THE APEX TIMES
NVIDIA adds a 64GB option to DGX Spark, aiming to push more agent development onto developers’ own hardware
The DGX Spark, a compact local-AI system, will be offered with 64GB of unified memory through major PC makers starting Oct. 23, with NVIDIA also expanding multi-device scaling tools and launch support for popular models.
NVIDIA is expanding its DGX Spark “personal AI supercomputer” line with a new 64GB unified-memory configuration designed for developers and researchers who want to run local AI agents and models without relying on cloud instances for every task.
In a post published Oct. 2, NVIDIA said local AI is becoming more useful “by the token,” as AI agents move from prototypes into day-to-day development workflows. The company’s pitch is that increasingly capable open models are shrinking enough to fit on more devices, enabling builders to experiment privately and iterate faster on their own data.
The new DGX Spark configuration will be available this month through top manufacturer partners, including Acer, ASUS, Dell, Gigabyte, HP and MSI. NVIDIA said it will ship with DGX OS and the company’s NVIDIA AI software stack “ready to use from day one,” and that it will run local agent workloads privately on the device, without a cloud dependency.
Hardware-wise, NVIDIA said DGX Spark combines the NVIDIA Grace Blackwell compute with unified memory, NVIDIA ConnectX-7 networking, and a CUDA-accelerated AI software stack in one system. Unified memory is the platform design that allows the system to treat memory as a shared resource for compute and acceleration, with the goal of simplifying how developers run and scale AI workloads.
NVIDIA’s 64GB SKU is positioned as an “accessible price point” while keeping the same core platform elements as the 128GB model. NVIDIA said the new option includes the GB10 Grace Blackwell Superchip, the DGX OS, and the full NVIDIA AI software stack, and it supports running up to 100-billion-parameter models and agentic applications “fully on device.” NVIDIA also tied the platform to common local-development tooling, saying the system ships ready with NVIDIA Agent Toolkit, CUDA-X AI libraries, and Nemotron open models, along with popular runtimes including Ollama, vLLM, and PyTorch with CUDA.
For developers whose projects outgrow a single box, NVIDIA said two 64GB DGX Spark units can be clustered together for higher capacity. Instead of requiring separate setup work, NVIDIA said the devices can connect directly and pool their memory to 128GB, expanding model support to up to 200 billion parameters while delivering what it describes as twice the memory bandwidth. In NVIDIA’s own testing using a Qwen 3.8 27B benchmark, NVIDIA said a two-unit 64GB cluster delivered up to 1.7x performance compared with a single system, and it suggested performance headroom as workloads demand.
NVIDIA also highlighted built-in networking. It said every DGX Spark ships with a ConnectX-7 network interface card (NIC) out of the box, and that two units can connect directly using a QSFP cable. The cluster assistant is designed to configure multi-node scaling by detecting connected units, validating device configuration, and setting up the ConnectX-7 network, with the goal of keeping developers focused on model and agent work rather than infrastructure.
NVIDIA said DGX Spark ships “ready for agent development” from power on, including tools meant to reduce setup time. The company referenced Agent Toolkit and CUDA-X AI libraries, and it said developers can get from “power-on to running models in minutes.” It also said Blender is among the first major creator application providers to support the platform, with a prebuilt installer coming soon.
In software, NVIDIA said it is also working on a “Sync Model Launcher” that will arrive at the end of the month. NVIDIA described this launcher as a simplified way to download and launch Qwen3.8 27B on a single DGX Spark or a cluster, with NVIDIA Sync configuring how the model runs across connected devices and making it accessible from users’ laptops. NVIDIA added that the launcher will set up OpenCode to use the model so developers can begin coding in a browser.
NVIDIA said the DGX Spark 64GB system will be available exclusively through its manufacturer partners starting Friday, Oct. 23, with a starting price of $4,999. For software examples, NVIDIA pointed developers to agentic AI playbooks for systems such as “NemoClaw,” “OpenClaw,” “Hermes Agent” and “OpenShell,” and it said additional playbooks are coming soon specifically to 64GB devices.
Why It Matters
- By adding a lower-memory SKU at a defined price point, NVIDIA is trying to widen access to local agent development, potentially reducing the need for continual cloud compute.
- DGX Spark’s built-in multi-node scaling approach, using ConnectX-7 networking and NVIDIA’s sync tools, could make it easier for developers to move from experimentation to larger local workloads.
- The emphasis on ready-to-run runtimes and launch tooling suggests NVIDIA wants to compress the time from purchase to active model deployment on a developer’s own hardware.
- If local AI use continues to expand beyond demos, offerings like DGX Spark could benefit from higher demand for turnkey “on-prem” AI developer platforms rather than only high-end data-center builds.
Sources
Key Facts
- NVIDIA will offer a new DGX Spark configuration with 64GB of unified memory through partners including Acer, ASUS, Dell, Gigabyte, HP and MSI.
- The 64GB DGX Spark is expected to run agent workloads fully on device without a cloud dependency and is positioned to support up to 100-billion-parameter models.
- DGX Spark combines NVIDIA Grace Blackwell compute, unified memory, NVIDIA ConnectX-7 networking, and a CUDA-accelerated NVIDIA AI software stack.
- Two 64GB units can be clustered to pool memory to 128GB, with NVIDIA saying up to 200-billion-parameter model support and reporting up to 1.7x performance on a Qwen 3.8 27B test.
- NVIDIA said the DGX Spark 64GB will start at $4,999 and become available through partners starting Oct. 23.
Technology Related
Oracle’s AI cloud pricing risk: analysts say “fixed margin” deals could limit benefit as GPU costs rise
A market commentary points to how the timing and structure of Oracle’s large AI cloud commitments may affect profitability if GPU rental prices keep climbing. The concern is less about demand and more about margin mechanics.
Oracle shares slide as report points to more debt plans tied to AI chip spending and OpenAI revenue disclosure
A Yahoo Finance report linked Oracle’s stock weakness to outlines that the company may use additional borrowing to fund AI chip purchases, with investor attention also drawn to new disclosures about OpenAI revenue.
US pauses Microsoft’s PERM green-card program after political push on hiring priorities
A reported government action is freezing Microsoft’s ability to use PERM for certain employment-based permanent residency cases, drawing public criticism from Vice President JD Vance that the company must prioritize hiring American workers.
Amazon rolls out redesigned Alexa tablet lineup with Alexa+ built in, prices starting at $229
The new tablets combine updated hardware, an Alexa+ driven home experience, and direct access to Google Play apps, as Amazon continues pushing its assistant deeper into everyday devices.
Amazon’s “About You” feature raises questions about how shopping data is turned into a consumer profile
A Yahoo Finance segment examined Amazon’s “About You” feature and how it appears to translate purchases and activity into a personalized view of consumers’ spending habits, highlighting the growing scrutiny of personalization and privacy.
Adobe shares rebound on Thursday as investors step back into the stock after a prolonged slide
Adobe Inc. (NASDAQ:ADBE) climbed on Thursday, with investors buying shares following a long period of weakness even as major U.S. indexes posted declines.
Chipotle rises after report Starbucks considered takeover talks, spotlighting M&A speculation and sector sentiment
A market wrap highlighted Chipotle shares moving higher on a report that Starbucks had looked into a possible acquisition of the burrito chain. The move underscored how takeover speculation can quickly spill over into consumer and restaurant equities, even when details remain unclear.
Apple investors brace for possible services risk ahead of earnings, according to Citi note highlighted by Yahoo Finance
A recent market discussion cited a Citi analyst’s latest take on Apple’s services business, framing a “stealth” concern that could matter to investors as the company prepares to report results.
Labor Department action on green-card sponsorship for tech workers rattles investors, dragging Infosys toward a six-year low
A U.S. Labor Department move tied to H-1B hiring and green-card sponsorship is raising fresh concerns about demand stability for offshore IT services providers, with Infosys sliding toward its weakest level in about six years.
Trump administration bars Microsoft from US green-card program for foreign workers, report says
A Trump administration action reportedly prevents Microsoft from participating in a US immigration channel used by companies to bring in foreign workers, citing concerns about staffing American roles with overseas hires.