OpenAI launches Jalapeño AI chip with Broadcom

OpenAI announced the Jalapeño ASIC—its first proprietary AI chip—on June 24, 2026, marking the company’s move from relying solely on third‑party GPUs to designing its own inference silicon, a shift that could reshape its cost structure and supply chain dynamics [1].

Custom silicon for inference

OpenAI and Broadcom co‑designed Jalapeño as an application‑specific integrated circuit (ASIC) focused exclusively on inference—the repeated, low‑latency execution of trained models. The chip pairs a large compute block with six high‑bandwidth memory stacks, a layout that “cuts the power needed to run the complex, multi‑step workloads” compared with general‑purpose GPUs, according to Broadcom’s own testing [2]. While Broadcom claims roughly a 50 % cost reduction versus typical AI GPUs, the figure is self‑reported and not independently verified, underscoring the strategic rather than purely economic motive: tighter control over the hardware stack that powers ChatGPT and related services [2].

Market impact and competitive context

The move follows OpenAI’s rapid escalation in compute demand, which its president described as “cannot get compute fast enough” and which Broadcom’s CEO said is “simply insatiable” across its six customers [1]. By securing a custom chip, OpenAI reduces reliance on Nvidia’s GPUs—the dominant supplier since the 2022 generative‑AI boom—and diversifies its hardware portfolio, joining existing partnerships with AWS’s Trainium, AMD, and Cerebras [1]. Broadcom, meanwhile, has seen its share price rise 10 % in 2026 and nearly seven‑fold since the end of 2022, buoyed by a $73 billion order backlog that it ties to a $100 billion annual AI‑chip revenue goal by 2027 [1][2]. The Jalapeño launch therefore reinforces Broadcom’s position as a “toll booth” for AI silicon, serving multiple labs while capturing recurring revenue.

What to watch

• Prototype delivery – A physical Jalapeño sample is slated for delivery on June 26, 2026; its performance will be the first concrete test of OpenAI’s cost and efficiency claims.
• Scale‑up timeline – Broadcom expects small‑scale prototype production in late 2026, with a broader ramp in 2027 and full‑tilt deployment in the first half of 2028 [1].
• Supply‑chain constraints – Both firms depend on Taiwan Semiconductor for advanced fabrication and packaging, a capacity that is already booked through 2026, potentially limiting how quickly the chip can replace existing GPU workloads.

OpenAI’s Jalapeño represents the first step in a longer journey toward a vertically integrated AI stack; its success will hinge on whether the custom ASIC can deliver the promised efficiency gains at scale while navigating a tightly constrained semiconductor supply chain.