DGX Spark 64GB at $4,999: local AI now costs more for less
Nvidia's DGX Spark 64GB starts at $4,999 on Oct. 23, more than the 128GB box launched at. Size the model before you buy local AI hardware.
Nvidia announced a DGX Spark 64GB today, starting at $4,999 from Acer, ASUS, Dell, Gigabyte, HP and MSI on October 23. It has half the memory of the original DGX Spark, which launched at $3,999 with 128GB. So the cheaper-tier local AI box now costs a thousand dollars more than the bigger one did a year ago. That is what the memory shortage looks like on an invoice.
What actually happened
From Nvidia's announcement:
- Same chip, less memory: the GB10 Grace Blackwell Superchip, DGX OS and the full Nvidia AI software stack, with 64GB of unified memory.
- Model size: Nvidia says one unit runs models up to 100 billion parameters. Two clustered units (128GB total) run models up to 200 billion.
- Clustering: a new Sync Cluster Assistant links units over the built-in ConnectX-7 networking. Nvidia reports two clustered 64GB units ran Qwen 3.8 27B at up to 1.7x the performance of one.
- Software: Ollama, vLLM, PyTorch, Nemotron models and the Nvidia Agent Toolkit ship ready to use.
The price context: Tom's Hardware reported in February that Nvidia raised the 128GB Founders Edition from $3,999 to $4,699, citing memory supply. Its coverage of the 64GB model calls it a lifeline for buyers who can work with less.
Why local AI hardware pricing matters for your business
Size the model first, then the box. A dense 27B model fits in about 32GB. If that is what your workload needs, 64GB is enough and paying for 128GB buys nothing. If you need a 70B+ model with a long context, 64GB gets tight fast.
Do the break-even math. $4,999 up front versus API spend. If your agents burn a few hundred dollars a month in tokens, the box takes years to pay back, and API prices keep falling. Local wins on privacy, fixed cost and always-on agents, not on raw price.
Hardware prices now move like cloud prices. Two increases in one year. If you plan a local AI rollout, lock the quote and the delivery date in writing.
Keep the stack portable. Ollama and vLLM run the same open models on a Spark, a Mac Studio or a rented GPU. Build on those and the hardware choice stays reversible.
Key takeaways
- DGX Spark 64GB starts at $4,999 from six OEMs on October 23
- The original 128GB Spark launched at $3,999; memory shortages pushed prices up
- One 64GB unit runs models up to 100B parameters; two clustered run up to 200B
- Pick the model you need first, then buy the smallest box that runs it
- Build on Ollama or vLLM so the hardware decision stays reversible
Local box or API? Run your monthly token spend through our ROI calculator, then talk to us. We build model-agnostic agents that run on your hardware or anyone's cloud.
Sources: Nvidia, Tom's Hardware.
- #nvidia
- #dgx-spark
- #local-ai
- #hardware
- #memory-prices
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Strands Decider 2B: AWS open-sources a 115ms decision model
AWS Strands Labs open-sourced Strands Decider 2B, a decision model that picks from fixed options with a confidence score. Use it to route tickets and gate tool calls.
Read itMeta Muse Gadgets: open ESP32 and Linux SDK for agent hardware
Meta open-sourced Muse Gadgets, an Apache 2.0 ESP32 firmware and Linux SDK that puts its Muse AI agent on cheap hardware. What it means for shop-floor devices.
Read it