The data is out. OpenRouter processed 100 trillion tokens in their latest study. The finding? Open-weight models now account for nearly 60% of API traffic. Llama, Mistral, DeepSeek, Qwen — they aren't just catching up. They're consuming.
But before you frame this as another 'decentralization wins' narrative, pause. I've seen this movie before. It was 2017, and I was auditing ICO smart contracts in Mumbai. Found reentrancy vulnerabilities in three major projects. Shorted their tokens after launch. 40% ROI in 72 hours. The lesson? Leverage doesn't create wealth. It creates the illusion of liquidity until the margin call.
That same lens applies here. Open-weight models are leveraging low costs and open access to attract developers. The 100 trillion token spike looks like adoption. But is it real liquidity — or just the illusion of it?
Context: The OpenRouter Study
OpenRouter is an API aggregation platform. It acts as a router, connecting users to dozens of models — both open-weight and closed-source. Their study claims that over the past 12 months, open-weight models have grown from a niche share to dominating traffic. The data is internally gathered, not independently audited. Methodology? Undisclosed.
What we know: OpenRouter has a financial interest in promoting any model that drives volume. Low-cost open-weight models attract price-sensitive developers. Those developers generate traffic. Traffic makes OpenRouter look valuable. It's a circular incentive: more usage equals more market share, but not necessarily more sustainable revenue.
Bull markets hide structural flaws. Bear markets execute them. In crypto, we learned that during DeFi Summer 2020. High APYs drew liquidity, but the underlying vaults were unsustainable. I analyzed Yearn's early vaults and warned about liquidity traps. The music stopped. The same pattern is emerging here: open-weight models offer cheap inference today, but can they maintain quality, safety, and innovation velocity without the revenue that closed models generate?
Core: The Tokenomics of AI Models
Let's apply the framework I use for crypto tokenomics to AI models. Every model has a 'tokenomic' structure — supply (compute), demand (inference), and value accrual (revenue). Closed models like GPT-4o and Claude have high margins, captive enterprise customers, and ongoing R&D funded by subscription fees. Open-weight models have lower margins, rely on ecosystem contributions, and monetize through hosting services (Together AI, Perplexity) or hardware partnerships.
Tokenomics is just game theory with real money. The game here: open-weight models are playing a volume-over-value strategy. They sacrifice per-token revenue to capture mindshare. This works in a bull market for AI — total spend is rising. But when the next capital efficiency squeeze hits, investors will ask: who actually makes money?
My analysis of the 2020 DeFi liquidity trap taught me that high usage doesn't mean high value. Uniswap V4's hooks introduced programmability, but the complexity spike scares off 90% of developers. Similarly, open-weight models offer flexibility, but the devil is in the deployment complexity. Enterprises still pay for reliability, not just price.
Three core insights from the OpenRouter data:
First, the majority of open-weight model calls are likely from low-value use cases — experimentation, academic research, personal projects. These users have low customer lifetime value. Closed models retain high-spending enterprises.
Second, the 100 trillion figure masks a concentration of winners. Llama 3.1, Mistral Large, and Qwen 1.5 dominate. Just as in crypto where a few Layer 1s capture most TVL, a few open-weight models capture most traffic. The rest are ghosts.
Third, the real battle is not model vs model — it's ecosystem vs ecosystem. Open-weight models enable a flywheel: more models → better tools → more developers → better models. This mirrors the Ethereum vs 'Ethereum killers' dynamic. The killer isn't a single chain; it's the entire L2 ecosystem. Similarly, open-weight isn't a single model; it's a constellation of fine-tuned variants.
Contrarian: The Decoupling Thesis That Everyone Misses
The headline 'open-weight models are eating the market' implies a zero-sum game. I disagree. Total AI token consumption is exploding so fast that both open and closed can grow in absolute terms. The real disruption isn't cannibalization — it's commoditization of the model layer.
The protocol isn't the product. The liquidity is. In crypto, we learned that the value isn't in the smart contract; it's in the liquidity pools. For AI, the value isn't in the model weights; it's in the data moats and application interfaces. Open-weight models lower the barrier to entry, but they also compress margins for everyone. This is a classic race to the bottom.
Consider: If every startup can use Llama for free, where is the competitive advantage? It shifts to unique data, user experience, and brand. The same happened with blockchain — permissionless access commoditized ledgers; value flowed to applications (Uniswap, Aave) and infrastructure (Lido, Chainlink).
My contrarian take: The OpenRouter study is a marketing document, not a neutral observation. Just like Crypto Briefing's coverage often favors decentralization narratives, OpenRouter wants to attract more model providers and users to its platform. The 100 trillion figure is impressive, but without segmentation by revenue, it's noise.
Moreover, closed models are fighting back. OpenAI dropped prices 80% in 2024. Anthropic introduced Haiku for cheap inference. Next-generation models (GPT-5, Claude 4) could widen the performance gap again. Decentralization is a spectrum. Most projects are on the centralized end. Open-weight models still depend on centralized data centers and corporate sponsors (Meta, Microsoft). They aren't permissionless in practice.
Takeaway: How to Position for the Next Cycle
We're entering a phase similar to 2022 bear market consolidation for AI. The hype around open-weight models will peak, then reality check. The winners will be those who build sustainable unit economics — not just traffic.
From my experience restructuring our research framework during the 2022 crash, I know that resilience metrics matter more than growth metrics. For AI, look at: model hosting margins, customer churn, and the cost of serving one query. For crypto AI projects (e.g., Bittensor, Render, Akash), the open-weight trend is a double-edged sword: more demand for decentralized inference, but also more competition from centralized and cheaper alternatives.
Most 'communities' are just coordinated bag-holders. Don't get caught in the meta. The smart money will rotate from model providers to infrastructure and application layers. Watch for projects that provide differentiated inference (privacy, specific verticals) rather than generic compute.
My final takeaway: The OpenRouter study is a signal, not a verdict. The market is growing, but the structure is fragile. Leverage doesn't create wealth. The next margin call will come when investors realize that 100 trillion tokens don't equal 100 trillion dollars in value. Position accordingly.
— Avery Wilson