Forty million six hundred thousand transactions per second.
I read that at 2:14 a.m., and my first instinct wasn't to tweet it. It was to open a terminal. Numbers that large don't describe blockchains. They describe press strategies. I have watched this movie three times now — the Shanghai withdrawal queue in May 2023, the Solana congestion panic in February 2023, the Arbitrum Nitro latency runs in July 2023. Every one of them taught me the same lesson: the interesting part of a performance claim is never the number itself. It is the methodology the number is hiding.
So let's do the boring part. Let's talk about what was actually measured, by whom, and under what conditions.
The claim, stripped to its bones. Sui says its "programmable offchain tunnels" reached 40.6M TPS during a live stress test. The test simulated high-frequency activity between AI agents. That is the entire disclosure. Four information points, every one of them sourced to Sui itself. No third-party reproduction. No independent audit. No onchain transaction hash a stranger like me can pull up and verify.
That sourcing detail is not a footnote. It is the spine of everything below. When a single party is the sole witness to its own record, you do not discard the claim — you discount it, and you keep the discount visible in every sentence that follows.
Context: what a "tunnel" actually is. The word is doing a lot of unearned work. Sui never defines it. So I have to infer, and I will flag that I am inferring. Take the three adjectives the announcement leans on — "programmable," "offchain," "high-frequency" — and the closest industry analogue is a state channel or a payment channel. Two or more parties open a channel off the main chain, exchange transactions as fast as their hardware allows, and settle only the net final state back to L1.
This is not new technology. Bitcoin's Lightning Network shipped the concept years ago. Ethereum's state channel research, Raiden, the entire lineage — a decade of engineering, most of it focused on exactly the two problems I will raise later. What Sui has wrapped around the channel is a "programmable" layer plus an AI-agent use case. That is differentiation. It is not a paradigm shift, and it should not be priced like one.
I want to be fair to Sui's base layer, because it is genuinely good. Move is a well-designed language with real safety properties. The object-centric data model is an architectural choice, not a marketing slide. Sui's consensus design is fast by construction. None of that is in dispute. But the tunnel announcement is not about the base layer. It is about a test that ran alongside it, in a controlled environment, and produced a number.
A promise the channel model keeps making and keeps breaking. It is worth remembering how this story usually ends, because the ending is old. Lightning Network promised instant, near-free Bitcoin payments and delivered a network that works beautifully for large, well-capitalized nodes and poorly for everyone else. Raiden promised Ethereum scalability and quietly faded. The recurring failure mode is never the raw speed — channels are fast. The failure mode is routing, liquidity fragmentation, and the capital cost of keeping funds locked in the right place at the right time. Every channel-based scaling attempt has collided with the same wall. Sui has not shown a reason why this time is different. It has shown a number.
Core: the arithmetic nobody published. Here is the forensic math that dissolves 40.6M.
Offchain throughput is a product, not a measurement. It equals the number of parallel channels multiplied by the per-channel processing rate. A single channel, in most implementations, clears tens of thousands of transactions per second — that is a bandwidth and hardware ceiling, and it is roughly fixed. To reach 40.6 million, you need on the order of a thousand or more channels running concurrently at peak. That is entirely achievable. It is also entirely meaningless as a comparison to L1.
Why? Because offchain channels do not consume L1 resources. They do not compete for block space. They do not pay gas per interaction. That is the whole point of them. So their aggregate TPS can be inflated without limit simply by opening more channels — the figure scales with the size of the test harness, not with the capacity of the network. Comparing "40.6M offchain TPS" to "Solana's thousands of onchain TPS" is like comparing the throughput of a thousand private phone calls to the throughput of one public radio station. Different substrate, different physics, different claim.
The three parameters that would make this real. An honest engineering disclosure would ship three numbers alongside the headline. Sui published none of them, and the absence is the actual finding.
Start with the channel count. Without it, 40.6M is uninterpretable — it could be ten thousand channels doing forty thousand each, or a hundred thousand channels doing four hundred each. Those are completely different engineering stories.
Then the settlement model. Was value genuinely moving across channels, or did each channel run as a closed loop? A closed loop is a benchmark artifact; cross-channel settlement is the hard, unsolved, interesting problem.
Then the bottleneck everyone avoids: the final settlement latency and cost back to L1. This is where every offchain scaling design lives or dies. If you can clear 40.6M interactions offchain but it takes an hour and a fortune in gas to settle the net result onchain, you have not scaled anything. You have moved the queue and hidden it. The silence around these three numbers is more informative than the number itself.
The settlement bottleneck is the whole game. Let me dwell on that third parameter, because it is the one the industry systematically under-weights. An offchain system is a promise to settle later. The value of that promise depends entirely on how cheap and how fast "later" is. If final settlement is expensive, channels become a way to defer costs, not eliminate them — you just batch the pain. If settlement is slow, the channels carry counterparty risk for the duration, and capital sits locked and idle. Lightning solved throughput and got stuck on exactly this: the liquidity to route payments was expensive to hold and hard to predict. A 40.6M TPS test that never discloses settlement latency and cost is a car commercial that never mentions the price.
How I would actually verify this claim. Since Sui won't, here is the test I would run — and I have run versions of it. I would stand up a controlled harness, open a defined number of channels, and measure per-channel throughput in isolation to establish the baseline. Then I would force cross-channel settlement under load and watch for the routing failures that always appear. Then I would batch the net results back to L1 and measure the real cost and latency per settled unit. Raw logs, published, reproducible by anyone. That is what I did for the Nitro migration — a thousand transactions, before and after, charts and gas comparisons, no adjectives required. When I listened to the Shanghai withdrawal contracts in May 2023, I didn't trust the aggregators; I cross-referenced raw block data against gas-price spikes. A 42-second arbitrage window was the reward for not taking anyone's word for it. That discipline is exactly what is missing here.
Synthetic load is not production load. The test simulated AI-agent activity. Read that slowly. It was a synthetic benchmark — a script written to exercise the exact path the engineers optimized for. I have built these. In July 2023 I stood up a latency harness for the Arbitrum Nitro migration and fired a thousand transactions before and after the upgrade to measure finality time. That harness told me something real about a controlled variable. It told me nothing about what happens when ten thousand unrelated wallets compete for the same sequencer during an airdrop.
Synthetic tests measure the ceiling under controlled conditions. They systematically avoid the adversarial conditions that define production: malicious participants, network partitions, settlement congestion, griefing, the ordinary chaos of open access. Sui's test almost certainly never saw any of it. The word "live" in "live stress test" is doing heavy lifting — live means "running," not "open to the public under real stakes."
The tradeoff the announcement skipped. State channels are not free. They buy throughput by selling two things: composability and capital efficiency.
Composability is the first casualty. Assets locked inside a channel cannot interact with external DeFi protocols. Your money sits in a room with the door closed until settlement. Capital efficiency is the second. Participants must lock funds to open the channel, and that capital sits idle by definition, earning nothing, doing nothing, until the channel closes. This is the exact cost Lightning has wrestled with for years. It is why channel-based scaling has never captured the general-purpose compute market despite a decade of effort.
Sui's framing implies these costs do not exist. They do. They are structural, not incidental, and they are the reason a high-frequency, directed, small-value transfer between two agents is the perfect channel use case — precisely because it needs none of the things channels sacrifice.
Where the token story quietly breaks. Everyone reads 40.6M as bullish for SUI. I read it as neutral-to-bearish for the token's demand curve, and here is the chain.
SUI's value capture depends on onchain gas consumption and staking. Offchain tunnels exist for one purpose: to reduce onchain transactions. If the tunnels work as advertised and get adopted, they remove gas demand from L1. The success of the scaling technology is, in a narrow and uncomfortable sense, a headwind for the token. That is not a fatal flaw — it is a structural tension every L1 scaling strategy carries. But you will not find it in the announcement, because the announcement only sells the "TPS goes up" half of the equation.
I have seen this reflex before, and it wears a different costume every cycle. It is the same reflex that turns liquidity mining into a TVL subsidy — a metric engineered to look like traction, funded until the incentives stop, at which point the "users" evaporate overnight. The number was never measuring demand. It was measuring the subsidy. A synthetic TPS benchmark measures the harness, not the network. Same category of illusion, different outfit.
The honest classification of this milestone is "ecosystem capability proof" — a narrative asset, not a value-capture event. It does not create token demand. It does not route fees to holders. It creates a slide for a pitch deck and a headline for a bull market.
The competitive frame, honestly drawn. Let me put the comparison on the table with a warning welded to it.
Sui, by this claim: 40.6M offchain, aggregated, synthetic — comparable to nothing.
Solana: thousands of onchain TPS, real, verified, backed by a mature and heavily used ecosystem.
Ethereum L2s, both ZK and Optimistic: thousands to tens of thousands of onchain aggregate, secured by Ethereum's settlement guarantees, holding the deepest TVL in the market.
Placing Sui's offchain figure in the same column as those onchain figures is not a comparison. It is a category error with a chart attached. I show the table only to demonstrate the dimensional mismatch, not to rank anyone. The instant you flatten offchain and onchain into one "TPS" column, you have laundered a marketing claim into a fact.
The contrarian angle: this is a cognition risk, not a technology risk. The dangerous thing about 40.6M TPS is not that it is false. It may be perfectly true — offchain. The danger is that it is built to be misread. Headlines will say "Sui hits 40.6M TPS." Readers will file it next to Solana and Ethereum and conclude Sui is a hundred times faster. That conclusion is wrong, and the announcement is structured to encourage it without ever stating it.
I have spent eleven years watching this exact maneuver. In February 2023, when Solana stalled, the feeds screamed that Solana was dead. I pulled validator logs through a private RPC endpoint and found congestion from a failing validator cluster — not a consensus failure. The panic narrative and the technical reality were two different documents. The same gap exists here. The reality is "a stress test hit a synthetic offchain throughput." The narrative is "the fastest chain on earth." Those are not the same claim, and the distance between them is exactly where retail capital gets extracted.
The self-report problem. One more forensic point, and it is the one I keep circling back to. The sole source is Sui. In my FTX work, I learned to treat self-reported numbers the way a prosecutor treats a defendant's alibi — as a starting hypothesis, never a conclusion. I traced $2.1 billion in Alameda flows not by reading statements but by following onchain transfers across wallets. The chain does not lie. Announcements do, or at least they select.
Here, there is no chain to follow. The 40.6M happened offchain, in a test environment, disclosed by the party that benefits most from the disclosure. That does not make it a lie. It makes it unverified, and unverified numbers deserve a discount — a steep one, applied without apology. When the source field of every data point reads "Sui," you are not reading a report. You are reading a press release with a chart.
The AI-agent narrative is the real product. Strip away the TPS and ask what Sui is actually selling. The answer is positioning. Sui wants to be the AI-agent-native chain — the place where autonomous agents transact at machine speed. That is why the test simulated AI agents and not NFT mints or DEX swaps. The AI-agent frame is the most transmissible narrative in the market right now, and pairing it with a shocking number is a deliberate double-amplification: a story plus a statistic, each making the other louder.
The demand side of that story is unproven. There is no economy of AI agents transacting onchain at scale today. I built a prototype last year that wired an LLM to a multi-sig wallet to run a simple DeFi strategy — impressive, fragile, and nowhere near a market. The 40.6M TPS is a supply-side claim. It says "we could handle it." It says nothing about whether anyone needs it. Capability is not demand. A benchmark is not a market.
The DeFi blind spot. Here is a downstream consequence almost nobody is discussing. If offchain tunnels succeed, they pull activity offchain, and offchain activity is invisible to DeFi composability. The more successful the tunnel, the more onchain liquidity gets bypassed. GameFi and micro-payments benefit — they need exactly the high-frequency, low-value, low-latency profile tunnels provide. But general-purpose DeFi, which generates the deepest liquidity and the most fees, is precisely what channels cannot serve. This is a double-edged technology being sold as a one-directional win.
Regulation, briefly, because it is coming. A pure performance claim does not touch securities law. But the AI-agent framing opens a door nobody is standing in yet. If autonomous agents begin executing high-frequency financial strategies without human sign-off, you create a genuine regulatory vacuum: who is liable for an agent's trades, and how do you run AML surveillance on a machine that never sleeps and never files a report? Today's compliance regimes are already theater — a determined wallet hops through a few addresses and the KYC wall becomes a suggestion. Point that same weakness at autonomous agents and the monitoring gap stops being a nuisance and becomes structural. That is a future problem, not today's. But the test Sui ran is the first rough sketch of the world that creates it.
What I would watch instead of the number. Forget 40.6M. It is a dead-end metric. Here is the signal set that would actually tell me whether any of this is real.
Real onchain TPS on the Sui explorer — not offchain, onchain, and sustained over weeks, not a single block.
A third-party reproduction of the benchmark, ideally with the three missing parameters disclosed.
A published technical spec and open-source code for the tunnel, so the channel count and settlement model can be inspected by people who are not on Sui's payroll.
Genuine adoption: high-frequency applications deploying on Sui, with measurable transaction volume that is not subsidized by incentives.
And the tell that always matters — whether competing L1s rush out their own astronomical TPS numbers within the next quarter. If they do, you will know this was a marketing salvo, not an engineering milestone. Arms races are about narrative share, not capability.
What a credible milestone would look like. For contrast, imagine the announcement Sui did not make. It would read something like this: "We opened 4,200 channels; each sustained a measured 9,600 TPS under adversarial load; cross-channel settlement cleared 98% of transfers; net settlement back to L1 cost a median of 0.0004 SUI and finalized in 1.8 seconds; raw logs and the harness are open source." That is a milestone. It has numbers that can be checked, parameters that can be attacked, and a methodology that can be reproduced. What we got instead was a single aggregate figure with no denominator. The difference between those two announcements is the difference between engineering and advertising.
The takeaway. 40.6M TPS is a number that is probably true and almost certainly misleading. It is an offchain aggregate from a synthetic test, self-reported by the project that profits from the headline, with the three parameters that would make it meaningful left unpublished. That is not a scandal. It is a marketing event wearing an engineering costume, and the costume is tailored for a bull market that rewards shock over substance.
The thing I actually take from this is smaller and colder: the TPS war has entered its desensitization phase. Numbers are getting bigger and meaning less. The market is beginning to price these claims near zero, and it should. When a benchmark this large produces this little verifiable information, the correct response is not excitement. It is a terminal window and one question: what did you actually measure, and who checked?
Sui may well be the AI-agent chain it is auditioning to be. But you will not learn that from a stress test. You will learn it from onchain traffic that shows up whether or not anyone is watching the announcement.