Sui’s programmable offchain tunnels reached 40,614,180 transactions per second during a live stress test on Wednesday at Sui Basecamp in Singapore, the layer-1 blockchain said, marking a sharp jump from the roughly 6.1 million TPS recorded in a similar exercise in July.

The figure reflects throughput inside Sui tunnels rather than raw onchain execution. Tunnels are programmable channels that process activity offchain and settle to the Sui blockchain when closed. During the demonstration, more than 10,000 tunnels were opened on Sui mainnet, with workloads spanning payments, games and chat applications.

Sui said the architecture is built for high-frequency activity, including interactions between AI agents, without forcing every transaction onto the base layer. The test was designed to simulate the kind of machine-scale demand that developers expect as autonomous software agents begin transacting with one another, buying data, paying for compute, coordinating tasks and settling micro-payments.

Key facts from the live test

• Peak throughput: 40,614,180 TPS.
• Previous test: about 6.1 million TPS in July.
• Location: Sui Basecamp in Singapore.
• Mainnet tunnels opened: more than 10,000.
• Workloads: payments, games and chat applications.
• Independent auditor: CertiK, with a report expected in the coming days.

What the 40.6 million TPS milestone actually shows

The headline number is eye-catching, but it measures a specific part of the transaction lifecycle. Sui tunnels are not the same as the Sui mainnet. They are programmable offchain channels that can absorb rapid-fire interactions between participants and then submit a consolidated settlement to the blockchain when the channel closes. That design allows the tunnel layer to handle repeated actions at machine speed while relying on Sui for final settlement and dispute resolution.

For AI-agent use cases, that distinction matters. An autonomous agent may need to pay for a stream of data, request a compute job, bid in an auction, send a chat message or execute a game action hundreds or thousands of times in a short window. Putting every one of those actions directly onchain would add latency and cost. Tunnels attempt to preserve the security and settlement guarantees of the underlying chain while moving the high-frequency traffic into a more flexible environment.

The live test put that idea under load. Sui said more than 10,000 tunnels were opened on mainnet during the exercise, and the activity was not limited to a single application category. Payments, games and chat workloads were all part of the demonstration, suggesting the network was testing a mix of transaction patterns rather than a narrow synthetic loop. The peak of 40,614,180 TPS came from that combined environment.

Why AI agents are driving the throughput race

The timing reflects a broader shift in blockchain planning. For years, capacity discussions focused on human users: how many payments a network could handle, how quickly a decentralized exchange could settle a trade or how cheaply a user could mint an asset. AI agents introduce a different demand profile. Software agents can operate continuously, respond to events in milliseconds and transact at volumes that humans would never generate.

That does not mean millions of AI agents are already transacting onchain every day. The test simulated high-frequency activity; it did not prove that a live market of autonomous agents already exists at that scale. But infrastructure providers are preparing for a future in which machine-to-machine payments, API metering, data access and coordination services become routine. In that environment, throughput is not only about human convenience. It becomes a prerequisite for applications that cannot function if every action waits for a block.

“Sui is built for the moment when millions of AI agents are transacting every day,” said Adeniyi Abiodun, co-founder and chief product officer at Mysten Labs, the original contributor to Sui. He added that no market in the world needs 6 million transactions per second, let alone 40 million, because people do not move that fast. Agents do.

The quote captures the core argument behind the test. Traditional payment networks are sized for human behavior. Even the busiest global card networks process transactions at rates that are orders of magnitude below the peak shown in the Sui tunnels demonstration. If agents begin to act as economic actors, they could generate demand patterns closer to web traffic than to human commerce.

How Sui tunnels fit into the network’s design

Sui has positioned itself around an object-centric data model and parallel execution, features intended to let independent transactions proceed without unnecessary contention. Tunnels extend that philosophy by allowing repeated interactions to occur in programmable offchain channels. When a tunnel closes, the final state is settled to Sui, reducing the number of onchain operations required for a long-running exchange.

The model has similarities to payment channels and state channels, but Sui describes its tunnels as programmable. That means developers can define rules for how participants interact, what conditions must be met before settlement and how disputes are handled. For AI agents, programmable rules could be used to enforce spending limits, verify service delivery, manage escrow or coordinate multi-agent workflows without each step becoming a separate onchain transaction.

There are trade-offs. Offchain systems must manage liveness, data availability and dispute resolution. If a participant disappears or a channel is not closed properly, the design needs mechanisms to protect funds and finalize state. The security properties of a tunnel depend on its implementation, the collateral backing it and the guarantees provided by the underlying chain. A throughput record does not answer those questions on its own.

That is why the audit matters. CertiK served as an independent auditor during the demonstration and is reviewing the test data, with a report expected in the coming days. Until that review is published, the 40.6 million TPS figure should be understood as a reported peak from a controlled live test, not as an independently verified production benchmark.

Competitive context and what comes next

The result adds to a wider capacity race across blockchains and scaling systems. Layer-1 networks, rollups and specialized execution environments are all trying to demonstrate that they can support applications that go beyond today’s decentralized finance and consumer crypto use cases. AI agents have become a useful framing for that effort because they imply sustained, high-frequency demand rather than occasional human clicks.

For Sui, the test is also a marketing and developer-recruitment signal. A throughput milestone gives builders a reason to experiment with agent-based payments, gaming economies and chat applications that might otherwise be constrained by onchain throughput. The fact that the demonstration involved more than 10,000 tunnels on mainnet, rather than only a closed test network, gives the result more operational relevance, though the audit will determine how much weight it carries.

The July test, which reached about 6.1 million TPS, established a baseline. The new peak is more than six times that figure. If the tunnel architecture can sustain similar performance under real-world conditions, it could offer a settlement model for applications that need to process many small actions without congesting the base layer. If the performance depends on tightly controlled test conditions, the result will be remembered as a proof of concept rather than a production capacity claim.

Sui’s broader argument is that the next wave of blockchain demand will not come only from people. It will come from software that transacts on behalf of users, applications and other software. Tunnels are one attempt to give that activity a place to happen without forcing every interaction through a global consensus process. The live test was designed to show that the model can handle extreme throughput while still settling to Sui.

Whether that vision translates into durable usage is a separate question. Throughput benchmarks can demonstrate engineering ambition, but they do not prove that developers will build applications at that scale or that users will pay for them. The AI-agent economy is still emerging, and the standards for identity, payments, reputation and dispute resolution are far from settled. Infrastructure that anticipates the demand can be useful, but it does not create the demand by itself.

CertiK is reviewing the test data, and a report is expected in the coming days.

Source:Cointelegraph News