Gigabyte has brought three machines to the Bangladesh market aimed at running AI models locally rather than in the cloud, distributed through Smart Technologies (BD) Limited.
What is on sale
The AI TOP 100 W880 pairs an Intel Core Ultra 9 processor with an Nvidia RTX 5090 graphics card, at Tk 14,00,000.
The AI TOP ATOM uses Nvidia’s Grace Blackwell superchip with 128GB of memory and 4TB of storage, at Tk 10,00,000.
Alongside them is a new Eagle series laptop: AMD Ryzen 5, Nvidia RTX 3050, 16-inch display, at Tk 1,47,000.
Gigabyte says the desktops support running AI models directly on the device, and that the range was brought in with research, data analysis and gaming demand in mind. The company’s country manager, Khwaja Md Anas Khan, told Prothom Alo that having AI work run directly on the machine would reduce dependence on the cloud, making AI-based work, research and data analysis easier on a personal computer.
The claim is true, and incomplete
Reduced cloud dependence is a real benefit, and for some buyers here it is the only thing that matters. If you are working with patient records, financial data, legal documents or anything else that cannot leave the country or the building, a local machine is not a cost decision at all. It is the only lawful option. The same applies if your connection is unreliable enough that a timed-out upload wastes an afternoon.
But “reduces cloud dependence” is not the same as “costs less”, and the two often get merged in a showroom.
The arithmetic nobody puts on the price tag
Tk 14 lakh is roughly $11,500 at current rates. Renting a comparable high-end GPU by the hour from a cloud provider runs, very broadly, in the low single-digit dollars per hour — so somewhere in the order of a few thousand hours of rental before the purchase pays for itself on compute alone.
A machine running eight hours a day, six days a week, is busy about 2,000 hours a year. On that usage the workstation is defensible inside two years. A machine running a few hours a week is not: it is a Tk 14 lakh asset depreciating while idle, and the cloud bill it replaced would have been smaller.
Then add what the sticker leaves out. Electricity at continuous load. Cooling, which in Dhaka is not a rounding error. Backup power, because a 575W-class card mid-training does not tolerate load-shedding. Warranty and parts availability — a point worth asking the distributor directly, since an RTX 5090 that fails locally is not a next-day replacement.
Which machine is for which job
The two desktops are genuinely different tools, and the prices invert what you might expect. The ATOM, at Tk 10 lakh, is the Grace Blackwell machine with 128GB of unified memory — the one that can hold a large model in memory at all. The W880, at Tk 14 lakh, is the faster gaming-lineage card. For fine-tuning or running a large model locally, memory capacity is usually the binding constraint, not raw speed.
The Eagle laptop at Tk 1,47,000 is in a different category entirely. An RTX 3050 with a Ryzen 5 is a competent laptop that will run small models and quantised versions of larger ones. It is not a training machine, and buying it expecting to do serious local AI work would be a mistake.
One thing to check before buying
Local component prices have been volatile. RAM in Dhaka rose again through September on a memory shortage, and graphics card prices have moved with it. Ask for the price in writing with a validity date, and confirm whether the figure includes VAT — a Tk 14 lakh quote that becomes Tk 16 lakh at invoice changes the payback arithmetic considerably.
And if the work is in Bangla, test before you commit. Model performance in Bangla is still constrained by how little training data exists, and no amount of local GPU fixes that.
