Interview, Fireside Chat
Factory's Matan Grinberg: The Coming ‘Dark Factory’ Where Software Builds Itself
- Matan Cohen, co-founder and CEO of Factory, rejects Amazon's "customer obsession" as an input metric, arguing that true success is an output metric where customers become obsessed with the product itself.
- Factory was founded in April 2023 with a vision of fully autonomous agents, a strategy Cohen describes as being "two years too early," resulting in a period where the market and enterprise procurement teams were not yet ready.
- In a decisive move driven by the principle of building "obsessed customers," Factory proactively refunded all revenue from early enterprise customers despite reaching nearly $2 million in revenue, acknowledging the product was not yet delivering sufficient value.
- The company identifies "model independence" as its core competitive differentiator, allowing enterprises to avoid vendor lock-in with single providers like Anthropic or OpenAI and enabling the hot-swapping of models for cost or performance optimization.
- Factory launched the "Droid CLI" on September 26, 2025, marking a pivot from fully autonomous agents to tools that meet developers where they are, aligning with established workflows like autocomplete and chat interfaces.
- Cohen observes a market shift from "token maxing" (unrestricted AI usage) to "cost rationalization," where enterprises dynamically route tasks to different models based on complexity rather than using a single frontier model for all tasks.
- Factory's "Model Router" now handles dynamic task routing, utilizing open-weight models (e.g., GLM 5.2) for lower-stakes tasks while reserving frontier models (e.g., OpenAI o1, GPT-5.6) for critical decision-making.
- Token distribution statistics show a rapid increase in open model adoption: open models accounted for less than 1% of tokens at the start of the year, single digits in Q1, and have now crossed into double-digit percentages.
- Factory is moving toward an outcome-based business model where tasks are priced by results rather than usage, creating a marketplace where model providers bid to complete specific tasks within validation criteria.
- The "software factory" concept involves codifying tribal knowledge and assembly-line processes, aiming to shift from vague "vibe-based" decision-making to mathematical optimization of human labor versus token allocation.
- Successful enterprise transformations often rely on company-wide hackathons that allow teams to fail safely, whereas top-down mandates from boards often lead to poor adoption rates.
- Future projections indicate a shift from synchronous AI usage to asynchronous autonomous "dark factories," where agents independently detect signals, fix issues, and build solutions without human initiation.
- Cohen predicts a turbulent 12-24 month correction phase where misallocated engineering resources are reallocated, but foresees a long-term net positive where engineers solve previously unsolved problems in sectors like government software and pharmaceutical research.
- The company's internal token economics show that open models are increasingly competitive with "frontier minus one" models, offering comparable performance at lower costs and faster speeds.
- Factory's architecture ensures that all automations and artifacts created by agents reside within the customer's codebase, preventing the vendor from retaining "tribal knowledge" or creating dependency on proprietary assets.
- Cohen notes that model-harness co-design is inferior to building a robust, model-agnostic harness that generalizes performance across multiple models, avoiding overfitting to a single provider's nuances.