Sui's programmable offchain tunnels processed 40,614,180 transactions per second during a live stress test on Wednesday, a roughly sevenfold increase over the 6.1 million TPS the layer-1 blockchain recorded in a similar test in July.

The demonstration took place at Sui Basecamp in Singapore and simulated high-frequency activity between AI agents, according to the network. Transactions ran through Sui tunnels — programmable channels that settle to the Sui blockchain only when closed, keeping individual messages off the base layer.

More than 10,000 tunnels were opened on Sui mainnet during the test, carrying activity across payments, gaming and chat applications. The network said the architecture targets workloads where agents transact continuously and where pushing every message onchain would be operationally impractical.

What did the test actually measure?

The headline figure covers throughput inside the tunnels rather than base-layer finality. Tunnels let counterparties transact offchain and post a single settlement to Sui at close, meaning the 40.6 million TPS number reflects aggregate message volume across the open channels. The July benchmark, which peaked near 6.1 million TPS, used the same construction, giving a like-for-like comparison between the two runs.

CertiK served as an independent auditor for the demonstration and is reviewing the test data. A report is expected in the coming days and will determine whether third-party verification supports the claimed throughput.

Why is Mysten Labs positioning this around AI agents?

Adeniyi Abiodun, co-founder and chief product officer at Mysten Labs, the original contributor to Sui, framed the result as preparation for machine-driven transaction volume rather than human payments.

"Sui is built for the moment when millions of AI agents are transacting every day," Abiodun said, adding: "No market in the world needs 6 million transactions per second, let alone 40 million: people just don't move that fast. Agents do."

The operational claim matters for infrastructure buyers. Agent-to-agent workloads — settlement between automated services, in-game economies, machine-to-machine payments — generate transaction patterns that overwhelm conventional onchain execution. Offchain channels with periodic settlement shift the cost structure away from per-transaction fees and toward channel management, an approach with clear trade-offs in liveness assumptions and dispute handling.

What comes next?

The test result is a benchmark, not a production deployment at scale. CertiK's forthcoming report on the test data represents the next verification checkpoint, and the tunnels' viability for sustained agent-driven workloads will depend on real applications adopting the channel model beyond demonstration conditions.