Liquidity isn't where the volume is. It's where the value is. And the latest Vercel data on AI model token consumption screams a divergence that every crypto trader betting on AI tokens needs to internalize.
We didn't need another report telling us open-source models are eating the world. But the numbers from Vercel's platform โ a proxy for real-world developer usage โ expose a brutal truth: token volume is not revenue. And the market is pricing AI tokens as if volume equals value.
Context: What Vercel's Data Actually Tracks
Vercel is the default deployment platform for frontend-heavy web apps. Their AI SDK logs every model call made by developers building on their infrastructure. That's not a perfect proxy for the entire AI market โ it skews toward web devs, code generation, and content workflows. But it's a real-time ledger of where developers are actually sending their queries.
Here's the headline: open-source models now account for 62% of all tokens consumed in February 2025, up from 28.4% in December. DeepSeek โ the Chinese open-source upstart โ passed Google in token volume, becoming the second-largest model provider behind Anthropic. But here's the kicker: open-source models generated only 8.6% of total spending. Anthropic, with 30% of tokens, captured 65.1% of the expenditure.
Core: The Order Flow Tells a Different Story
I've run quant models on API pricing data for two years. The Vercel figures confirm what I've seen in my own backtests: the market is bifurcating into two distinct liquidity pools.
Pool A: Open-source models (DeepSeek, Llama, Qwen) โ high-frequency, low-margin, commodity-like. They're used for bulk tasks: translation, classification, code completion. The unit economics are brutal. At 8.6% of spend for 62% of tokens, the average revenue per token is roughly 1/15th of Anthropic's. That's not a sustainable business โ it's a race to the bottom subsidized by VC funding or state backing.
Pool B: Closed-source high-end models (Anthropic, OpenAI) โ lower frequency, high margin, premium pricing. Developers pay for reliability, reasoning, and safety. Anthropic's 65.1% spend share on 30% token share means its users are willing to pay 4-5x the market average per token. That's the kind of pricing power that builds a moat.
The real signal isn't open-source gaining share. It's that the value capture is concentrating in fewer hands. Open-source is winning the usage war but losing the revenue war.
Contrarian: The Retail Narrative Is Wrong
Retail traders see "DeepSeek surpasses Google" and think open-source AI tokens are the next big thing. They're buying the narrative that open-source will democratize AI and eat closed-source margins. But the data says the opposite: closed-source models are capturing an increasing share of the economic value, not less.
Smart money is watching the divergence. The token volume growth (59% quarter-over-quarter) is driven by cheap open-source queries that generate near-zero profit per call. Meanwhile, Anthropic and OpenAI are raising prices โ and seeing no pushback. Why? Because their customers are building high-value workflows where a 5x price premium is noise compared to the cost of a mistake.
Here's the contrarian trade: short the open-source AI token basket, long the closed-source premium plays. Open-source projects like DeepSeek may have massive usage, but their tokenomics are structurally weak. They rely on below-cost pricing to buy market share. That's not a moat โ it's a burn rate. When the subsidy ends, so does the volume.
In the chaos of the sprint, speed wasn't the only factor. It was the ability to hold value under pressure. Closed-source models have that. Open-source models are still proving they can generate sustainable revenue.
Takeaway: Actionable Levels
If you're trading AI tokens, watch these levels:
- Anthropic's spend share: If it drops below 60%, the premium thesis weakens. Currently at 65.1%, it's a buy signal on any dip.
- Open-source token price per million tokens: Current implied price is ~$0.15 vs Anthropic's ~$2.50. A gap below $0.10 means the race to zero accelerates.
- DeepSeek's token volume stability: If their volume declines after the initial hype, it confirms the subsidy-driven pump.
My portfolio is short the open-source AI token proxies and long the infrastructure plays that benefit from total volume growth regardless of provider. Because liquidity isn't where the volume is. It's where the value is. And right now, value is hiding in plain sight โ in the 8.6% that controls 62% of the flow.