Equilibrium Infra Bulletin #53: AMMs, CLOBs, and the evolution of capital-efficient trading on-chain
Equilibrium designs, builds, and invests in core infrastructure for the decentralized web. We are a global team of ~30 people who tackle challenges around security, privacy, and scaling.
🔍 The evolution of capital-efficient trading on-chain
⚡️ Topic Summary
Hyperliquid’s recent success has reignited the discussion of DEXs and how onchain venues can be competitive against CEXs (or exchanges in traditional finance).
Price discovery doesn’t happen on-chain today for spot assets, and particularly in derivatives trading. Centralized exchanges such as Binance or Coinbase are where prices are “made”. A big portion of the onchain trading flow, apart from long-tail assets not supported by CEXs, is “downstream” action with the on-chain world catching up via various types of arbitrage.
The dynamic of on-chain playing catch-up is definitionally true as long as there’s a significant performance difference. Milliseconds matter; the difference between real-time and anything else is massive.
The history of onchain trading has been adaptation by necessity, particularly to the needs of market makers. The best market-making environment wins because user trades will ultimately be routed to that liquidity.
On-chain market structure development can be seen as the combination of (i) minimum required and (ii) maximum allowable actions in creating a “liquid-enough” trading venue. Minimum required refers to innovations that allow for a market to function, and maximum allowable is scaling in various forms (off-chain, latency, throughput, etc).
X * Y = K (fixed) AMM
“Market-makers” have 2 available actions: place or pull liquidity (in anticipation of beneficial or adverse flow).
In hindsight, it’s surprising how well this model worked given the average volatility of crypto tokens. DeFi summer did important lifting with token incentives.
Model can’t be iterated on (definitionally).
Trading against a pool
“Market-maker (i.e., pool)” essentially trades against users and doesn’t want to end at a loss. The main restriction for the pool is not to offer too good pricing to ensure no/manageable adverse flow.
Pioneered by Synthetix, with iterations like GMX and Drift’s vAMM. The limitation is that this model can only lag better-priced venues and offer something worse (but it can do so quite well).
As external markets get tighter, the pricing capability of virtual liquidity venues improves as well.
Concentrated fixed AMM
Market-maker has the same 2 available actions as in X * Y = K, but offers a different, more purpose-built, ready-made strategy.
Curve’s stablecoin pools are the most well-known example. The risk is the strategy’s assumption (e.g., a stablecoin depegging) failing.
This model can be innovated on by offering other purpose-built curves. An example of this is Orbital by Paradigm, which just adds dimensions to accept many more assets for a stablecoin strategy.
Conceptually, if equity indices become a big thing on-chain, one could offer time-offset fixed (concentrated or not) to "move” liquidity to match the S&P500’s 9% annual return rate before accepting trades. Then, market-maker losses are lessened if the assumption holds that price drifts in the same direction as the time-offset bias.
Similarly, in a manipulated market (e.g. Japan’s central bank won’t allow rates over X%), one could offer “wall-off” liquidity in a specific range — this works well until it doesn’t.
Concentrated custom AMM
Market-maker technically has all options available to them because placing liquidity in the single tick is the same thing as a limit order.
Concentrated liquidity AMMs offer both fixed AMM features (i.e., low minimum required actions) and take advantage of all scaling improvements.
Any custom “concentrated” fixed strategy can be implemented on top.
Pioneered by Uniswap v3 and remains the most market maker-friendly option on high-throughput L1s like Solana.
Definitionally, has the widest possible surface area for strategies that can be implemented in a gas-efficient fashion, as liquidity can be placed across a high number of "limit order” slots in a single transaction.
Order books
Market-maker has all the options available to them because a limit order can be placed at all possible slots.
However, has a very high minimum required volume threshold to create an efficient trading venue. If done fully on-chain, transactions have to be placed in a high number of slots to create a liquid venue.
Gas is prohibitive even on a fast chain. This is because limit order placements and cancellations tend to outnumber executed orders by at least an order of magnitude.
Attempts like Serum or Phoenix haven’t been quite uncompetitive even on Solana. Even sub-second consensus is still an order of magnitude above CEX ping times.
Costs are driven higher for all market makers since spamming the chain is the best strategy to get orders in.
Market makers don’t control transaction ordering, which leads to adverse selection when they can’t get their cancels or placements prioritized.
Hyperliquid is the most high-profile attempt to fix these problems.
With a small validator set, most of these problems are fixed essentially by having an architecture that mimics running an order book off-chain.
Efficiency-enhanced order books / concentrated AMMs
Euler’s EulerSwap and Fermi’s JIT liqudity services enchance the collateral value for order books and AMMs by keeping the trade colleteral “active” until a trade takes place. With this method, liquidity can sit on lending markets until it’s needed for trade execution.
Similar benefits can be derived just by having, e.g., stETH as the primary trading pair for USDC swaps. To an extent, it’s surprising this hasn’t happened yet.
Conceptually, a “final” evolution for on-chain trading venues, where incremental yield (and risk) can be increased by stacking more collateral options or by looping for leverage.
Privacy is also an efficiency enhancer in the form of “hidden” limit orders — market makers can’t be hit with toxic flow if no one knows where their orders are placed.
🤔 Our Thoughts
Efficiency-enhanced order books or concentrated custom AMMs are the on-chain trading end game. This takes advantage of optimal market structure (granularity) and maximizes the yield qualities of the assets themselves.
What’s left is pushing the (i) scaling and (ii) ordering properties of the underlying infrastructure to maximally benefit market makers.
Here, a few battles that are important to note:
Open block building vs. tight integration of trading flow
If you don’t have full control of execution guarantees, that’s a super concerning risk for leveraged (e.g., perpetuals) markets because liquidations can’t provably happen safely. With something like Hyperliquid, you can show that the system will always be solvent as long as orders are executed in a certain sequence.
Additionally, order placement and cancellation being free on Hyperliquid make it much more manageable to do an extremely high number of transactions. Prioritizing those orders instead of trader/taker ones changes the market structure to be much more beneficial to the market makers. Protection against toxic flow allows tighter quotes and better execution for users.
Single sequencer vs. distributed validator sets
In practice, the fact that ordering rules from a single source are better for traders pushes towards centralization. From there, you can work on how to improve trust in such a system — an example of this is Unichain’s block building inside a TEE.
An open and distributed validator set where the block builder chooses how to construct that block can, at any time, create conflicts for market makers. That’s why the beneficial ordering guarantees depend on, e.g., L2 sequencers, because they can be “trusted” and choose various options in passing down ordering-based profits (or MEV) to applications and users.
Solana is trying to solve this problem with Multiple Concurrent Proposers. If a leader doesn’t accept a cancellation, a market maker can submit to another validator. Applications are given a new feature that enables them to read the fees that are being paid to interact with them, and they can reject trader (taker) orders beyond a certain fee level while accepting market maker orders above that threshold.
A drawback to this model is that it still doesn’t allow free cancellations like a CEX can.
Co-location vs. consensus overhead
A single sequencer model allows for decreasing latency via co-location (traders being near the box where block building happens).
Despite this, the engineering work for decreasing latency in an L1 has good returns in making markets more competitive. If 1 block is produced per day and the BTC price has moved 20%, the block builder can pick off a bunch of stale orders. But if blocks happen every 150ms, not much has moved, and there are just fewer exceptions to take advantage of.
A technical tradeoff is that the faster blocks are built, the more difficult it becomes to do fancy stuff like MCP within one. That’s an engineering concern that can and will be figured out to whatever degree it can.
There’s likely a point where incremental benefits to latency don’t help that much — maybe somewhere in the ~50ms ballpark, or ~10x improvement from the fastest chains is sufficient. This is only achievable if multiple rounds of communication are not required for strong inclusion guarantees.
Conclusion
Based on the current state of the research, it’s likely that fast L1s can be fast enough from a latency (and throughput) perspective to run efficient markets in the next 12 months. However, from the ordering perspective, there still seem to be some fundamental limits that can’t be worked around.
This problem can be fixed almost in its entirety by running the limit order placement and cancellations off-chain. Conceptually, this is not so different from what single sequencer L2s do right already.
The technology is available (or close to it) to build competitive order books on fast L1s/L2s by combining hybrid on-chain and off-chain approaches. What matters more is nailing down the feature set to prioritize market makers, and trading volume can follow.
💡 Research, Articles & Other Things of Interest
🤓 MEV and the limits of scaling from the Flashbots team. Efficiency can be viewed with both a compute and market structure lens. In this article, you will find a great breakdown of the incentives to spam onchain transactions and how that increases costs for all users.
🎧 Solana’s path to decentralized NASDAQ from Lightspeed, featuring Max Resnick. On adverse selection for market-makers, multiple leader consensus, and other constraints that L1s have.
🔥 News From Our Partners
Euler launches EulerSwap, a new DEX built with just-in-time liquidity from Euler’s lending markets and on top of Uniswap v4 hooks.
MetaDAO announces the second ICO on its platform. MetaDAO offers a token launchpad but with investor protections and futarchy built in. If a founder stops executing, investors can withdraw their funds.
🤌 Personal Recommendations From Our Team
📚 Reading: The Danger of Delegating Education to Journalists, K. Anders Ericsson. Ericsson’s work was the premise of what eventually became Malcolm Gladwell’s 10,000-hour rule. He became frustrated by the misrepresentation of his work in the mainstream, leading to this article.
💡 Other: Why “no one plays defense in the NBA” by Thinking Basketball. A great breakdown of how a sport has developed over time and how that has changed strategy. This video combines video analysis and advanced analytics in a cool way.