10decodersMost teams buy an NVIDIA developer kit before they know their workload. Here is where to start in 2026.
Blueprints, NIM, NeMo, Jetson, DGX Spark. The tooling is broad enough that access is no longer the hard part. Picking the right layer to start at is.
September 17, 2026 · 8 min read
The expensive way to start
The usual story goes like this. An engineering lead gets budget for "a GenAI box", orders the most capable NVIDIA developer kit that fits the purchase limit, and then spends two months working out what to run on it.
By the time the team has a use case, they find the kit was designed for robotics when they needed a desktop for fine-tuning. Or that everything they wanted to test was already available as a hosted API they could have called on day one.
NVIDIA's GenAI tooling is now broad enough that access is not the problem. Picking the right layer to start at is. This article maps the stack, compares the developer kits on price and purpose, and lays out the starting path we use ourselves.
The NVIDIA GenAI stack in four layers
Stop reading NVIDIA as a GPU vendor. Read the stack top-down, from what your application calls to what it runs on.
| Layer | What it is | When you touch it |
|---|---|---|
| Blueprints | Reference applications for RAG, video search and agents, shipped with Docker Compose and Helm charts | First. Start from a working system and change it |
| NIM microservices | Optimised model containers with OpenAI-compatible APIs: LLMs, embedding, reranking, OCR, speech | When you move from hosted trial endpoints to self-hosted inference |
| NeMo | Tools to customise, retrieve, guard and evaluate: NeMo Retriever, NeMo Guardrails, fine-tuning and evaluation | When off-the-shelf models are not accurate enough on your data |
| Runtime and hardware | CUDA, TensorRT-LLM and the GPU: Jetson, DGX Spark, RTX PRO workstations, data-centre GPUs | Last. Size it once you know models and concurrency |
Read the stack top-down. Choose the blueprint, then the models, then the box.
Which developer kit fits which job
On a spec sheet the kits look similar. They are aimed at very different work. Prices below reflect NVIDIA's July 2026 Jetson increases, which were driven by memory costs.
| Option | Price (USD) | Built for | Good GenAI fit |
|---|---|---|---|
| Hosted NIM endpoints | Free to try | Prototyping against current models without hardware | Every project's first two weeks |
| Jetson Orin Nano Super | $399 | Entry edge AI, learning, small vision and language models | Small language models, on-device assistants, lab work |
| Jetson AGX Orin | $3,499 | Established edge and robotics deployments | Existing Orin fleets. For new builds, compare against Thor |
| Jetson AGX Thor | $5,499 | Physical AI: Blackwell GPU, 128 GB memory, 130W envelope | Vision-language models on robots, cameras and machines |
| DGX Spark | $4,699 | Desktop AI lab: GB10 Grace Blackwell, 128 GB unified memory | Local inference and fine-tuning on sensitive data, RAG prototypes |
The practical split is simple. If the model lives inside a device that moves or sees, look at Jetson. If the model lives on a developer's desk and the output is software, look at DGX Spark. If you cannot say which, you are not ready to buy either.
A 30-day starting path
Days 1 to 5: pick one workflow, not a platform
Choose a task with a measurable output, such as answering policy questions from 500 internal PDFs. Write down what a correct answer looks like before any code.
Days 6 to 12: prototype on hosted endpoints
Use the NVIDIA API catalogue through its OpenAI-compatible interface. Your code changes very little when you move to self-hosted NIM later.
Days 13 to 20: deploy the closest blueprint
For document Q&A that is the RAG Blueprint. Run it on your own data and log where answers fail: extraction, retrieval or generation.
Days 21 to 26: decide what must be local
Data residency, latency, or cost at volume. Those are the good reasons to self-host. If none apply, stay on managed inference.
Days 27 to 30: size hardware from evidence
You now know model sizes, token volumes and concurrency. That is what a hardware quote should be built on.
What a first build on DGX Spark looks like
Take a healthcare or financial services team that cannot send documents to a public API. A sensible first build is a local document assistant:
- NIM containers for an embedding model, a reranker and a mid-sized LLM, all on the one unit
- NeMo Retriever extraction to pull text, tables and charts out of scanned PDFs
- A vector store such as Milvus or Elasticsearch running alongside
- NeMo Guardrails to keep answers on topic and filter unsafe content
This build is not production. It proves whether answer quality on your documents clears the bar, with data never leaving the room. Production usually moves to data-centre GPUs or managed cloud once concurrency grows, and we tell clients to plan for that from day one.
Not sure which layer to start at?
Our two-week GenAI architecture assessment ends with a working prototype on your data and a hardware and hosting recommendation you can take to procurement.
Book a Free AI Assessment →Where NVIDIA's stack is not the right answer
- You have a few thousand documents and no data residency rule. A managed LLM API with basic retrieval is cheaper and faster to ship.
- You are standardised on one hyperscaler's AI services and have no appetite for running containers. The operational cost of self-hosting will outweigh the performance gain.
- Nobody can own it. Self-hosted GenAI needs someone who reads GPU memory errors and container logs. Budget that skill before the hardware.
- The use case is still "explore GenAI". A kit does not create a use case. Hosted endpoints cost nothing to explore with.
What to do this week
- Write one sentence describing the workflow and how you will score a correct answer.
- Create an NVIDIA developer account and call one LLM and one embedding endpoint from the API catalogue.
- List your constraints: data residency, target latency, expected daily queries.
- Only then choose between hosted, DGX Spark, Jetson or data-centre GPUs using the table above.
Frequently asked questions
Do I need NVIDIA hardware to start building GenAI applications?
No. NVIDIA's API catalogue at build.nvidia.com lets developers call current NVIDIA and open models through an OpenAI-compatible interface without owning a GPU. Buy hardware once you know your models, data residency needs and query volumes.
What is the difference between DGX Spark and Jetson AGX Thor?
Both have 128 GB of unified memory and a Blackwell GPU. DGX Spark is a desktop AI system for building, fine-tuning and running models locally. Jetson AGX Thor is an edge computer for robots and machines that process sensor data in real time within a 130W power envelope.
What is NVIDIA NIM?
NIM is NVIDIA's set of pre-packaged, optimised model containers. Each exposes a standard API for an LLM, embedding, reranking, OCR or speech model, so the same application code works against hosted endpoints and self-hosted deployments.
Is the Jetson Orin Nano Super good enough for GenAI?
It suits small language models, learning and on-device prototypes. It is not suited to running large LLMs or serving many users. NVIDIA raised its developer kit price to $399 in July 2026.
