This compact personal cloud pairs a 24GB RTX 5090 Laptop GPU with an open-source operating system and one-click AI apps. The hardware is capable. The $4,999 price asks buyers to put a large value on convenience.
Home AI has a software problem disguised as a hardware hobby. Buying a fast GPU is easy enough. The less charming part comes later: Linux drivers, containers, model servers, storage permissions, remote access, updates and the sudden realization that your “private AI” needs six browser tabs and a small collection of GitHub bookmarks.
The Olares One tries to sell the finished version instead. It is a compact, always-on AI server with Olares OS already installed, a browser-based desktop and a market of more than 200 deployable apps.[1][2] The current US store listing shows it in stock at $4,999, with shipping promised within 24 hours.[3]
That is expensive. It is also one of the clearer attempts to turn a powerful local-AI machine into something a household or small team could use without appointing a part-time systems administrator.
The hardware is powerful, but read the GPU name carefully
Inside the 3.5-liter chassis are an Intel Core Ultra 9 275HX, 96GB of DDR5 memory, a 2TB NVMe SSD and Nvidia graphics with 24GB of GDDR7 VRAM. There is 2.5Gb Ethernet, Wi-Fi 7, HDMI 2.1, USB-A and Thunderbolt 5. The whole computer weighs 2.15kg, although the external power supply brings the traveling weight to 3.1kg.[3]
One specification deserves bold type: this is a GeForce RTX 5090 Laptop GPU, sometimes described by Olares as the RTX 5090 Mobile. It is not the 32GB desktop RTX 5090.
Nvidia lists the laptop part with 10,496 CUDA cores, 24GB of GDDR7 and 896GB/s of memory bandwidth.[8] Its desktop namesake has 32GB and 1,792GB/s, exactly twice the bandwidth.[9] Olares runs the mobile GPU at up to 175 watts in performance mode, according to its own product FAQ.[3] The shared name does not make them peers.
That caveat does not make the mobile chip weak. Twenty-four gigabytes of fast VRAM remains a useful tier for home AI. Quantized 7B to 14B language models fit easily. Many 20B to 32B models also fit at 4-bit precision, although context and cache memory reduce the neat capacity suggested by parameter math. Image models and many ComfyUI workflows benefit from both the VRAM and Nvidia’s broad CUDA support.
Larger models expose the dividing line between VRAM and ordinary system RAM. Olares has 96GB of DDR5, but it is not one unified high-bandwidth pool. Models too large for the GPU must spill work into slower system memory. In Olares’ own benchmark, GPT-OSS-120B exceeded the 24GB VRAM capacity and used CPU offloading; reported throughput fell from 36.16 tokens per second for one request to 4.44 tokens per second at eight concurrent requests.[7] A big model may run, in other words, without running the way the “5090” badge encourages you to imagine.
The smaller-model results are much more convincing, though they remain company tests rather than independent HighVRAM measurements. Olares reported 139 tokens per second for GPT-OSS-20B and 38 tokens per second for a 4-bit Gemma 3 27B model using vLLM at single-request concurrency. Its Qwen3 30B mixture-of-experts test reached 157 tokens per second.[7] Different architectures and quantizations make those figures poor material for a simple ranking, but they show where this machine should be happiest: fast 20B-to-30B inference, local agents and CUDA-heavy image or video generation.
A reviewer at DB Tech, who disclosed that Olares supplied the machine, reported roughly 125 to 130 tokens per second with GPT-OSS-20B. He also generated four 1,024-pixel images in about 30 seconds and a short Wan 2.2 video in a little over seven minutes.[6] Those observations are encouraging, but they are not a substitute for a broad independent benchmark suite on retail hardware.
Olares OS is the reason to buy it
A conventional workstation can match or beat these components. Olares One’s case rests on what happens after it arrives.
Olares OS wraps Kubernetes, storage, private networking, app deployment and GPU allocation in a desktop-like web interface. The company says users can install local models, Open WebUI, ComfyUI and other self-hosted tools from its app market rather than assembling each service manually. Apps run in isolated environments, and the system can give one workload exclusive GPU access or share the GPU between services.[4][5]
That combination suits a home better than a benchmark chart suggests. One box can answer private questions about family documents, transcribe meetings, generate images, host a media library and serve an AI interface to laptops or tablets around the house. The computer does the heavy work while the devices people already own become clients.
The ownership story is unusually credible, too. The Olares repository is public under the AGPL-3.0 license, and the company supports installing the OS on compatible Ubuntu or Debian hardware. Its documented minimum is four CPU cores, 8GB of available RAM and 150GB of SSD storage; a discrete GPU is optional.[5] A sponsored hands-on review also confirmed that Olares OS can be replaced or used alongside Windows.[10] That does not prove the system is perfectly private or painless, but it gives buyers an exit route that proprietary AI appliances often lack.
There is useful room to grow. Olares says the memory can be upgraded to 128GB, a second internal M.2 slot is available, NAS shares can be mounted over SMB and Thunderbolt 5 can attach an external GPU enclosure. Windows can run in a virtual machine, as a dual-boot installation on a second drive or as a complete replacement for Olares OS.[3]
The price buys integration, not maximum VRAM
At $4,999, the Olares One costs about $139 a month when spread over three years. Its one-year warranty and 2TB base drive look modest beside that number.[3] The store’s subscription-savings examples also deserve skepticism: replacing several premium cloud plans only saves money if a buyer already pays for them, and a local model does not automatically reproduce every service’s quality or features.
The current price is also far above the $2,999 figure used in Olares’ November 2025 benchmark and launch-period coverage.[6][7] At the earlier price, the compact hardware was unusually aggressive. At $4,999, buyers should compare it with a desktop workstation, a high-memory Mac or AMD unified-memory system, and cloud API costs based on their actual usage.
Olares One makes the most sense for a privacy-conscious creator, developer, family or small office that wants one CUDA-capable machine available all day and values a maintained interface more than component-by-component value. It is less compelling for someone who mainly chats with a 7B model, enjoys building PCs, needs a 70B model entirely in fast memory or expects desktop RTX 5090 performance.
There is a sensible test before spending five thousand dollars: install Olares OS on spare compatible hardware. If its app market, remote access and management layer remove the friction that has kept local AI out of daily use, the Olares One becomes easier to justify. If they do not, the expensive part of the package will not fix the problem.
Sources
1. Olares home page 2. Olares One product page 3. Olares One store listing 4. Olares OS product page 5. Olares GitHub repository 6. DB Tech Olares One review 7. Olares local AI benchmark 8. NVIDIA GeForce RTX 50 Series laptop specifications 9. NVIDIA GeForce RTX 5090 desktop specifications 10. Mickhat Olares One review