Thoughts on the current (un)profitability of the AI ecosystem (johnsillings.com)

The piece argues that widespread claims of an AI “house of cards” — a fundamentally unprofitable ecosystem destined for a dramatic collapse — are overstated. Instead, the current unprofitability resembles earlier scale-driven industries (ridesharing, streaming, cloud, semiconductors) where investors funded aggressive growth to capture share, many players failed, and a few incumbents emerged. The takeaway for the AI/ML community: unprofitability during rapid expansion is normal and often resolves through consolidation rather than catastrophe. Technically, the more important story is falling input costs: cheaper and more efficient model architectures, better model orchestration and serving, faster/denser GPUs, and training-amortization across larger deployments all improve unit economics. The most likely outcome is incremental and uneventful—some price increases, continued reductions in backend costs, and winners that capture dominant share—rather than a single explosive unwind. For practitioners and startups, that implies focusing on cost-efficiency, scalable model operations, and defensible distribution rather than assuming the market will suddenly collapse.