Conference Presentation, Panel
RAISE Summit 2025: The AI Evolution Open Source, Fast Inference, and the Agentic Revolution
- AI agents are projected to enter production faster than larger models due to easier verification, with a wave of consumer personal assistants for tasks like accounting expected to emerge within the next year, potentially starting immediately.
- While proprietary models currently hold a performance lead of six to 12 months over open source variants, cost and accessibility are anticipated to level the playing field as technology stabilizes, fostering a coexisting ecosystem where open source models offer long-term advantages in cost, latency, and privacy.
- The industry narrative is shifting toward a superintelligence race involving the entire stack from compute to agents, driving huge sustained investment and a proliferation of silicon where controlling destiny requires presence at both the frontier model and compute layers.
- Infrastructure constraints are becoming critical, with energy availability and gigawatt-scale data centers acting as bottlenecks, prompting efficiency battles focused on multi-tenancy and hardware footprint reductions from 140 kilowatts to 10 kilowatts to democratize access.
- Enterprises will increasingly adopt open source models to train private data and retain ownership in perpetuity, while companies at the capability frontier continue prioritizing cutting-edge proprietary models for tasks like coding where performance accuracy outweighs cost initially.
- The market is expected to transition from a pre-training regime to post-training and test-time reasoning, with smaller models around three to four billion parameters growing to handle edge data wrangling and enable robotics capabilities like generalization.
- As agentic AI democratizes access, the fleet size for inference compute is expected to grow by orders of magnitude, requiring infrastructure providers to reduce computing footprints and enabling data center players to enter the space as agent factories within 60 days.
- Economic dynamics will shift as "price per token" becomes a primary metric, with many SaaS companies expected to convert into agentic applications or face extinction, while specialized silicon proliferates for device inference in humanoids and physical AI.
- Competitive landscapes will evolve as general knowledge diffusion narrows the gap between top players, though tensions between closed and open approaches will persist, with open source ecosystems likely solidifying into ubiquitous platforms similar to Linux or Windows.
- Future risks include a potential slowing or bottlenecking of frontier improvement rates that could negatively impact the industry, data privacy breaches materializing within the next one to two years, and social concerns arising as AI adoption outpaces government and civil society adaptation.