Pular para o conteúdo
Menu
Mix Vale

Production of Meta’s own processors begins in September to reduce costs

Maria
Por 5 min de leitura
Compartilhe
Meta
Foto: Meta - Poetra.RH / Shutterstock.com

The giant parent company of Facebook and Instagram will begin production of its latest processors focused on artificial intelligence from September, as shown in an internal document leaked on July 9th. This strategy seeks to reduce financial dependence on external video cards at a time of serious global shortage of components, allowing the company to optimize investments worth billions and achieve the long-awaited independence in the hardware sector.

Publicidade

Strategic partnerships for semiconductor development

According to internal records, at least one of the new components passed the testing phase in an agile period of approximately one and a half months. To make this complex supply chain viable, the corporation established alliances with giants in the global technology sector:

chip
Chip – Gorodenkoff / shutterstock.com/

Publicidade
  • Broadcom works directly on the architectural design of the processors.
  • TSMC (Taiwan Semiconductor Manufacturing Company) assumed responsibility for the physical manufacturing of the parts.
  • South Korean Samsung will supply the high-performance RAM modules.
  • SanDisk was chosen to deliver the data storage solutions.
  • Sumitomo Electric will send essential fiber optic equipment for server communication.

The technical specifications for these four new devices, which are part of the project known as Meta Training and Inference Accelerator (MTIA), were initially revealed to the public in March this year. A portion of these parts are already in operation on the company’s servers, while the rest will be activated between the end of this year and the course of next year. The modular format adopted in parts engineering demonstrates preparedness for rapid changes in computational demands, ensuring adaptation until mass production is fully achieved.

How the MTIA project architecture works in practice

During the official presentation of the project, executives highlighted that each new version of MTIA uses the previous generation as a foundation, employing small blocks of independent circuits. This methodology allows you to quickly incorporate the latest discoveries about data processing in artificial intelligence, resulting in a hardware update cycle that is much faster than the technology industry standard.

The main financial objective of this venture is to save resources that would otherwise be used to purchase graphics cards from dominant suppliers, such as Nvidia and AMD, whose high-end models can cost tens of thousands of dollars per unit due to high market demand. Despite still predicting bulky purchases from these brands, the processors themselves will be directed to training recommendation algorithms on social networks and performing inference tasks — the process where the already trained AI responds in real time to user requests. The corporation has been working on creating its own circuits since 2023.

Billion-dollar investments in infrastructure and processing capacity

To support all these innovations, the company has been injecting massive amounts into expanding its digital infrastructure. In April, a financial statement projected that the company’s capital expenditures are expected to reach a staggering range of between $125 billion and $145 billion this year alone, with the overwhelming majority of that sum stamped solely on advancing machine learning technologies.

On a global scale, the corporation has signed several contracts focused on building new data centers and providing electricity, committing tens of billions to ensure there is no shortage of processing power for the Muse Spark family of language models. The internal report points to an aggressive goal of installing 7 gigawatts of computing capacity by the end of December, with concrete plans to multiply this volume by two in the following year.

Publicidade

The industry’s movement to escape the hardware monopoly

The company’s recent trading history shows a diversification of partners, including a large agreement reached last year with ARM to support the processing of its content feeds. At the same time, large-scale contracts were signed with AMD for the acquisition of Instinct line accelerators, in addition to a multibillion-dollar partnership with Amazon aimed at using the e-commerce giant’s central processors in specific cloud operations.

The search for alternatives to Nvidia’s expensive equipment has become a widespread race in Silicon Valley. Recently, the creator of ChatGPT, OpenAI, revealed that it is designing a chip focused on inference together with Broadcom, while Anthropic is negotiating the production of its components with Samsung. Giants like Google and Amazon have already had their own lines of silicon running on their servers for years, paving the way for dozens of startups to try to grab a slice of this lucrative sector. This entire movement highlights a paradigm shift, where large corporations demand total control over their physical infrastructure to survive in an extremely predatory and expensive innovation market.

Notícias relacionadas