PC build for AI and ML in India: VRAM, RAM, real prices

VRAM decides which models you can run

A local model runs fastest when all of its weights sit in the graphics card's memory. The download size of a model is a good first estimate of the VRAM it needs, and the conversation context (the KV cache) needs more on top. Ollama's documentation notes that memory use grows with context length and with the number of parallel requests.

The sizes below are Ollama's default downloads for each model, which are already compressed (quantised). Running at higher precision needs more memory.

If a model does not fit, it can still run. llama.cpp, the open-source engine behind many local AI apps, supports CPU and GPU hybrid inference, and Ollama can likewise split a model between GPU and CPU: part of the model lives in system RAM. That works, but it is much slower, so treat it as a way to try a model rather than to use it every day. Run ollama ps and check the Processor column: 100% GPU means the model is fully on the card.

Image and video generation (Stable Diffusion, Flux, video models) is just as VRAM-hungry as chat models. If you plan to do both, size for the bigger of the two.

NVIDIA CUDA vs AMD ROCm vs Intel: the software reality

Hardware specs matter less than whether your tools run on the card. As of October 2026:

• For inference only (chatting with local models), AMD and Intel are workable. Ollama uses ROCm on supported Radeon cards and Vulkan for other AMD and Intel GPUs.
• For training and fine-tuning, NVIDIA is still the safe pick. AMD lists a known issue where training with Llama-Factory or Unsloth can crash or reset the GPU on the RX 9070 series under Windows.
• If you use Linux, a 16 GB Radeon such as the RX 9070 XT (₹82,799 in our catalogue on 5 Oct 2026) is a cheaper route to 16 GB than NVIDIA. It uses 304 W typical board power, per AMD.

Three example AI builds with today's prices

We ran our PC builder's AI and ML profile at three budgets on 5 October 2026. At ₹1,50,000 the builder reports that the budget is below its cheapest AI build and returns a ₹2,09,038 build instead (₹59,038 over). The ₹3,00,000 and ₹7,00,000 budgets fit.

SwapMyTech prices on 5 Oct 2026; prices move, check the builder. Every part can be swapped there, and the total updates as you go. You can build it in the PC builder.

Reading the three builds

• Entry. The RTX 5060 Ti 16 GB is the cheapest 16 GB NVIDIA card in our catalogue, and it is the part doing the AI work. The processor is modest on purpose; inference barely uses it once the model is on the GPU. Memory is the second-largest cost because DDR5 prices rose sharply in 2026.
• Mid. Same 16 GB of VRAM, so the same models fit, but the RTX 5070 Ti has 8,960 CUDA cores and a 256-bit memory bus against the 5060 Ti's 4,608 cores and 128-bit bus, per NVIDIA. That means faster token generation and image renders, not bigger models. The Z890 board is more than a non-K Core Ultra 5 needs; a B860 board (the builder lists a Gigabyte B860M Eagle WiFi6 at ₹14,399) saves ₹16,591.
• Workstation. The RTX 5090's 32 GB GDDR7 on a 512-bit bus is the step that unlocks 30B-class models with room for long context. NVIDIA rates it at 575 W and asks for a 1000 W system minimum; the builder picked 1500 W for an estimated 1,099 W load. The 16-core processor and 64 GB of RAM suit data preparation, compiling and running several tools at once.

Our catalogue had no 24 GB card on 5 October 2026, so the jump from 16 GB to 32 GB is a big one in price. If 30B models are the goal and the workstation is out of reach, a 16 GB card with hybrid CPU offload is the stopgap.

System RAM, storage and power

• System RAM. 32 GB is the floor for AI work. A useful rule of thumb is at least as much system RAM as VRAM, and more if you plan to offload part of a model to the CPU, because the offloaded layers live in RAM. 64 GB is worth it for a 32 GB card.
• Storage. Models are big: Ollama lists Llama 3.3 70B at 43 GB and Qwen3 235B at 142 GB. Add image checkpoints and datasets and 1 TB fills fast. Plan for 2 TB of NVMe, or add a second drive later.
• Power supply. Follow the GPU maker's system power figure, not the card's own rating. NVIDIA lists 600 W required system power for the RTX 5060 Ti, 750 W for the RTX 5070 Ti and 1000 W for the RTX 5090. AI jobs hold the card at full load for hours, so pick a Gold-rated or better unit with headroom.
• Cooling and noise. Long runs heat the room as well as the case. An airflow cabinet and a UPS for India's power cuts matter more here than in a gaming PC.

Lowering the cost

The graphics card is close to half or more of every build above, so it is where trade-ins help most. An older card with 8 GB or more still has resale value. At SwapMyTech you can get a trade-in estimate and set it against the new build. If you only want to try local AI first, an 8 GB gaming card runs 7B-8B models today; see our gaming PC under ₹75,000 guide.