Ten Thousand TPS, Three Point Two Billion in Testnet TVL, and the One Number Nobody Published

The Number in the Deck and the Number Missing From It

There is a number in the deck and a number missing from the deck. The number in the deck is $3.2 billion — testnet total value locked, reported after an incentive campaign, screenshotted, and reproduced across a dozen newsletters inside forty-eight hours. The number missing is retention. No active address count. No seven-day return rate. No cohort curve. No median deposit size. No split between the recursive capital that loops through three lending markets to farm points and the capital that actually intends to transact.

The rest of the stack reads the way these stacks always read. A mainnet date in Q2. A throughput claim of 10,000 TPS. A commitment to EIP-4844 adaptation. Eighteen people on the team, three of them out of the Ethereum Foundation research pipeline. A $20 million Series A led by Paradigm. A release plan for the token: team locked twelve months, then thirty-six months linear; early investors locked six months, then eighteen months linear.

We didn't read the press release twice. We read the release schedule once, and then we opened the contracts.

That habit is the whole method. Silence is the loudest audit trail in the market. Everything a launch announcement omits is a decision someone made with a spreadsheet open, and the omissions cluster in the same three places every time: the cost side of the network, the composition of the demand, and the locus of control.

What follows is an audit of the announcement archetype, using this one as the specimen. Not a verdict on a team. A schema for reading the next twenty of these, because there will be twenty of them before the year turns.

What an L2 Announcement Is Actually Made Of

An L2 launch is not a single document. It is a stack of four, and each one has a different author, a different audience, and a different tolerance for precision.

The blog post is written by communications and reviewed by engineering. It carries the numbers that sound large. The pitch deck is written by the founders and aimed at funds. It carries the numbers that sound defensible. The technical documentation is written by engineers and aimed at other engineers. It carries the numbers that are true, buried in the middle of a page nobody links to. The roadmap is written by nobody in particular and is the only document with no author to hold accountable when a quarter slips.

Compare the four and you get the delta. The delta is the announcement's real content.

For context on why the sequencer question dominates this category: an L2 is not a smaller blockchain. It is a settlement relationship. Execution happens somewhere, data is posted somewhere, and disputes or proofs are resolved somewhere. The interesting part of the design space moved, over roughly four years, from 'can we execute off-chain and settle on-chain' to 'who gets to decide the order of transactions, and what happens when that party misbehaves.' The first question is solved engineering. The second is a governance problem wearing engineering clothes.

That is the frontier this class of project claims to be attacking. The claim is plausible. The claim is also, in almost every roadmap I have read since 2023, a research program with a date stapled to it.

10,000 TPS Is a Claim About Ethereum, Not About the Rollup

Start with the headline number. Throughput without a workload definition is not a measurement. It is a unit-free adjective. 10,000 of what? A token transfer is not a swap. A swap is not a proof verification. A proof verification is not a state-diff update on a large contract with heavy storage.

The number that actually constrains a ZK-rollup is not its prover's throughput. It is the aggregate data budget of the layer it settles to.

Post-EIP-4844, blobs are the settlement medium. Each blob is 128 KB. The protocol targets three per block with a ceiling of six, on a twelve-second slot. Do the arithmetic and the entire network — every rollup on Ethereum, all of them, competing — has a maximum of roughly 786 KB of blobs per block, which is on the order of 64 KB per second at the theoretical ceiling, and about half that at the target. That is a shared resource. It is not owned by anyone, and it is priced by a fee market that is separate from the execution base fee.

A 10,000 TPS claim is therefore not a statement about a prover. It is a statement about how much of Ethereum's blob space this one rollup expects to consume, and at what price, while every competitor stays asleep.

Compression changes the arithmetic. ZK-rollups post state diffs rather than raw calldata, and a well-built one can push a large effective transaction count through a compressed blob. So the ceiling is not sixty transactions per second. It is a number that moves with the workload, the compression ratio, and the fee the operator is willing to pay.

Which means the honest unit is not TPS. It is dollars per transaction at a stated blob base fee. Publish that curve and you have published a business. Publish a TPS number and you have published a slogan.

Second constraint, equally ignored: latency and throughput trade against each other in the proving pipeline. Cutting a proof every twelve seconds for small batches costs more per transaction than proving every ten minutes for a large batch. But ten-minute finality is unusable for a bridge and unpleasant for a DEX. So the operator sits on a moving optimum, and the optimum moves every time the fee market moves.

$3.2 Billion on a Testnet Is a Measurement of Scripts

Now the demand number. Testnet tokens are free. A testnet TVL figure is a count of scripts that ran, weighted by the size of the balances the scripts were handed. It is not a dollar figure in any economic sense, and everyone in the industry knows this, and everyone repeats it anyway.

I have run this decomposition before, at a smaller scale, with real money. In the 2020 DeFi Summer I put $50,000 of my own capital into Uniswap V2 and Curve, not to trade but to instrument. I wrote Python harnesses to backtest liquidity provision across volatile pairs, and the finding that stuck with me was that a disciplined rebalancing rule cut impermanent loss by roughly 15% in volatile pairs. The number is not the point. The point is who was running the script. Provision that survives is provision that is automated, and automated provision exits the instant the spread between emissions and the cost of capital plus gas inverts.

Apply the same lens. Three questions decompose any testnet figure:

Recursive share. How much of that $3.2 billion is the same capital cycling through a lending market, a vault, and back into the same deposit contract to multiply points per dollar? In my experience with incentivized campaigns, this share routinely exceeds half, and it is trivially detectable on-chain by looking at deposit addresses with high in-degree from protocol contracts.

Withdrawal half-life. After the last emission epoch, what fraction of the balance remains after seven days, thirty days, ninety days? A campaign that retains 12% at ninety days produced a smaller network than a campaign with one-tenth the TVL and 60% retention. Nobody publishes this because the retention number is always worse than the TVL number.

Distribution. A Gini coefficient over depositor balances, or simply the top-ten depositor share. A network with $3.2 billion concentrated in twenty sybil clusters is not a network. It is a customer list with twenty names on it.

Real TVL is roughly headline TVL multiplied by (1 minus recursive share) multiplied by the persistence factor. Every input to that formula was available to the team on day one. None of it was in the post.

The Sequencer Is the Last Centralized Machine in the Room

Here is where the announcement spends its most confident adjectives and its least specific nouns.

The sequencer orders transactions. That is the job. Ordering is not a commodity function; it determines who gets filled first, who gets sandwiched, who gets censored, and who captures the extractable value that ordering creates. A single-operator sequencer is a single point of censorship, a single point of MEV capture, and a single point of liveness failure. Every rollup shipping today has one.

Decentralizing it splits into three sub-problems that are usually collapsed into one in marketing copy.

Leader election. Some process must pick who orders the next batch, and that process must be fast, because users will not accept a swap that takes two seconds to feel settled. Consensus among N validators is slower than a single node with a database. The entire difficulty is in that sentence.

Ordering fairness. Even with many sequencers, ordering rights concentrate. You can measure this in real time: take the sequencer address set and compute the Gini of blocks produced. If four operators produce 90% of blocks, you have an oligopoly wearing a quorum.

Verifiability. The strongest version of the claim is not that many parties order, but that a single party can order and still be forced to include, and can be proven to have ordered correctly. That is the escape hatch: a force-inclusion path on L1, with a defined delay parameter and a bonded cost to the operator. If a user can post to L1 and have the transaction included within a bounded window, then the sequencer is censurable in practice but not in principle. That is a meaningful property. It is also the exact property the announcement did not describe.

So when a Q2 roadmap lists 'decentralized sequencer,' read it as a research agenda, not a delivery milestone. The honest formulation, the one I have seen exactly twice, reads: a permissioned operator set secured by a threshold signature, with a published path to permissionless participation and a forced-inclusion fallback live at launch. If the word 'decentralized' appears without an operator count, a key policy, and a delay parameter, it is decoration.

The Prover Is Where the Business Model Lives or Dies

This is the section that never makes the newsletter, and it is the only one that determines whether the network exists in three years.

A ZK-rollup operator has four cost lines. Proving compute. Data availability in blobs. L1 settlement — the gas to verify the proof, which lands in the low hundreds of thousands of gas per batch depending on the proof system. And operations. Revenue is L2 execution fees plus whatever the token subsidy adds.

The structural problem is that both the cost side and the revenue side are denominated in a market the operator does not control. Blob fees are set by aggregate demand from every rollup. L1 verification gas is set by the same base fee that prices everything else. And the revenue per transaction is measured in fractions of a cent, because that is the product: cheap execution.

A ZK-rollup is a business that buys wholesale compute and sells retail throughput, with both prices set by a third party. The margin is not a strategy. It is a spread.

Now add the capex treadmill. Prover hardware depreciates. GPUs turn over in roughly eighteen to twenty-four months against a moving proving system. FPGAs and ASICs are closer to the metal, which means they are closer to obsolescence the moment the proof system changes — and proof systems change, because the entire research frontier is measured in constraint count per transaction. Change the arity of the curve, change the recursion strategy, and the hardware resets.

The high-end configuration makes it worse. The closer a rollup gets to bytecode-level and consensus-level equivalence, the larger the circuit each transaction must be traced through, and the more constraint rows a simple swap consumes. Language-level equivalence — compile Solidity to a custom backend — is cheap to prove and expensive to adopt, because every existing contract has to be redeployed and every integration rebuilt. Bytecode-level equivalence is the reverse: cheap for developers, expensive for provers. There is no free corner on that frontier.

My position has not changed since I first ran the numbers. Unless gas returns to bull-market levels, the operators are bleeding, and the token is not a governance instrument — it is a line item in the cost of goods sold. Most L2 tokens are not governance tokens. They are electricity tokens, and the vote is a rounding error attached to the subsidy.

The Unlock Schedule Is the Only Honest Document in the Stack

Back to the schedule, which is the part I read first.

Team: twelve-month cliff, thirty-six-month linear tail. Investors: six-month cliff, eighteen-month linear tail.

Run the ratio. If the investor allocation is I and the team allocation is T, monthly investor release is I divided by eighteen, and monthly team release is T divided by thirty-six. The ratio of release rates is 2I divided by T. If the two allocations are equal, investors unlock at exactly twice the team's rate — and they keep that rate for eighteen months, long after the team's tail has ground on for years.

That is the design. It is not a scandal; it is a disclosure. The unlock schedule is the only document in the stack that tells you who the exit was engineered for. Anything faster than the founder lock is by definition an exit path, and the only open question is the size of the allocation it applies to.

Superimpose the two curves and the shape becomes obvious. Supply expands linearly after a cliff. Narrative demand does not expand linearly; it arrives as a burst and decays. So the crossing point is a math problem, not a sentiment problem. If the token generation event lands with mainnet in Q2, the investor cliff lands roughly in Q4 — the same quarter the narrative needs to be at its loudest to absorb the first tranche.

And here is the correlation nobody models: funding windows are correlated across the sector, which means unlock windows are correlated across the sector. Twenty rollups raised in the same eighteen months will unlock in the same eighteen months. The supply curves stack. A sideways tape does not absorb stacked supply curves quietly.

Eighteen People, One Upgrade Key

Eighteen people for execution, proving, sequencer networking, governance, ecosystem development, and security is lean bordering on structural. Lean is correct for a research lab. It is a load-bearing failure for an operator.

The governance question is sharper and takes thirty seconds to resolve. Who holds the upgrade key? How many signers, and who are they? Is there a timelock, and how long? A network that calls itself a decentralized sequencer layer while its upgrade path is a three-of-five multisig has decentralized the marketing and centralized the machine. The safe upgrade path is not a philosophical preference; it is the single highest-leverage attack surface in the entire architecture. Whoever holds that key can rewrite the state transition function.

I spent part of 2025 on the technical verification side of three pilot projects under a Proof of Decentralization framework drafted with a small independent team of legal engineers for a state blockchain council. The output was unglamorous: a set of measurable indices — node distribution, Nakamoto coefficient, governance participation concentration, upgrade-key threshold and timelock duration, sequencer operator count. The lesson generalizes. Decentralization is not a vibe. It is a vector of numbers, and every one of them can be published before a token exists.

Publish them before the token exists. After the token exists, every disclosure becomes a negotiation with a counterparty that has a lawyer and an incentive to characterize the same fact in the opposite direction.

The Press Release as a Legal Document

The regulatory reading is mechanical. The four prongs of the test are investment of money, a common enterprise, an expectation of profit, and that expectation deriving from the efforts of others. The press release is the evidence for prongs three and four. A Series A headline naming a lead investor, a release schedule that creates a distinction between early and late holders, and a token whose functional role in the cost structure is subsidy — that is a document set, not an accident.

The industry has a habit of describing the fourth prong in its own marketing copy, then expressing surprise at the consequences. The fix is not better lawyering. It is a different ordering: publish the upgrade-key policy and the sequencer operator list before the token, quantify the decentralization claim while it is still a fact and not a defense.

Four Buyers, Four Price Discovery Mechanisms

Step back and sequence the capital. A research narrative attracts a funding round. The funding round funds a testnet incentive program. The incentive program manufactures a TVL headline. The headline supports a mainnet launch. The launch supports a token event. The token event creates a marginal seller eighteen months later.

Four different buyers, each with a different valuation method. The fund prices a thesis. The points farmer prices a yield. The momentum trader prices a chart. The market maker prices flow.

The announcement is priced by the first buyer. The protocol is funded by the second. The chart is made by the third and broken by the fourth.

This is also where the industry's favorite narrative, liquidity fragmentation, does its work. Fragmentation is not a technical condition. It is a marketing requirement for the next product, and it needs a shortage story to sell against, which is why every cycle produces one.

In a sideways tape, none of that resolves. Chop is for positioning. Flow follows fear, but only if the protocol holds — and the only positioning signals that survive a flat market are the ones with a cost attached: the blob base fee trend, the withdrawal half-life of the incentive cohort, the real revenue to subsidy ratio, and the first upgrade-key change after mainnet.

The Contrarian Cut: The Risk Is an On-Time Launch

Everyone is modeling delay. Delay is the comfortable risk, because it is visible and because it can be blamed on research.

The uncomfortable risk is an on-time launch. If the team ships in Q2 with a single-operator sequencer and brands it decentralized — exactly what the industry did with the word immutable in 2017 — the price may not care, but the word will. And the word is the asset. In a sector whose entire social proposition is verifiability, a term that cannot be audited is a liability carried by every project that uses it.

Auditing isn't about finding intent. I have no idea what this team intends, and it is irrelevant. What matters is the observable output at launch: an operator count, a key policy, a force-inclusion delay. Those are facts. If the marketing and the facts disagree, that is not a fraud claim; it is a schema mismatch. Schema mismatches are what actually kill ecosystems, because every integration built on the assumption of shared sequencing or permissionless ordering has to be rebuilt when the assumption fails. That rebuild cost lands on developers who never read the deck.

Second cut, in the other direction: $3.2 billion of scripted testnet capital is genuine evidence of competence. Assembling an incentive system that moves that much liquidity requires real skill in mechanism design, sybil resistance, and distribution. That skill is underrated. Just do not confuse it with a protocol capability. Dismissing the campaign is a mistake. Mislabeling it is worse.

What to Watch Instead

Six observables, all public, all measurable without a relationship with the team.

One: the force-inclusion path — does it exist, what is the delay parameter, who pays the bond. Two: the blob base fee, because it is the true price of the product's input. Three: the withdrawal half-life of the incentive cohort at thirty and ninety days. Four: the first upgrade-key change after mainnet, and whether the timelock held. Five: the ratio of real execution revenue to subsidy spend, quarter over quarter. Six: the cost of a proof, in dollars per transaction, at a stated blob base fee.

That sixth number is the only one that classifies the project. It is also the one that never appears in the deck.

Code is the only law that doesn't take a meeting, and the ledger doesn't care which of these numbers looks better in a slide. So a question for whoever writes the next announcement in this category: if the marginal cost of a verified state transition is the single figure that determines whether your network is solvent, why is it the one figure you have never published?