Razer AVA Local vs Cloud: What Razer Has Confirmed About Where Its AI Runs

01.What Razer has confirmed

In March 2026, ahead of GDC, Razer announced that AVA would become an "agentic" assistant able to chain tasks across your apps, and that it would run locally on the PC and in the cloud. To split the work it introduced the Razer Inference Control Plane, which according to Razer "routes requests intelligently between local and cloud models" for lower latency and smoother multi-step tasks. We cover it in detail in our article on AVA as an agent.

Quyen Quach, Razer's vice president of software, explained it to Variety:

Three other official facts help put it in context:

• AVA needs a Windows PC. According to Razer's FAQ, it connects over USB-C, and that link lets "PC Vision Mode" analyze the screen with low latency.
• It uses third-party language models. Razer says AVA "currently utilizes state-of-the-art LLMs" and that its architecture is meant to support other platforms, including a dedicated Razer AI. At CES it ran on xAI's Grok. In The Verge interview, Min-Liang Tan said Razer hadn't decided which model it will use.
• The beta runs through Razer Cortex and Razer ID. Razer says beta data is wiped at the end of the program. The Early Access launch is planned for early 2027.

02.What Razer hasn't said

Nor should local processing be read as a privacy promise. Quach justified it on speed and cost, not privacy, and as long as the conversation goes through a cloud LLM, what you say will leave your PC. What a companion keeps from what you tell it, and how to delete it, is covered in our guide to AI companion memory.

03.AIKit and Akash: why they get mixed up with AVA

Razer AIKit is an open-source kit for developers who want to run and fine-tune models on their own GPUs. It runs in Docker inside WSL 2, bundles vLLM, Ray and LlamaFactory, and requires an NVIDIA GPU from the Volta generation or later. Razer hasn't presented it as AVA's local side. We explain it in our article on AIKit, Forge AI and Tenstorrent.

The only time Razer has used AIKit alongside AVA was AVA Mini, its April Fools joke. Behind it was a real image generator running on RTX 4090 and 5090 GPUs rented on Akash Network. According to Razer, between March 31 and April 4 it produced over 11,000 images, averaging 3.24 seconds and $0.01 per image, against the $0.03–0.15 charged by typical cloud APIs.

Those figures come from an image generator on rented servers, not from the assistant running on a user's PC. They say nothing about AVA's latency, the GPU it will need or where its servers will be.

04.Do you need an RTX 4090 to use Razer AVA?

There's no basis for saying so. For now, Razer only requires a Windows PC with a USB-C connection. If the local side ends up needing a dedicated GPU, Razer will have to publish the requirements. Until then, any specific card is a guess. What is confirmed is in Razer AVA's requirements.

05.Our take

Splitting work between the PC and the cloud makes sense: asking the time or turning the volume up shouldn't depend on a server, and handling that locally makes each answer cheaper. But Razer has described it as something it's testing, with no details, so we wouldn't treat it as a feature to count on when buying.

If you're deciding today, the safe assumption is that AVA will need internet to hold a conversation, because its brain is still a cloud LLM. Once Razer publishes the final spec sheet, we'd check three things before buying: which features work offline, what is processed on the PC, and whether there are GPU requirements.