Token Inventory: The Invisible Cap on Arbitrage Reach

Most Cycles Start with a Token I Don't Hold

I'm staring at the path discovery output again. The bot has surfaced thousands of theoretically profitable cycles, but the executable subset — the cycles I can actually fire transactions for — is a small fraction of that. The pattern becomes obvious once I see it: nearly every promising cycle starts and ends with a token I don't own.

The wallet holds SOL and USDC. That's it. Those are the canonical tokens for atomic execution — you start with one, hop through several DEXs, end back at the same one. But the cycles the math is finding involve BONK, JUP, mSOL, RAY, and a long list of other SPL tokens. Each of those cycles is structurally unreachable from where I currently stand. The path search engine isn't broken. My token inventory is the bottleneck.

Why a Cycle's Starting Token Matters

For anyone new to this: cyclic arbitrage works by buying token A, swapping it through a series of liquidity pools, and ending up back at token A with slightly more than you started with. The cycle must close at the starting token — that is the definition of a closed loop in token space. If the cycle is A → B → C → A, you need A to begin.

Now imagine a graph. Tokens are nodes, liquidity pools are edges between them, and the exchange rates plus fees are the edge weights. The bot's job is to find loops in this graph where multiplying the edge weights yields a number greater than one — a profitable cycle. The math doesn't care which token the loop is rooted at. But the executor very much does. If the cycle starts at JUP and the wallet has zero JUP, the executor can't initiate the first swap.

So the bot's real opportunity space isn't "every profitable cycle in the graph." It's "every profitable cycle that begins and ends at a token I actually hold." That's a dramatically smaller set, and the gap between those two sets is exactly the opportunity that token inventory unlocks.

The Math of Token Coverage and Path Count

There's a 2024 academic paper titled "An Improved Algorithm to Identify More Arbitrage Opportunities on Decentralized Exchanges" that quantifies this relationship more cleanly than anything else I've seen. The authors compare a traditional Moore-Bellman-Ford approach with a Modified MBF that uses a line graph construction to expand token graph coverage.

Per the 2024 academic study, the MMBF method discovered 23,868 arbitrage paths worth more than $1,000 each. The same dataset, analyzed with the traditional MBF approach, surfaced only 19 such paths. That is a difference of more than three orders of magnitude on the same underlying liquidity. The single variable that changed was how many tokens the algorithm considered as candidates for cycle endpoints.

The path length distribution reinforces the point. According to the same paper, MMBF-discovered cycles cluster in the 7-to-11-hop range. Traditional methods, constrained to a narrower token set, surface cycles of 3 to 4 hops. Maximum detected arbitrage profit jumped from around $100,000 (traditional method) to roughly $1 million (MMBF). Longer cycles, made executable by broader token coverage, are where the larger profits hide.

This is the closest thing to a peer-reviewed proof that what I'm staring at on my screen is a real phenomenon and not a quirk of my particular setup. The relationship between tokens covered and paths executable is structural.

The Solana Reality: SOL and USDC Dominate, But Don't Cover

The research that's most relevant to my immediate situation comes from a 2025 analysis of Solana DEX flow. According to that 2025 analysis, SOL appears as a pair token in 42% of all trades on Solana DEXs, and USDC accounts for 30% of pair appearances. Together they touch the majority of liquidity-weighted volume. So my SOL/USDC-only strategy isn't pointing at a backwater of the ecosystem. It's covering the canonical pair.

But covering the majority of volume is not the same as covering the majority of cycles. Most three-hop cycles can be expressed as SOL → token X → USDC → SOL or similar — but that requires only that SOL and USDC appear somewhere in the cycle, not that they are the start and end. Four-, five-, and seven-hop cycles increasingly route through mid-tier tokens like LSTs (liquid-staked SOL variants), governance tokens, and meme tokens, and many of those cycles never touch USDC at all.

There's a useful analogy here. Think about Walmart versus a corner deli. Walmart stocks the canonical goods — milk, bread, eggs — and captures the majority of foot traffic. But the deli a few blocks away survives by selling the niche stuff Walmart doesn't carry. My bot, with only SOL and USDC, is like a store that stocks two products. The customer who walks in wanting an mSOL-to-JUP-to-RAY-to-mSOL cycle leaves empty-handed, regardless of how good my service is on SOL/USDC trades.

The Long Tail Is Where the Arbitrage Growth Is

The Solana arbitrage landscape has shifted dramatically in roughly fifteen months. According to a 2025 market analysis, cyclic arbitrage on Solana aggregators rose from 2.5% of volume in August 2024 to over 40% in October 2025 — a roughly 16x increase. That growth did not come from bots competing harder for the same SOL/USDC cycles. That pair has been saturated by professional teams for years. The growth came from bots discovering and executing cycles involving tokens that simply weren't being arbitraged before.

As more SPL tokens listed on more DEXs, the token graph expanded. As the graph expanded, executable cycles multiplied. Bots that could adapt their inventory to the new graph captured the new flow. Bots that stayed locked to SOL/USDC kept seeing the same opportunities everyone else was racing for.

Per a 2026 market report, Solana DEX volume reached $1.5 trillion in 2026, a 57% increase year over year. Within Jito's bundle infrastructure, a 2026 market report records more than 90 million successful arbitrage transactions generating roughly $142.8 million in combined profits. That kind of throughput doesn't come from two-token strategies. It comes from a population of bots collectively covering a broad token graph at scale.

A 2025 market analysis, citing analyst Eekeyguy, notes that "arbitrage trading on Solana comprises atomic and bundled types" — meaning both the single-transaction kind and the bundled inclusion kind contribute. Both, however, rely on having the right tokens on hand at the moment the opportunity opens.

The Cost Side: Holding More Tokens Isn't Free

There's a counter-pressure to all of this that I have to take seriously. Holding more tokens isn't a free lunch. Each token position carries:

• Capital tied up that could otherwise be in SOL accruing staking rewards or in stablecoins earning yield somewhere.
• Inventory risk — the price of BONK or RAY can move while sitting in the wallet, eroding the implicit value of holding it even if no arbitrage uses it.
• Operational overhead — more Associated Token Accounts to keep funded, more transaction account references to manage, more chances for a stale or empty account to break a cycle execution at the worst possible moment.
• Rebalancing cost — every time the bot converts surplus from one token back to base, it pays fees and may impact prices in thin pools.

So the right question isn't "should I hold every SPL token on Solana?" Obviously not. The question is "which tokens give me the most cycle expansion for the smallest capital commitment and the smallest operational footprint?"

This turns out to be an optimization problem with no closed-form answer. The token set that maximizes executable cycles today may not be the same set tomorrow, because pool TVLs shift, new DEXs launch, and the dominant token pairs evolve. The data is constantly changing under your feet.

Automation Becomes Mandatory

Manually picking which tokens to hold, in what amounts, and rebalancing as the ecosystem shifts is not a task a human can keep up with. There are thousands of SPL tokens with meaningful volume on Solana. A human looking at a dashboard, deciding to add a position in PYTH because they saw it appearing in a few cycles last week, has already lost the race by the time they execute.

The bot needs to figure this out on its own. Without revealing implementation specifics, the conceptual loop has three phases.

First, a signal layer: which tokens are appearing most frequently in cycles that pass the bot's profitability filters? This is a frequency analysis over the path search output, ideally weighted by something more meaningful than raw count — expected profit, pool depth, or cycle reliability.

Second, a budget layer: the total capital allocated to non-base token inventory is a portfolio-level constraint, not a per-token decision. The total non-base position might be a small fraction of working capital — enough to enable cycles, not enough to take serious directional exposure.

Third, an execution layer: when an opportunity to opportunistically acquire a target token arises (for example, during a profitable cycle that ends with surplus in that token), the bot can let the surplus stay rather than converting it back. When inventory gets too high in a token whose cycle frequency has dropped, the bot can rebalance back to base during the next favorable swap.

The whole thing is a feedback loop. A new token added means new cycles become executable, which generates more frequency data, which adjusts the portfolio composition, which feeds back into which cycles are surfaced. Over time, the portfolio drifts toward whatever set of tokens actually matters for executable arbitrage in the current market state.

This is the same kind of architecture that a food truck operator uses when deciding where to park. They don't pick one corner forever; they read foot traffic, weather, events, and rotate to the locations where demand is currently. The bot's portfolio works on the same principle, just on a token-by-token basis and at sub-second response time.

Inventory Versus Flash Loans

Whenever this topic comes up, the same alternative gets raised: flash loans. On Ethereum-style chains, a bot can borrow $10 million in a single atomic transaction, execute the arbitrage, repay the loan, and pocket the difference — all without holding the borrowed token beforehand. Per a 2025 technical analysis of multi-surface arbitrage routing, this lets bots "execute, say, a $10 million arbitrage without having $10 million in upfront capital, as long as it is paid back by the end of the transaction."

So why not just use flash loans for every cycle and avoid the token inventory problem entirely?

On Solana specifically, flash loans add real complexity. They consume more compute units, increase transaction size against the strict per-transaction byte limit, introduce additional cross-program-invocation hops, and require that a flash loan provider have liquidity available in the target token at the moment of execution. For small cycles — sub-thousand-dollar profits, which are the bulk of what a multi-token portfolio is unlocking — the overhead can eat the entire edge.

Holding small token positions, by contrast, is cheap on Solana. The rent for an Associated Token Account is a fraction of a SOL, and the marginal compute cost of using an existing token account is minimal. Holding $50 worth of BONK has trivial opportunity cost if it unlocks even one profitable cycle per day. The math flips compared to Ethereum.

This means inventory and flash loans aren't mutually exclusive — they target different regimes. For large cycles where capital efficiency matters and the per-cycle profit can absorb the overhead, flash loans win. For the bulk of mid-tier cycles where speed and simplicity dominate, direct inventory wins. The auto-rebalancing portfolio is built for the latter, which on Solana is the larger volume of opportunities.

It's also worth noting that Solana doesn't have a public mempool in the same sense Ethereum does. Transactions follow different propagation paths, which changes the front-running dynamics but doesn't change the atomicity constraint that makes inventory valuable.

The Frequency Multiplier from Diversification

There's a compounding effect worth naming explicitly. According to a 2024 guide on triangular arbitrage in DeFi, monitoring around 3 tokens yields 5 to 10 trades per day, while monitoring 20 to 50 paths yields 50 to 100 or more trades per day. That's roughly a 5x to 10x increase in opportunity frequency from diversification alone, before any improvement in edge detection, latency, or execution quality.

For a bot operating on thin margins, 10x more attempts per day is not 10x more profit. Margins compress as competition tightens per slot, and not every attempt lands. But it is meaningfully more chances to land, more empirical data per day to refine pricing models, and more available cycles when the obvious SOL/USDC opportunities have already been claimed by faster bots in the same block.

The frequency math is the second-order benefit. The first-order benefit is even more important: simply having access to cycles that were structurally invisible before.

Inventory as Competitive Infrastructure

The framing that keeps making sense to me is this: token inventory is competitive infrastructure on the same level as latency.

Latency infrastructure is well understood. Per a 2026 analysis of Solana MEV operations, "infrastructure decisions can create 30 to 50 percentage point differences in landing rates between identical strategies." Colocation, dedicated streaming, optimized transaction submission paths, accurate tipping models — all of this is the visible layer of the Solana MEV arms race. It gets discussed at conferences and in blog posts. Numbers get quoted.

Token inventory is the invisible layer. Two bots with identical latency profiles, identical fee strategies, and identical path search algorithms can have wildly different P&L just because one holds the right ten tokens and the other holds two. The first bot sees and executes cycles the second can't even attempt. The differential never shows up in a latency metric because it's not a speed problem — it's a reachability problem.

This is the part that doesn't show up in talks. Everyone debates RPC providers and tip ratios in public. Nobody publicly lists which tokens they hold and why. But as the 2024 academic paper makes clear, the cycle count differential from token coverage can be three orders of magnitude. The asymmetry between what's discussed and what actually matters tells you where the current edge lives.

It's a bit like NASCAR teams talking about tire compounds and pit strategy while keeping their fuel mixture proprietary. The visible parts are commoditized; the quiet parts are where the wins compound.

What This Changes About the Bot's Architecture

Adding an auto-rebalancing token portfolio is not a small change. It ripples through the entire system:

• Path search must allow cycles to start at any token currently held, not just SOL or USDC. The graph traversal logic widens.
• Execution requires that the on-chain program accept variable starting tokens, and the wallet needs Associated Token Accounts pre-created for every token that might appear in a cycle.
• Risk management has to account for inventory positions in mid-tier tokens, which introduce directional exposure that didn't exist with a SOL+USDC-only setup. Position limits become per-token, not just per-cycle.
• Accounting and P&L get tricky when one cycle ends with surplus in BONK and the next cycle wants to start with USDC. Profit denomination across a multi-token portfolio requires a consistent valuation framework, applied consistently across runs.
• Failure modes multiply: a token account that's not yet rent-funded can break a cycle execution at the last instruction. Edge cases that didn't exist with two tokens become routine with twenty.

None of these are insurmountable. But they're all real engineering. The reason most retail-scale bots don't take this route is the same reason most amateur traders stick with spot positions in two coins: complexity scales faster than expected, and the simple version mostly works.

The question, for any bot operator, is whether "mostly works" is good enough. In a market where cyclic arbitrage volume is up 16x in fifteen months per the 2025 market analysis, and the discoverable path count varies by three orders of magnitude based on token coverage per the 2024 academic paper, "mostly works" is leaving the long tail entirely on the table.

What Expanding the Path Set Implies for the Broader MEV Direction

Looking at the ecosystem trajectory, the direction of travel is consistent. The number of liquid SPL tokens is growing. The number of DEXs is growing — recommended coverage now spans constant-product AMMs, concentrated-liquidity pools, dynamic-bin protocols, order books, and proprietary AMMs that don't expose public pools at all. Each new DEX adds edges to the token graph. Each new token adds nodes.

A bot strategy fixed at "two base tokens, a handful of DEXs" is structurally bounded as the ecosystem expands. A strategy that adapts its token inventory and DEX coverage to the ecosystem's current shape can keep growing into that expansion. The bot that learns which tokens are appearing in cycles today, and adjusts its holdings accordingly, has a structural advantage over the bot that picked its base tokens a year ago and never revisited the decision.

This isn't a permanent moat. The same automation pattern, once the concept spreads, will become standard infrastructure rather than a differentiator. Inventory management will get commoditized the way latency optimization is being commoditized — first the cutting edge, then the standard kit. The window during which holding the right tokens is a quiet edge will close, just as the window for cheap colocation closed.

But for the moment, "what tokens do you hold?" is one of the questions a sophisticated operator doesn't share, while "what's your tip ratio?" gets discussed openly on panels. That asymmetry — between what's debated publicly and what actually moves P&L — is the reliable signal for where the current edge actually lives. When everyone is talking about latency, the move is to also be working on the thing nobody is talking about yet.

Key Takeaways

• An arbitrage bot's executable opportunity space is bounded by which tokens it holds: cycles must start and end at a token already in the wallet, regardless of how profitable the underlying graph math says they are.
• A 2024 academic study demonstrates that expanding token graph coverage can increase discoverable arbitrage paths by more than three orders of magnitude on the same dataset.
• SOL and USDC dominate Solana DEX pair appearances (42% and 30% respectively per a 2025 ecosystem analysis), but a vast space of mid-tier-token cycles sits outside that coverage, with less competition and longer, more profitable paths.
• Token inventory is competitive infrastructure on the same plane as latency — invisible from the outside, decisive for which cycles a bot can actually execute, and currently underdiscussed in public MEV conversations.
• Manual portfolio management at the required scale and update frequency is impractical; an auto-rebalancing approach — frequency-driven signal, portfolio-level capital budget, opportunistic acquisition during profitable cycles — is the practical implementation path, and the engineering complexity is real but bounded.