The DGX Station starts around $80,000 and tops $120,000. Here is what else buys you serious local AI capacity, and the uncomfortable reason most of the cheap alternatives got expensive too.
A friend sent me a screenshot on Tuesday. NVIDIA's own marketplace, RTX PRO 6000 Blackwell, 96GB. Sixteen thousand dollars.
He had built his hardware budget in a spreadsheet eleven months earlier, when that same card listed at $8,565. His whole plan, the one where he skipped the DGX Station and bought two workstation cards instead, had quietly doubled in price while he was busy getting approvals.
That is the actual state of this decision in August 2026, and it is why most DGX Station alternative comparisons you will find are wrong. They were written against 2025 prices. The memory shortage moved every single number on the board, and it did not move them evenly.
So here is the honest version. What the DGX Station costs, what the six real alternatives cost right now, and the case for not buying any of them.
Stay with me on that last one. It is not the answer you expect from a page like this.
What a DGX Station is, and what it actually costs
The DGX Station is NVIDIA's desktop workstation built around the GB300 Grace Blackwell Ultra Desktop Superchip. A 72-core Grace CPU on Arm Neoverse V2, fused to a Blackwell Ultra GPU over NVLink-C2C.
The number that matters is memory. Up to 784GB of coherent memory in some configurations, with 252GB of that being HBM3e running at 7.1 TB/s. For reference, an H200 does 4.8 TB/s. It also carries a ConnectX-8 SuperNIC at up to 800 Gb/s, which lets you link two Stations together, and it supports MIG partitioning into as many as seven isolated instances.
NVIDIA pitches it at the trillion-parameter model class. That is not marketing inflation. It genuinely holds models nothing else on a desk can hold.
Now the price, which NVIDIA has never officially published.
There is no Founders Edition this time. NVIDIA supplies the integrated motherboard and lets partners build the workstation around it, so pricing comes from ASUS, Dell, GIGABYTE, MSI, and Supermicro, with HP arriving later. An MSI XpertStation WS300 showed up on CDW at $96,995.99. Distributors currently quote a range from just under $100,000 to about $123,000 depending on storage, add-on GPU, and on-site service. One tracker puts the entry point closer to $80,000. Lead times as of August 2026 run four to thirteen weeks.
Dell, notably, declined to quote a price for its GB300 variant while cheerfully publishing $4,757 for its smaller GB10 machine. Make of that what you will.
When a vendor will not print the price, the price is the feature they are least proud of.
Wait. Do you mean the Station or the Spark?
This trips up a surprising number of people, and getting it wrong costs you either $92,000 or a machine that cannot do the job.
DGX Spark is the small one. GB10 Superchip, 128GB of LPDDR5X unified memory, roughly 273 GB/s of bandwidth, about 1 petaflop of FP4 compute. It launched at $3,999 in October 2025 and was raised to $4,699 in late February 2026, with NVIDIA citing memory supply.
DGX Station is the workstation. GB300, up to 784GB, HBM3e at 7.1 TB/s, and a price roughly twenty times higher.
They are not competitors and they do not share an alternatives list. A Spark buyer is asking "can I prototype locally instead of burning cloud credits." A Station buyer is asking "can I hold a trillion-parameter model in my building." If you landed here but you actually meant the smaller machine, our DGX Spark alternatives breakdown covers that comparison properly.
Here is where it gets messy. Almost everything positioned as a "DGX alternative" online is really a Spark alternative. Nothing under $20,000 replaces a Station on capacity. What the cheaper options do is let you ask whether you needed that capacity at all.
The DGX Station alternatives worth pricing, ranked by cost
Prices below reflect August 2026 street reality, not launch MSRPs. Verify before you buy, because these have been moving monthly.
| Option | Memory | Bandwidth | Price (Aug 2026) | Launch price |
|---|---|---|---|---|
| NVIDIA DGX Station (GB300) | up to 784GB coherent | 7.1 TB/s (HBM3e) | $80,000 to $123,000 | never published |
| RTX PRO 6000 Blackwell 96GB | 96GB GDDR7 | 1.79 TB/s | $13,998 to $16,000 | $8,565 |
| Mac Studio M3 Ultra | 96GB max (was 512GB) | 819 GB/s | $5,299 | $3,999 |
| NVIDIA DGX Spark | 128GB LPDDR5X | 273 GB/s | $4,699 | $3,999 |
| Strix Halo 128GB mini PC | 128GB unified, ~96GB GPU-addressable | 256 GB/s | $3,449 to $4,349 | ~$1,999 |
| RTX 5090 multi-GPU build | 32GB per card | 1.79 TB/s | $3,000 to $5,000+ per card | $1,999 |
| Cloud rental (B200 / B300) | data center class | full HBM | $5.29 to $17.80 per GPU-hour | n/a |
Look at that "launch price" column for a second. Every owned option on this list is between 40 and 90 percent more expensive than it was at introduction. The rental column is the only one that did not personally punish you for planning ahead.
RTX PRO 6000 Blackwell 96GB: the alternative that stopped being obvious
For most of 2025 this was the easy answer. Ninety-six gigabytes of GDDR7 with ECC in a single slot, 24,064 CUDA cores, 1.79 TB/s, and it slotted into a workstation you already owned.
Then it went from $8,565 at launch to $13,250 in June 2026 to $16,000 on NVIDIA's marketplace this month. Roughly 87 percent up from launch. Newegg has listed it near $13,998 and B&H around $15,499, so shop around, but the direction is one way.
The cause is the GDDR7 shortage. This card uses a 96GB clamshell design, the largest VRAM pool on any discrete GPU, which makes it the single most supply-sensitive product in the category.
Worth knowing: it shares the exact same 1.79 TB/s bandwidth as an RTX 5090. You are buying capacity and slot count, not speed.
Mac Studio M3 Ultra: great bandwidth, amputated ceiling
A year ago this was the memory capacity champion outside the data center. 512GB of unified memory at 819 GB/s, roughly three times the bandwidth of a DGX Spark, for around $10,000.
That machine is gone. Apple pulled the 512GB configuration in March 2026 and the 256GB tier in May, both blamed on memory supply, then raised prices on June 25. The M3 Ultra now tops out at 96GB and costs $5,299, up from $3,999.
Read that again if you are working from an older comparison table. The Mac Studio is now in the same capacity tier as a single RTX PRO 6000, not the tier above it. Used 512GB units have appeared on eBay above $25,000, which tells you exactly how the secondary market feels about scarcity.
The M5 Ultra refresh has slipped to roughly October 2026. If you can wait, wait. We compared the two ecosystems in more depth in our piece on Apple Silicon versus NVIDIA for AI workloads if the platform question is the real one for you.
DGX Spark: the CUDA parity play
At $4,699 the Spark is not fast. Its 273 GB/s bus is the bottleneck, and dense 70B decode lands in single digits with common tooling.
What it buys is software. The full CUDA and TensorRT-LLM stack, matching what your cloud instances run, so what works on your desk works in production. NVIDIA's CES 2026 software update delivered up to 2.5x improvement over launch performance through TensorRT-LLM work and speculative decoding, which materially changed its value.
Buy it as a development box. Do not buy it as an inference server.
Strix Halo 128GB mini PCs: the capacity bargain that also inflated
The AMD Ryzen AI Max+ 395 puts 128GB of unified LPDDR5X in a machine the size of a lunchbox, with roughly 96GB addressable as GPU memory on Linux. That is more usable model space than an RTX 5090 at a fraction of the power draw.
In June you could get a 128GB GMKtec EVO-X2 for about $2,199, or a Framework Desktop at $1,999. As of August 2026 the 128GB tier across Framework, GMKtec, and Beelink runs $3,449 to $4,349.
Bandwidth is 256 GB/s, so this is a capacity-first box, not a speed box. Run big models slowly and privately, and be comfortable on Linux, because the Windows path costs you 20 to 30 percent throughput.
RTX 5090 multi-GPU: the DIY route
Same 1.79 TB/s as the PRO 6000, 32GB per card, and it launched at $1,999. Street price now sits between roughly $3,000 and $5,000 depending on the week.
Four of them gets you 128GB of pooled VRAM for the price of one PRO 6000, plus a power supply problem, a thermal problem, a chassis problem, and tensor parallelism configuration to debug. Genuinely good value if your time is cheap and your electrician is friendly.
Renting DGX-class hardware instead of buying it
Here is the math nobody in a hardware sales conversation wants on the whiteboard.
A B200 rents for $5.29 to $7.05 per GPU-hour, with Lambda at the floor of that range across a six-provider survey. B300 capacity runs $7.10 to $17.80. H200 sits around $3.70, and H100 starts near $2.01 for PCIe.
Take the middle of the DGX Station range, call it $100,000. At $7 an hour on a B200, that is more than 14,000 GPU-hours. Nineteen months of literally never turning the thing off. And you are renting current data center silicon rather than owning a desktop part that will be a generation behind by the time the DRAM crunch eases.
Buying wins on three specific things. Data that legally cannot leave your building. Workloads that run genuinely continuously at high utilization. And latency-sensitive interactive work where round trips to a region hurt.
Everything else is a preference dressed as a requirement. We ran the same exercise for the smaller machine in DGX Spark cloud rental versus buying, and the shape of the answer was identical.
Now the part I actually want you to sit with.
A lot of people pricing a DGX Station tell me they need it "for AI agents." Almost none of them do. An autonomous agent spends its life doing orchestration, tool calls, memory lookups, and API round trips. The model inference is usually a hosted API call. The bottleneck is scheduling and context management, not local FLOPs.
Local silicon matters when the weights must stay in the room. It does not matter for most agent work, and conflating the two is how hardware budgets get approved for problems they do not solve.
Picking a DGX Station alternative by what you are actually doing
Training or fine-tuning trillion-parameter-class models on premises, with data that cannot leave: buy the Station. Nothing else on this page does it. Accept the four to thirteen week lead time.
Running 70B to 120B inference locally for a small team: the RTX PRO 6000 at 96GB, or a Strix Halo box if throughput matters less than capacity and privacy. The Mac Studio is now in this bracket too rather than above it.
Developing against CUDA so your local work matches your cloud deployment: DGX Spark, and skip the rest.
Bursty research workloads that spike and idle: rent. Every time. The utilization math almost never favors ownership unless you are above roughly 60 percent duty cycle.
Building AI agents rather than training models: you are shopping in the wrong category. Start with the orchestration layer and buy hardware later, if ever. If you are running open models behind agents specifically, our notes on local model hardware for agent workloads get into what actually gets used versus what gets bought.
The most expensive hardware mistake in 2026 is not buying the wrong box. It is buying any box for a workload that was never compute-bound.
If your work is the agent layer and not the model layer, start free on BetterClaw. The free plan gives you one agent with every feature and no credit card. Pro is $49 a month for five agents, unlimited connectors, and 90-day memory, with 20 percent off annually. Full pricing here. Fifty plus companies run agents on it including Carelon, Grainger, KeHE, Premier, and Robert Half, and none of them bought a workstation to do it.
What I would tell a friend with the purchase order open
Do not benchmark this decision against last year's prices. That is the whole lesson of 2026 so far.
The DGX Station is a legitimately remarkable machine and there is a narrow set of teams who should buy one without hesitation. If you hold regulated data and you need a trillion-parameter model in your own building, the alternatives are not alternatives, they are compromises with a different shape.
For everyone else, the shortage has done something useful. It has forced a question that cheap hardware let people avoid for years. Do you actually need to own the compute, or did owning it just feel more serious than renting it?
My friend with the spreadsheet ended up renting for six months while he found out. He has not gone back to the purchase order.
Frequently Asked Questions
What is the best DGX Station alternative?
It depends on what you are replacing. If you need the raw memory capacity, nothing under $20,000 genuinely substitutes for a GB300 Station, and renting B200 or B300 capacity in the cloud is the closer functional match. If you need 70B to 120B local inference rather than trillion-parameter work, the RTX PRO 6000 Blackwell at 96GB or a 128GB Strix Halo mini PC covers it for a small fraction of the price.
How does the DGX Station compare to the DGX Spark?
They are different machine classes despite the shared branding. DGX Spark uses the GB10 Superchip with 128GB of unified memory at roughly 273 GB/s and costs $4,699, while DGX Station uses the GB300 with up to 784GB of coherent memory including HBM3e at 7.1 TB/s and costs somewhere between $80,000 and $123,000. Spark is a development box, Station is a workstation-class compute machine.
How much does an NVIDIA DGX Station cost in 2026?
NVIDIA has never published an official price and does not sell a Founders Edition, so pricing comes from OEM partners. An MSI XpertStation WS300 listed on CDW at $96,995.99, and distributor quotes generally run from just under $100,000 up to about $123,000 depending on storage and service, with some trackers putting the floor nearer $80,000. Lead times in August 2026 range from four to thirteen weeks.
Is renting DGX-class hardware cheaper than buying?
For most teams, yes, and by a wide margin. At $5.29 to $7.05 per GPU-hour for a B200, a $100,000 purchase equals well over 14,000 GPU-hours, which is more than a year and a half of continuous use before you break even. Ownership wins mainly when data cannot leave your premises or when utilization is genuinely sustained above roughly 60 percent.
Do I need a DGX Station or any local GPU to run AI agents?
Almost certainly not. Agent workloads are dominated by orchestration, tool calls, memory retrieval, and API round trips rather than local model inference, so the bottleneck is scheduling and context management rather than FLOPs. Unless your model weights are legally required to stay in your building, a managed agent platform with your own API keys costs nothing to test and removes the hardware question entirely.
