A test of local artificial intelligence on a $7,000 Asus ProArt P16 RTX Spark generated 220,202 tokens in just under 13 minutes. The run used OpenAI’s 120-billion-parameter gpt-oss model to examine 403 candidates as part of a broader research project.
Instead of using a chatbot for individual questions, the experiment had the model process hundreds of potential data points, assess them in different ways, challenge its own conclusions and produce extensive analysis. Because the model ran entirely on the laptop, the additional run did not create a separate API charge.
That difference changed how the testing was approached. With cloud-based AI, each additional experiment raises questions about token costs, including whether to analyze every candidate, add another challenger or repeat the process when information changes. Running locally made it easier to conduct broader searches, repeat analyses and test ideas that might not produce useful results.
The experience also exposed the costs that remain associated with local AI. The laptop itself is expensive, and electricity and testing time are not free. Anyone buying hardware to avoid cloud AI bills still needs to account for those expenses.
The main benefit identified in the test was not simply replacing cloud workloads with a cheaper local option. Once the computer had already been purchased, the incremental cost of launching another large analysis became relatively small. That made it practical to attempt computations that might otherwise be abandoned because of per-token charges.
The approach does not eliminate the value of efficiency. Faster models can save time and power, while smaller models may be more practical for some tasks. Producing large volumes of AI output also does not guarantee useful results. However, the testing suggests that local AI’s broader advantage may be enabling experiments and repeated analyses that users would not consider if every token had to be paid for separately.
Sources: PCWorld.