NVIDIA's personal AI supercomputer is $6,950 since 2 October, and Linux-only. Here is what else buys you 128 GB of local AI memory at every budget, with prices checked on 17 November 2026 and benchmarks dated per row.
| Machine | US price (config) | Unified memory | Memory bandwidth | gpt-oss-120b decode | Notes |
|---|---|---|---|---|---|
| NVIDIA DGX Spark (baseline) | $6,950 (128 GB / 4 TB, NVIDIA); $4,999 (64 GB, OEMs only) | 128 GB (or 64 GB) | 273 GB/s | 41.9 tok/s Ollama, 50.5 tok/s SGLang (LMSYS) | CUDA, Linux only; fastest prompt processing in this class |
| ASUS Ascent GX10 | $5,999 (1 TB) at Newegg and B&H, often out of stock | 128 GB | 273 GB/s | Same as Spark (same GB10 chip) | Support through ASUS, not NVIDIA |
| Other GB10 clones (Lenovo PGX, Dell, HP, GIGABYTE, MSI, Acer) | Dell from $6,213; HP and GIGABYTE about $6,000 (1 TB); Lenovo listed a little lower at some resellers | 128 GB | 273 GB/s | Same as Spark | Buy on price, stock and SSD size |
| AMD Strix Halo, 128 GB (Bosgame M5, GMKtec EVO-X2, Framework Desktop) | Bosgame $3,099; GMKtec $3,499.99 (1 TB) / $3,649.99 (2 TB); Framework $3,449 | 128 GB | ~256 GB/s | Similar to Spark on token generation (The Register) | 2 to 3x slower prompt processing; ROCm, not CUDA; runs Windows |
| Apple Mac Studio M5 Max / M5 Ultra | M5 Max from $2,499 (36 GB), $5,399 at 128 GB; M5 Ultra from $5,499 (96 GB) | 36-128 GB / 96-512 GB | 460-614 GB/s / 1.2 TB/s | Faster than Spark on token generation (bandwidth) | No CUDA |
| RTX 5090 desktop | Card from $6,249.99 new (Newegg, 14 Nov; MSRP $1,999) | 32 GB VRAM | 1,792 GB/s | Doesn't fit without CPU offload | 3x+ faster than a Spark on models that fit in 32 GB |
| RTX Spark (laptops; Surface Dev Box desktop) | Laptops from $2,599; 128 GB builds about $5,900 to $6,999; Surface RTX Spark Dev Box $5,999 (November) | Up to 128 GB | 300 GB/s | Close to Spark | Windows on Arm with CUDA; about 7% slower than a Spark on agent work (Artificial Analysis) |
Benchmarks are single-user decode speeds from LMSYS (October 2025), The Register (December 2025) and Artificial Analysis (October 2026); exact figures vary by runtime and quantisation. Prices are the lowest US listing we found on the date checked; the clone and Strix Halo prices move weekly, so confirm at checkout.
By budget:
- Under $3,500: Bosgame M5 or GMKtec EVO-X2 (Strix Halo, 128 GB). Watch the listing: $2,199.99 buys the 64 GB EVO-X2, not the 128 GB one.
- $3,500 to $6,000: Framework Desktop if you want repairability; a GB10 clone if you need CUDA; Mac Studio M5 Max 128 GB if you want speed.
- $6,000 and up: DGX Spark only if you need DGX OS and NVIDIA's software from NVIDIA itself; Mac Studio M5 Ultra if you need more than 128 GB.
- $0: a cloud API plus an agent platform, if you want agents, not a GPU box.
This page covers hardware alternatives at every budget, the GB10 clones, used units, and the route that needs no hardware at all.
DGX Spark price today, and the cheapest way to get the GB10 chip
$6,950 for the NVIDIA Founders Edition with 128 GB of unified memory and 4 TB of storage, direct from NVIDIA. It launched at $3,999 and NVIDIA raised it twice, both times citing memory supply. The same day as the second increase NVIDIA announced a 64 GB model at $4,999, which went on sale on 23 October through partners only.
| Date | DGX Spark 128 GB price | What changed |
|---|---|---|
| Launch (October 2025) | $3,999 | Founders Edition, 4 TB |
| February 2026 | $4,699 | First increase, memory supply |
| 2 October 2026 | $6,950 | Second increase, memory supply |
| 23 October 2026 | $6,950 (128 GB), $4,999 (64 GB) | 64 GB model on sale through Acer, ASUS, Dell, GIGABYTE, HP and MSI |
Over three years the 128 GB model is about $193 a month for the hardware, or roughly $218 a month once you add electricity at the full 240 W rating of its power supply. Every alternative below is really a comparison against that number.
64 GB or 128 GB? The 64 GB model keeps the GB10 chip, DGX OS and the full software stack with half the memory, and NVIDIA rates it for models up to about 100 billion parameters. Per gigabyte it is the worse deal (about $78 per GB against $54), and two clustered 64 GB units cost $9,998 to reach what one $6,950 unit already has.
The cheapest way to get the chip. Usually not from NVIDIA. GB10 clones from the usual OEMs are the same silicon in a different case, and the NVIDIA developer forum's rule of thumb is that they perform within about 2% of each other, so buy on price, stock and SSD size. At our latest checks the ASUS GX10 listed at $5,999 (1 TB) and Dell's Pro Precision with GB10 at $6,213 (1 TB, in stock on dell.com), and Lenovo's ThinkStation PGX was listed a little lower than both at some US resellers. If 64 GB is enough, the 64 GB Spark at $4,999 is the cheapest way into the platform. If you need much more than a Spark, the step up is a GB300 workstation; check the current DGX Station price before assuming it is in reach.
RTX Spark GPU equivalent
Short answer: for graphics, the RTX Spark GPU performs like a laptop RTX 5070. For local AI it is the DGX Spark's sibling, not a graphics card: it lands within about 7% of a DGX Spark on agent work, because it is essentially the same Grace Blackwell design running at lower power, on Windows.
The RTX Spark superchip pairs a 20-core Grace Arm CPU with a 6,144-core Blackwell RTX GPU and up to 128 GB of LPDDR5X at 300 GB/s, slightly above the DGX Spark's 273 GB/s. Laptops have been on sale since 16 October 2026 from $2,599 (Surface Laptop Ultra, ASUS ProArt P14); 128 GB builds run about $5,900 to $6,999. The Surface RTX Spark Dev Box desktop is $5,999 and ships in November. By default a 128 GB machine gives its GPU 64 GB; you can raise that to fit larger models.

For graphics: in Notebookcheck's 3DMark Steel Nomad run it scored 3,134 against 3,075 for a laptop RTX 5070, and Windows Central calls it RTX 4070, 5060 or 5070-class in most non-AI work. The difference is memory: a laptop RTX 5070 has 8 GB of VRAM.
For local AI: Artificial Analysis ran the same 168-turn local agent workload on several machines (4-bit models, tool time excluded):
| System | Qwen 3.8 27B | Qwen 3.6 35B-A3B | Ling 3.0 Flash (124B MoE) |
|---|---|---|---|
| GeForce RTX 5090 desktop | 4.9 min | 2.0 min | Does not fit |
| DGX Spark | 17.3 min | 7.3 min | 14.9 min |
| RTX Spark laptop | 18.7 min | 7.6 min | 15.3 min |
| AMD Ryzen AI Halo (Strix Halo) | 34.5 min | 11.7 min | 25.1 min |
| MacBook Pro M5 Pro, 64 GB | 37.6 min | 11.9 min | Does not fit |
Lower is better. Source: Artificial Analysis AgentPerf-Local, 15 October 2026.
Treat the RTX Spark as a DGX Spark with Windows, a battery and 10% more bandwidth; treat its GPU as a laptop RTX 5070 with 64 GB of VRAM. The full specs, memory split and software comparison are in our RTX Spark vs DGX Spark comparison.
Best DGX Spark alternatives ranked by price
Cheapest first, with DGX Spark at the bottom for reference.
| Option | Price | Memory | What you give up vs. DGX Spark |
|---|---|---|---|
| Ollama on hardware you already own | $0 | Whatever you have | Capacity. Realistically about 30B models on 32 GB |
| Cloud inference (OpenRouter, Groq, Together) | $0 upfront, ~$10-50/mo typical | Unlimited | Data sovereignty and offline use |
| Managed agent platform (BetterClaw) | $0 on the free plan, plus your model costs (BYOK, zero markup) | N/A | Direct GPU access. You don't host the model at all |
| Cloud GPU rental (Vast.ai, RunPod) | From about $1/hr | 80 GB A100 and up | Data sovereignty and 24/7 economics |
| Bosgame M5 (Strix Halo) | $3,099 | 128 GB | CUDA, and about half the agent speed on dense models |
| Framework Desktop (Strix Halo) | $3,449 | 128 GB | CUDA. Gains repairability |
| GMKtec EVO-X2 (Strix Halo) | $3,499.99 (1 TB), $3,649.99 (2 TB) | 128 GB | CUDA. Gains the most community testing |
| Mac Studio M5 Max | $5,399 at 128 GB (from $2,499 at 36 GB) | 36-128 GB | CUDA. Gains up to 614 GB/s |
| Mac Studio M5 Ultra | From $5,499 | 96-512 GB | CUDA. Gains 1.2 TB/s, about 4x the bandwidth |
| GB10 clones (ASUS GX10, Lenovo PGX and others) | From $5,999 (1 TB) | 128 GB | Nothing architecturally. Same chip, OEM support |
| NVIDIA DGX Spark (reference) | $6,950 (128 GB) | 128 GB | - |
Three things stand out:
- The cheapest 128 GB box is $3,099, not $6,950. The Bosgame M5 matches the Spark's 128 GB in the same bandwidth class (~256 GB/s against 273 GB/s) for $3,851 less. You lose CUDA, and on long agent runs you lose speed.
- CUDA plus 128 GB starts below the Founders Edition. A GB10 clone at $5,999 saves about $950, and a 128 GB RTX Spark laptop costs about $5,900 with a screen and a battery.
- The RTX 5090 has left the conversation on price. The cheapest new card on Newegg was $6,249.99 on 14 November, more than three times its $1,999 MSRP, for 32 GB of VRAM. That is more than a 128 GB Strix Halo box. It is still the fastest option by far for anything that fits in 32 GB; above that, it is the wrong tool.
DGX Spark vs the alternatives: head to head
| DGX Spark vs | Price difference | Who wins on speed | Pick the alternative if |
|---|---|---|---|
| RTX Spark laptop | About $1,050 to $4,350 cheaper | DGX Spark, by about 7% on agent work | You want CUDA on Windows, or a laptop |
| AMD Strix Halo (Bosgame M5, EVO-X2, Ryzen AI Halo) | $2,950 to $3,850 cheaper | DGX Spark, about 2x on long agent runs; similar on raw token generation | You don't need CUDA and price matters most |
| Mac Studio M5 Max (128 GB) | $1,551 cheaper | Mac on raw tokens per second (614 GB/s vs 273 GB/s) | You want the best out-of-box experience and don't need CUDA |
| Mac Studio M5 Ultra | $1,451 cheaper at base | Mac, by a wide margin (1.2 TB/s) | You run large dense models |
| RTX 5090 desktop | The card alone costs about the same ($6,249.99) | RTX 5090, more than 3x on models that fit in 32 GB | Your models fit in 32 GB of VRAM |
| GB10 clones | $0 to about $950 cheaper | Identical chip | A clone is in stock for less from a vendor you trust |
For the bandwidth reasoning behind the Mac results, see our DGX Spark memory bandwidth breakdown.
The alternatives in detail
Cloud inference ($0 up front)
If your goal is running AI agents, not running local hardware, cloud inference gives you every model without buying anything. OpenRouter, Groq and Together.ai all offer BYOK-compatible endpoints, and they reach frontier models (Claude, GPT) that no local hardware can run. At $6,950 the Spark amortises to about $193 a month over three years; most agent workloads cost $10 to $50 a month in API fees. You'd need to run inference 8+ hours a day at high volume before local hardware breaks even. We ran the full DGX Spark vs local GPU vs cloud API cost comparison over 12 months.

Ollama on your own machine ($0)
Before spending $6,950, check what you already own. Ollama runs on Mac, Windows and Linux. With 16 GB you can run Gemma 4 12B or Qwen 3.6 35B-A3B (3B active parameters); with 32 GB, Qwen 3.8 27B and most open models that matter for agent work.
brew install ollama && ollama run gemma4:12b
To put those models behind always-on cloud agents, the free BetterClaw desktop app connects Ollama, LM Studio or llama.cpp on your machine to your BetterClaw agents, with no ports to open.
Apple: Mac Studio M5 Max and M5 Ultra
The Mac Studio was refreshed on 25 August 2026 (there was never an "M4 Ultra"; the previous generation was M4 Max and M3 Ultra):
- M5 Max, from $2,499 with 36 GB, configurable to 128 GB. 460 GB/s on the 32-core GPU, 614 GB/s on the 40-core. The 128 GB build needs the 40-core GPU and costs $5,399 with 1 TB, $1,551 less than the Spark, with roughly 2.25x its bandwidth.
- M5 Ultra, from $5,499 with 96 GB, up to 512 GB, at 1.2 TB/s: roughly 4x the Spark's bandwidth. The fastest way to generate tokens from a large local model on a desk.
The M5 GPU also has dedicated matrix hardware, so the old "Macs are slow at prompt processing" line is less true than it was. The trade-off is still CUDA: your local stack won't mirror an NVIDIA cloud deployment. More in Apple Silicon vs NVIDIA for AI agents.
GB10 clones: every box with the Spark's chip
NVIDIA licensed the GB10 superchip to its usual OEM partners. Each ships the same 128 GB, the same 273 GB/s and the same DGX OS in a different case, with a different warranty, storage option and price.
| Machine | Price (US) | Storage | Checked |
|---|---|---|---|
| NVIDIA DGX Spark Founders Edition | $6,950 | 4 TB | 17 Nov |
| ASUS Ascent GX10 | $5,999 (Newegg, B&H); ASUS eShop $5,999 / $6,999 / $7,999 for 1 / 2 / 4 TB, out of stock | 1 TB base | Oct to Nov |
| Lenovo ThinkStation PGX | Listed from a little under $6,000 at US resellers; check lenovo.com | 1 TB or 2 TB | Nov |
| Dell Pro Precision with GB10 | From $6,213 (1 TB), $6,593 (2 TB), $9,043 (4 TB), in stock on dell.com | 1 TB to 4 TB | 27 Oct |
| HP ZGX Nano G1n | From $5,999.99 (1 TB) | 1 TB to 4 TB | 28 Sep |
| GIGABYTE AI TOP ATOM | $5,999.99 (1 TB, out of stock) | 1 TB or 4 TB | 25 Oct |
| MSI EdgeXpert | Varies by retailer and SKU | 1 TB to 4 TB | - |
| Acer Veriton GN100 | Varies by retailer | 4 TB | - |
The practical rule: these are the same silicon, so the only variables are price, stock and who answers the phone when it breaks. If a clone is in stock under $6,950 from a vendor your procurement already approves, buy the clone. If you'll keep several large models on disk, pay for the bigger SSD. Stock, not price, is usually the real constraint; the cheapest 1 TB versions are frequently sold out.
AMD Strix Halo: Bosgame, GMKtec, Framework and the rest
Every box here uses AMD's Ryzen AI Max+ 395 with 128 GB of unified LPDDR5X at about 256 GB/s:
| Machine | 128 GB price | Notes |
|---|---|---|
| Bosgame M5 | $3,099 (2 TB) | Cheapest 128 GB box we found |
| Framework Desktop | $3,449 | Repairable and modular; launched at $1,999 |
| GMKtec EVO-X2 | $3,499.99 (1 TB), $3,649.99 (2 TB) | The most community-tested; $2,199.99 buys the 64 GB model |
| Minisforum MS-S1 Max | $3,919 (2 TB) | |
| AMD Ryzen AI Halo Developer Platform | $3,999 | AMD's own box, Micro Center exclusive, Linux or Windows |
| HP Z2 Mini G1a | From about $4,000 | Enterprise warranty, Ryzen AI Max+ PRO 395 |
| Beelink GTR9 Pro | $4,349 (2 TB) |
What they share: 128 GB, about 6% less bandwidth than the Spark, ROCm and Vulkan instead of CUDA, and Windows or Linux. What they lack: prompt-processing speed on long agent runs.
Cloud GPU rental
RunPod, Vast.ai and Lambda rent GPUs by the hour, from about $1 an hour for an 80 GB A100 to $6 to $8 an hour for a B200. At 8 hours a day, 20 days a month, an A100 costs roughly $150 to $290 a month with far more compute than a Spark. Owning wins if you'd rent more than about 4 hours a day, every day.
A managed agent platform
If you want agents, not a GPU box, the model is a component, not the product. BetterClaw connects to 28+ model providers with your own keys at zero markup, handles agent logic, 500+ connectors, scheduling and memory, and can also run local models through the desktop app. The free plan costs $0; you pay your model provider directly.
The AMD DGX Spark competitor
A 128 GB Strix Halo box is AMD's direct answer to the DGX Spark: the same memory, about 6% less bandwidth, at roughly 45 to 60% of the price.
- Token generation is close. The Register put HP's Z2 Mini G1a (same AMD chip) against a Spark: in single-batch inference the AMD box "trades blows with the Spark on token generation, but falls well behind on time to first token". The Spark was 2 to 3 times faster at prompt processing.
- Long agent loops favour NVIDIA. Artificial Analysis measured AMD's Ryzen AI Halo at 34.5 minutes on a Qwen 3.8 27B agent workload the Spark finished in 17.3. "Same tokens per second, half the speed on long agent loops" is the honest summary.
- Software is the real trade-off. Ollama, llama.cpp and vLLM run on ROCm or Vulkan; NVIDIA NIM, TensorRT-LLM and NemoClaw do not run at all.
Is there an AMD equivalent to DGX Spark? Yes: a Strix Halo mini PC such as the Bosgame M5 ($3,099) or GMKtec EVO-X2 ($3,499.99), or AMD's own Ryzen AI Halo Developer Platform ($3,999) for official support and Windows. Both match the Spark's 128 GB. Neither supports CUDA.
What owners report lines up with the benchmarks:
- People who own both an EVO-X2 and a Spark find token generation nearly identical, with the Spark well ahead on prompt processing and the better pick for PyTorch and vLLM work.
- Researchers treat the Spark as a CUDA prototyping box, not an inference value play.
- In side-by-side tests, tokens per second tracks memory bandwidth, so a maxed-out M5 Mac leaves both behind.
Strix Halo boxes are also a popular way to run local models for agents; see local model hardware for agent workloads.
Used and refurbished DGX Spark: what to pay
The used market is thin and mostly eBay, and asking prices swing week to week, so check completed (sold) listings rather than asking prices before you bid. The useful ceiling is simple: a new GB10 clone such as the ASUS Ascent GX10 lists at $5,999 with the same chip, so a used 128 GB / 4 TB Founders Edition only makes sense meaningfully below that. Sealed new units sometimes list above NVIDIA's $6,950 because stock is tight, and used clones rarely save much against new ones.
Rules for buying used:
- Get the invoice. Hardware support usually routes through the original seller.
- Check the SSD. Founders Editions have 4 TB; many clones have 1 TB.
- Skip zero-feedback sellers on high-value listings; scam listings for in-demand hardware are common.
- Compare with new clones. Anything at or above $5,999 used is worse value than a new GB10 clone with a fresh warranty.
The verdict: buy used only if a clean 4 TB Founders Edition with an invoice comes in clearly under the $5,999 new-clone price; otherwise buy a clone new.
What is still coming, and why waiting hasn't paid
- RTX Spark desktops (November 2026), including the Surface RTX Spark Dev Box at $5,999: the first Windows desktops with CUDA and 128 GB of unified memory.
- LPDDR6 systems (2027). Roughly 2.25x the per-die bandwidth of LPDDR5X. When it lands, the 273 GB/s ceiling that limits the Spark and every Strix Halo box moves.
- No DGX Spark successor yet. NVIDIA's Computex 2026 Spark roadmap (Vera Rubin, then Rosa Feynman) covers RTX Spark PCs and names no new DGX Spark desktop, with no dates or prices.
Every unified-memory box got dearer in 2026: the Spark went from $3,999 to $6,950, the 128 GB Framework Desktop from $1,999 to $3,449, and the RTX 5090 from $1,999 to over $6,000. Patience saved money in previous hardware cycles; in this one, buy when you have a specific need, not on speculation.
Which models fit
Budget 10 to 15 GB of headroom above the weights for the KV cache and the OS.
| Model | Size at 4-bit | On a DGX Spark |
|---|---|---|
| Qwen 3.8 27B | ~17 GB | About 11 tok/s, 24-26 tok/s with MTP speculative decoding in vLLM. Best default for agent work |
| gpt-oss-120b (MXFP4) | ~60 GB | 41.9 tok/s Ollama, 50.5 tok/s SGLang (LMSYS) |
| Qwen 3.6 35B-A3B | ~20 GB | Fastest in its class: only 3B active parameters |
| Llama 3.3 70B (Q4_K_M) | ~43-49 GB | Fits; about 4-5 tok/s, slow for interactive use |
| gpt-oss-20b | ~13 GB | Runs almost anywhere |
| DeepSeek V4, Llama 3.1 405B | Hundreds of GB | Does not fit |
The constraint is bandwidth, not capacity: prefer MoE models on a Spark. If you only have a small GPU, see the best coding model for 12 GB of VRAM, and check which Ollama models support tool calling before you pick one for an agent.
Who should buy a DGX Spark, and who should buy none of these
Four kinds of buyer get their money's worth from a Spark:
- The CUDA portability developer. Code built on a Spark runs on NVIDIA's cloud and data centres without a translation step.
- The data-sovereignty buyer who also needs NVIDIA's software. If sovereignty is the only requirement, a $3,099 Strix Halo box meets it.
- The fine-tuner. NVIDIA rates the 128 GB Spark for fine-tuning models of up to 70 billion parameters.
- The cluster builder. Two 128 GB units over a QSFP cable handle models up to about 400B parameters, by NVIDIA's figures.
By use case:
- Lowest cost for AI agents: cloud inference plus BetterClaw's free plan.
- Local privacy under $3,500: Ollama on what you own, or a Bosgame M5.
- 128 GB and Windows: AMD Ryzen AI Halo ($3,999), or an RTX Spark laptop if you also need CUDA.
- CUDA laptop: RTX Spark, 128 GB builds from about $5,900.
- Cheapest NVIDIA stack: the 64 GB Spark at $4,999, or a GB10 clone at $5,999 for 128 GB.
- Best Mac: Mac Studio M5 Max for value, M5 Ultra for speed.
- I want a used one: see what to pay for a used DGX Spark.
Who shouldn't buy any of these. If your agent runs fewer than a few thousand tasks a day, or the models you want are closed (Claude, GPT), no box on this page helps. A cloud API plus an agent platform does the job, and if you want agents, not a GPU box, BetterClaw's free plan runs them on your own keys with nothing to install. For the honest view of what runs locally, see what you can actually run locally in 2026.
Also note: a Spark is a desktop, with no managed uptime, failover or replacement SLA. If an agent is a production dependency serving customers around the clock, it belongs on cloud infrastructure.
Frequently Asked Questions
What is the best DGX Spark alternative in 2026?
It depends on your budget. Under $3,500, a 128 GB Strix Halo mini PC such as the Bosgame M5 ($3,099) or GMKtec EVO-X2 ($3,499.99), without CUDA. Around $6,000, a GB10 clone such as the ASUS Ascent GX10 ($5,999) for the same chip. At the Spark's price, a 128 GB Mac Studio M5 Max ($5,399) for more than twice the bandwidth. For AI agents, cloud APIs plus a managed platform cost $0 up front.
How much does the DGX Spark cost?
The 128 GB NVIDIA DGX Spark costs $6,950 since 2 October 2026. It launched at $3,999 and was raised to $4,699 in February 2026, both increases citing memory supply. A 64 GB model costs $4,999 and has been on sale since 23 October 2026 through Acer, ASUS, Dell, GIGABYTE, HP and MSI.
What is the cheapest way to get a DGX Spark?
Buy a GB10 clone or go smaller. The ASUS Ascent GX10 lists at $5,999 with 1 TB, about $950 under the Founders Edition, and some resellers list Lenovo's ThinkStation PGX a little lower. The 64 GB Spark is $4,999. A used Founders Edition is only worth it clearly below the $5,999 new-clone price.
What is the RTX Spark GPU equivalent?
For graphics, a laptop RTX 5070: in Notebookcheck's 3DMark Steel Nomad test the RTX Spark scored 3,134 against 3,075. For AI, it is the DGX Spark's sibling, about 7% slower on agent work in Artificial Analysis's test, with 300 GB/s of bandwidth and up to 128 GB of unified memory, on Windows.
Is there a Chinese or AMD DGX Spark alternative?
Yes on AMD: any 128 GB Ryzen AI Max+ 395 (Strix Halo) box, from $3,099. Most of those are made by Chinese mini PC brands (GMKtec, Beelink, Minisforum, Bosgame) and ship worldwide. For the GB10 chip itself, xFusion's FusionXpark is sold on quotation through resellers with no US retail listing, and Orange Pi's AI Studio Pro uses Huawei Ascend chips with CANN rather than CUDA and has no US retail channel.
Is a used DGX Spark worth it?
Only at the right price. A new GB10 clone such as the ASUS Ascent GX10 lists at $5,999 with the same chip, so a used 128 GB / 4 TB Founders Edition needs to come in clearly below that. Check sold listings, not asking prices, get the invoice, check the SSD size and avoid zero-feedback sellers.
Do I need local hardware to run AI agents?
No. Most AI agents run on cloud APIs and never touch local hardware. Platforms like BetterClaw connect to 28+ model providers with your own keys, at zero markup. Local hardware only makes sense when data sovereignty, offline use or very high-volume inference justifies it, and the BetterClaw desktop app can connect those local models too.
Should I wait to buy a DGX Spark or an alternative?
Not for prices. Every unified-memory machine in this category got more expensive in 2026, including the Spark itself, twice. The one launch still ahead is RTX Spark desktops in November, if you want a Windows desktop with CUDA and 128 GB. NVIDIA has named no DGX Spark successor.




