Interview, Fireside Chat
The Open-Source AI Reality | How Token Costs Will Fall 10X & Usage Will Explode 100X | Lin Qiao
- The industry is projected to transition from a focus on pure code generation to AI-assisted collaboration over the next 12 months, characterized as "The Year of Cowork."
- Token costs are expected to decrease drastically, with a prediction of a 10x reduction over the next three years to drive a 100x increase in usage, while current token counts are anticipated to grow 20 to 100 times by the end of next year.
- By the end of this year, the company aims to at least double its current revenue, having recently reached $800 million in ARR, a growth pace described as unparalleled compared to previous SaaS eras where $1 to $10 million in 18 months was exceptional.
- Strategic focus remains exclusive to infrastructure and models, explicitly excluding the application layer at this time, though entry into data centers is a possibility contingent on operational readiness.
- General intelligence advancement is expected to occur in step functions every year or three quarters, whereas specialization and the proliferation of millions of specialized models for specific use cases will accelerate much faster.
- Market dynamics are predicted to shift from "token maxing" to "ROI maxing" over the next couple of years, necessitating rigorous monitoring of spend versus return as models enter production.
- Supply chain constraints regarding infrastructure are expected to resolve within two to three years, leading to compressed costs and the resolution of issues surrounding US-native open models.
- Physical infrastructure construction in China is noted for its rapid velocity, and model superiority is predicted to become transient, with the dominant model changing frequently within a three-year horizon.
- The greatest current bottleneck is identified as the lack of systems designed for very large models, specifically those with 10 trillion parameters, while hardware innovation is considered the fastest-moving sector.
- A transition is expected from hyper-growth and rapid expansion to a focus on margin optimization once systems are stabilized and fully defined for large-scale deployment.
- Enterprise workloads will increasingly rely on customized models rather than off-the-shelf solutions, with the majority of traffic expected to come from these specialized instances as companies move toward owning their own intelligence.
- Geopolitical considerations may lead to the emergence of sovereign models owned by large nations or blocks to mitigate risks associated with infrastructure dependency and supply chain disruptions.
- Future development will likely involve self-evolving automated systems where routing and tuning are handled dynamically, while chip design will only be pursued once specific workloads are fully stabilized.