We shipped an order-book DEX in the early 2020s. The matching engine worked, the market structure was honest, professional traders said kind things — and AMMs won anyway. The constraint was never engineering. It was who supplies liquidity when nobody is paid to care, and the math of that question decided the market before our code compiled.
- Order books need active market makers whose quoting must be profitable per venue; AMMs turned liquidity into a passive commodity anyone can supply. On the long tail, passive beat professional.
- The maker math on a thin venue is a doom loop: low volume → negative quoting PnL → makers leave → wider spreads → lower volume. AMMs never enter the loop.
- CLOBs won where volume density pays professionals: perpetuals. Hyperliquid — a fully onchain order book on its own L1 — cleared over $432B in March 2026, roughly 70% of onchain perp volume.
- What we'd build today: AMM liquidity as the base layer, order-book precision only where margin trading demands it, and intent-based flow above both.
- The transferable lesson: mechanism work is liquidity economics first, matching engines second. We learned it the expensive way.
Yellow Research — Hyperliquid perp dominance (2026)
Zpoken shipped-work index (incl. earlier trading-infrastructure protocols)
What we built, and why it looked right
Price-time priority, a proper book, no impermanent loss, limit orders that mean what they say — everything a trader raised on real exchanges recognizes as market structure done correctly. The pitch wrote itself: AMMs were a workaround for chains too slow to host real matching; we would build the real thing. The engineering held up its end. The economics did not, and the reason generalizes.
The maker math that killed it
A market maker's day on any venue nets out as
Maker PnL = spread capture − adverse selection − inventory risk − opportunity cost
Every term except the first scales with being on a thin venue. Work the toy case: a pair clearing $500k/day at 10bps of realized half-spread grosses $500 across all makers — against inventory that must be hedged elsewhere, flow that is sharpest exactly when quoting is riskiest, and capital that could be earning fees passively in an AMM pool instead. The rational maker leaves; spreads widen; the remaining flow gets worse; volume follows it out. That loop has no stable point on the long tail — and a venue's long tail is most of its listings. The AMM sidesteps the entire structure: liquidity is supplied passively, fees accrue pro-rata, nobody's job is to care about any single pool, and a pool with $50k of depth quotes continuously for free. Passive liquidity is worse liquidity — and infinitely more available, and availability is what composability pays for.
Where CLOBs won anyway
Order books did not lose; they retreated to where their economics work. Perpetuals concentrate volume into few margined markets whose funding and liquidation mechanics demand price precision an AMM curve cannot give. There, volume density pays professional makers handsomely — and the numbers are public: Hyperliquid, running a fully onchain order book on its own purpose-built L1, cleared more than $432B in perp volume in March 2026, roughly 70% of the onchain perpetuals market, with open interest above $9B. The Solana lineage (Serum, then Phoenix) tells the same story on the spot side's dense head. The rule that survives: CLOBs where volume density pays professionals to quote; AMMs where it does not; and the crossover is measurable per market, not a matter of taste.
What we'd build today
Not a venue — a structure. AMM pools as the permissionless base layer every asset gets for free; order-book or auction precision only for the markets whose density earns it; and intent-based routing above both, where solvers compete on execution rather than users choosing venues. The 2020s taught the industry that these are layers of one system, not competing religions — and taught us specifically that the mechanism design must start from who supplies liquidity at what profit, with the matching engine as a consequence.
What the failure taught us
Three lessons we now bring to every mechanism engagement. Liquidity economics precede mechanism design — model the supplier's PnL before writing the market's rules. Elegant structure loses to available structure — composability compounds availability, and availability compounds everything else. And honest retrospectives are cheaper than repeated tuition: this post exists because the lesson cost us real engineering years, and the fastest way to keep its value is to state it plainly.
If you're working on something similar, we'd like to hear about it. Talk to a founder — 30 minutes.