NEW YORK/SAN FRANCISCO – Meta Platforms plans to start manufacturing its artificial intelligence chips in September as part of a plan to increase overall computing capacity to 14GW in 2027, according to an internal memo seen by Reuters.
The tech company’s data center chip, codenamed “Iris,” is part of a fourth-generation project of in-house designed meta-training and inference accelerators.
The plan is to use custom-built silicon to improve the AI that powers Facebook and Instagram’s social media platforms.
Testing the chip took only six weeks and found no major problems, according to the memo.
This relatively quick progress signals positive momentum for an internal effort that has struggled since its launch more than 50 years ago.
Meta is tailoring the chip to its needs, working with Broadcom to help with design and with Taiwan Semiconductor Manufacturing Co to help with manufacturing.
This approach is likely to help the company reduce massive computing costs and increase its independence from chip suppliers such as Nvidia and Advanced Micro Devices.
The chip is intended to augment the large number of graphics processing units (GPUs) used for AI applications that Meta has purchased from Nvidia and AMD.
But adopting the latest GPUs for a company of Meta’s size was “a significant and time-consuming undertaking,” the memo said.
Shares fell after the news, but recovered after the company announced developer access to its AI coding model, which competes directly with OpenAI and Anthropic. Shares were up 4.6% in late afternoon trading.
“You can’t become an AI giant if you rely on other companies for chips,” said Mike Gualtieri, vice president and principal analyst at research firm Forrester. “Hyperscalers and even SpaceX are planning chips because that’s the only way to compete on price for model usage.”
Meta announced the technical name Iris in March along with three other AI processors.
Typically, companies release AI chips at intervals of a year or more, but the company plans to launch a chip about every six months through 2027.
According to the memo, Meta plans to deploy 7GW of computing infrastructure in 2026. To achieve this total, Meta expects to add 1GW in the first half of 2026 and another 2.5GW by the end of 2026, the note said.
One gigawatt of energy is enough to power approximately 800,000 homes.
The company plans to double its capacity again in 2027, reaching a total of 14GW in 2027, according to the memo.
The company expects to spend up to US$145 billion (S$187 billion) on AI infrastructure in 2026, a significant chunk of Big Tech’s projected spending of more than US$700 billion on the technology.
According to the memo, Meta has entered into long-term, multi-year supply agreements to expand its computing infrastructure.
These include agreements with Samsung Electronics for memory chips, SanDisk for flash storage, and Sumitomo Electric for fiber optic equipment.
These long-term contracts are critical to data center expansion goals as a memory chip shortage prompts companies like Apple to raise prices.
Demand for components such as memory and AI chips is surging as tech companies race to expand data centers to keep up with AI’s thirst for computing power.
Analysts at Morgan Stanley said “chipflation” has become a macroeconomic concern, with the prices of memory and other chips rising rapidly and significantly. Reuters
