AI infrastructure · 01
Data-centre GPUs: Blackwell, Hopper and Instinct, matched to your workload.
Most GPU enquiries start with a model name. The better question is how much GPU memory your model needs, how far the job has to scale, and what your racks can power and cool. We help you answer that, then source the right accelerator in the right form factor.
B300 · B200 · H200 · H100 · L40S · RTX PRO
NVIDIA
Instinct MI355X · MI325X · MI300X
AMD
SXM · OAM · PCIe
Form factors
OEM servers or qualified baseboards
Supplied as
Overview
Which GPUs are in demand, and why
In 2026 most buyers choose between six or seven chips. Memory capacity and bandwidth have grown fastest: from 80 GB on the H100 to 141 GB on the H200, around 180–192 GB on the B200 and 288 GB on the B300 and MI355X.
Blackwell (B200, B300 and the GB200/GB300 rack systems) is the top tier for frontier training and large-scale inference. It is also the hardest to get: hyperscalers hold most of the allocation, and enterprise lead times run from a few months to most of a year depending on channel and order size.
Hopper (H200 and H100) remains the workhorse for fine-tuning and production inference. The H200's 141 GB of HBM3e lets larger models fit on fewer GPUs, and supply is far more predictable than Blackwell. AMD's Instinct MI300X and MI325X offer even more memory per GPU and are a strong choice where the software stack runs on ROCm.
For inference, visualisation and smaller models, PCIe cards such as the L40S and RTX PRO 6000 fit standard air-cooled 2U and 4U servers at a fraction of the power. Rule of thumb: pick the smallest GPU whose memory fits your model, then check what's actually available.
SXM and OAM modules are not sold like add-in cards. They come on 8-GPU baseboards inside OEM servers, connected by NVLink or Infinity Fabric. We therefore quote them as complete, OEM-qualified systems, which is also how the manufacturer warranty stays valid.
At a glance
- Training
- B200 / B300, H200, MI325X / MI355X on 8-GPU baseboards
- Inference
- H200, H100, L40S, MI300X; PCIe for smaller models
- Workstation & visual
- RTX PRO 6000 Blackwell, L40S
- Minimum order
- One 8-GPU node or one PCIe server; cards in OEM-supported configurations
- Lead time
- Quoted per model after upstream confirmation
How we work
We source from India's established brands — we do not manufacture
Akarta Exports is a sourcing and export house. We do not own factories and we do not claim to. Our value is that we know which well-known Indian brands and manufacturers make what, to what standard, and at what price — and we buy from them to the exact specification you give us.
01 — Your specification, written down
We start from your specification, not from a catalogue. Grade, dimension, tolerance, finish, test standard, pack, artwork and labelling are put in writing and confirmed with you before a single enquiry goes out.
02 — Matched to the right Indian brand
We place that specification with established, well-known Indian brands and manufacturers in the cluster that genuinely makes the product — Morbi for tiles, Rajkot and Ludhiana for machinery, Meerut and Jalandhar for sports goods, Kerala and Gujarat for spices.
03 — Offers compared on merit
Because we own no plant and owe no factory a minimum offtake, we compare offers openly on quality, consistency, capacity and price — and tell you where a lower price will not hold up in your market.
04 — Inspected, consolidated, shipped
Our own inspectors clear the goods at the brand's plant and again before loading. Multiple brands and multiple divisions consolidate into one container, one invoice and one document set.
Brand names are disclosed to you on the offer and named on the inspection report, so you always know whose production you are buying. Where a buyer already has a preferred Indian brand, we are happy to source that brand and handle only the inspection, consolidation, documentation and shipping.
Are you an Indian manufacturer or brand owner in this sector? Register your plant with our export desk.
The range
GPUs we source
- NVIDIA B200 & B300 (Blackwell)
- HGX 8-GPU baseboards in air- or liquid-cooled 8U/10U servers; GB200 and GB300 NVL72 rack systems for large clusters.
- NVIDIA H200 & H100 (Hopper)
- SXM5 on HGX 8-GPU baseboards; H200 NVL and H100 PCIe for 4-GPU and 8-GPU PCIe servers.
- AMD Instinct MI300X, MI325X, MI355X
- OAM modules on 8-GPU universal baseboards in OEM servers, for teams running ROCm, PyTorch and vLLM.
- NVIDIA L40S, L4 & RTX PRO
- PCIe cards for inference, video and visual computing in standard air-cooled servers.
Technical data
Headline specifications
Vendor-published figures per GPU for the SXM/OAM versions unless noted. Use them for shortlisting; the quotation confirms the exact SKU.
Swipe table to see all columns →
| GPU | Memory | Bandwidth | Power (TDP) | Interconnect | Best for |
|---|---|---|---|---|---|
| NVIDIA B300 | 288 GB HBM3e | ~8 TB/s | ~1,100–1,400 W | NVLink 5 | Largest models, long context |
| NVIDIA B200 | 180–192 GB HBM3e | ~8 TB/s | ~1,000 W | NVLink 5 (1.8 TB/s) | Frontier training & inference |
| NVIDIA H200 | 141 GB HBM3e | 4.8 TB/s | 700 W | NVLink 4 (900 GB/s) | Fine-tuning, high-throughput inference |
| NVIDIA H100 SXM | 80 GB HBM3 | 3.35 TB/s | 700 W | NVLink 4 | General training & inference |
| AMD MI325X | 256 GB HBM3e | 6.0 TB/s | 1,000 W | Infinity Fabric | Memory-heavy LLMs |
| AMD MI300X | 192 GB HBM3 | 5.3 TB/s | 750 W | Infinity Fabric | Large-model inference |
| NVIDIA L40S (PCIe) | 48 GB GDDR6 | 864 GB/s | 350 W | PCIe Gen4 | Inference, vision, graphics |
Figures from NVIDIA and AMD datasheets; final configurations vary by OEM server.
Customs classification
HS codes for this range
Harmonized System codes most commonly used when clearing these goods. The final 8-digit tariff line is confirmed on your commercial invoice according to the destination country's tariff schedule.
| HS code | Classification |
|---|---|
| 8473.30 | GPU accelerator cards and baseboards (parts of 8471 machines) |
| 8471.50 | When supplied installed in complete servers |
Quality assurance
Nothing leaves India until our own inspectors clear it
Akarta runs an in-house quality department with inspectors assigned to each division. Goods are checked at the factory during production and again before loading, against the written specification agreed on your purchase order — not against the factory's own grading.
- Model, form factor and memory size matched to the quoted OEM server
- Server confirmed on the GPU maker's qualified-system list
- Availability or allocation confirmed upstream before payment
- Warranty territory and support route stated on the quotation
ISO 9001:2015
Certified quality management system
Govt. approved
Registered Indian merchant exporter
In-house docs
Export paperwork prepared by our team
Packing & shipping
Packed the way your market receives it
Tell us the pack, carton print and shipping marks you need and we will quote against them. Loading plans are prepared before dispatch so the container arrives the way your warehouse expects it.
- GPUs shipped installed in OEM servers, or on sealed baseboards where supported
- Anti-static, shock-indicated packaging
- Insured air freight to your site or forwarder
Global reach
Quoted in your currency, shipped from your nearest Indian port
Enquiries reach us from importers, distributors, contractors and retail chains across five continents — Africa, the Middle East, Europe, the Americas and Asia-Pacific. We quote FOB, CFR or CIF, load from Mundra, Kandla, Nhava Sheva or Pipavav depending on your routing, and keep one commercial contact accountable for the consignment from purchase order to bill of lading.
Africa
Kenya, Tanzania, Uganda, Ethiopia, Nigeria, Ghana, Côte d'Ivoire, Senegal, Egypt, Morocco, Algeria, Sudan, Mozambique, Zambia, South Africa
Middle East
UAE, Saudi Arabia, Qatar, Oman, Kuwait, Bahrain, Jordan, Iraq
Europe
Netherlands, Germany, Spain, Italy, Poland, United Kingdom
Americas
United States, Canada, Mexico, Brazil, Chile
Asia-Pacific
Australia, New Zealand, Singapore, Malaysia, Vietnam, South Korea, Japan
Load ports
Mundra · Kandla · Pipavav · Nhava Sheva
Incoterms
FOB · CFR · CIF · EXW on request
Correspondence
English, with documentation to buyer-country requirements
Buyer questions
What importers usually ask us first
Can I buy loose SXM GPUs?
SXM and OAM modules are normally sold on baseboards inside OEM servers. Buying them that way keeps the warranty and support valid, so that's how we quote them.
H200 or B200?
If your model fits in 141 GB per GPU and you need hardware soon, H200 is usually the faster route. B200 gives roughly double the performance, but lead times are longer and most systems need far more rack power.
Do you supply AMD Instinct?
Yes. MI300X and MI325X systems come from several OEMs, and availability is often better than for top-end NVIDIA parts.
Common enquiries
Specifications buyers most often ask us to quote
These are the enquiries that reach this desk most weeks. If yours is on the list, we already have the grades, packing and loadability worked out — send the quantity and destination port and we will price it.
Also in this division
Related ranges
Export desk
Request a quotation
Send your product type, grade, quantity and destination port. Our export desk replies with available grades, packing options, loadability and indicative pricing — usually within one working day.
AI infrastructure desk
Buying GPUs, server memory, GPU server nodes or data-centre networking? These have their own two-step enquiry so the right specialist sees it first.
Open the AI hardware enquiry →