Gavin Baker Predicts Hybrid Multi-Model AI Future with Open-Weight Dominance

The most likely future for AI is a hybrid, multi-model landscape where open-weight tokens constitute the overwhelming majority of usage, benefiting infrastructure and applications through increased fragmentation and competition. While frontier labs may see reduced margins compared to earlier expectations, lower costs at the model layer mechanically drive higher compute demand, provided training remains funded.

Compute Economics and Financing

As below-market compute contracts from 2024 and 2025 roll off and reprice higher, cloud and neocloud revenues are expected to accelerate, potentially marking the peak of external financing needs for the current buildout. Data from 3Fourteen Research indicates that B200 clusters could generate a 37% IRR even if rental prices decline by 40% per year, a scenario Baker considers highly fundable given that prices are currently rising.

Regarding recent controversies over Memorandums of Understanding (MOUs), Anthropic’s reported $413 billion in contractual commitments translates to roughly $31 billion in annual cloud spend, which is internally financed by their ARR. The remaining $161 billion in spend at Broadcom will be financed by Broadcom and private credit.

Strategic and Safety Implications

Baker argues that fragmentation at the model layer is the best outcome for human values, safety, and agency, as it prevents a small group of individuals from centralizing control over AI direction. He posits that slowing the buildout is net positive long-term because the market is structurally short of watts, wafers, and capital, reducing the risk of an overbuild or bubble. The most plausible near-term bear case remains a severe cyber event, though AI cyberdefense is ramping up quickly.