DGX Spark vs RTX 5090 vs RTX Pro 5000: Best Hardware for Local AI in 2026

Which one to buy

RTX Pro 5000 Blackwell 48GB (PNY)

Best for most AI builds: 48 GB, 300 W, runs 70B models

Check price

DGX Spark: ASUS Ascent GX10 128GB

Complete 128 GB system for 100B-plus models

Check price

GeForce RTX 5090 (ASUS ROG Astral)

Fastest for models under 32 GB, images and games

Check price

AMD Ryzen AI Halo

128 GB for about $4,000, without CUDA

Check price

NVIDIA's DGX Spark Founders Edition. Photo: Daniel Lu, CC BY-SA 4.0.

The best local AI hardware for most people in November 2026 is the RTX Pro 5000 Blackwell 48GB, as long as you already own a desktop to put it in. On Amazon it costs about the same as an RTX 5090, holds 16 GB more, draws about half the power, and runs a 4-bit 70B model entirely in its own memory. If you don't have a suitable PC, or you want to run 100B-plus models, buy a DGX Spark instead. It is a complete computer with 128 GB of memory, and NVIDIA sells its own version for about $7,000. The GeForce RTX 5090 is the fastest of the three on any model that fits in 32 GB, and the only one that doubles as a high-end gaming card.

A year ago these weren't in the same price bracket. The RTX 5090 listed at $1,999 and the DGX Spark launched at $3,999. The memory shortage has since pushed all three to roughly $7,500 to $8,500 on Amazon, so the choice now comes down to which one does the most for the same money.

Our picks

• RTX Pro 5000 Blackwell 48GB is the best card for most local AI builds, with 48 GB of ECC memory, a 300 W power draw, and room for a 70B model. Its 72GB version runs gpt-oss-120b several times faster than the Spark.
• NVIDIA DGX Spark, linked here in its ASUS Ascent GX10 version, is the best complete system and the only one of the three that holds 100B-plus models like gpt-oss-120b.
• GeForce RTX 5090 is the fastest for models under 32 GB, for image and video generation, and for games.
• AMD Ryzen AI Halo has the same 128 GB as the Spark for about $4,000, without NVIDIA's CUDA software.

DGX Spark vs RTX 5090 vs RTX Pro 5000 specs

Specs come from NVIDIA's DGX Spark, RTX Pro 5000 and RTX 5090 pages, with the Pro 5000's core count from PNY's retail listing at Scan. The AI TOPS figures are NVIDIA's peak ratings, useful for comparing compute between these three but not for predicting chat speed. Amazon prices were checked on November 1, 2026. Most of the listings came from third-party sellers with a handful of units in stock, so expect them to change daily.

Memory decides which models you can run

A model has to fit in memory to run at full speed, so start with the models you want and work back to the hardware. At 4-bit quantization, the usual format for running models at home, a model takes a little over half a gigabyte per billion parameters, plus room for the conversation or document it is working on (the context).

The gpt-oss and GLM file sizes come from Hardware Corner's DGX Spark benchmarks, and the context limits from its RTX 5090 and RTX Pro 5000 tests.

A model that doesn't fit can still run. llama.cpp can split a model between the graphics card and system RAM, but the layers in system RAM run at system RAM speed, and a 70B model split across a 5090 and DDR5 loses most of the card's advantage.

Speed: RTX 5090, RTX Pro 5000 and DGX Spark benchmarks

Once a model fits, memory bandwidth sets how fast it writes, because each new token means reading the model's active weights from memory again. That's why the 5090, at 1,792 GB/s, is the fastest of the three, and why the Spark, at 273 GB/s, is the slowest on any model the cards can also hold. As a rough guide, anything above about 10 tokens per second is comfortable to read as it streams, while coding agents that write thousands of tokens per task benefit from far more.

Hardware Corner ran the RTX 5090 and the RTX Pro 5000 48GB through the same llama.cpp tests in March 2026. This is text generation at a 16K context, in tokens per second:

The 5090 is 15 to 28 percent faster on every model both cards can hold. The Pro 5000 keeps going after the 5090 runs out of memory. It ran Llama 3.3 70B at 24 tokens per second with a 4K context, and it held Qwen3 30B-A3B all the way to a 256K context at 37 tokens per second.

LMSYS tested the DGX Spark against the RTX 5090 and the RTX Pro 6000 with gpt-oss-20b in Ollama:

On a model that small, the 5090 writes about four times faster than the Spark and reads a long prompt about four times faster too.

The Spark looks better on large mixture-of-experts models, which use only a slice of their weights for each token. gpt-oss-120b has 117 billion parameters but uses about 5 billion per token, and in a llama.cpp test on the Spark it generated 60.6 tokens per second with an empty context and 47.5 at 16K. Of the cards here, only the 72GB Pro 5000 can hold that model, and it runs it much faster:

Dense models are the Spark's weak spot. Ollama measured Llama 3.1 70B at 4-bit at 4.4 tokens per second on the Spark, against 22 to 24 for the 48GB Pro 5000 in Hardware Corner's test. It measured Qwen3 32B at 9.4, where Hardware Corner measured the 5090 at 51 to 61.

Best for most AI builds: RTX Pro 5000 Blackwell

Check the RTX Pro 5000 48GB price on Amazon

The RTX Pro 5000 Blackwell. Image: NVIDIA.

The RTX Pro 5000 is NVIDIA's workstation card one step below the RTX Pro 6000. It has 14,080 CUDA cores and 48 GB of GDDR7 with error correction on a 384-bit bus, for 1,344 GB/s. It draws 300 W through a dual-slot card with a blower-style fan that pushes heat out the back of the case, so several can sit side by side in one workstation. NVIDIA's MIG feature can split one card into two isolated GPUs, which helps when two people or two jobs share a machine.

At launch prices, the 5090 was the obvious buy for AI at home, at less than half the Pro 5000's price. In November 2026 the two cost about the same on Amazon, and the Pro 5000's extra 16 GB is the difference between running a 32B model and a 70B one. Hardware Corner measured Llama 3.3 70B at 4-bit at 22 to 24 tokens per second, which is fast enough for chat and coding help. On smaller models the 5090 is still 15 to 28 percent faster, at nearly twice the power.

It needs a host PC with a PCIe x16 slot. PCIe 5.0 gets the full bus speed, though single-card chat barely uses the bus once a model is loaded. NVIDIA doesn't sell it as a consumer product, so Amazon's listings come from third-party resellers. The PNY retail card (part number VCNRTXPRO5000B-PB) was $7,985 from a seller called AzITStore on November 1. It's also on eBay, where IT resellers list sealed workstation cards. Check the part number and the seller's feedback before you pay.

Buy it if you have a desktop with room and power for a 300 W card, and you want 70B models at a usable speed without the 5090's power draw.

Check RTX Pro 5000 prices on eBay

RTX Pro 5000 72GB for gpt-oss-120b

NVIDIA's 72GB version has the same chip, the same 1,344 GB/s and the same 300 W as the 48GB card, with 24 GB more memory. It's no faster on anything that fits in 48 GB, so the extra money only buys room for models between about 48 and 70 GB. The most useful of those is gpt-oss-120b, the model the DGX Spark handles best. Its 59 GiB of weights fit on the 72GB card with space left for the context.

Once the model fits, the card's memory bandwidth, about five times the Spark's, takes over. HostKey ran gpt-oss-120b on the 72GB card in Ollama, with the context window set to 128K, at 156.5 tokens per second. That's nearly four times the 41 tokens per second in Ollama's own Spark test, and two and a half times the best llama.cpp result on the Spark. The tests are almost ten months apart on different software versions, so treat the gap as approximate. HostKey also ran a 70B DeepSeek-R1 model at 25 tokens per second, where Ollama measured the Spark at 4.4 on Llama 3.1 70B at 4-bit.

The catch is the price. Amazon's only listings on November 1 were $9,400 from a single reseller, and the PNY-branded one had a 2.2-star average from three ratings. That's about $1,400 more than the 48GB card, and about $2,400 more than NVIDIA's own Spark before you count the PC around the card. Look on eBay or at a workstation reseller first, and check for PNY's 72GB part number, VCNRTXPRO5000B72-PB.

Buy it if gpt-oss-120b or another 50 to 70 GB model is the one you'll run every day, you already have a desktop, and speed matters more to you than the Spark's lower price and extra 56 GB.

Check RTX Pro 5000 72GB prices on eBay

Best complete system: NVIDIA DGX Spark

Check the ASUS Ascent GX10 (DGX Spark) price on Amazon

The DGX Spark's rear panel. The two ports on the right link a second Spark. Photo: Daniel Lu, CC BY-SA 4.0.

The DGX Spark is a 150 mm square box built around NVIDIA's GB10 chip, which puts a Blackwell GPU and a 20-core Arm CPU in one package with 128 GB of LPDDR5x that both share. The Founders Edition has a 4 TB SSD, 10 Gb Ethernet, Wi-Fi 7, four USB-C ports (one of them for power), HDMI, and a ConnectX-7 network card whose two QSFP ports link a second Spark at 200 Gb/s. The whole system runs from a 240 W power supply, and NVIDIA rates it at 35 dB of sound power under full GPU load.

Its 128 GB is the most memory you can get anywhere near this price. NVIDIA says it holds models up to 200 billion parameters and can fine-tune models up to 70B, and two linked Sparks handle up to 400B. It ships with DGX OS, an Ubuntu-based Linux with CUDA, PyTorch and NVIDIA's AI tools set up, and it works the same way as NVIDIA's data center systems. For developers who will later run the same code on DGX servers or cloud GPUs, that matters more than speed.

The weak point is its 273 GB/s of memory bandwidth, about a fifth of the RTX Pro 5000's. Mixture-of-experts models like gpt-oss-120b run at a usable speed, but dense models get slow as they grow, down to 4.4 tokens per second on Llama 3.1 70B at 4-bit in Ollama's test. The Arm CPU and Linux also rule out Windows, and any software without an Arm Linux build won't run. llama.cpp, Ollama and SGLang all do.

NVIDIA sells the 128 GB Founders Edition through its own marketplace for about $7,000. That's after a jump from $3,999 at launch to $4,699 in February, and another increase in October that HotHardware put at $6,995. Amazon didn't carry the Founders Edition on November 1, only partner versions with the same GB10 chip. The ASUS Ascent GX10 with 128 GB and a 2 TB SSD was $8,429 from a third-party seller. Skip listings from resellers that open the box to upgrade it. One Amazon listing for a Gigabyte AI TOP ATOM, sold under a reseller's own brand, said the manufacturer's warranty was void because of a hardware upgrade.

A 64 GB version went on sale October 23, only through NVIDIA's partners, from $4,999. The 64 GB ASUS Ascent GX10 is the one listing in this article sold by Amazon itself ($5,599, temporarily out of stock on November 1). NVIDIA rates it for models up to 100B parameters, but gpt-oss-120b's 59 GiB of weights leave almost no room in 64 GB, so buy the 128 GB model if that's the model you want.

Buy it if you're starting without a desktop, want 100B-plus models, or are building software that will run on NVIDIA's data center hardware later.

Check DGX Spark prices on eBay

Fastest under 32 GB: GeForce RTX 5090

Check the RTX 5090 price on Amazon

The RTX 5090 Founders Edition going into a desktop. Still: ZMASLO, CC BY 3.0.

The RTX 5090 is NVIDIA's top gaming card, with 21,760 CUDA cores and 32 GB of GDDR7 on a 512-bit bus for 1,792 GB/s. It is the fastest of the three on any model that fits. LMSYS measured 205 tokens per second on gpt-oss-20b, four times the Spark's speed, and it beat the Pro 5000 by 15 to 28 percent in Hardware Corner's tests. Image and video generation speed depends more on raw compute than on memory bandwidth, and at 3,352 TOPS by NVIDIA's rating against the Pro 5000's 2,064, the 5090 is the card for Flux, Stable Diffusion and local video models. It's also the only one of the three you'd buy for games.

Its 32 GB covers 30B-class models with room for a long context, but not 70B models at 4-bit, and in Hardware Corner's test Qwen3 32B ran out of memory past a 32K context. The card draws up to 575 W, and NVIDIA recommends a 1,000 W power supply. It takes a single 600 W PCIe Gen 5 cable, or four 8-pin cables through the included adapter.

The 5090 still lists at $1,999, but you can't buy one for that. NVIDIA has passed GDDR7 cost increases on to board partners, and Micro Center switched many 5090 models to "Not Carried" in October. On November 1, Amazon's cheapest new listings were about $7,680 for the ASUS ROG Astral, sold by Amazon's global store, and $7,700 for the ASUS TUF Gaming model from a third-party seller. The Astral is a 3.8-slot card, so check your case clearance. Used and open-box 5090s are listed on eBay too. Pick a seller with a long feedback history and a return policy.

Buy it if your models fit in 32 GB, you generate images or video, or you'll also play games on the same PC.

Check RTX 5090 prices on eBay

Budget for the rest of the PC

The RTX 5090 and RTX Pro 5000 prices above are for the card alone. That's fine if you already have a recent desktop with a free x16 slot and a strong power supply. If you don't, the PC around the card costs more than it used to, because the same shortage hit desktop memory: TechRadar reported 64 GB of DDR5 going from the low $200s to over $1,000 in six months. A 5090 also needs a 1,000 W power supply and a case with room for a card up to four slots thick.

The DGX Spark's price covers the whole computer, including the CPU, 128 GB of memory, storage, networking and the power supply. Starting from nothing, a Spark at NVIDIA's price costs well under either card plus a new PC to run it. It uses less power too. The entire Spark runs from a 240 W supply, while a 5090 can draw 575 W by itself.

Other local AI hardware to consider

AMD Ryzen AI Halo

Check the AMD Ryzen AI Halo price on Amazon

AMD's own mini PC uses the Ryzen AI Max+ 395 with 128 GB of LPDDR5x at 256 GB/s, close to the Spark's bandwidth, plus a 2 TB SSD and 10 Gb Ethernet. It was $3,999.99 on Amazon on November 1, sold by Micro Center and shipped by Amazon. In Hardware Corner's comparison, a system with AMD's Strix Halo chip, the family the Ryzen AI Max+ 395 belongs to, generated gpt-oss-120b text at 34 tokens per second against the Spark's 39, but read prompts at 340 tokens per second against the Spark's 1,723, so long documents and large codebases take far longer to load. It runs AMD's ROCm software instead of CUDA, and tools built only for CUDA won't run on it.

Surface RTX Spark Dev Box

RTX Spark is NVIDIA's chip for Windows PCs, with the same 20-core Arm CPU and 6,144 CUDA cores as the Spark's GB10. Microsoft's Surface RTX Spark Dev Box puts it in a desktop with 128 GB of memory and a removable 2 TB SSD. It costs $5,999 from the Microsoft Store, and the first orders are due to ship by November 24. It's the one to look at if you want the Spark's memory under Windows.

RTX Pro 6000 Blackwell

The 96 GB RTX Pro 6000 has the 5090's bandwidth with three times its memory, and LMSYS measured it slightly ahead of the 5090. It also costs about twice as much as anything above. Amazon listings ran $17,000 to $18,500 on November 1, and NVIDIA's own price had already climbed to $13,250 by June, from about $8,500 at launch.

DGX Spark, RTX 5090 and RTX Pro 5000 FAQ

Is the DGX Spark faster than the RTX 5090?

Not on models that fit in the 5090's 32 GB. LMSYS measured the 5090 at 205 tokens per second on gpt-oss-20b and the Spark at 49.7. The Spark is the faster choice only for models too big for the card, such as gpt-oss-120b, which the 5090 can't hold at all.

Can the RTX 5090 run a 70B model?

Not in its own memory at 4-bit, where a 70B model needs about 42 GB. You can offload part of it to system RAM with llama.cpp, or squeeze it in at 2-bit quantization, but you lose a lot of speed with the first and a lot of quality with the second. For 70B models, the RTX Pro 5000 48GB or the DGX Spark is the better fit.

Is the RTX Pro 5000 better than the RTX 5090 for AI?

At November 2026 prices, yes, for most AI work. The two cost about the same on Amazon, and the Pro 5000 has 48 GB instead of 32 GB, ECC memory, and a 300 W power draw instead of 575 W. The 5090 is 15 to 28 percent faster on models that fit in both, and it's the better card for image generation and games.

Is the RTX Pro 5000 72GB worth it over the 48GB?

Only if you'll run models between about 48 and 70 GB, such as gpt-oss-120b. It has the same chip and memory speed as the 48GB card, so it's no faster on anything smaller, and it cost about $1,400 more on Amazon in November 2026. On gpt-oss-120b, HostKey measured it at 156.5 tokens per second, several times the DGX Spark's speed.

Can the DGX Spark run Windows?

No. It runs DGX OS, NVIDIA's Ubuntu-based Linux, on an Arm CPU, and it isn't a gaming PC. For the same kind of chip under Windows 11, look at RTX Spark PCs such as Microsoft's Surface RTX Spark Dev Box.

Should I buy the 64GB or the 128GB DGX Spark?

Buy the 128 GB model if you want gpt-oss-120b or anything larger, since its 59 GiB of weights barely fit in 64 GB. The 64 GB version, from $4,999 through partners, makes sense for 70B-class models and smaller. NVIDIA says two linked 64 GB units handle models up to 200B.

Why is the RTX 5090 so expensive in 2026?

GDDR7 memory has become much more expensive as AI data centers buy up memory supply, and NVIDIA has passed those costs on to its board partners. The card still lists at $1,999, but Amazon's cheapest new 5090 was about $7,680 on November 1.

What is the best GPU for running LLMs at home?

Among cards under $10,000 in November 2026, the RTX Pro 5000 48GB has the best balance of memory, speed and power, and the RTX 5090 is the fastest if your models fit in 32 GB. The 96 GB RTX Pro 6000 is faster and holds more, at about twice the price. If you'd rather buy a complete system, the DGX Spark holds bigger models than any of the cards.

Verdict

Buy the RTX Pro 5000 48GB if you have a desktop to put it in and want to run models up to 70B at a good speed. At November 2026 prices it costs about the same as a 5090 and gives you more memory at half the power. Buy the DGX Spark if you're starting without a PC, want gpt-oss-120b and other 100B-plus models, or are writing software that will move to NVIDIA's data center hardware. NVIDIA's own store had the lowest price on it. If gpt-oss-120b is the model you'll use most and you have a desktop, the 72GB Pro 5000 runs it two and a half to four times as fast as the Spark, for about $2,400 more before the PC. Buy the RTX 5090 if your models fit in 32 GB, you generate images or video, or you also play games, and compare eBay with Amazon before paying close to four times its list price.

Where to buy