A few days ago, MP Mariana Bezugla solemnly introduced NVIDIA DGX as a local alternative to cloud AI services. Now she explained why she brought it back and why she «didn’t declare anything.»
A few days ago, MP Mariana Bezugla solemnly introduced NVIDIA DGX as a local alternative to cloud AI services. Now she explained why she brought it back and why she «didn’t declare anything.»
Let us remind you: in early May, Mariana Bezugla announced the purchase of an AI computer based on the NVIDIA GB10 Grace Blackwell Superchip with 128 GB of memory for approximately $4,000. The device was named Athena and was to become a local platform for processing sensitive data — so that personal requests from citizens and sensitive materials «would not quietly merge into ChatGPT.» Bezugla planned to declare the purchase.
However, the deputy’s plan failed. «Athens returned. So, there’s nothing to declare. Rozetka, thanks for your patience,» the deputy wrote on Telegram.
Bezugla gave eight reasons for returning.
The first are technical: crude firmware, slow operation, stationary form factor that limits use. Added to this is the fact that the device requires daily configuration and actually requires either knowledgeable personnel or the deputy’s personal time — which, according to her, «is not her core area of work.»
Then there are the specific details that were alarming. The default interface offers to interact with the AI assistant via a Telegram bot, and among the languages, Russian is present. Bezugla wrote that this is «not pleasing»: «Immediately, thoughts arose that such unexpected support for Russian could be about some cooperation in the corporate environment and about using technologies for everyday local solutions and bypassing the sanction, security restrictions of cloud subscriptions… It’s embarrassing. The chip and operating system are American,» she says.
Separately, there is the issue of encryption. The system relies on the drive’s built-in hardware encryption, which is almost beyond the user’s control. Installing your own encryption, as is usually possible on regular Linux, is much more difficult here. According to reviews on the forums, it is «quite problematic.»
Before that, the firmware of a specific model of Asus Ascent GX10 turned out to be modified and buggy. The computer simply did not start once. An attempt to flash to the original version from NVIDIA via BIOS has not yet been successful.
Architecturally, according to the deputy, there are also limitations: Linux on ARM narrows the choice of third-party software. And despite the fact that the system is built on Linux, the key architecture remains the closed property of NVIDIA — not open source.
Bezugla formulated the conclusion succinctly: the price-quality ratio is not justified.
Despite all the difficulties, Bezugla conducted a full-fledged test. She downloaded GPT-OSS-120B — an open large language model from the developers of ChatGPT with 120 billion parameters. She downloaded 16 documents, several dozen pages each — publications on the topic of military reform, complex in content.
«GPT-OSS-120B 'digested' everything, made generalizations, answered specific questions. Then I asked to create a website and received hundreds of lines of code without any freezes,» the MP said.
The speed was mediocre: I set a task and went to make tea. But the quality did not sag: the model did not get lost, did not repeat itself, and gave equal answers of the same quality — both at the beginning and at the end of the chat.
«A person would not have to spend 30 minutes doing this kind of work, but a couple of days,» Bezugla said. The device, she said, is quiet, holds the load stably, and doesn’t overheat. The theoretical usage model that Bezugla saw: a human moderator plus an AI computer as a full-time assistant. You launch a chain of tasks, leave, and you have the result in a few hours.
To run even basic operations — download Ollama, load the model, install the graphical interface — Bezugliya had to work through the console every time. The device gave errors and required clarifications on the NVIDIA and Reddit forums. «Of course, the help of cloud AI as IT consultants for infrastructure deployment :) … We must not forget about the balance of time, effort, money and result,» she noted.
Separately, Bezugla noted that most of the tools on this platform work through Firefox — and added that it was convenient for her personally, because she has been using it as her main browser «since its appearance.»
The MP does not abandon the idea of a local AI infrastructure. According to her, experience has helped to finally determine priorities: work with non-sensitive data — legislative and public activities — can be improved through cloud solutions. For truly sensitive documents and citizens' appeals, «simpler technical solutions» will be suitable.
Among the options she is exploring: AMD Ryzen AI (Strix Halo) or Apple Mac Mini — as a «stable consumer option.»
