matan grinberg

Model providers (API businesses) are incentivized to maximize token usage, while independent application layers align with enterprise cost/quality/speed optimization. Value accrual is time-dependent and shifts between layers; forced bundling leads to vendor lock-in and stagnation like cloud.

Enterprises fear dependence on single model providers due to pricing volatility and lab instability; a routing layer that dynamically selects models per task (caching, compaction, tool use) delivers better performance than co-designed model-harness pairs by avoiding overfitting to model-specific quirks.

Models like GLM 5.2 now match last-gen closed models (Opus 4.7, GPT 5.5) at lower cost and latency; Factory already routes 50% of tokens to open models and expects asymptotic shift toward open-model dominance for implementation tasks, reserving frontier models for high-leverage decisions.

Current AI coding tools are synchronous (human-initiated); the shift to asynchronous agents that autonomously detect signals and generate solutions will unlock true 'dark factory' software development, with 90% of tokens becoming async.

Three-phase adoption: board pressure → token maxing (AI at all costs) → hangover (ROI scrutiny). This drives demand for intelligent routing, cost controls, and nuanced resource allocation per team/individual.

Open models are a critical counterbalance for cost/quality/speed tradeoffs. Enterprises will route most tasks to cheap open models, reserving expensive frontier models for high-leverage planning/decision tokens. This caps frontier pricing power.

Factory's router creates a de facto marketplace where model providers bid on tasks with validation criteria; this dynamic shifts pricing from token consumption to verified outcomes, aligning incentives and enabling enterprises to allocate capital between headcount and tokens based on measured leverage.

CIOs will soon decide marginal dollar allocation between headcount and tokens per org unit using quantitative feedback loops; software factories that codify tribal knowledge and close the loop on feature outcomes will replace 'vibes-based' resource allocation, making software development as measurable as model training.