Mark Zuckerberg’s Meta plans to design its own artificial intelligence chips in-house starting in September – part of an industry-wide effort by the biggest names in AI to start making their own chips amid ongoing high demand.
Zuck’s initiative, known internally as “Iris,” centers on developing custom silicon to supercharge the AI systems behind Facebook and Instagram, Reuters reported Thursday.
The social media giant — which expectes to spend up to $145 billion on AI infrastructure this year — is working with Palo Alto, Calif.-based Broadcom on design and Taiwan Semiconductor Manufacturing on production.
Meta joins a growing list of technology companies seeking to handle more of their chip development internally to cut costs and reduce dependence on Nvidia, which has dominated the AI chip business with its ultra-powerful semiconductors.
Even as Meta and other companies are launching their foray into chip building, the semiconductor industry remains under tremendous demand strain, and AI companies’ efforts to become more autonomous provides no silver bullet to the supply chain conundrum.
Demand for manufacturing, packaging and other chip production resources continues to outpace supply, while several specialized chip-making processes are controlled by a small number of companies already operating at capacity even as they invest mountains of capital to expand.
Meta’s latest project builds on a long-running effort to develop its own chips. Its Training and Inference Accelerators program, launched more than five years ago, has focused on in-house chip development, though progress has been slow.
Development of the new chip has reportedly moved much more rapidly. Testing took just six weeks and faced no major problems, according to Reuters. Meta plans to introduce a new chip roughly every six months through 2027, compared with the typical annual-or-longer release cycle for AI chips.
Meta is aiming to double its computing infrastructure in 2027, according to Reuters.
The custom product is intended to complement the large number of graphics processing units, or GPUs, that Meta buys from Nvidia and AMD for AI workloads.
But bringing the newest GPUs online at Meta’s scale “has been a heavy lift, and it has cost us time,” according to a company memo reviewed by Reuters.
Developing custom chips can potentially lower costs and diversify supply chains, Axios noted.
“I want something in my pocket when I’m sitting across the table from Jensen negotiating,” Bernstein senior analyst Stacy Rasgon told the outlet, referring to Nvidia CEO Jensen Huang.
In addition to Meta, Amazon, Google and Microsoft all have in-house chip programs. OpenAI recently introduced its first custom inference chip with Broadcom, while Anthropic is reportedly in talks with Samsung about developing its own chip.
Apple announced this week that it plans to spend more than $30 billion with Broadcom over the next five years, helping the chipmaker expand a manufacturing facility in Fort Collins, Colo.
The consumer tech giant already designs its own chips for the iPhone, iPad and Mac, and is reportedly developing separate processors for AI servers.
Samsung manufactures advanced chips for both its own products and outside customers, while Intel is working to expand its contract manufacturing business after its production technology fell behind in recent years, Axios noted.
Showing the complexity of attaining chip autonomy, those manufacturers rely on lithography equipment from Dutch company ASML — the only supplier of the most advanced machines used to produce AI chips, per the news site.
The Post has sought comment from Meta, Broadcom and Taiwan Semiconductor Manufacturing.