Broadcom shares jumped as much as 9% premarket after the Financial Times reported OpenAI placed a roughly $10 billion order for custom AI chips, a move that would lessen reliance on Nvidia’s GPUs. The rally, if it holds, would be Broadcom’s biggest daily gain since April and adds roughly $125 billion to its market value. Nvidia slipped about 1% in early trading as investors weighed the read-through for GPU demand. Broadcom’s earnings beat and a stronger AI outlook set the stage for the move.
The stock’s surge follows Broadcom’s fiscal third-quarter results and guidance that topped estimates, with CEO Hock Tan flagging a fourth major customer for its custom AI silicon business. AI semiconductor revenue of about $5.2 billion and solid margin execution reinforced the message that Broadcom’s AI strategy is scaling. Year-to-date, the shares were up 32% through Thursday’s close; today’s pop extends that outperformance and marks the sharpest single-session move since an April spike of nearly 20%. Retail interest has spiked alongside the institutional bid, with message traffic and bullish sentiment rising across trading platforms. The tape is treating the OpenAI order as validation that Broadcom’s custom chip bet is moving from pipeline to production.
People familiar told the FT the OpenAI chips are set to enter series production next year, with shipments targeted for 2026. On Broadcom’s call, Tan said the new customer brings immediate and substantial demand and lifts growth prospects into 2026. He did not name OpenAI, but the contours align with the report: a large inference-focused program at a hyperscale buyer, sized in the billions, moving to production on an accelerated timetable. Broadcom’s backlog now tops $110 billion, giving visibility across hardware and software lines, and Tan signaled he intends to remain CEO through at least 2030. That continuity matters when ramping complex AI silicon programs that require long lead times, tight supplier coordination, and packaging capacity.
OpenAI’s push into custom chips raises an obvious question: is this the beginning of a material shift away from Nvidia? The near-term answer is more nuanced than a simple yes. Nvidia’s leadership in training remains intact, supply is still tight in 2025, and most hyperscalers run mixed estates. But OpenAI designing and deploying an ASIC with Broadcom for inference is consistent with what Amazon, Google, and Meta have already done to lower cost per token, reduce power draw, and control their roadmaps. The incremental pressure lands on Nvidia’s longer-term share of inference workloads and its pricing power at the margin. That explains the modest dip in NVDA shares on the headline even as the broader AI complex remains bid.
Wall Street is treating the OpenAI win as a step-change for Broadcom’s AI narrative. Goldman Sachs called the customer a major validation and noted management now sees “material” upside to prior 2026 AI semiconductor growth expectations of around 60%. JPMorgan said the quarter was driven by accelerating AI demand and puts Broadcom on track to approach $20 billion in AI revenue for fiscal 2025. Bernstein highlighted strength in both semis and software, while Bloomberg Intelligence flagged better-than-expected ramps at Broadcom’s custom ASIC customers. Guidance beat across revenue and margins supports the rerate. Layer on the Tomahawk 6 and Jericho 4 networking products, and the company has multiple shots on goal in AI infrastructure — custom compute plus merchant networking.
For OpenAI, building an internal chip platform for inference is about control and economics. Custom silicon tuned to its models and workloads can cut unit costs, reduce energy usage, and ease supply bottlenecks. The move mirrors peers’ strategies: Amazon’s Inferentia, Google’s TPU, and Meta’s in-house accelerators all exist to handle exploding AI workloads without ceding leverage to merchants. The FT reported OpenAI plans to use these chips internally, not sell them broadly, underscoring the target use case: scale inference for ChatGPT and enterprise products at lower cost. If performance per watt and cost per token improve materially, OpenAI can deploy more features and manage gross margins even as usage grows. Success here doesn’t require displacing Nvidia wholesale — it requires carving out the right slices of the workload.
The partnership also threads through Microsoft. OpenAI runs on Azure, and Microsoft has its own custom AI silicon in Maia. Adding a Broadcom-built ASIC to the mix suggests Azure will support a more heterogeneous fleet: Nvidia GPUs for training and many inference tasks, plus Microsoft and partner ASICs for targeted workloads. For hyperscalers, this is the operating model — diversify suppliers, optimize for specific use cases, and keep total cost of ownership trending down. Broadcom’s job is to meet those requirements without sacrificing time-to-market. If it can, the customer list behind OpenAI could grow, especially as inference volumes outstrip training spend over the next cycle.
The risk is execution, not demand. Custom ASICs at this scale require leading-edge process nodes, advanced packaging, and reliable HBM memory supply. 2026 is both soon and far — soon for a clean tape-out to packaged volume, far enough for macro and supply wrinkles to show up. Packaging capacity, including CoWoS-like solutions, remains tight industry-wide. Any slip in foundry or packaging timelines can push ramps, and cost targets will be tested by component inflation. Still, Broadcom’s deep ASIC history and merchant networking scale are advantages. The $110 billion backlog implies customers are booking capacity early, and that visibility should help manage the supply chain. Investors will want milestones: silicon in-house next year, pilot deployments, and signs the cost curve beats top-tier merchant alternatives.
The market is paying for scarcity. Few companies can ship custom AI silicon at hyperscale, and fewer still can do it with networking moats and software cash flows supporting the spend. Today’s move compresses time value — investors are bringing forward 2026 AI upside into current multiples. That leaves a narrow path: deliver on production timing, defend margins as mix shifts toward AI, and show that software remains resilient. For Nvidia holders, the signal is not existential but directional: inference share will fragment over time as custom chips proliferate. For Broadcom, the signal is more immediate: its AI strategy is paying off in real bookings and a larger, more durable role in the AI stack. Next catalysts include any confirmation from OpenAI, capacity updates from key suppliers, and Broadcom’s next checkpoint on AI revenue run rate.
