OpenAI's first AI chip unveiled! Collaborating with Broadcom and TSMC to take on Nvidia, does the name hide a secret?

OpenAI releases its first self-developed AI chip, Jalapeno, built by Broadcom and TSMC, with performance rivaling Nvidia. This move reduces external dependence and helps the company push toward its $1 trillion IPO valuation.

OpenAI's first AI chip debuts, partnering with Broadcom and TSMC

OpenAI, the AI giant preparing for an IPO, officially unveiled its first self-developed AI chip last night (June 24), named Jalapeño.

This chip was co-developed by OpenAI and semiconductor giant Broadcom (ticker: AVGO), positioned as an inference processor optimized for large language models (LLMs).

According to an official OpenAI article, Jalapeño took only 9 months from initial design to tape-out, a progress made possible by OpenAI leveraging its own AI models to assist in optimization. Currently, chip samples are running machine learning workloads, including the GPT-5.3-Codex-Spark model, at target frequencies and power levels in the lab.

Per a report from Reuters, the chip is manufactured by Taiwan's TSMC (NYSE: TSM, TWSE: 2330), while Canadian electronics manufacturer Celestica is building the server systems.

OpenAI President Greg Brockman told CNBC that the AI model's ability to accelerate chip development was surprising. Broadcom CEO Hock Tan stated, he has delivered physical samples to OpenAI CEO Sam Altman, with initial deployment expected by the end of 2026.

Following the announcement of its first AI chip, Broadcom's stock edged up to $382.08, rising to $389.55 in after-hours trading at the close.

Source: Google Finance. Stock price trend of semiconductor giant Broadcom (ticker: AVGO).

Why is OpenAI's chip called Jalapeño?

Jalapeño is a Mexican chili pepper with a strong, spicy flavor. Tech media Engadget noted that for this AI company's entry into chip production, it's a fiery start; BusinessKorea added that the naming also hints that OpenAI is directly targeting the Nvidia-dominated AI computing infrastructure market.

Additionally, in the semiconductor and software industries, R&D teams are very accustomed to using food, plants, or place names as codenames for projects during confidential phases. For example, many tech giants in the past have used desserts or fruits to name core products.

During its 9-month secret R&D period, OpenAI's hardware team chose Jalapeño as the codename for this inference chip, ultimately keeping it upon release, perfectly aligning with the humorous conventions of Silicon Valley engineers.

Source: Public Domain Pictures, photo by Sheila Brown. Jalapeño is a Mexican chili pepper with a strong, spicy flavor.

OpenAI chip claims performance rivaling Nvidia Blackwell

Tech giants are facing challenges of insufficient computing power supply. Broadcom CEO Hock Tan emphasized that Jalapeño's performance is comparable to Nvidia's Blackwell chip and Google's Tensor Processing Unit (TPU).

OpenAI hardware head Richard Ho pointed out that the chip is tailored for inference workloads, efficiently executing critical tasks. Official early testing shows significantly better performance per watt than current competing products on the market.

Self-developed chips can reduce reliance on Nvidia. Per a CNN report, Ben Barringer, technology research head at investment firm Quilter Cheviot, analyzed that tech giants are actively diversifying chip sources to avoid being constrained by a single supplier.

Besides partnering with Broadcom, OpenAI has also reached an agreement with Amazon Web Services (AWS) to use Trainium chips, and is cooperating with AMD and Cerebras, which went public in May.

OpenAI's president stated that OpenAI's demand for computing power is extremely urgent; Hock Tan also confirmed that customer demand for computing power has surged and will continue through 2028.

High-bandwidth memory challenges gross margins, full-stack strategy aids trillion-dollar valuation

Hardware manufacturing still faces supply chain cost challenges. Hock Tan admitted that due to the large amount of high-bandwidth memory (HBM) required for AI chips, current profit margins for custom chips are lower than for products like network switches, with memory primarily supplied by South Korea's SK Hynix and Samsung Electronics.

However, this plan is the first step in the long-term strategy of OpenAI and Broadcom. Both parties previously announced plans to deploy data centers with a scale of 10 gigawatts (GW) over the next few years, in partnership with Microsoft and other collaborators.

OpenAI's chip announcement comes at a critical time as it prepares for its first IPO, with market expectations of a potential $1 trillion valuation, placing it under greater profit pressure.

Jalapeño helps build a full-stack technical advantage. OpenAI's president stated that by designing more underlying infrastructure in-house, the company can deliver services with higher efficiency, allowing users to access advanced AI technology at lower and more reliable costs.

Further reading:
OpenAI files for IPO! But its cash burn rate is alarming, with positive cash flow unlikely within four years.
Huang Renxun's GTC 2026 speech summary: From RTX Spark to backplane stocks, five key points at a glance.