Block launches Dessa local AI voice model

For investors, the key point is cost and deployment. Running on-device rather than in the cloud can reduce inference costs and ease latency, which is especially important for Block’s Square ecosystem, where voice tools are already being used to handle ordering for merchants. If Block can turn Dessa into a practical product, it could broaden its software mix and deepen engagement with small businesses without leaning as heavily on expensive external AI infrastructure.

Block said the model was trained entirely on human speech data using just a single rack of Nvidia GPUs, a detail that points to the company’s interest in leaner model development rather than brute-force scaling. That also keeps the story linked to the broader AI hardware trade, where efficient training and local inference are becoming a selling point as companies look for ways to control costs.

The company is not opening Dessa to the public yet and is instead inviting security researchers to test how the system could be used to mislead people, underscoring the reputational and regulatory risks around increasingly lifelike voice agents. That caution matters for enterprises weighing adoption, especially in customer-facing settings where authentication, fraud and disclosure rules are likely to tighten.

For Block, the launch is another move to turn AI into a direct product advantage inside Square and its broader financial services platform. For the market, it adds to the race among AI developers and chipmakers to make voice interfaces faster, cheaper and more convincing.