Fireside Chat, Interview, Conference Presentation
"Is there an AI bubble?” Gavin Baker and David George
- Gavin Baker predicts the current AI environment is unlikely to resemble a 2000-style telecom bubble, citing the lack of "dark GPUs," while forecasting U.S. data center capacity additions of three to four trillion dollars over the next five years.
- He expects positive returns on invested capital for major spenders on Blackwell GPUs despite ongoing debates, and notes that reasoning capabilities have fundamentally altered frontier model economics to unlock flywheels for companies with large user bases.
- Baker views claims that scaling laws are ending as incorrect regarding GPT-5, clarifying the model was designed for economy rather than superior capability, and asserts AGI arrival is likely on shorter timelines than the ten-year estimate.
- Application SaaS companies serving fragmented SMBs may succeed by accepting declining gross margins similar to the Microsoft and Adobe cloud transition, while public coding firms can compete with Cursor by running AI products at breakeven alongside profitable legacy businesses.
- A significant portion of high-profile ASIC programs are forecast to be cancelled within the next three years, particularly if Google externalizes TPU sales, leading to a market battle primarily between Google's TPU and NVIDIA where most ASICs will ultimately fail.
- Baker suggests AI-native companies launching their own browsers risk missing the opportunity to leverage Chrome's established 5 billion user base, and anticipates a shift toward pricing models based on task resolution where verified rewards are measurable.
- Future labor structures may evolve to pay humans for outcomes rather than hours as AI augments or replaces labor in outcome-based services, with affiliate fees becoming standard revenue streams for AI personal assistants acting as shopping portals to eliminate ad impression inefficiencies.
- He deems the concept of work becoming optional as not implausible given technological potential and identifies humanoids as the inevitable winner in robotics due to their ability to learn from video content, effectively ending the debate on robot morphology.