Video · 5 October 2026 · 2 min read

Running AI Models Locally: What Hardware Is Needed?

Abdul Osama (عبدول أسامة) ran an AI model on his gaming PC without internet access and got replies, but found it slow and limited. He says running a much larger open-source model yourself could require costly hardware; renting hardware to host a model is another option.

Key takeaways

Abdul ran an AI model offline on a gaming PC, but found it slow and limited.
His hardware figures describe a much larger model, not every local model.
Renting hardware can avoid a server purchase, but it is not offline computing.
An open-source model alone does not guarantee data privacy on a rented server.

You can run an AI model offline on your own computer, but that does not mean a large model will run well on it. Abdul Osama (عبدول أسامة) says he ran a model on his gaming PC, spoke to it and received replies without internet access. In his experience, it was slow and limited.

“I managed to run an AI model on my gaming PC without internet, talked to it, and it replied.” — Abdul Osama

Why might a larger model need more hardware?

Running a model locally means using hardware you control rather than a remote service. An open-source model, in this context, is one you can run on hardware you choose, subject to its license. Larger models can place heavier demands on memory, storage and processing. A GPU, or graphics processing unit, is hardware commonly used to speed up that work.

After his test, Abdul looked into running a much larger model with capabilities closer to the online services people use. He gave these figures as estimates from his research, not universal requirements:

In numbers

Around 8 GPUs.
Around 500 GB of RAM.
Between 2 and 4 TB of SSD storage; an SSD is a fast storage drive.
He said a machine like that could cost more than $100,000.
He plans to compare 3 hardware-rental platforms in a later video; he did not name them here.

What can you do instead of buying a server?

Abdul suggests renting hardware from a hosting platform and choosing an open-source model to run on it. He mentions OpenAI and Anthropic—companies that offer AI models and services—to make the point that their services are not the only route.

| Option | What the video establishes | |---|---| | Local gaming PC | Worked offline in Abdul’s test, but the model was slow and limited. | | Server you own | A large model could call for expensive hardware. | | Rented hardware | Lets you choose a model without buying the server; costs and privacy terms depend on the provider. |

Background

Hosting an open-source model on rented hardware does not, by itself, guarantee that your data stays private or is never used for another purpose. Before sending sensitive data, check where it is processed, how long it is retained and what the provider’s terms allow. A remote rented server also does not meet an offline-use requirement.

What should you do next?

Decide which model you want to run and whether your priority is working offline or avoiding a hardware purchase. Compare that model’s requirements with your own machine and hosting options, then review the provider’s privacy terms.

Questions and answers

Can I run an AI model without internet access?

Yes. Abdul says he ran one on his gaming PC and received replies offline, though he found it slow and limited.

Does every local AI model need an expensive server?

No. Abdul’s figures were estimates he found while researching a much larger model. He had already run a model on a gaming PC.

What is the alternative to buying hardware for an open-source model?

Abdul proposes renting hardware to host a model you choose. He says he will compare platforms later, but names none in this video.

Will renting hardware keep my data private?

Not automatically. Abdul presents privacy as a reason to consider this route, but you still need to check the host’s data-handling terms and configuration.

Sources and tools mentioned

Full transcript

تشغيل نموذج ذكاء اصطناعي على جهازك لوكال مش مستحيل، بس في مشكلة. أنا فعلا قدرت أشغل نموذج ذكاء اصطناعي على جهاز الجيمينج بتاعي من غير انترنت وتكلمت معاه ورد علي. بس بصراحة كان بطيء وإمكانياته محدودة. فسألت نفسي طب لو عايز أشغل نموذج مفتوح المصدر بقوة قريبة من 4.8 أوبس أو تشات جي بي تي اللي احنا بنستخدمه كودكس، ولما دورت إيه الإمكانيات اللي أنا محتاجها في الجهاز عشان أعرف أستخدم حاجة زي كده لوكال، اتفجعت من الإمكانيات المطلوبة. إحنا مش بتكلم على باي دو وي لابتوب ولا ماك ميني مثلاً برامات كتير، إحنا بنتكلم على سيرفر ممكن يحتاج حوالي 8 كروت شاشة 16 جيجا بايت رام، عايز رامات في حوالي 500 جيجا بايت رام وعايز لك SSD ما بين 2 ل 4 تيرا بايت عشان موديل زي مثلاً أوبس 4.8 يشتغل عندك. جهاز زي كده ممكن يعدي الميت ألف دولار. وده السبب طبعاً إن شركات الزكاء الصناعية بيستثمروا في حاجتين، في النماذج نفسها وفي الديتا سنترز. لأنه الصناعة دي قائمة على أنهم يبنوا إنفستركتر حلو وبعد كده يأجروا للناس ويطوروا في الموديل بتاعتهم ودي الخدعة اللي هم شغالين علينا بيها دلوقتي. الخبر الحلو إنك مش محتاج تشتري جهاز بميت ألف دولار ولا مضطر إنك تستخدم أنتروبك أو أوبن إيه آي. الفيديو الجاي هعمل لكوا فيه مقارنة ما بين 3 منصات أوبن سورس بيأجروا لك الهاردوير وإنت تختار الموديل بتاعك سواء أنت هتجيبه وهم بيجيبوا لك جاهز ويسطبوه لك على يعني أنت تقدر تقول حاجة كده الشيرد هوستنج وتقدر إن أنت بعد كده تكون بتوفر فلوس وعارف إنه على الأقل أنت بتستخدم موديل مفتوح المصدر والبيانات بتاعتك ما بتتسربش وما بتتبعتش وما بيتطوروش بيها. أنا عبدالبسامع تابعونا عشان تعرفوا كل جديد في زكاء الصناعي والتسويق.

This text was written with AI from the video and may contain small transcription errors. · Markdown version

Watch next

Are AI tools becoming more expensive?

Abdul Osama (عبدول أسامة) expects some AI tools to become more expensive and predicts fewer benefits in cheap or free plans. He suggests running open-source models on owned or rented hardware as an alternative, while stressing that it takes learning and technical work.

5 October 2026 · 2 min read

WisperFlow for Dictation and Meeting Notes

Abdul Osama (عبدول أسامة) uses WisperFlow to dictate text instead of typing and to capture and summarize meetings or conversations. He demonstrates multilingual dictation and says the note taker can also be started manually without a detected meeting.

6 October 2026 · 1 min read

Codex vs Claude for Coding: Abdul Osama’s Take

Abdul Osama (عبدول أسامة) prefers Claude over Codex for coding and calls Sam Altman’s account of using Codex to generate app ideas marketing. The video offers an opinion, not a tested comparison of the two tools.

5 October 2026 · 1 min read