Who Eats the Spread? The Microstructure of Cross-Exchange Arbitrage in 2026

Abstract

1. The Visible Spread Is a Mirage

Open any crypto arbitrage dashboard in 2026. You will see spreads. Bitcoin at $64,980 on Coinbase, $65,034 on Binance, a 0.083% gap. ETH at $1,917 on Kraken, $1,920 on Bybit. Solana at $76.55 on OKX, $76.70 on Deribit.

A naive observer sees free money. A professional sees a data artifact.

The spread between two exchanges is the difference between the highest bid on one venue and the lowest ask on another at a specific microsecond, at the top of the order book. The moment a trade moves beyond top-of-book depth (and any trade of meaningful size does exactly that) the effective spread collapses. The 0.083% gap on BTC across Coinbase and Binance exists for roughly $200,000 of depth on each side. A $5 million trade walks the book down on the buy side and up on the sell side, consuming whatever spread existed and leaving behind slippage, fees, and regret.

"The gross spread on the dashboard is not the net spread on your P&L," notes Ali Hajimohamadi in a July 2026 industry survey. "Most visible price gaps between Binance, Coinbase, Kraken, OKX, Bybit, Deribit, and major DEX aggregators are now captured by professional market makers, low-latency bots, internal exchange systems, and liquidity providers before retail traders can act."

The dashboard spread is the starting gun. By the time you see it, the race is over.

2. The Anatomy of a Spread: Where Edge Actually Lives

To understand who captures the spread, we must first decompose it into its constituent parts.

A $100,000 BTC arbitrage between Exchange A and Exchange B does not capture 100% of the displayed spread. It captures:

Top-of-book spread × executable depth - (taker fees on both sides) - (withdrawal or bridge costs) - (slippage beyond top-of-book) - (execution risk from latency) - (inventory financing cost)

In 2026, for a BTC/USDT pair across top-tier CEXs, the breakdown looks roughly like this:

The spread that survives all frictions is measurable in basis points, not percentage points. And it only survives for order sizes within the overlapping depth of both order books, typically the first $100K to $500K of liquidity on major pairs.

This is the structural reality that explains why "simple arbitrage" stopped working for anyone without institutional infrastructure circa 2023. The edge is too thin, the required speed is too high, and the capacity is too small for anything other than fully automated, pre-funded, co-located systems.

3. Self-Impact: When the Arbitrageur Becomes the Market

The most important concept in arbitrage execution is also the most counterintuitive: the act of capturing a spread consumes the spread.

This phenomenon, known as self-impact or slippage, is the mathematical ceiling on arbitrage capacity. When a trader simultaneously buys BTC on Exchange A and sells on Exchange B, they are not interacting with a static spread. They are moving both order books:

• The buy order on Exchange A lifts offers, pushing the price up.

• The sell order on Exchange B hits bids, pushing the price down.

The spread that existed before the trade is smaller during the trade, and gone after it. If the trade is large enough relative to order book depth, the spread inverts mid-execution: the buy-side price rises above the sell-side price, and the arbitrageur locks in a loss.

This is why every arbitrage strategy has a deployable capacity ceiling. Beyond that ceiling, additional capital does not generate additional returns. It generates losses.

Consider a simplified order book model for BTC/USDT on a major exchange:

• Top 10 price levels on the ask side: ~$2M of liquidity

• Top 10 price levels on the bid side: ~$1.8M of liquidity

• Average spread across top 5 levels: ~0.04%

A $500,000 arbitrage trade consumes perhaps 0.01% in slippage. A $5,000,000 trade consumes the entire top of the book and walks 5-10 levels deep on both sides, generating 0.05-0.10% in slippage, more than the spread itself.

In traditional finance, this is well-understood. Market impact models (Almgren-Chriss, Kyle's lambda) quantify the relationship between order size and price impact. Crypto markets, despite their 24/7 operation and apparent depth, are shallower than they appear. The top-of-book liquidity on a major CEX represents a small fraction of total order book depth, and the visible depth is increasingly dominated by market-making algorithms that cancel and replace quotes in microseconds.

The practical implication is that cross-exchange arbitrage at scale is not about finding spreads. It is about rationing access to spreads across multiple venues, asset pairs, and time windows: executing thousands of small, sub-capacity trades in parallel rather than a single large trade that collapses under its own weight.

4. The Infrastructure Arms Race

If spreads are captured in microseconds, the competition is not between traders; it is between infrastructure stacks.

The modern arbitrage infrastructure in 2026 operates across four layers:

4.1 Market Data Layer

Professionals do not poll REST APIs for prices. They consume normalized, timestamped order book data via WebSocket streams and FIX protocol feeds from 15-30 exchanges simultaneously. The quality of the data feed, its latency, its normalization, its gap-detection, determines whether a spread is seen before it is gone.

A 2025 study from holysheep.ai documented the asymmetry: "CEX matching engines publish top-of-book and depth snapshots at 100-1000 Hz. DEX venues publish state changes only when a new block lands: every 12 seconds on Ethereum, every 250ms on Arbitrum/Base." The CEX-DEX arbitrageur operates across two fundamentally different temporal regimes, and the CEX side always moves first.

4.2 Execution Layer

Co-location, FPGA-accelerated order routing, and direct market access (DMA) are no longer optional. The leading arbitrage firms maintain physical servers in Equinix NY4, LD4, and TY3, the primary interconnection hubs for crypto exchanges. Round-trip latency from market data receipt to order submission is measured in single-digit microseconds.

4.3 Custody and Treasury Layer

Cross-exchange arbitrage requires pre-funded inventory on every target exchange. A firm that wants to capture a BTC spread between Binance and Coinbase must hold BTC (or USDT) on both venues before the opportunity appears. This creates a treasury management problem: idle inventory on exchanges that do not produce a spread is an opportunity cost. The most sophisticated operators optimize not just trade execution, but the allocation of idle balances across venues using predictive models of spread frequency and size.

4.4 Risk and Compliance Layer

Cross-border arbitrage introduces jurisdictional complexity. A spread between a Korean won market (Upbit, Bithumb) and a USD market may appear large, but the fiat on/off-ramp, capital controls, and KYC requirements consume the edge. The "kimchi premium", where BTC has historically traded at a premium on Korean exchanges, is not a pure arbitrage. It is a capital control premium that reflects the difficulty of moving fiat across borders, not a failure of market efficiency.

5. Where Edge Still Exists in 2026

If top-tier CEX-CEX arbitrage is a game for the infrastructure elite, where does executable edge remain?

5.1 Perpetual Futures Basis Trades (Funding Rate Arbitrage)

The most persistent spread in crypto is not between spot exchanges but between spot and perpetual futures. Funding rates on perpetual swaps create a predictable, recurring premium or discount relative to spot. Capturing this spread (going long spot, short perps, and collecting funding) does not require microsecond timing. It requires capital efficiency, inventory management, and attention to liquidation risk. This is the core strategy that BHLE and similar execution engines target.

5.2 Cross-Chain DEX Routing

Fragmentation across L1s, L2s, and sidechains creates persistent price discrepancies that are gated not by speed but by bridge latency and gas costs. An asset that trades at a premium on Arbitrum relative to Ethereum mainnet represents a genuine spread, but capturing it requires navigating bridge finality times (minutes to days), gas fee volatility, and slippage on both sides. The edge goes to those who can pre-position inventory across chains and time bridge transactions during low-congestion windows.

5.3 New Token Listings

When a token lists simultaneously on Binance and Coinbase, the first 30 seconds of trading produce the widest spreads, and the highest volatility. Capturing this edge requires not speed alone, but exchange relationships that provide advance notice of listing times, pre-configured trading infrastructure, and risk models that account for extreme initial volatility.

5.4 Regional Fiat On/Off-Ramp Arbitrage

Fiat-to-crypto pairs in emerging markets (NGN, TRY, ARS, INR) often trade at persistent premiums or discounts relative to USD pairs. These spreads exist not because markets are inefficient but because capital mobility is constrained. Local banking rails, FX controls, and limited exchange connectivity prevent the spread from being arbitraged away. This is not a technology problem. It is an access and compliance problem.

6. From Execution to Architecture: The BHLE Paradigm

The BHLE (Basis High-Latency Execution) engine, developed by Base58 Labs as the execution core of BASIS, was designed around these structural realities.

Its architecture acknowledges three truths:

1. Spreads are capacity-constrained. Every arbitrage opportunity has a maximum executable size before self-impact turns the edge negative. Deploying more capital than the available depth can absorb is not aggressive, it is self-defeating.

2. Spreads are temporally sparse. A given asset pair on a given venue pair may produce an executable spread for 0.3% of trading minutes. Idle capital waiting for the next spread is dead weight. The solution is not faster execution of one spread, but parallel execution across hundreds of spread sources: spatial arbitrage across exchanges, funding rate arbitrage across perpetual markets, and DeFi lending arbitrage across protocols.

3. Spreads are execution-dependent. The theoretical edge on a backtest is not the realized edge on a live trade. Slippage, queue position, exchange downtime, network congestion, and gas spikes consume a portion of every spread. An execution engine must price execution risk explicitly into its opportunity selection, rejecting trades where the risk-adjusted edge is negative even if the nominal spread appears positive.

BHLE implements these principles through a deterministic state machine: Normal operation (continuous scanning and execution), BSCB circuit breaker (automatic halt at 0.001% loss threshold), and DMM drawdown management (position-level risk limits). The engine does not chase every spread. It selects only those where depth, latency, cost, and timing align to produce a positive risk-adjusted return at the specific order size being deployed.

This is the difference between arbitrage as a trading strategy and arbitrage as an execution architecture. One hunts spreads. The other builds the machinery that makes spread capture possible at scale.

7. Conclusion: The Spread Belongs to the Prepared

The popular narrative, "arbitrage is dead in 2026", is both true and misleading.

True, in that the simple, visible, retail-accessible spreads of 2017-2021 are gone. The gap between two top-tier CEXs for a major pair is captured by machines before a human can blink. The dashboard spread is a ghost.

False, in that spreads do not exist at all. They exist in the interstices: between spot and derivatives, between chains, between fiat rails, between listing moments. They are smaller, faster, and more deeply embedded in infrastructure than they were a decade ago. But they are real.

The difference between those who capture these spreads and those who do not is not intelligence, strategy, or even speed. It is capacity infrastructure: the ability to maintain pre-funded inventory across dozens of venues, to normalize market data across temporal regimes, to price execution risk into every order, and to route capital only where the edge survives its own consumption.

The spread, in the end, goes to whoever built the better machine.

Data and analysis as of August 2026. References: Hajimohamadi, A. "Cross-Exchange Arbitrage in 2026: Why Most Spreads Are Dead," Startupik (Jul 19, 2026); Adadurov et al. "Second Thoughts: How 1-second subslots transform CEX-DEX Arbitrage on Ethereum," arXiv:2601.00738 (Jan 2, 2026); "Arbitrage trading between decentral and central cryptocurrency exchanges," ScienceDirect (Jul 1, 2026); "CEX Order-Book Depth vs DEX On-Chain Events in HFT Arbitrage," holysheep.ai (Jun 25, 2026); BASIS Whitepaper: Deployment Capacity, docs.basis.pro.