Jack Collier, Chief Growth & Marketing Manager, IO.Net: Why Open Models Need Open Infrastructure

The Paradox of Open Models and Infrastructure

• Current AI development faces a paradox where superior open-source models are losing ground to proprietary alternatives despite their technical capabilities.
• The speaker attributes this adoption gap not to model quality, but to a restrictive infrastructure ecosystem that hinders open model usage.
• Centralization of AI control and compute power among a few providers is identified as a primary risk, leading to potential data abuse and power disparities.

Barriers to Open Source Adoption

• Compute Supply Constraints: Global demand for compute is growing so rapidly that supply must double every six months just to maintain parity.
• Market Concentration: Sixty-five percent (65%) of global compute capacity is controlled by only three providers, creating an artificial bottleneck for smaller entities.
• API Dominance: Seventy percent (70%) of current model usage relies on closed-source APIs, compared to only 30% for open-source models.
• Developer Intent vs. Reality: A recent survey indicates that 83% of developers prefer building with open-source models but are deterred by a lack of usability and tooling.
• Security Risks: Reliance on closed systems forces users to surrender data ownership and lack transparency regarding how their information is utilized.

Addressing the Three Fallacies of Open Models

• Cost Fallacy: The misconception that open models are expensive is countered by distributed compute networks that aggregate idle GPUs globally.
  • Cost comparison indicates open infrastructure can reduce pricing from approximately $15 per million output tokens (closed) to roughly $3.50.
• Quality Fallacy: The belief that open models cannot match proprietary performance is challenged by LM Arena data showing open models competing effectively across various verticals.
  • This scoring system relies on over 3 million user votes across more than 200 different models.
• Usability Fallacy: The perception that open models require complex, disparate tools is being resolved by unified platforms offering end-to-end environments.
  • These solutions enable seamless fine-tuning, inference, and RAG implementation without gated access or permission checks.

io.net Infrastructure and Capabilities

• Network Scale: io.net operates a distributed compute network spanning 1,38 countries with over 1,000 GPUs available on demand.
  • The platform supports diverse environments including Ray, MegaRay, bare metal, and containers.
• Transparency Metrics: All network supply and earnings data are recorded on a blockchain for verification.
  • The network has generated approximately $15 million in earnings with over 10,000 active GPUs currently operational.
• Intelligence Layer Usage: The infrastructure currently processes over 1 billion output tokens per week for more than 5,000 users.
• Service Expansion: A new "Training as a Service" feature allows users to link Hugging Face models with datasets for fine-tuning.
  • Launching later this week, the service offers the first five trainings free of charge.

Strategic Incentives and Future Outlook

• Ecosystem Benefits: Open infrastructure grants users full data ownership, enabling innovation without gatekeeping by large hyperscalers.
• Developer Incentives: io.net is currently hosting an inaugural hackathon with QR code registration and offers startup cloud credits.
• Market Trend: The speaker forecasts a shift toward permissionless, distributed infrastructure that democratizes access to high-performance AI compute.