The first custom AI inference chip from OpenAI, named Jalapeño, was unveiled this Wednesday, with Wedbush Securities viewing it as merely the initial step in a broader series of proprietary chip developments.
In a research note, Wedbush analyst Matt Bryson stated that historical precedent shows a mature application-specific integrated circuit (ASIC) typically requires multiple generations and several years of refinement to reach commercial readiness.
He further noted that this development is likely a positive for both Broadcom (AVGO.US) and OpenAI, provided the chip ultimately delivers on its promised performance. However, it would not be surprising if initial shipments are limited and require a second, third, or even fourth design iteration before achieving widespread adoption, the analyst added.
According to reports, Jalapeño is an ASIC specifically engineered for large language model (LLM) inference. OpenAI handled the underlying architecture design, while its partner Broadcom was responsible for silicon implementation and network hardware. Celestica (CLS.US), a Canadian electronics manufacturing services provider, managed the integration of the board and rack systems.
The project, from initial design to tape-out, was completed in just nine months. OpenAI credited its own AI models for accelerating the entire development process.
OpenAI described Jalapeño as the first product in a multi-generation computing platform. The company plans for the initial deployment of this platform by the end of 2026, with ongoing expansion in the subsequent years.
In-House Chips as a Strategic Lever for IPO
By venturing into custom chip development, OpenAI joins the ranks of companies like Google, Amazon, and Meta Platforms, a move that could potentially lessen reliance on GPU market leader Nvidia.
Independent technology analyst Patrick Moorhead commented that when chip suppliers capture over 75% of design profits, companies are inevitably compelled to explore alternatives. He noted that all major hyperscale data center operators and two leading LLM developers are now pursuing their own chip designs, a path where not everyone will succeed but most participants will gain some benefit.
Successfully demonstrating to the public markets an ability to manage infrastructure costs could significantly strengthen OpenAI's position for an initial public offering (IPO).
However, recent reports suggest that, due to market volatility, OpenAI is leaning towards postponing its IPO timeline from this fall to 2027, while still maintaining its $1 trillion valuation target.
Separately, it is reported that rival Anthropic, which is also preparing for a potential public listing, is exploring the development of its own chips.
Meanwhile, Broadcom has established itself as the partner of choice for companies aspiring to design custom silicon.
William Blair analyst Sebastien Naji stated that this collaboration further solidifies Broadcom's leadership position in the ASIC domain, helping the company continue to expand its roster of top-tier clients and diversify its AI-related revenue streams. The analyst maintained an Outperform rating on Broadcom.
Comments