The number arrived on a Tuesday morning in Singapore, wrapped in the polished cadence of a keynote: 40,614,180 transactions per second. By every measure of public relations, it was a triumph. Sui had broken its own July record by 6.1 million TPS, and the charts were immaculate โ clean ascending lines, the kind that make a conference hall lean forward in its seats. The applause, I have no doubt, was genuine.
And yet, reading the release from my desk in Sydney, the number would not sit still. It refused to mean what it appeared to mean. Somewhere between the comma-separated digits and the word "throughput," a category had been quietly exchanged โ the way a magician swaps a coin while you watch the wrong hand. The silence between the digits holds the truth. And in this case, that silence is measured in the distance between ten thousand channels and forty million transactions. It is a distance of several orders of magnitude, and almost no one reporting the story paused to measure it.
I have spent the last several years auditing the plumbing beneath financial systems โ first the internal risk models of an Australian bank, where I watched management fail to price an emerging asset, and now the settlement architecture of a central bank digital currency. That work instills one discipline above all others: always ask which layer the number describes. A throughput figure is meaningless until you know what is being pushed through it, and at what cost, and who is doing the counting.
Off-chain channels are not a new idea, though the marketing often presents them as one. Bitcoin's Lightning Network formalized the pattern years ago: two parties open a channel by committing funds to a shared on-chain contract, then exchange an arbitrary number of signed updates between themselves without ever touching the base layer. Only when the channel closes does the final balance settle to the main chain. Ethereum's Raiden, Celer, and a dozen state-channel projects extended the concept, each promising to move the world's transaction load into a private ledger shared by two participants and anchored, eventually, to a public one.
Sui's contribution โ and it is a contribution, if a modest and incremental one โ is programmability. Because Sui's smart contracts are written in Move, the channel logic can be generalized, composed, and pointed at applications rather than mere payments. The pressure test, according to Sui, covered payments, gaming, and chat: three of the highest-frequency, lowest-value, most latency-sensitive categories that exist. Mysten Labs co-founder and CPO Adeniyi Abiodun framed the exercise as groundwork for an economy of autonomous agents โ a network "built for millions of AI agents transacting daily."
That framing is seductive, and I want to be fair to it. Payments, micro-transactions inside games, and machine-to-machine messaging are precisely the workloads that choke a base layer. They are also precisely the workloads that channels were invented to absorb. So the choice of scenario is not arbitrary; it is coherent. The team knows its engineering, and the pedigree is real: Mysten Labs was founded by veterans of Meta's Diem and Novi projects, engineers who spent years trying to build a high-performance payment rail inside a corporate and regulatory straitjacket. When a Diem alumnus builds a channel, it is worth taking seriously.
But coherence of scenario is not the same as coherence of claim. And the claim โ 40.6 million TPS, presented in the grammar of a public blockchain's performance โ is where the trouble begins. To understand why, we have to separate two things the industry has spent years deliberately conflating.
Start with the mechanics of a channel update. When two parties exchange a signed state update inside an open channel, no consensus is reached. No block is produced. No validator votes. No global ordering is established. The update is a message โ cryptographically signed, yes, but a message between two endpoints that already agree on the channel's construction. Its speed is therefore bounded not by the consensus algorithm, not by block time, not by validator count, but by the hardware of the machines exchanging it and the bandwidth of the wire between them.
This is the heart of the matter. A channel's throughput ceiling is a property of computers, not of a blockchain. You can push a channel to ten million or forty million updates per second precisely because you have removed every constraint that makes a blockchain a blockchain โ the global agreement, the replication, the adversarial environment, the Byzantine fault tolerance. What remains is a very fast private ledger shared by two parties. That is a useful thing. It is not a public chain's performance, and presenting it as one is a category error dressed in the language of a record.
The correct comparison, then, is not Sui versus Visa, or Sui versus Solana. The correct comparison is Sui's channels versus Lightning versus Raiden versus a PostgreSQL table sitting behind an API gateway. Against those, forty million updates per second is unremarkable. A single modern server can sustain millions of simple state transitions per second in memory. The number is large because the problem was made small. Visa processes roughly sixty-five thousand transactions per second in production, under adversarial conditions, with chargebacks and fraud and settlement finality โ a figure that is smaller only because it is measured against a harder question. Comparing the two is like comparing the speed of a private conversation to the speed of a parliamentary vote. Of course the conversation is faster. It is not doing the same work.
Now look at where the real constraint lives โ and here the announcement inadvertently reveals its own answer. During the test, Sui states, more than ten thousand channels were opened on the mainnet. That figure is the tell. Ten thousand channels is the actual quantity of on-chain work the system performed. Everything else โ the forty million transactions โ happened in the spaces between those channels, invisible to consensus, invisible to the chain, invisible to the security model that supposedly protects the funds locked inside them.
Think about what opening and closing a channel actually requires. Each open is an on-chain transaction: it commits funds to a contract, registers the channel, and consumes block space and validator attention. Each close is another on-chain transaction, and โ critically โ a settlement event that must be honored by the base layer. If the test opened ten thousand channels, then the base layer processed on the order of ten thousand channel-management operations. The forty million figure describes updates inside those channels. The base layer never saw them. It counted to ten thousand and called it a record.
So the meaningful throughput question is not "how many updates can a channel carry?" It is "how fast can the base layer open and close channels, and what happens when many of them settle at once?" Sui did not answer that question. It answered the easy one.
This is not a trivial omission, and I say that from experience. Several years ago, while analyzing DeFi's explosive growth during the 2020 summer, I spent six months correlating stablecoin issuance against global M2 money supply. The conclusion I reached โ that much of DeFi's apparent activity was a reflection of fiat liquidity rather than the creation of new value โ was unpopular, but it came from asking where the actual constraint lay. The same instinct applies here. The dramatic number describes the unconstrained layer; the constrained layer is the one being kept quiet. When I published that paper, three crypto hedge funds cited it and almost no traditional finance peer did. The lesson was not that the crowd was wrong, but that the crowd prefers the number that flatters.
Consider the settlement risk directly. If ten thousand channels close in a short window โ say, because a game session ends, or a market event triggers a mass exit โ the base layer must absorb ten thousand settlements at once. Channel designs exist precisely because base layers cannot cheaply handle every transaction; the entire premise is that settlement is deferred and batched. But deferral concentrates risk in time. A pressure test that opens channels, runs millions of off-chain updates, and then quietly settles them is testing the easy half of the system. The hard half is the stampede at closing time, and the announcement does not mention it.
This is precisely the problem I worked on last year, advising on the design of a digital currency for a central bank. The tempting architecture was hybrid: let retail transactions settle on a Layer-2 solution to reduce energy and cost, and let the base ledger handle only finality. It looked elegant on a whiteboard. It also created exactly the deferred-settlement exposure I am describing โ a hundred thousand small interactions compressed into a handful of settlement events, all arriving under stress, all demanding the base layer's undivided attention. We spent more time modeling the closing stampede than modeling the steady state, because the steady state is the part that flatters and the stampede is the part that breaks.
A credible benchmark would disclose the topology. How many nodes participated, and where were they located? Were they co-located in a single data center, where latency approaches zero and bandwidth is effectively infinite, or distributed across continents under realistic network jitter? Was the load adversarial โ did anyone attempt to spam the channels, force disputes, or trigger mass exits? Was the transaction mix representative, or synthetic and uniform? None of these parameters appear in the announcement, and without them the figure is not a measurement so much as an assertion. I have audited systems where an honest benchmark produced a number one-tenth the size of the marketing figure โ not because the engineers lied, but because the marketing number was generated under conditions no real user would ever experience.
There is a second, subtler omission: the trust model of the channels themselves. Off-chain channels require either monitoring โ someone must watch the chain for fraudulent close attempts, as Lightning demands โ or trusted operators. The announcement does not disclose how disputes are resolved, who custodies funds mid-channel, or whether a channel operator can become a single point of failure. Structure cannot contain the chaos of human hope, but neither can it ignore the ordinary adversarial behavior that channels invite. Until those mechanisms are audited, the forty million figure rests on an architecture whose failure modes remain entirely unexamined.
Which brings us to verification โ the third and most uncomfortable gap. Every number in this announcement is self-reported. The independent audit from CertiK, we are told, is "expected in the coming days." In other words, at the moment of maximum publicity, the data carries no third-party validation whatsoever. We do not know the node count, the geographic distribution, the hardware specification, or whether the transactions were synthetic or real. A pressure test conducted under idealized conditions โ co-located nodes, generous hardware, uniform load, no adversaries โ can produce almost any number its designers wish. We measured the shadow, mistaking it for the form.
I have seen this pattern before, from the other side of the table. In 2017, while auditing cross-border liquidity risk models at a Sydney bank, I filed a report arguing that our regulatory capital requirements failed to account for the emergent volatility of Bitcoin, then trading above fifteen thousand dollars. Management dismissed the finding: crypto was a novelty, not a macro force. The lesson I took was not that I was right โ though I was โ but that institutions systematically mistake the convenient metric for the meaningful one. A bank measures capital against the assets it chooses to see. A blockchain measures throughput against the layer it chooses to count. The psychology is identical, and it is not confined to either side of the crypto divide.
And then there is the question no one at the keynote asked, because it is impolite: does the success of off-chain channels actually help the SUI token? Follow the value. Channel transactions, by design, do not consume on-chain gas. They do not pay base-layer fees. They do not reward validators. If Sui succeeds in pushing the world's high-frequency activity off-chain, the demand for on-chain block space โ and therefore for the token that prices it โ may fall, not rise. This is the value-capture paradox: technical triumph at the settlement layer can hollow out the economic layer that funds it. The more efficient the escape hatch, the less traffic remains to pay the toll.
This is not a prediction; it is a structural question the announcement leaves entirely open. If channels carry fees, who collects them? Do they accrue to SUI holders, to validators, to channel operators, to no one? The report is silent. A network can be blazing fast and its token can be worthless; those two facts are not in tension โ they are, historically, frequently correlated. Liquidity is a ghost that haunts the ledger, and so is value that moves where the ledger cannot see it.
The consensus reading of this announcement is that Sui has demonstrated a leap in capability, and that the AI-agent economy is now within reach. I want to invert that. The most likely outcome is that this becomes a footnote โ a number remembered by no one outside a conference hall, and a capability that changes nothing about Sui's competitive position.
Here is why. Performance records in this industry are not engineering milestones; they are marketing cycles. Solana did it. Every Layer-2 rollout did it. The pattern is ritualized: announce a figure that cannot be compared to anything, let the press repeat it without a denominator, and let the community treat the resulting ambiguity as achievement. We built castles on the tidal data of sentiment. The tide recedes; the castles survive in the retelling but not in the fundamentals. The choice to unveil the number at Sui Basecamp โ an ecosystem gathering โ confirms the audience. This was built to reassure builders and holders, not to inform analysts.
The deeper contrarian point is that the AI-agent framing is a wager on demand that does not yet exist. There is no large-scale economy of autonomous agents transacting daily. Building infrastructure for it is not wrong โ someone must build first โ but it is unverifiable. You cannot falsify a claim about a market that has not arrived. And when a narrative depends on a future that never quite materializes on schedule, the infrastructure sits idle and the token drifts toward its fundamentals, which are quiet. The record from July, broken again in the same manner, suggests a rhythm: a performance announcement every few months, each one a little larger than the last, each one measuring the same unconstrained layer.
The truly counterintuitive conclusion is this: a blockchain that can push transactions off its own ledger has solved a scaling problem by refusing to solve it. The achievement belongs to the channel, not the chain. Sui's base layer is no faster today than it was in July. Only the place where the counting happens has moved.
So what should a careful observer watch, rather than applaud? Three signals. First, the CertiK report: if it discounts the figure or discloses idealized test conditions, expect the narrative to reverse within days. Second, the settlement stress test: observe what happens when ten thousand channels close together, because that is where the base layer's real limit lives, and it is the test Sui has not yet run. Third, and most telling, the SUI token's own reaction over the following seventy-two hours โ if the market prices the news at all, it will tell you whether anyone believes the number means what it says.
The question worth holding is not how many transactions per second a channel can carry. It is whether a chain that hides its traffic off-ledger can still capture the value of moving it. The archive remembers what the algorithm forgets โ and the archive, in this case, holds a rather different figure than forty million. It holds ten thousand, and it is asking who pays.