I enjoyed reading the president of the Mozilla Foundation’s op-ed in today’s FT: Open source could pop the AI bubble — and soon.
I don’t think open source software will be the catalyst for popping the AI bubble[^1], however I do think (a) it is one of the challenges to the current narrative that is fueling AI valuations, (b) the capital cycle will turn, and (c) we’re closer to the peak of the bubble than the beginning of it.
A non-exhaustive list:
- Limits to productive capital absorption
- Peak scarcity pricing for GPUs (supply catches up to demand)
- Technology companies transitioning from asset-light businesses to capex-heavy businesses
- Grid access constraints driving datacenter operators toward behind-the-meter generation, where turbine and transformer supply are severely constrained
A few random thoughts/prognostications:
- Hardware/infrastructure layer: NVIDIA GPUs and Google TPUs become commoditized; hyperscaler compute (Azure, AWS) faces margin compression but enjoys enterprise lock-in with high switching costs
- Systems software: CUDA (NVIDIA) lock on the market weakens as hardware-agnostic options mature; potentially 18-36 months before the marginal buyer has options and the ability to deploy across manufacturers
- Data layer: Strategic / trapped value lies in companies’ proprietary unstructured data lakes; question: does being a custodian of data (e.g., Box, Snowflake) translate to capture rights for AI training / inference, or do enterprises maintain separation b/w storage and intelligence layers?
- Frameworks / tooling: Have no thesis
- Model layer: OpenAI and Anthropic are very strong, but will face stiff competition from open-source models per the FT piece; my core thesis is that focus shifts from prioritization of compute to inference optimization (cost, latency, energy efficiency), and that open source models deliver ‘good-enough’ performance for general use cases within 18 months; as model quality converges, switching costs drop; OpenAI / Anthropic don’t capture full TAM, particularly ex-US where local models and data sovereignty concerns matter [^2]
- Serving / APIs: Solid moat for serving the latest models to technology-forward enterprises and startups willing to pay for API access and managed infrastructure
- Application layer: Opportunity for differentiation through brand affinity and UX, but unclear whether application-layer companies can capture value or will be marginalized as commoditized front-ends to model APIs
I am pondering three questions (not financial advice!):
- Do I own enough Alphabet? (I don’t think I do.)
- Is Apple the dark horse in the competition? I think so but haven’t acted on it. (Apple Silicon enables on-device inference, privacy positioning, and ecosystem lock-in, potentially inverting the centralized API model; retains flexibility on model layer, can partner with different providers or deploy proprietary models with each OS release)
- When do training datasets get exhausted? I am over my skis on this, but am enjoying the Dwarkesh discussion with Andrej Karpathy
[^1]: I am speaking to the hardware and model layers. I think there will be many valuable businesses built on AI tooling that solve acute pain points within verticals. I also think there will be an explosion of software necessitating a greater variety of cybersecurity solutions.
[^2]: I’ve run multiple local models on a Mac; they’re slower and less refined than frontier models, but improving rapidly. I can envision free local access to today’s best-in-class capabilities by year-end 2027.