DEVOURED
Meta stopped releasing Muse Spark as open weights entirely, and Kimi now requires apps over 100M MAU to display 'Kimi K2.6' branding in their UI, signaling the end of open model licensing as a check on frontier lab pricing.
What: Martin Alderson argues that open weights models (Llama, DeepSeek, Qwen, Gemma) have kept frontier lab pricing in check by offering inference at under 10% of frontier model token costs. Recent trend: Meta no longer releases Muse Spark weights, Alibaba releases models API-only, Kimi K2.6 requires 'Kimi K2.6' branding for apps over 100M MAU or $20M/month revenue, Mistral added commercial restrictions. Only DeepSeek moved toward more permissive licenses.
Why it matters: Open weights models function as a pricing floor similar to generic pharmaceuticals—their mere availability prevents frontier labs (OpenAI, Anthropic, Google) from extracting the full consumer surplus developers would pay. If training costs force consolidation into a handful of Western and Chinese 'superlabs' with restrictive licenses, the AI market could shift from competitive pricing to oligopoly pricing, with labs capturing the gap between cost and what users would actually pay (author notes he'd pay 10x current prices).
Deep dive
- Open weights models (where trained model weights are public but not necessarily training data) have been a 'load-bearing assumption' underneath AI pricing, allowing users to run models privately on-prem or via cheap hosted providers at <10% frontier token costs
- Three core advantages: privacy/compliance for sensitive data that can't leave the network, flexibility for fine-tuning and quantization, and dramatic cost savings
- Licensing trend is tightening: Meta's Muse Spark not released as open weights at all, Alibaba prioritizing API-only releases, Kimi K2.6 requiring UI attribution above scale thresholds ($20M/month revenue or 100M MAU), Mistral imposing commercial restrictions
- DeepSeek is the notable exception, moving toward more permissive terms
- Economic mechanism: open weights provide contestable market pressure even without high adoption—the threat of switching disciplines frontier lab pricing behavior, similar to generic pharma forcing brand-name price cuts
- Author's concern: as training costs rise, consolidation into 3-5 players (big Western labs + state-backed Chinese 'superlab' mergers, similar to CRRC rail consolidation) would eliminate this price ceiling
- Consumer surplus at risk: author personally would pay 10x current prices, professional/agentic use cases have even wider value-cost gaps that an oligopoly could capture
- Distillation as release valve depends on having strong base models in the first place, which is exactly what's eroding
- Counter-argument: faster hardware and easier 'good enough' model training could maintain competition despite consolidation in frontier training
- Hardware market (GPUs/TPUs) shows intense competition despite few manufacturers, suggesting oligopoly isn't inevitable
- Core thesis: open weights erosion has 'enormous implications for the wider economy' if frontier labs gain pricing power to extract currently-unrealized consumer value
Decoder
- Open weights: AI models released with trained parameters public but not training data or methodology. Distinct from 'fully open' or 'reproducible' models (which include all training artifacts) and traditional open source software.
- Distillation: Training smaller AI models on outputs from larger frontier models to transfer capabilities at lower computational cost, though still requires access to a capable base model.
- Consumer surplus: Economic term for the gap between what consumers would pay for something versus what they actually pay—the 'unrealized' value that competitive markets leave with consumers rather than extracting as profit.
Original article
Open weights models allow anyone to run the model on their own hardware. This allows model use to be private, flexible, and low-cost. Many companies will run open weight models at significantly less cost than the cost of frontier models per token, and the performance gap isn't that big. However, the increasing costs of training mean that more and more models are being released under tighter license conditions. The competitive open weights ecosystem may soon be at an end, and this has enormous implications for the wider economy.
Explore all 73 articles in today's digest
Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi