Broadcom reportedly in early talks to finance OpenAI's custom chip purchases

Broadcom is no longer selling silicon — it's selling balance sheets. The $60B debt facility for Anthropic was the pilot; the OpenAI financing talks are the replication play. AI labs are shedding capex from their own books, converting chips into compute-as-a-service, while Broadcom simultaneously occupies the roles of architect, lender, and integrator. On the technical axis, custom ASICs will capture over 40% of inference workloads from general-purpose GPUs within 18 months. But the real moat has shifted from transistor density to financial structure: whoever controls financing pricing controls compute allocation. Compliance risk is materially underpriced. Once debt instruments are contractually linked to chip delivery, export-control scrutiny extends from goods to obligations. Two of the world's leading AI labs now have their compute lifelines tethered to a single supplier — concentration risk is approaching systemic thresholds. NVIDIA's CUDA moat faces a dimensional attack. The competition isn't happening at the silicon layer; it's being restructured at the capital-structure layer. AMD, lacking both custom-design capability and a financial license, is being squeezed from both flanks. 12–24 month outlook: expect two to three additional chip-financing deals. The AI infrastructure market is transitioning from a semiconductor cycle to a capital-markets cycle.