OpenAI is making its own AI chips with Broadcom's help (www.engadget.com)

OpenAI has expanded its hardware push by partnering with Broadcom to design custom "AI accelerators" and rack systems that Broadcom will deploy in OpenAI’s infrastructure and partner data centers starting in H2 2026, with rollout targeting completion by the end of 2029. The agreement covers roughly 10 gigawatts of chips and is worth “multiple billions” of dollars; the companies began collaborating about 18 months ago. This deal joins recent supply pacts with NVIDIA (10 GW and a reported $100 billion strategic relationship), AMD (6 GW, for which OpenAI is paying tens of billions and may take a stake), and Oracle (4.5 GW of data-center capacity), as OpenAI races to secure large-scale compute. Technically and strategically, the Broadcom tie-up signals OpenAI’s move toward vertical integration and hardware–software co-design: custom accelerators and engineered racks can be optimized for OpenAI’s model architectures, memory, interconnect and energy profiles, improving performance per watt and scaling efficiency. It also diversifies suppliers beyond NVIDIA/AMD, reducing single-vendor risk as OpenAI pursues CEO Sam Altman’s ambitious 250 GW target over eight years (versus ~2 GW today)—a goal that would require unprecedented capital (estimates as high as $10 trillion) and new financing models. For the AI/ML community, this underscores an arms race in bespoke silicon, infrastructure scale, and the economics of training ever-larger models.