Interview, Fireside Chat
Factory's Matan Grinberg: The Coming ‘Dark Factory’ Where Software Builds Itself
- Enterprises are expected to transition from no routing to active, opinionated model allocation within the next six to 12 months.
- Forecasts indicate that within 12 to 24 months, 90% of AI tokens in software development will be asynchronous tokens generated by autonomous agents rather than human-initiated interactions.
- A critical allocation decision for CIOs regarding incremental spending between human headcount and AI tokens will soon require quantitative feedback loops.
- The industry is projected to face a necessary correction for current resource misallocation and bloat within the next few years.
- Current "token maxing" adoption patterns will shift toward cost rationalization and dynamic routing as organizations curb usage regardless of cost.
- By the 2030s, the market for software development tools is anticipated to move from usage-based metrics to outcome-based pricing models.
- Future token consumption is expected to be directed primarily toward open-source models, though frontier intelligence will remain essential for high-stakes decision-making.
- The market share of open models is projected to asymptote toward the vast majority of token volume due to cost-effectiveness and optionality.
- Long-term reallocation will move engineers from low-value tasks to solving complex, previously unsolvable problems in sectors like government systems and pharmaceutical research.
- The current era of software development lacking rigorous accounting for token and human labor is predicted to appear archaic compared to standards 10 years in the future.
- Organizations adopting a "fail fast" mentality and company-wide hackathons are expected to navigate AI transformation successfully, whereas those insisting on perfection will struggle.
- Software development is predicted to evolve into "dark factories" where autonomous agents operate continuously with minimal human intervention.
- Model-agnostic harnesses will be superior to model-specific co-design to avoid overfitting to individual model nuances.
- Competitive dynamics will require providers to demonstrate they are not a single point of failure, emphasizing the necessity of model independence to prevent vendor lock-in.
- The performance distinction between open and frontier models is expected to blur, with open models reaching parity with frontier models from a prior generation, such as GLM 5.2 matching Opus 4.7 or GPT 5.5.