Meta to Mass Produce Custom Iris AI Chip in September
Meta Platforms will start manufacturing its new custom artificial intelligence chip in September. An internal company memo shows the tech giant aims to double its total data center computing capacity to 14 gigawatts next year.
The data center chip is codenamed "Iris" as it belongs to the Meta Training and Inference Accelerator project, an in-house effort to build custom silicon for the company's social media applications.
Meta completed bug testing for Iris in six weeks without finding any significant technical issues. The company designed the chip alongside Broadcom and contracted Taiwan Semiconductor Manufacturing Company to handle the actual fabrication.
The custom silicon will run the AI workloads behind Facebook, Instagram, and WhatsApp. Meta currently buys large numbers of graphics processing units from Nvidia and Advanced Micro Devices to handle these exact tasks.
According to the internal document, adapting the latest third-party graphics processors at a firm as large as Meta:
"has been a heavy lift, and it has cost us time"
The new production schedule establishes an aggressive timeline for the social media firm. Meta plans to launch a new version of its custom silicon roughly every six months through 2027.
Typical semiconductor firms take a year or more between major processor releases. The rapid release cycle represents an attempt to catch up after the in-house hardware program struggled for several years.
To secure its data center supply chain, Meta signed multi-year purchase contracts with multiple hardware vendors. The company locked in deals with Samsung Electronics for memory chips, SanDisk for flash storage, and Sumitomo Electric for fiber-optic gear.
The heavy infrastructure spending caused some anxiety among public investors. Meta shares fell more than two percent following the disclosure of the data center expansion plans.
The company expects to spend up to 145 billion dollars on artificial intelligence infrastructure this year. This budget forms a massive portion of the 700 billion dollars that major technology corporations plan to spend on infrastructure globally.
Rising demand for data center components continues to push hardware prices upward. Morgan Stanley analysts noted that rapid price increases for memory and processors created a macroeconomic trend they call:
"chipflation"
Wall Street analysts maintain mixed views on whether custom chips will hurt dominant suppliers. Bank of America analysts pointed out that custom chips coexisted with commercial processors for a decade without stopping Nvidia's market growth.
