Comparison
How Locai One compares.
Locai One against NVIDIA DGX Spark and the enterprise AI factories from Dell, HPE and Lenovo, using each vendor's published specifications.
In short
Every option on this page runs NVIDIA. The choice is not which silicon vendor to back, it is which NVIDIA platform your workload needs and how much of the stack around it you assemble yourself. DGX Spark is NVIDIA's desktop development platform. Dell, HPE and Lenovo build validated data-centre architectures. Locai One is the full stack in one purchase: production-tier NVIDIA Blackwell hardware, Locai OS, Juno models, the Locai App your team uses, and an OpenAI- and Anthropic-compatible API, already together, with no per-token bills.
Three ways to buy on-premise AI
They are not competing versions of the same product. They are three different amounts of work left for you to do.
Development platforms
NVIDIA DGX Spark, Lenovo ThinkStation PGX
Built on NVIDIA's GB10 Grace Blackwell Superchip. A large unified memory pool in a machine that fits in a bag, designed for prototyping, fine-tuning and local agents at a desk. NVIDIA positions it as a development platform and it is excellent at that.
Enterprise AI factories
Dell AI Factory, HPE Private Cloud AI, Lenovo Hybrid AI Advantage
Validated NVIDIA reference architectures wrapped in vendor services. Enormous ceiling, from hundreds to tens of thousands of GPUs, and the right answer at data-centre scale. They assume a rack, a procurement cycle and a platform team.
Appliances
Locai One, Locai One Pro
One purchase at a fixed price. Production-tier NVIDIA Blackwell silicon chosen for serving rather than prototyping, with Locai OS, the models, the applications and the API already on it. Plug it in, serve the organisation.
Side by side
Scroll the table sideways to see every column.
| Locai OneAppliance | Locai One ProAppliance | NVIDIA DGX SparkDevelopment platform | Enterprise AI factoryDell · HPE · Lenovo | |
|---|---|---|---|---|
| Silicon | ||||
| GPU | 1× NVIDIA Blackwell | 2× NVIDIA Blackwell | NVIDIA GB10 Grace Blackwell | NVIDIA data-centre GPUs |
| GPU memory | 96 GB GDDR7 | 192 GB GDDR7 | 128 GB LPDDR5x, unified with the CPU | Configurable |
| Memory bandwidth | 1,792 GB/s | 1,792 GB/s per GPU | 273 GB/s | Configurable |
| Largest model | Up to 120B parameters | Up to 300B parameters | Up to 200B | Configurable |
| Form factor | Desk-side tower | Desk or floor | 150 × 150 × 50.5 mm, 1.2 kg | Rack |
| Power | 1,000 W | 1,500 W | 240 W | Data-centre power and cooling |
| What ships with it | ||||
| Arrives with a model installed | — | — | ||
| Operating system | Locai OS | Locai OS | NVIDIA DGX OS | Vendor stack with NVIDIA AI Enterprise |
| Serving API | OpenAI & Anthropic compatible REST | OpenAI & Anthropic compatible REST | Bring your own serving layer | Bring your own serving layer |
| Applications for end users | — | — | ||
| Identity, access and audit | Built in, on the appliance | Built in, on the appliance | Bring your own | In the vendor platform |
| Runs air-gapped | ||||
| Commercials | ||||
| Starting price | £29,950 + VAT | £49,950 + VAT | $4,699 MSRP | By quotation |
| Per-token bills | None | None | None | None |
| Best for | A team or department serving real users | A whole organisation, regulated industries | Developers prototyping and fine-tuning | Data-centre scale across many workloads |
Third-party figures are taken from each vendor's published specifications as of August 2026 and may change. DGX Spark memory, bandwidth, dimensions and power are from NVIDIA's DGX Spark specification and user guide; its MSRP is NVIDIA's published Founders Edition price. Dell, HPE and Lenovo platforms vary by configuration, so their rows describe the approach rather than a fixed build. Locai One performance depends on model, configuration and workload.
What you actually buy with Locai One
A GPU server is one layer of an on-prem AI system. The rest is an operating system, a model that fits the hardware, a serving API, identity and audit, and something non-engineers can open. Locai One is that stack as a single product, on NVIDIA Blackwell, at a fixed price.
Hardware
A desk-side NVIDIA Blackwell appliance, 96 GB of GDDR7 in Locai One and 192 GB across two GPUs in Locai One Pro, at 1,792 GB/s per GPU. That is the workstation-tier part NVIDIA publishes for production serving, not the development Superchip. Air-cooled, on a wall socket, no rack required.
Locai OS
The appliance boots into Locai OS. Nothing to install. From one screen you manage users, models, GPU health and the local audit trail. Email and TOTP are built in. Every request is recorded on the machine itself, not in a vendor's cloud.
Models
Locai Juno ships with the unit: a family of models compressed and specialised with our SPACE algorithm, then tuned to this hardware. Open-weight models from Hugging Face, public or private, install alongside. You choose in the portal and it is ready to serve in a few minutes.
Locai App
Your team talks to the appliance from a Mac or Windows app downloaded from the machine on your own network. About fifteen minutes from power-on to a first useful answer, from data that never left the building.
The API
An OpenAI- and Anthropic-compatible REST endpoint on the same box. Chat, embeddings, tool calling and streaming, with per-team keys you issue. Point Cursor, Claude Code, GitHub Copilot, n8n, LangChain or anything else that takes a base URL at the local address and it resolves to the machine in your building.
One purchase, the rest of the stack included
That is the difference the table is really measuring. NVIDIA DGX Spark is an outstanding development platform: CUDA-native, quiet, cheap to own, and the right machine if you are prototyping. Dell, HPE and Lenovo give you validated NVIDIA infrastructure at data-centre scale. In both cases you still stand up the model, the serving layer, identity, audit, the applications and the update path, then keep them running.
Locai One is the opposite shape of product. You buy the hardware once, from £29,950 + VAT, and the software that makes it useful to a whole organisation is already on it. Unlimited inference inside your network. No per-token bills as usage grows. Air-gapped if your policy requires it. The weights, the logs and the documents your team indexes all live on the appliance.
The full product story is on Locai One, with the model family on Juno.
Which one you should actually buy
Buy a DGX Spark
If you want a CUDA-native machine on your desk to prototype against, fine-tune with LoRA and run local agents, DGX Spark is purpose-built for that and nothing here matches it on size, power draw or price. Its 128 GB unified pool is larger than a single Locai One's 96 GB, it draws 240 W, and two units link over ConnectX-7 for bigger models. NVIDIA built it as a development platform and it is a very good one.
Buy an enterprise AI factory
If you are standing up shared infrastructure for thousands of users across many workloads, the validated NVIDIA architectures from Dell, HPE and Lenovo scale further than any appliance and come with the services to match. HPE quotes deployment in hours for Private Cloud AI; Dell's is a reference architecture plus integration. Both assume you have a data centre and a platform team.
Buy a Locai One
If you want a model serving your organisation, inside your network, this quarter, buy the appliance. Memory bandwidth sets how fast a loaded model decodes, and serving many people at once is a bandwidth problem before it is a capacity problem, which is why Locai One uses the workstation-tier Blackwell part at 1,792 GB/s. More importantly, the OS, Juno models, Locai App and compatible API are already on it. One purchase, the full stack, no assembly.
If the question is really about cost rather than hardware, the AI cost calculator puts an owned appliance against what you pay per token today. To choose between the two Locai units, see Locai One vs Locai One Pro.
Frequently asked questions
What is included in a Locai One purchase?
The hardware, Locai OS, Locai Juno models, the Locai App for Mac and Windows, and an OpenAI- and Anthropic-compatible REST API. Users, access, model updates, GPU health and a local audit trail are all in the Management Portal. You pay once, from £29,950 + VAT, and run unlimited inference inside your network. Open-weight models from Hugging Face can be installed alongside Juno.
Is Locai One an alternative to NVIDIA DGX Spark?
They are built for different jobs, and both run NVIDIA Blackwell. DGX Spark is NVIDIA's development platform: a 128 GB unified-memory machine at 273 GB/s for prototyping and fine-tuning at a desk. Locai One uses the workstation-tier Blackwell GPU at 1,792 GB/s, which is the bandwidth you want when a model is serving a team concurrently, and it ships with the OS, models, app and API already on it. If you are developing, buy the Spark. If you are serving, buy the appliance.
Does Locai One run NVIDIA hardware?
Yes. Locai One is built on NVIDIA Blackwell, with 96 GB of GDDR7 in the standard unit and 192 GB across two GPUs in Locai One Pro. NVIDIA silicon is the foundation of every option on this page.
How does it compare to Dell AI Factory or HPE Private Cloud AI?
Those are data-centre platforms built on NVIDIA reference architectures, sold by quotation and sized from a rack upwards. They scale far beyond a single appliance. They give you validated infrastructure to run your own model layer on. Locai One arrives as a finished system: hardware, Locai OS, Juno models, the Locai App and a compatible API, already installed, at a fixed price.
Could I just buy the hardware and build this myself?
Yes, and plenty of teams do. You would be assembling the inference server, the model and its quantisation, the serving API, identity and access control, audit logging, the end-user applications and the update path, then maintaining all of it. An appliance is that work already done and supported.