AI and Open Source: Shaping Each Other's Future | RAISE Summit 2026

Panel Composition & Context:

• The session features leaders from Grafana (Raj), Mozilla (Marc), Ivan (Oskari), Comet (Gideon), and Together AI (Charles), moderated by Olivier (Comet's founding partner).
• The core theme explores the symbiotic relationship between AI and open source, specifically focusing on definitions, licensing, and economic models.

Definitions: Open Source vs. Open Weight:

• Marcelo (Mozilla) distinguishes "open weights" (e.g., Mistral, GLM) as models allowing usage, modification, and fine-tuning but lacking transparency into pre-training data and full code.
• True "open source" models, per the panel, require full transparency of training data, code, and the ability to pick up at any checkpoint, with Nvidia's Nemotron cited as a closer approximation to this standard.
• Gideon (Comet) and Raj (Grafana) note that for many use cases (like agent observability), the specific label matters less than the practical ability to run models privately and evaluate data without sending it to a provider.
• Oskari (Ivan) insists on strict adherence to the OSI definition: software must be usable, modifiable, and buildable without reliance on external entities, arguing that "open weight" models should not be mislabeled as "open source."
• Charles (Together AI) counters that economically, "open weights" often serve the same function as open source by allowing derivatives and unique IP generation (e.g., Decagon extending base models), even if the upstream pipeline is closed.

Licensing Strategies & Commercial Viability:

• Raj (Grafana) explains the decision to switch from a permissive Apache license to the AGPL to prevent cloud hyperscalers from offering the software as a managed service without contributing back, thereby protecting the company's value capture.
• The shift to AGPL was driven by the need to balance "value creation" (open source altruism) with "value capture" (commercial sustainability) against competitors who previously exploited permissive licenses.
• Oskari (Ivan) contrasts AGPL with "source-available" licenses like SSPL, arguing that true open source must allow for viable forks and community control if a project becomes proprietary, citing the eventual return to open source by projects like Elasticsearch and Redis.
• The panelists agree that while permissive licenses (MIT/Apache) facilitate rapid adoption, restrictive licenses (AGPL/copyleft) may be necessary for startups to survive against well-funded hyperscalers.

Economic Models & Value Proposition:

• Gideon (Comet) identifies the primary challenge for open source companies as balancing the free community version with monetized features, noting that successful firms must provide enough value to the free tier to drive adoption while restricting critical enterprise capabilities.
• Charles (Together AI) argues that their competitive moat is not IP protection but a continuous research pipeline that outperforms hyperscalers in model quality, inference speed, and cost, decaying only by innovation cycles of 6–9 months.
• Charles emphasizes that users often pay for "trust" and the optimization of complex infrastructure rather than just feature unlocks, creating a "freemium" model based on reliability rather than coercion.
• Olivier (Comet) notes that investors seek founders who can sustain a system where the company creates more value than it captures, ensuring long-term R&D funding through a balance of open contribution and closed commercialization.

Impact of AI on Developer Activity:

• AI acts as a net positive for project velocity by filtering automated PR noise and testing environments, though it creates an "identity crisis" for maintainers who miss the craft of manual coding.
• Panelists observe a cultural shift where developers move from direct contribution to reviewing agent-generated code, requiring new skills in filtering and validation.
• There is a noted rise in "entitlement" from large corporate contributors (e.g., Microsoft, IBM) who expect maintainers to address issues, causing burnout in non-commercial projects.
• AI lowers the barrier to forking and maintaining derivatives, potentially revitalizing the ecosystem by enabling more users to run custom, production-grade forks, though it also increases the volume of low-quality branches.

Future Outlook (5-Year Horizon):

• The consensus predicts open source and AI will remain "best friends," as AI models trained on open source code increasingly recommend and generate open source solutions.
• Marcelo notes that models trained on open source data are more likely to suggest open source infrastructure, creating a self-reinforcing cycle of adoption.
• Oskari envisions a "hybrid ecosystem" where open source dominates for flexibility and cost, while closed models persist for specific, high-end use cases.
• Gideon highlights a potential tension: while open source is the technological best friend for AI infrastructure, it poses a significant economic threat to large AI players by commoditizing the underlying stack.
• Charles adds that for future developer infrastructure, open source remains the primary discovery mechanism, as AI agents will be trained to explain and use open source tools to millions of developers.