Amid speculation about its chip strategy, OpenAI has denied plans for large-scale use of Google's TPUs, saying it is only experimenting. The company aims to develop its own chips, though industry insiders believe a clear path will take time as it evaluates long-term AI service and computing needs.
Recent rumors suggested OpenAI might be using Google's TPU for computing, but OpenAI clarified it has no plans for large-scale use of Google's chips and is only experimenting. The company intends to focus on developing its own chips moving forward.
OpenAI weighs custom chip development as CSPs deepen ASIC investments
ASIC industry insiders believe that it may take some time before OpenAI's chip usage strategy becomes clear. This is because OpenAI must first confirm its overall AI service and functional computing requirements before determining if developing proprietary ASICs is cost-effective or if continuing to use standard chips—or even adopting Google's TPU—is more practical.
Industry sources point out that major US cloud service providers (CSPs) invest heavily in ASIC development because their core businesses require specialized chips tailored to unique network service models to enhance AI computing efficiency.
For example, Google has been developing TPUs for years and possesses extensive experience in this field. Its products deliver strong computational performance, making it the largest buyer among CSPs. This explains why some companies outsource their AI algorithm training workloads to Google's TPU chips.
Consequently, OpenAI is inevitably considering collaboration possibilities amid this broader trend.
Building custom chips poses resource and coordination challenges for OpenAI
Although OpenAI's GPT and Google's Gemini are direct competitors, startups like OpenAI already need substantial resources to advance AI algorithms and related services. Developing ASIC chips would further increase resource demands.
Moreover, "chip design" is a completely different expertise and ecosystem from "AI algorithms." Even if much of the work can be outsourced to professional ASIC vendors, OpenAI still needs to carefully consider how much to invest and how large a team to establish for coordinating and communicating with these vendors.
Semiconductor supply chain players note that leading ASIC manufacturers today can provide highly advanced design and manufacturing services. For OpenAI's projects, market leader Broadcom is likely to handle the chip production. Even without extensive chip development experience, OpenAI should be able to produce viable products.
However, whether these products fully meet OpenAI's actual needs would require having an internal team to facilitate communication. Without this, repeated back-and-forth adjustments could significantly delay chip launch schedules and potentially lead to the conclusion that developing proprietary chips is not worthwhile.
OpenAI explores multiple chip paths as it seeks clarity on long-term AI needs
Currently, OpenAI's clearer chip adoption strategies include using mainstream Nvidia solutions. Based on founder Sam Altman's recent high-profile endorsement at an AMD event, OpenAI is expected to adopt some AMD solutions in the future.
As for whether to use Google's TPU, develop in-house chips, or rely on ASICs, this decision hinges on OpenAI's future AI service development plans. If certain unique functions cannot be efficiently executed with general-purpose chips, then independent chip development becomes imperative.
OpenAI faces a wide range of options. The key is to have a clear direction for product development to avoid wasting resources after chip development is completed.
Article translated by Charlene Chen and edited by Jingyue Hsiao
