aravind srinivas

Models are becoming utilities; the durable moat is orchestrating across models, tools, local and server compute to maximize valuable output tokens per unit of energy per user. The orchestrator wins, not the model builder.

Physical constraints — power generation, grid interconnection, cooling, and local permitting — are far exceed digital infrastructure. Resistance to data center build-outs will continue and worsen, limiting supply growth regardless of capital availability.

Reselling raw model tokens is a commodity business with no moat. Value accrues to the layer that grounds models in context, orchestrates across multiple models and tools, and delivers a unified product experience — the 'conductor' of the AI orchestra.

High-bandwidth memory is the binding constraint for GPU cluster scaling; the supplier of the bottleneck component captures disproportionate economic value, making memory vendors more valuable than some hyperscalers.

Perplexity's model-agnostic orchestration — auto-routing to GPT, Claude, Gemini, open-source — minimizes token spend via RAG and avoids context-window bloat, yielding positive gross margins on all revenue.

Denied access to NVIDIA GPUs and HBM, Chinese labs (DeepSeek/Huawei) innovated on KV cache compression, SSD-based inference, and low-interconnect training — a different architectural bet that could prove more efficient and scalable.

Running continuous agent loops purely on server-side tokens is economically infeasible. The winning architecture routes workloads to local models on device for routine tasks, calling frontier models only for high-value reasoning, dramatically lowering cost per useful token.

Perplexity's Mac Mini-based local orchestration layer keeps sensitive data on-prem while bursting to cloud for heavy tasks; enterprises save >$100M via Computer automation, driving faster enterprise than consumer revenue growth.