Sui says its programmable offchain tunnels hit a peak of 40,614,180 transactions per second during a live stress test built to mimic high-frequency agent traffic. That is a big number, and it's Sui's own. The independent audit is still pending. The useful question for builders isn't "is 40M real?" It's "what does a tunnel actually count, and does my agent need one?"
What was claimed
Per Cointelegraph's report of Sui's announcement, the test ran on Wednesday at Sui Basecamp in Singapore. More than 10,000 tunnels were opened on Sui mainnet. Activity spanned payments, games and chat apps. The peak beat a previous Sui test from July that reached about 6.1 million TPS.
The record comes from Sui itself, via a Sui blog post and a post from the SUI Network account on X. Mysten Labs co-founder and CPO Adeniyi Abiodun framed the point this way: "Sui is built for the moment when millions of AI agents are transacting every day."
What a tunnel is
Tunnels, as described in the coverage, are programmable channels. Transactions are processed offchain between participants, then settled to the Sui blockchain when the tunnel closes. Sui pitches them for high-frequency activity, including agent-to-agent interactions, without pushing every transaction through the base layer.
If that sounds like a payment channel with a programmable twist, the shape is familiar. Open a channel, exchange many small updates, settle once. The appeal for agents is obvious: an agent paying another agent per API call, per inference or per tick doesn't want to wait on consensus each time.
The number, read carefully
Here is the part worth a raised eyebrow. 40.6M TPS is offchain tunnel throughput, not base-layer throughput. Settlement happens when tunnels close. So the figure isn't directly comparable to the onchain TPS numbers other networks publish. Compare it to those and you're comparing a channel's message rate to a chain's block rate. A headline number can be true and still answer a different question than the one you're asking.
What the coverage does not say matters too. It gives no details on how many agents took part, transaction sizes, test duration or hardware. Without those, 40.6M is a peak, not a sustained rate, and not a spec you can plan capacity around.
Claimed vs. verified
Here's the status as of Oct 7:
- ▸Claimed by Sui: a 40,614,180 TPS peak across 10,000+ tunnels on mainnet.
- ▸Pending: CertiK served as independent auditor during the demonstration and is reviewing the test data. A report is expected "in the coming days," per the article. It had not been published when Cointelegraph ran the story.
- ▸Not stated: agent count, transaction size, duration, hardware.
Until the report lands, "Sui hit 40.6M TPS" is best read as "Sui says it did, and an auditor is checking." Credit where due: naming an independent reviewer up front is a better habit than most TPS claims manage. Now we need to see what the review covers and publishes.
What it means for agent builders
Abiodun's argument, in the article, is that people don't transact at millions per second but agents might. That's a reasonable thesis, and it points at a real design question: where does your agent's payment traffic live?
- ▸Low-frequency, high-value: settle onchain. Channels add complexity you don't need.
- ▸High-frequency micropayments between known counterparties: a channel-style design starts to make sense. You trade instant updates for the need to open, hold and close state correctly.
- ▸Pay-per-call over HTTP: x402-style flows settle per request today. Whether channels make that cheaper is an open question worth testing, not assuming.
The prompt is the cute part. The payment rail is where the work lives. Whatever rail you pick, check the failure modes: what happens to funds in an open tunnel if a counterparty disappears, who can force-close, and what your wallet controls allow an agent to lock up.
What to watch
- ▸The CertiK report. Does it cover throughput, tunnel settlement correctness, or both? What scope does it state?
- ▸The missing details. Agent count, duration, transaction size. Sustained numbers beat peaks.
- ▸Real agent traffic. A stress test simulating agents is not agents paying for something they need. Watch for teams shipping on tunnels, not just benchmarking them.
- ▸Tooling. SDKs, wallet support and skills that let a builder open and settle a tunnel without hand-rolling channel logic.
The ambition is real, and it's a sensible one for agents. The receipt isn't in yet. Sato tracks evidence, not vibes: when the report lands, we'll look at what it says. Building an agent that moves value? Map the stack, wallets, payment rails and skills at satohub.ai/build.