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

An open ecosystem draws many contributors and alternative providers, which accelerates innovation and reduces the chance that any single company will monopolize the infrastructure.

Prioritizing strict margin targets too early forces resource constraints that curb experimentation and speed, which slows innovation during a hypergrowth phase.

Fast model releases and frequent new hardware SKUs make older chips lose value quickly, which shortens the effective economic life of infrastructure and makes build-versus-buy timing much more risky.

New technology adoption starts with hackers who demand control, but as the market broadens most users lack deep expertise and need simpler, automated tools, which forces products to trade knobs for strong defaults.

Even with product market fit, exploding inference and deployment costs can sink a company, so businesses must control model and infrastructure economics or risk 'scaling to bankruptcy.'

Leaders need constant ground-level awareness because information degrades across organizational layers, and fresh, high-volume feedback is what lets them make correct rapid decisions in fast-moving environments.

Custom chips only pay off after workload patterns stabilize because tape-out is costly and hard to change, so premature hardware bets risk being mismatched to rapidly evolving applications.

Each company solves unique problems with unique data and workflows, so owning and evolving in-house intelligence becomes necessary to optimize alignment and performance in ways a rented generic service cannot.

High inference prices attract competitors and engineering effort, and as chips, serving software, and model efficiency improve, tokens per task fall and overall costs compress, enabling much larger usage.

Splitting reinforcement learning into a trainer that produces new weights and distributed rollout fleets that collect rewards lets teams sync fresh models across many data centers, which avoids the need for one ultra-dense cluster while keeping feedback timely.

Products, regions, and companies have unique tastes, policies, and proprietary data, so intelligence will fragment into millions of specialized models tuned to each context rather than one universal AGI.

At production scale, per-request API costs and unpredictable billing make renting models financially risky, so owning models gives companies the control needed to optimize cost, quality, and deployment behavior.