Two point five nine basis points. That is what a swap costs on a public Solana AMM in a calm minute — no headlines, no liquidation cascade, no whale lifting the book. On a professional market maker's pool, same order, same second, same reference frame: 0.26 basis points. A tenfold gap, measured against a Bybit microprice converted through the USDC/USDT mid. A preprint from ETH Zurich and Category Labs put both figures in one table last week, and much of the coverage that followed read like a coronation for professionalized market making.
Hold on. The same table carries two-second markouts: plus 0.37 basis points for the professional pool, minus 0.22 for the public one. Cheaper for the trader, and still profitable for the maker. Markets are not charities. Something paid for that spread.
That something is the story almost nobody is filing.
Context: a vending machine, and a doorman
A public AMM — Raydium, Orca, the usual Solana names — is a vending machine with no bouncer. Anyone deposits, anyone swaps, and the curve quotes the same to a bored retail trader and to an arbitrage bot that has already watched the price move on three other venues. Passive LPs absorb that asymmetry. The academic name is adverse selection; the quantified version is LVR, loss-versus-rebalancing, the drag an LP eats relative to a strategy that rebalances on every tick. LVR is not a bug in a public pool. It is the invoice for being permissionless.
A propAMM — Jump Crypto's BisonFi being the reference case — is the same machine with a doorman. Pricing adjusts against inventory, quote freshness, and what Jump's own writeup calls “the quality of inbound flow.” Bouncers, in plain English. The mechanism is not new in kind; every serious trading desk has routed around toxic flow for decades. What is new is that it now happens on-chain, behind a public settlement layer, with the router as its instrument.
The method deserves credit before the critique. The study separates two things this industry habitually blends: an execution-cost proxy, meaning a fill's deviation from a reference price, and markout, meaning whether the price moves favorably after the fill. The authors explicitly warn against using the first to argue the second — you cannot convert a cheap fill into a profitability claim. I spent 2020 inside Compound's governance mechanics at an audit shop in Warsaw, and the discipline generalizes: measure the thing you measured, and name the thing you didn't.
One caveat before the numbers harden into folklore. The dataset covers September 2025 through August 2026 — a window extending past the plausible date of any published analysis. Either a typo for 2024–2025 or a genuine provenance problem. The reference price is a single source, Bybit's microprice. One venue, one book, and every downstream conclusion inherits its biases.
Here is why this lands on my desk rather than in a forum thread. I spend my days bridging institutional capital and crypto natives, and the pitch I keep hearing is that decentralized venues will win on transparency. This paper is a quiet counterargument. Allocators price disclosure. They will ask what the effective execution cost is for their flow, and a venue average will not survive the meeting. A market that cannot tell an LP its attributable adverse-selection cost is not a market institutions can underwrite — only rent.
Core: the filter is the product
Here is the part that matters. The 0.26-versus-2.59 gap was not measured across all flow. It was measured on quiet flow — fills where the reference price moved less than one basis point in the five seconds before and the one second after execution. That filter is the study's cleverest move and its most loaded one, because it isolates precisely the trades where nobody held an information edge.
On that subset, the professional pool wins twice: better price for the taker, positive markout for the maker. Which means the edge is not superior pricing over uninformed traders. It is selection. It reads the order, decides the order is benign, and quotes tight. The public pool, lacking the doorman, quotes wide against everyone and eats the difference on the informed minority.
Now watch how the labeling happens, because this is where decentralization quietly changes hands. Jupiter, the dominant router, uses a dedicated signer to identify front-end flow and tag it “retail and non-toxic.” BisonFi prices against inventory, freshness, and flow quality. The authors land the sharpest line in the paper right here: identifying a flow's origin is not independently proving its harmlessness. A wallet signature is a claim about who clicked, not about what they know. Tag a trader as retail and you have tagged a person, not an intent.
And the tag is sticky. Once a router classifies flow, the classification travels with the order into every downstream venue. That is a market-structure primitive — the DEX cousin of payment for order flow, except the payment is rendered in price rather than rebates, and the payer is whichever pool still accepts the untagged remainder. In 2017 I audited more than forty whitepapers for a Baltic ICO platform and found eighty percent of them economically hollow. The failure mode then was fantasy tokenomics. The failure mode now is subtler: real efficiency, quietly financed by a group that never signed up for the job.
There is a social layer here that the market-structure crowd will skip. “Retail and non-toxic” is not a neutral technical descriptor. It is a class label, and the people it sorts into the screened pool are, on average, the ones with the least leverage and the least information. Everyone sorted out lands in the public pool, quoting wider, absorbing more adverse selection, earning less. In 2021, running an NFT marketplace campaign for fifty women creators, I learned how fast a community notices when it is being treated as residual demand. The same instinct applies to LPs.
Then there is Base. On Tessera, a professional maker there, execution prices averaged 1.08 basis points worse than a reconstructed prior-block-end quote — 0.56 basis points when weighted by volume. The researchers describe the block-timing pattern as “spoofing,” then state plainly that they have no direct evidence of operator intent. Read that twice. The word carries enormous weight while the disclaimer does the opposite. In traditional markets, spoofing is not a stylistic complaint; it is an enforcement category with decades of case law behind it. Importing the vocabulary without the intent evidence is either a warning shot or a liability, depending on which lawyer reads it first.
0x did not wait for the debate. The router said it would cut off the relevant sources until execution issues were resolved. That is the most consequential sentence in the whole affair, and it is not about market making at all. It is about who holds the off-switch. Routers do — and they are now de facto arbiters of venue legitimacy, with no shared standard for what consistent execution even means. The paper's own parity test checks a single snapshot. Whether a venue's quote-to-fill relationship holds across blocks — the thing that actually decides whether a taker got a fair deal — remains unverified.
And this is not a Solana problem wearing a Solana costume. The sample spans Solana, Base, and Monad, which tells you the quiet-flow dynamic follows the architecture rather than the chain. Base adds its own wrinkle: an L2's block timing and sequencer batching can widen the gap between the quote a router displays and the fill a taker receives. Any chain that inherits an aggregator inherits the classification layer underneath it.
Credit where it is due on integrity. The authors cite Jump Crypto's report while explicitly flagging Jump as an interested operator, and they separate their measurement from their interpretation at every turn. In a field where most “research” is a launch announcement with a PDF attached, that restraint is rare. It also means the warnings carry more weight, not less. The caveats are not hedging. They are the argument.
True ownership begins where the server ends. In this architecture the server did not disappear. It moved into the router.
Contrarian: efficiency was never the question
The consensus takeaway is that propAMMs make Solana cheaper. Granted. Cheaper for whom, funded by whom, and for how long are the questions nobody priced.
Here is the uncomfortable synthesis. A public AMM's publicness is not a feature competing against efficiency. It is a subsidy. Permissionless liquidity means accepting every order, including the ones that lose money. Strip out the benign flow and route it to a screened pool, and what remains in the public pool is a higher concentration of toxic flow. The LVR burden does not vanish. It concentrates. Public LPs keep identical withdrawal rights and a worsening book.
The authors are scrupulously honest about the limit of their evidence: venue averages cannot prove that professional pools caused aggregate passive LP losses, because the data lacks trade-level attribution. I would read that the other way. The loss may be real precisely because it is unmeasurable at the venue level — a slow bleed that never shows up as one bad day. If you cannot attribute it, you cannot defend against it.
And the mirror image should unsettle anyone who spent the last cycle arguing DeFi would eat Wall Street's lunch. Best-execution obligations, order-flow payment, fair-access rules, client segmentation — every one of those traditional institutions has now reappeared on-chain under a different name. We did not replace the market structure. We rebuilt it, faster, with worse disclosure and nobody accountable for the label on your order.
There is a governance hole underneath all of it. Router inclusion rules determine which venues see good flow, and LPs — the people whose capital makes liquidity real — have no seat at that table. They cannot vote on whether their pool gets tagged as residual. They cannot audit the classifier. The routing logic is the missing key, and the absent voice belongs to the depositors. Protocol designers love to invoke credible neutrality. Credibility is earned by publishing your inclusion rules, not by insisting that routing is technical and therefore apolitical.
Debate is the compiler for better consensus. Today that consensus is being compiled by whoever writes the router's inclusion rules, and almost nobody has read the source.
Takeaway
Watch three signals, not price. Whether routers adopt cross-block parity testing instead of single-snapshot checks. Whether public AMMs ship toxicity-resistant designs and thereby blur the line between public and professional pools. And whether quiet flow becomes a disclosed, priced asset class — or stays an invisible subsidy extracted from LPs who never got a vote.
Public liquidity is only public until the good flow stops arriving. After that, it is just the last seat at the table.