Fireside Chat, Interview
Fireside Chat with William Falcon & Ozan Kaya of Lightning AI | RAISE Summit 2026
- Voltage Park's Origin and Scale: The company began in early 2024 after a nonprofit acquired 24,000 NVIDIA H100 GPUs in 2023, leveraging a "Black Friday" procurement strategy to secure front-of-line access.
- Rapid Team Expansion: Ozan joined with a team of three, scaling operations to 160 employees to manage the complexity of deploying 24,000 GPUs across six distinct data centers by late 2024/2025.
- Infrastructure Strategy: Voltage Park operates at 25–70 megawatts of power across six locations (avoiding single-site concentration) to serve the "Gen AI" sector rather than hyperscalers requiring gigawatts, recently signing a lease in Canada to expand North American presence.
- Customer-Centric Service Model: The company targets a "Four Seasons" white-glove service standard, prioritizing 99.99% uptime to support high-stakes customers like Cursor, who moved their operations from other partners in late 2024.
- Direct-to-Enterprise Focus: Unlike competitors concentrating on hyperscalers, Voltage Park's thesis was to democratize compute access for Gen AI startups and enterprises, betting on the trend of companies wanting to own their full AI stack via open-source models rather than relying on closed APIs.
- Merger Rationale: Voltage Park merged with Lightning.ai to move "up the stack" from pure GPU-as-a-service to a full software platform, combining Voltage Park's compute infrastructure with Lightning's enterprise-grade tooling.
- Lightning.ai's Capabilities: Founded by Will (former Facebook AI researcher and creator of PyTorch Lightning), the software platform enables enterprises to train, fine-tune, and deploy models without requiring PhD-level engineering teams.
- Strategic Timing: The merger occurred after both founders identified a market gap: enterprises were outgrowing POCs and requiring dedicated compute for larger training runs, a need that hyperscalers like AWS and GCP were failing to meet cost-effectively for open-source workflows.
- Open Source and Digital Sovereignty: The partners argue that "owning intelligence" is critical to avoid censorship and dependency on single providers (e.g., Anthropic, OpenAI), citing recent restrictions on model access as a catalyst for the shift toward local and open-source models.
- Hybrid Ownership Model: The business advises a "rent then own" strategy: startups should use APIs like OpenAI for bootstrapping, but once product-market fit and proprietary data are established, they should fine-tune and own their own models.
- Long-Term Economic Thesis: The speakers predict that hardware costs will commoditize similar to the transition from mainframes to iPhones, shifting the primary economic value from compute power to the ownership of data, models, and customer relationships.
- Hardware Evolution: While GPUs (NVIDIA) remain the superior choice for training and current workloads, the speakers foresee a future where inference workloads for older models migrate to diverse hardware, including TPUs, IPUs, and CPUs.
- Decoupling Dependency: Customers retain the flexibility to host models on-prem or use their own compute providers, as the combined entity focuses on tooling and stack management rather than vendor lock-in.