The Open-Source AI Reality | How Token Costs Will Fall 10X & Usage Will Explode 100X | Lin Qiao

20 insights saved from this video by @technology

@technology
Companies will build the intelligence-stack components that create unique value and share the components that provide common knowledge.
@technology
One person can potentially turn an idea into a scalable product in weeks instead of dozens of engineers working quarters.
@technology
With small amounts of proprietary data, open models can outperform general-purpose models on a company’s unique problem.
@technology
AI’s next discipline is maximizing return on investment, not token usage. Cost and attribution become decisive.
@technology
Frontier APIs cannot be deeply customized. Open models can be tuned and optimized for specific workloads.
@technology
Workload-specific optimization can cut inference costs roughly five to ten times. At billions of users, even five percent matters.
@technology
The infrastructure bottleneck is not just HBM or chips. It is the missing integrated system for serving extremely large models.
@technology
Token prices should fall when shortages attract competition, lowering infrastructure costs and making AI usage explode.
@technology
Companies gain AI control by building on open models and converting proprietary data into their own intelligence.
@technology
The frontier may be a routing system that assigns each task to the model best suited to its complexity.
@technology
AI is expanding access to creativity for nearly everyone, creating unusually broad demand and unusually rapid company growth.
@technology
Rapid growth became manageable after hiring people with high velocity, extreme ownership, strong communication, and refusal to accept no.
@technology
Customized models tailored to each workload create value beyond one-size-fits-all deployment, even at higher prices.
@technology
Open models already run in production for recruiting, finance, coding, debugging, and internal agents.
@technology
Product-line optimization comes after the portfolio and use cases are established, not while creating continual differentiation.
@technology
Extreme early adopters reveal what a mature product must support, even when they are not the ultimate target.
@technology
Technology adoption moves from deep expert control toward accessibility and less control for broader users.
@technology
AI infrastructure remains constrained for another one to one-and-a-half years, while costs should compress over two to three years.
@technology
General intelligence should not belong to one company. Common intelligence and specialized intelligences serve different purposes.
@technology
Nvidia's model work addresses a U.S. supply-chain blockage: without a U.S.-native open model, the AI stack cannot flow.

More from 20VC with Harry Stebbings

See everything from 20VC with Harry Stebbings

Post on X

Keep what you found in The Open-Source AI Reality | How Token Costs Will Fall 10X &…

A free account saves any of these cards to your Collection, and Korva resurfaces them so you actually remember.