At GTC 2025, a quiet but consequential moment unfolded in the AI hardware market: Asus placed its Ascent GX10 beside Nvidia's DGX Spark and offered the same foundational Grace Blackwell architecture for a thousand dollars less. This is not merely a pricing skirmish — it is a signal that the tools of serious AI development are beginning to find competitive pressure, and with it, the possibility of broader access. Where Nvidia built density and prestige into its machine, Asus built portability and networking, betting that researchers and developers will value the freedom to move, scale, and clus
Asus Ascent GX10 emerges as lighter, cheaper DGX Spark alternative at GTC 2025
A thousand dollars less, and the machine feels lighter in your hands
Mark
So the Ascent GX10 is basically the same chip as the DGX Spark, just cheaper?
Mimi
Same Grace Blackwell processor, same 128GB of memory, same Tensor Cores. The real differences are in the chassis and what's built in. Asus went lighter with plastic instead of all metal, and they included networking hardware that would cost you $1,500 to $2,200 if you bought it separately.
Luke
But they cut the storage in half—4TB down to 1TB. That's not nothing for some workflows.
Mimi
True. If you're storing large datasets locally, that matters. But for inference and fine-tuning, 1TB might be enough, especially if you're pulling data from elsewhere.
Mark
The networking piece seems like the real story. Why does that matter so much?
Mimi
Because if you want to cluster multiple machines together to scale up your compute, you need fast, low-latency connections between them. Asus built that in. Nvidia's initial support is only for two units clustered together, but people are already thinking about how to chain more.
Luke
Do we know if that clustering actually works beyond two units? Or is that speculation?
Mimi
That's speculation from the analysts who saw it. The hardware is there, but Nvidia hasn't officially blessed anything beyond two units yet.
Mark
So for someone who just wants one machine to run models locally, does the cheaper price actually mean anything?
Mimi
You save a thousand dollars. Whether that's worth the lighter chassis and less storage depends on what you're doing. If you're moving it around or planning to add more machines later, it's a real advantage.
Luke
And we should note—this is Asus's take on the same chip. Dell and HP have their own versions too. So Nvidia's not the only game in town anymore.
Mimi
Right. The Grace Blackwell chip is the commodity now. The question is who packages it best for what you actually need.
O Pulso
- Nvidia arrived at GTC 2025 as the presumed authority on desktop AI supercomputing, pricing its DGX Spark at $3,999 — but Asus immediately disrupted that assumption with a $2,999 alternative built on the same core silicon.
- The $1,000 price gap is not explained by inferior processing power or memory — both machines carry 128GB unified memory and the same Blackwell GPU architecture — but by deliberate engineering trade-offs in materials and storage.
- Asus sacrificed 3TB of onboard NVMe storage relative to Nvidia's offering, yet embedded a dual-port ConnectX-7 networking card that retails separately for $1,500–$2,200, reframing the value equation entirely for users planning to cluster units.
- The lighter plastic chassis makes the Ascent GX10 meaningfully more portable than Nvidia's dense, metal-clad machine — a practical advantage for labs or developers who move hardware or build distributed local infrastructure.
- Analysts are watching closely: while Nvidia currently supports clustering of only two DGX Spark units, the Ascent GX10's built-in RDMA networking suggests users may push well beyond that ceiling, potentially reshaping how small-scale AI clusters are assembled.
At GTC 2025, a quiet but consequential moment unfolded in the AI hardware market: Asus placed its Ascent GX10 beside Nvidia's DGX Spark and offered the same foundational Grace Blackwell architecture for a thousand dollars less. This is not merely a pricing skirmish — it is a signal that the tools of serious AI development are beginning to find competitive pressure, and with it, the possibility of broader access. Where Nvidia built density and prestige into its machine, Asus built portability and networking, betting that researchers and developers will value the freedom to move, scale, and cluster over raw storage capacity.
Nvidia's DGX Spark debuted at GTC 2025 as a compact but serious AI supercomputer — Grace Blackwell architecture, 128GB unified memory, 4TB of NVMe storage, priced at $3,999. It was designed for developers, researchers, and data scientists who want to run complex AI models locally without depending on cloud infrastructure. Then Asus arrived at the same conference with the Ascent GX10, built on the identical Grace Blackwell foundation, priced at $2,999.
The Ascent GX10 is not a budget imitation. It carries the same 128GB unified memory and Blackwell GPU with fifth-generation Tensor Cores and FP4 precision support. The primary concession is storage — 1TB of NVMe versus Nvidia's 4TB — a trade-off many workflows can absorb in exchange for a thousand dollars in savings.
What distinguishes the Asus machine goes beyond price. Hands-on observers at GTC noted that its partial plastic chassis makes it noticeably lighter than Nvidia's dense, metal-bodied unit — more portable, easier to reposition or transport. The rear panel further separates it: alongside HDMI, four 40Gbps USB4 ports, and a 10GbE interface, Asus included a dual-port Nvidia ConnectX-7 NIC built for RDMA clustering. That card alone typically costs $1,500–$2,200 on the open market.
Nvidia currently plans to support clustering of up to two DGX Spark units. Analysts believe the Ascent GX10's built-in networking hardware and lower per-unit cost will encourage users to chain more units together — building local AI infrastructure at a scale that neither machine was originally marketed to serve alone. For research labs and smaller AI operations, that scalability, combined with the price advantage, positions the Ascent GX10 as a genuine alternative rather than a lesser substitute.
Nvidia's DGX Spark arrived at GTC 2025 as a desktop-sized AI supercomputer built around the Grace Blackwell chip—powerful enough to handle serious model development and inference work without offloading to the cloud, compact enough to fit on a desk. The price tag was $3,999. But Asus showed up at the same conference with its own answer: the Ascent GX10, built on the same Grace Blackwell foundation, priced at $2,999.
The Ascent GX10 is not a stripped-down knockoff. Both machines carry 128GB of unified memory and the Blackwell GPU with fifth-generation Tensor Cores and FP4 precision support—the same core architecture that makes the DGX Spark appealing to developers, researchers, data scientists, and students who want to work with complex AI models locally. The trade-off is storage: Nvidia gives you 4TB of NVMe SSD space; Asus provides 1TB. For a thousand dollars less, that's a reasonable compromise for many workflows.
What makes the Ascent GX10 interesting is not just the price. ServeTheHome, which examined the machine hands-on at GTC, noted that Asus chose plastic for parts of the chassis where Nvidia used metal. The result is a system that feels noticeably lighter—more portable, less like a dense brick. If you're thinking about moving the machine between locations or clustering multiple units together, that weight difference matters. The DGX Spark, by contrast, has the heft and density of an Apple Mac Studio.
The rear panel tells another story. The Ascent GX10 includes an HDMI port, four USB4 ports running at 40Gbps, a 10GbE network interface, and a dual-port Nvidia ConnectX-7 NIC designed for RDMA clustering. That last component is the real value play. A single ConnectX-7 NIC typically costs between $1,500 and $2,200 when purchased separately. Asus built one directly into a $2,999 system. For someone planning to cluster multiple units together—to scale up compute capacity beyond what a single box can deliver—that's a significant advantage.
Nvidia initially plans to support clustering of up to two DGX Spark units. But analysts watching the Ascent GX10 suspect users will quickly figure out how to chain more units together, especially given the built-in networking hardware and the lower per-unit cost. The machine is designed to be clusterable from the ground up, not as an afterthought. That positioning could appeal to research labs, smaller AI shops, and anyone who wants to build a local AI infrastructure without the capital outlay of a full data center.
The Ascent GX10 is not cheaper because it cuts corners on the processor or memory. It's cheaper because Asus made different engineering choices—lighter materials, less onboard storage, a focus on networking and portability. Whether those choices align with your needs depends on what you're actually trying to do. But for developers and researchers who want Grace Blackwell performance without the Nvidia premium, and who value the ability to move the machine or cluster it with others, the Ascent GX10 represents a genuine alternative.
Citações Notáveis
For some context here, a Nvidia ConnectX-7 NIC these days often sells for $1500–2200 in single unit quantities. At $2999 for a system with this built-in that is awesome.— Patrick Kennedy, ServeTheHome
Quer a matéria completa?
Leia o original em TechRadar ↗
Technology