ASUS and Poesis say they have run autonomous AI agents with live market capital for a week, using an NVIDIA-powered deskside supercomputer rather than cloud infrastructure. The experiment matters less as a trading tip than as a test of whether agentic AI can research investments, manage risk and execute trades locally under predefined rules.
NVIDIA Watch analysis
What happened
The companies say the workflow ran on ASUS’s ExpertCenter Pro ET900N G3, built on NVIDIA’s DGX Station platform with a GB300 Grace Blackwell Ultra Desktop Superchip. ASUS says the machine can deliver up to 20 petaflops of AI performance and 748GB of coherent memory.
According to the announcement, the agents handled investment research, risk management and trade execution using live capital while remaining within mandates set in advance. The models, agents and workflows ran locally, which is the point of the demonstration: sensitive financial operations need not depend entirely on a remote AI service.
The release describes the project as a limited-duration proof of concept. It does not provide a performance record, a comparison with human traders or evidence that the system produced durable returns. The experiment is therefore a deployment demonstration, not a new investment strategy waiting for your pension.
Why it matters
Running agentic systems locally could appeal to financial firms that want more control over data, latency and operational access. It also shifts the difficult questions closer to the hardware: how are mandates enforced, how are trades checked, and what happens when an agent encounters a market event outside its assumptions?
NVIDIA benefits here as the supplier of the computing platform, while ASUS is positioning a large local machine as an enterprise AI product. The more interesting commercial test is whether this kind of system becomes useful beyond a carefully bounded showcase.
Our read
This is a credible sign that autonomous AI workflows are moving from chat windows into consequential software systems. It is not proof that AI traders can beat markets, and the companies do not claim that it is.
The sensible takeaway is narrower and more useful: local, high-memory systems can now host multi-agent workflows that touch live operations. Before anyone calls that progress, the industry should publish the controls, failure cases and full accounting of what the agents actually did. “It stayed within its mandate” is a starting line, not a trophy.
What to watch
- Whether ASUS or Poesis publish the experiment’s returns, costs, drawdowns and human-intervention record.
- How the agents’ predefined mandates were tested against unusual or rapidly changing market conditions.
- Whether other asset managers adopt local agent systems for research and execution.
- Whether regulators treat autonomous trade execution as a new operational risk category or an existing one wearing a shiny jacket.
Discussion spark: Should autonomous trading agents be judged first by their returns, or by whether their controls and failure records are independently auditable?
Sources and evidence
- ASUS and Poesis Team Up on Autonomous Trading Agents Powered by NVIDIA Technologies – Financial Times (17 September 2026, 23:10 UTC)
Independent WittyWires tracker for public updates about NVIDIA. Not affiliated with or endorsed by NVIDIA; this is not an official account.