lin qiao

The AI stack is stratifying: chips (Nvidia) → infrastructure (Fireworks) → models (open/closed) → applications. Fireworks occupies the specialized intelligence platform layer, offering per-workload model tuning + inference optimization (quality, speed, cost) rather than commodity inference. This layer captures value as enterprises demand control over their intelligence.

Open models have crossed a quality threshold where they solve 90% of enterprise problems at 15x lower cost; they are far easier to tune with small proprietary datasets, letting companies "hill climb" to better-than-generalist performance on specific tasks — driving a structural shift from renting closed APIs to owning customized intelligence.

As AI moves to production scale, companies hit cost walls with closed APIs (CFOs block rollout). The solution: own open-weight models, customize with proprietary data, control inference. This mirrors the SaaS era where every company built its own software stack — now every company will own its intelligence stack.

Coding agents (Cursor et al.) proved the model; now "co-work" agents explode across verticals: general deep research, legal, finance, customer support, recruiting, sales, marketing, healthcare. Consumer-facing companies also adopt genAI for recommendation transformation. This diversification reduces platform concentration risk.

Token costs will drop 10x over three years as supply constraints (energy, chips, manufacturing) ease, triggering 100x usage explosion. Current 30-40% gross margins reflect hypergrowth reinvestment, not structural limits. The "capex bubble" narrative ignores the demand pull from drastically cheaper intelligence.

Next-gen inference requires heterogeneous hardware: flops-intensive GPUs for prefill + SRAM accelerators for generation. This demands unique data center design (liquid cooling, power, fiber, maintenance) — not commoditized. Nvidia's Groq acquisition validates this. Operational excellence in heterogeneous deployment becomes a competitive moat.

Just as countries need sovereign power grids, they need sovereign base model infrastructure. Lin Qiao warns that reliance on a single foreign provider (e.g., US labs for Europe, or Chinese models for US) creates existential cutoff risk. This drives investment in domestic open-model ecosystems and on-premise inference.

Nvidia now releases three GPU SKUs per year; model generations peak in value before the next release. This compresses effective hardware depreciation from 6 years to ~2-3 years, changing build-vs-buy economics for data centers and favoring rental/elastic capacity over ownership in early stages.