Liquidity doesn't.
Not in crypto. Not in AI inference. It follows the path of least resistance, pools where the cost is low, and evaporates when the faucet tightens. Last week, Anthropic raised Claude Code's weekly limit by another 50%. The market cheered. More tokens for everyone. But the auditor blinked; the market didn't.
I've seen this pattern before. In 2017, I audited 40+ ICO whitepapers, watching teams promise 'unlimited scaling' while running on a single AWS instance. In 2022, I mapped Terra's algorithmic stablecoin to shadow banking structures, predicting the collapse weeks before the market realized that 'yield' was just a tax on ignorance. Now, Anthropic is doing the same dance with compute credits.
The Context: Claude Code as a Macro Asset
Claude Code is Anthropic's AI coding agent, bundled into Pro ($20/month) and Max ($100/month) subscriptions. Unlike API calls, which are metered by token, Claude Code uses a weekly usage limit—a quota that caps how many tasks a user can run. In May 2025, Anthropic increased that limit by 50% for the first time. Then again in July. Now, they're extending the increase through August 31, with a promise to make it permanent.
But the language is telling: 'Demand has been strong, and compute will remain tight for the next few weeks.' This is not a product update. It's a supply chain confession.
Based on my experience analyzing cross-border payment flows, I've learned that any artificial constraint on a digital asset—whether it's a stablecoin redeemability ceiling or a weekly API quota—is a signal of underlying structural fragility. Claude Code's limit is a liquidity cap on inference compute. And Anthropic is running out of liquidity.
The Core: Inference as the New Collateral
Let's break down the economics. Claude Code sessions are heavy: long context windows, multiple tool calls (file edits, terminal commands, web browsing), and iterative code generation. Each session consumes order-of-magnitude more compute than a typical ChatGPT query. At scale, one active Claude Code user can burn through as much inference as 50 casual chat users.
Anthropic's subscription model is a fixed price for variable compute. The limit is their only hedge against a single user bankrupting their margin. By increasing the limit by 50% without raising the price, Anthropic is effectively subsidizing usage—buying time until their next compute infrastructure comes online (new data centers, more GPU contracts).
But here's the hidden signal: the increase is not uniform. The company hasn't said whether all tiers get the same boost. In my 2024 ETF arbitrage study, I found that institutional custody fees undercut traditional banking rails only when volume hit a certain threshold. Similarly, Claude Code's limit increase likely benefits high-usage users disproportionately, letting them test the ceiling while the company measures the cost.
The real story is not the 50%—it's the fact that 50% is still a cap. If Anthropic had unlimited compute, they would remove the limit entirely. The fact that they are incrementing the quota, rather than abolishing it, means their supply curve is still steeper than their demand curve. This is the exact opposite of the 'hypergrowth' narrative that VCs are selling.
The Contrarian Angle: Decoupling Thesis
The market consensus is that Claude Code's limit increase is a win for developers—more tokens, more productivity, more value. The contrarian view is that this is a canary in the coal mine for the entire AI compute ecosystem.
Anthropic's compute tightness is not a temporary blip. It's a structural bottleneck that will persist until either (a) dedicated inference hardware (like custom ASICs) goes mainstream, or (b) the model becomes dramatically more efficient. Neither is happening in the next 90 days. The August 31 'permanent' deadline is a soft promise. If the new data center contracts don't deliver, the limit will be extended again—or worse, prices will go up.
This mirrors the DeFi liquidity trap I analyzed in 2020. During DeFi Summer, protocols like Compound inflated their TVL with token emissions, creating the illusion of deep liquidity. When the incentives dried up, the TVL vanished. Anthropic's limit increase is a form of token inflation: they're injecting more 'compute credit' into the system, but the underlying resource (GPU time) is finite. The only difference is that Anthropic is not minting a token—they're exhausting their existing lease.
The decoupling is this: Claude Code's usage growth is decoupling from Anthropic's ability to serve it. The limit increase is a band-aid, not a cure. The market is pricing Anthropic based on user growth, but the real metric is margin per compute unit. And that margin is compressing.
I've seen this exact behavior in cross-border settlement systems. When a bank runs out of nostro liquidity, they start rationing settlement windows. The outward message is 'improving efficiency'; the inward truth is they're one wire away from a freeze. Anthropic's 'compute tightness' is the same—a liquidity squeeze that they're trying to manage through quotas.
The Takeaway: Positioning for the Squeeze
What does this mean for a crypto-native analyst? Three things.
First, the AI compute market is now a macro asset. Just as we track central bank balance sheets to understand liquidity, we should track major model providers' data center contracts and inference quotas. The next 'black swan' in AI won't be a model collapse—it will be a compute supply crunch that forces price renegotiations across the entire stack.
Second, the limit increase is a leading indicator of Anthropic's next funding round. If they are burning cash on compute subsidy to grow users, they need a capital injection before Q4 2025. Watch for a new round announcement at a valuation that assumes the limit becomes permanent—but with founders' revenue projections that don't account for the cost of that permanence.
Third, and most importantly, treat Claude Code's quota like a stablecoin peg. The limit is the peg. The demand is the market. If the limit becomes permanent without a corresponding price increase, the peg holds. If Anthropic has to raise prices or reduce limits post-August, the peg breaks. And broken pegs, as I learned from Terra, tend to cascade.
The auditor blinked. The market hasn't. Not yet. But when the August 31 deadline passes, watch the margin calls—not the user counts.