machines use the 2242 form factor—smaller than the standard 2280 SSDs most people encounter. Ziskind purchased a Corsair MP700 micro, a 4TB Gen 5 NVMe drive, for $484 from Amazon (he found it slightly cheaper at Micro Center). Add that to the $3,000 base ASUS GX10, and you're at $3,500 for a 4TB system—cheaper than most 1TB alternatives, let alone the premium 4TB models.

But there's a complication. You can't just swap drives without losing your operating system, settings, and installed software. This is where the hardware gets interesting. Ziskind uses a $58 NVMe duplicator—a device with two slots that clones one drive to another. Slot A for source, slot B for destination. Push a button, wait 15 minutes, done.

The process reveals something about how storage upgrades work in practice. The cloning preserves partitions, which means the OS and data transfer intact, but the extra space initially appears as unallocated. After the swap, Ziskind's system boots normally—same passwords, same desktop background, same files—but the terminal shows only 916GB available. The disk utility, however, recognizes the full 4TB capacity. A quick partition resize operation, and suddenly: "3.6 terb. Now I can fit a lot of models on there. A lot of models."

That storage matters for AI workloads. Large language models eat disk space—a 70B parameter model can easily exceed 100GB. If you're experimenting with multiple models, testing different quantizations, or keeping various versions for comparison, 1TB disappears fast. The Gen 5 upgrade also improves model loading times compared to the Gen 4 drives that ship in base configurations.

The Weird Experiment

Then Ziskind does something he explicitly tells viewers not to replicate. He has a 2TB WD Black drive in the standard 2280 form factor—physically longer than the 2242 drives these machines are designed for. What happens if you just... stick it in there and let it hang out the side?

"Do I recommend you doing what I'm about to try? No. Don't do this. That's why I have a YouTube channel so I can do stupid things."

He plugs it in, leaves the case open, flips the machine upside down, and powers on. It boots. Ubuntu loads. Everything works. The system recognizes the drive, partitions it, and operates normally despite the SSD literally protruding from the case. "I cannot believe that. That just works now. It's ugly. Very ugly."

The demonstration is less about recommending this approach (he doesn't) and more about exposing what's technically possible. If the system accepts a full-size drive this easily, upgrading to 8TB becomes theoretically feasible—though 8TB Gen 5 drives currently run $1,500-$1,600. Still, the option exists.

The ASUS Advantage

The ASUS GX10 differs from other GB10 machines in ways that matter for upgradeability. The case uses magnetic feet that allow units to stack—clever if you're building a small cluster. More importantly, it's heavier than competitors because ASUS added extra heat pipes and a seven-stage fan system instead of the typical five-stage setup. Better cooling means more thermal headroom for sustained workloads.

The ease of opening the case surprised me. No special tools, no warranty-voiding stickers placed strategically over screws. Just remove the bottom panel and you're in. Whether ASUS designed this intentionally for user upgrades or simply chose a practical engineering approach, the result is the same: accessibility.

What This Means

Ziskind's experiment highlights a tension in the AI hardware market. Manufacturers sell identical base platforms at various price points, with storage configurations as the primary differentiator. But storage is commodity hardware—the same components available to anyone. The premium for pre-installed capacity often exceeds what users would pay to upgrade themselves.

This creates an interesting market dynamic. Companies positioning these machines for AI development are targeting users who, by definition, are comfortable with technology. Many have software engineering or data science backgrounds. They know how to clone a drive. The premium pricing strategy assumes either that users value convenience enough to pay double, or that they won't realize there's an alternative.

The video also surfaces questions about 8TB configurations. Currently, 8TB drives in the 2242 form factor don't exist—only the larger 2280 size. But if demand grows for compact AI workstations with massive local storage, manufacturers will eventually produce smaller form factors. The market incentive exists; the engineering challenge isn't insurmountable.

For now, Ziskind's approach works: buy the cheapest configuration, upgrade it yourself, pocket the difference. "This right here is now the cheapest 4 TBTE GB10 device out there."

The demonstration suggests that in AI hardware, as in so much of tech, paying attention to what you're actually buying—and what you could build yourself—matters more than it used to.

—Marcus Chen-Ramirez