Quick Overview
In this demonstration video, Krish Naik showcases a custom desktop PC build referred to as the ASUS agent computer or AI workstation. The presentation explains the hardware components assembled for running local artificial intelligence workloads, evaluates local language models against cloud application programming interfaces, and demonstrates agentic workflow software.
Key Points
- 1.Krish Naik built an ASUS AI workstation featuring an ROG Strix GeForce RTX 5070 Ti graphics card with 16 GB GDDR7 memory to execute artificial intelligence inference locally.
- 2.Local inference with models such as Qwen 3.5 4B and Llama 3.2 eliminates per-token application programming interface charges that accumulate when running continuous multi-agent workflows.
- 3.In a side-by-side benchmark, a local agent running Qwen 3.5 completed a multi-step factual research task in 17 seconds at zero dollar cost, while the cloud model took 28 seconds and incurred usage fees.
- 4.When testing live web search and summarisation, the cloud agent failed to finish within the configured step limit, whereas the local model completed the assignment in under 30 seconds.
- 5.The workstation supports ASUS Zenni Claw, an application designed to turn text prompts into multi-step agentic workflows such as news summaries and presentation slide generation.
Summary
Krish Naik introduces his custom desktop build, designated as an ASUS AI workstation and agent computer. He explains that while developer access to cloud application programming interfaces like OpenAI, Anthropic Claude, and Google Gemini is widespread, autonomous agent fleets running all day across multi-agent pipelines generate large token volumes that rapidly increase operational expenses. By shifting inference workloads to dedicated local hardware, developers can eliminate per-token charges and create a hybrid infrastructure that relies on cloud models only when necessary.
The video tours the physical build and its hardware components. The system is assembled around an ASUS ROG Maximus Z890 Hero motherboard and powered by an ASUS TUF Gaming 1000W Platinum power supply. Graphics processing and inference acceleration are provided by an ASUS ROG Strix GeForce RTX 5070 Ti equipped with 16 GB of GDDR7 memory. Thermal management is handled by an ASUS ROG Strix LC III 360 ARGB liquid cooler to ensure system stability under sustained computing loads. Naik notes that the setup can be customised through ASUS partners such as EliteHubs depending on specific model footprints and memory demands.
Transitioning to software, Naik demonstrates running local models directly on the operating system using Ollama. After verifying the installation and viewing available models such as Qwen 3.5 4B, Llama 3.2, and Gemma, he initiates a command-line interaction with Qwen 3.5 to illustrate its reasoning capabilities and response speed. He then uses Llama 3.2 to generate Python source code for a playable Tetris game. Windows Task Manager and the Nvidia System Management Interface confirm that the computational workload and memory consumption are handled on the graphics card rather than an external network.
To assess real-world viability, Naik launches a Streamlit benchmark tool called Local vs Cloud Agent Arena, developed using Claude Code. The arena puts a local agent powered by Qwen 3.5 4B against an OpenAI cloud model. In a math and demographic query asking for the population of Japan and a 17 percent calculation, the local agent finishes in 17.0 seconds across two tool calls at zero cost, whereas the cloud model takes 28.0 seconds across four tool calls and costs $0.0194. In a second evaluation covering current weekly artificial intelligence news announcements, the cloud agent exhausts its permitted step budget and fails, while the local model completes five tool calls in 29.8 seconds with zero token fees.
Finally, Naik demonstrates ASUS Zenni Claw, a desktop productivity software application designed to orchestrate agentic workflows using either local models like Gemma 4 12B or external cloud services. He prompts the system to compile an industry news digest, showing how the agent inspects source materials, triggers web search tools, and formats structured outputs. He also requests a slide presentation outline on agentic AI, prompting Zenni Claw to execute a PPTX generation skill, install necessary software packages, resolve script errors automatically, and output the finished deck. Naik concludes by emphasising that local workstations provide a viable foundation for hybrid AI architectures.
Hardware Specifications of the ASUS Agent Computer
Krish Naik details the components used in the workstation build, including an ASUS ROG Maximus Z890 Hero motherboard, an ASUS ROG Strix GeForce RTX 5070 Ti graphics card with 16 GB GDDR7 video memory, an ASUS TUF Gaming 1000W Platinum power supply, and an ROG Strix LC III 360 ARGB liquid cooler. The hardware configuration is specifically targeted at hosting local language models and handling inference workloads without relying exclusively on cloud servers.
Running Local Language Models with Ollama
Using Ollama through the command line interface, Naik runs interactive sessions with local open models, including Qwen 3.5 4B and Llama 3.2. He tests conversational reasoning speeds and prompts Llama 3.2 to write a complete Python implementation for a Tetris game. Real-time resource monitoring via Windows Task Manager and the Nvidia System Management Interface confirms the workload runs directly on the local GPU memory.
Benchmarking Local Inference Against Cloud APIs
Through a custom Streamlit web interface named Local vs Cloud Agent Arena, Naik benchmarks a local multi-agent setup against OpenAI cloud models. In a multi-step demographic calculation task for Japan, the local Qwen model completed the task in 17 seconds with two tool calls, compared to 28 seconds and four tool calls for the cloud API. In a subsequent live news retrieval test, the cloud agent exceeded the step threshold while the local agent finished in 29.8 seconds at zero token cost.
Productivity Workflows with ASUS Zenni Claw
The presentation demonstrates ASUS Zenni Claw, a software tool that connects to local or cloud models to automate everyday tasks into structured workflows. Naik shows how the program uses a local Gemma model to execute multi-step routines, including parsing tech news sources into a formatted digest and creating a multi-slide presentation deck using automated tool calls and dependency installation.
The Bottom Line
The video establishes that a desktop workstation equipped with consumer-grade hardware and 16 GB of video memory can reliably execute multi-step agentic workflows and local reasoning models. It demonstrates that local inference provides significant speed and cost advantages for routine tasks when compared directly with cloud application programming interfaces. While the tests highlight that local models operate without recurring token charges, the demonstration leaves open how larger models requiring substantial memory beyond 16 GB would perform on comparable single-GPU consumer systems.
FAQ
What is an ASUS AI agent computer and what is it used for?
An ASUS AI agent computer is a desktop workstation built with dedicated GPU hardware to run large language models and multi-agent artificial intelligence workflows locally without relying on paid cloud APIs.
Which graphics card is installed in the ASUS AI workstation shown in the video?
The workstation is equipped with an ASUS ROG Strix GeForce RTX 5070 Ti graphics card featuring 16 GB of GDDR7 video memory.
How does the local agent compare to the cloud model in execution time during the population calculation benchmark?
The local agent running Qwen 3.5 4B completed the Japan population calculation task in 17.0 seconds, whereas the cloud agent running GPT-5 took 28.0 seconds.
What software tool is used in the video to run and interact with local language models from the command prompt?
The video uses Ollama to download, manage, and run local language models like Qwen 3.5, Llama 3.2, and Gemma directly in the command prompt.
What is ASUS Zenni Claw and what functions does it perform during the demonstration?
ASUS Zenni Claw is an application that converts everyday user prompts into guided agentic workflows, demonstrated in the video generating an industry news summary and building a slide deck.
Worth watching for
Developers, AI engineers, and tech enthusiasts seeking to build dedicated desktop hardware for running local language models and multi-agent workflows.
- asus
- local-llm
- ollama
- agentic-ai
- nvidia-rtx