‘Selling coffee beans to Starbucks’ – how the AI boom could leave AI’s biggest companies behind (techcrunch.com)

The article argues that foundation models — the massive pretrained models at the core of today’s AI — are becoming a commodity for many startups. With pre‑training scale showing diminishing returns, companies are increasingly winning by fine‑tuning, reinforcement learning, retrieval augmentation, and product/UI work rather than by pouring billions into bigger base models. At conferences and in practice, startups treat GPT, Claude or Gemini as interchangeable backends, focusing instead on domain specialization (coding tools, enterprise data, image apps) where post‑training customization and interface design drive value. That shift risks turning big labs into low‑margin “coffee‑bean” suppliers if application‑layer firms or open‑source alternatives capture customers and pricing power. For the AI/ML community this matters both commercially and technically. It redirects attention and investment toward fine‑tuning pipelines, instruction tuning, RLHF, embeddings/retrieval, MLOps for model switching, and UX that embeds models into workflows. It also reframes risk for hyperscalers — brand, infra and capital still matter, but they may not guarantee a durable moat. Breakthroughs toward AGI or domain‑specific wins could reverse this trend, but in the near term the competitive battleground has moved from raw pre‑training scale to how effectively teams adapt and productize models.