The Twenty Minute VCThe Open-Source AI Reality | How Token Costs Will Fall 10X & Usage Will Explode 100X | Lin Qiao
At a glance
WHAT IT’S REALLY ABOUT
Open-source inference, specialized models, and token economics reshaping AI adoption
- Qiao frames the overlooked opportunity as the “specialized/private intelligence” layer: most valuable data is private inside enterprises, so the winning approach is customizing models on proprietary data rather than relying solely on frontier APIs trained on public internet data.
- He predicts token costs will fall ~10× over three years due to competition and easing supply constraints, and that this reduction will unlock ~100× usage growth as AI becomes a default utility rather than a gated expense item.
- Open models are positioned as crossing a capability threshold: they are now “good enough” for many workflows and, crucially, far easier to steer/tune with small amounts of company-specific data to outperform general-purpose models on targeted evals.
- He challenges the idea that product-market fit automatically implies a durable business in AI, warning that many companies can “scale into bankruptcy” because inference COGS and infrastructure constraints make economics and control central from day one.
- The conversation highlights constraints and strategy across the stack—energy/chips/manufacturing bottlenecks, fast hardware/model depreciation, multi-model routing, and why Fireworks prioritizes agility and specialization over going full-stack into apps or chips (at least early).
IDEAS WORTH REMEMBERING
5 ideasMost economically valuable AI will be trained on private enterprise data, not the public web.
Qiao argues the majority of the world’s data is locked inside applications and enterprises and won’t be shared, making “private/specialized intelligence” the real frontier and a major source of competitive advantage.
Open-weight models change the operating model because they give customers true control.
With weights in hand, companies can tune, add guardrails, and deploy on their own terms; this is structurally different from renting a closed API that embeds another provider’s judgments and cannot be deeply customized.
AI breaks the SaaS rule that product-market fit implies scalability and durability.
In AI, inference costs can grow with usage; Qiao notes real cases of “scaling to bankruptcy,” especially for incumbents with large user bases who can’t afford to roll out expensive AI features without cost control.
Measure token economics by “cost per task,” not price per token.
Different models vary in verbosity and token usage; a cheaper-per-token model can cost the same per outcome if it uses more tokens, so optimization should focus on task-level efficiency and precision.
Customization + deployment rigor becomes a defensible moat (not just model access).
Fireworks emphasizes per-workload deployments (“one size fits one”), and even technical guarantees like bitwise equivalence between training and inference (“zero KLD”) so training spend translates into production quality.
WORDS WORTH SAVING
5 quotesWhat I don't want to see is there's only one company owns intelligence. That doesn't make sense to me.
— Lin Qiao
I do think the cost of token will go down drastically, 10X cost reduction in the next three years, and this 10X cost reduction will drive 100X usage.
— Lin Qiao
Public internet is very small corpus of data compared with world's data. Majority of world's data actually private data locked inside application, locked inside enterprise.
— Lin Qiao
If our future world is gonna be ruled by one standard, a taste dictate by one company, we turn ourself into an army of robots-
— Lin Qiao
Leadership is just judgment. It's not privilege, it's judgment.
— Lin Qiao
High quality AI-generated summary created from speaker-labeled transcript.
