Fireside Chat, Conference Presentation, Interview
Fireside Chat by SambaNova CEO, Rodrigo Liang & Leila Abboud of the FT | RAISE Summit 2024 | Paris
- More companies are expected to integrate AI into daily operations, driving a reinvigoration of the data storage market as enterprises leverage previously unused data to train accurate models.
- Revenue is projected to triple again in the coming year, following a similar tripling last year, while 100 trillion-parameter models are planned for release to Fortune 500 and Global 3000 entities within the current year.
- Samba Nova anticipates competing against NVIDIA by offering a "credible replacement" for pre-training, fine-tuning, and inferencing, with claims of a 10x reduction in cost, power, and energy compared to rivals.
- Market expectations include a stratification of AI infrastructure where hyperscale data center chips are reserved for major players, alongside anticipated chip development efforts by Intel, Amazon, Google, and Microsoft.
- The company intends to ship fourth-generation chips and release new models every few months, supported by a $1.1 billion raised over the first three and a half years to cover engineering and chip development without needing immediate additional funding.
- Strategic capabilities include the ability to host 500 models in a single eight-socket system to achieve multi-tenancy and reduced footprints, while modular model composition allows for asynchronous updates, such as modifying specific Mistral models without retraining the entire system.
- Customers will be able to configure models by adding, deleting, or modifying open-source components with potential changes from 54 to 84 or 24 models within a few weeks, and retain trained models in perpetuity even if subscriptions end.
- Inbound investor inquiries have increased significantly following the Samba 1 announcement, while open source community innovations are expected to reduce the company's need to spend $100 million on model training.