Meta has decided to take a step forward and invest in its own artificial intelligence chips to power the massive infrastructure that supports Facebook, Instagram, WhatsApp, and its new generative AI tools. This move responds to the enormous increase in computing power demand and the rising cost of relying almost exclusively on providers like Nvidia and AMD.
With this launch, Mark Zuckerberg's company aims to reduce its reliance on third parties, lower the operating costs of its data centers , and tailor hardware to the highly specific workloads within its platforms. It's not about selling processors to other companies, but about building custom silicon for internal use.
What is MTIA and why does Meta want its own AI chips
The program revolves around the Meta Training and Inference Accelerator (MTIA) family , a project the company has been working on since 2023 and which is now taking public shape with several chip models. The central idea is to design more compact and efficient processors, intended for a limited set of AI tasks rather than the large, general-purpose chips that dominate the market.
As the company explained, MTIA's architecture focuses on balancing computing power, memory bandwidth, and memory capacity to serve classification and recommendation models , as well as increasing inference workloads. In other words, they are optimized to decide what each user sees, how quickly AI assistants respond, and how millions of simultaneous requests are processed.
Meta emphasizes that it doesn't manufacture for the general market , so its chips don't need to incorporate all the components of a conventional GPU. This omission of unnecessary features allows them to cut costs and fine-tune the design to the actual needs of their services.
The new chips will be manufactured by Taiwan Semiconductor Manufacturing Co. (TSMC) , the world's largest semiconductor foundry and a regular partner of giants like Apple, Nvidia, and Qualcomm. The process, from design to delivery of the first batches, takes approximately two years.
Four generations of MTIA chips with differentiated functions
The current roadmap envisions four generations of in-house chips: MTIA 300, 400, 450, and 500 , with staggered releases through 2027. All of them are being developed in parallel, allowing Meta to tailor each iteration to the evolution of its AI workloads.
The MTIA 300 is already in production and is being used for classification and recommendation systems —that is, to decide which content, ads, and videos are shown to each user on Facebook and Instagram. This first model has been designed to process large volumes of requests in real time with low energy consumption.
Its successor, the MTIA 400 (also known internally as Iris), has passed laboratory tests and is being deployed in the company's data centers. Meta claims it quadruples the performance of certain operations compared to the MTIA 300 and significantly improves memory access speed, resulting in faster responses and greater capacity to handle concurrent traffic.
The next models, MTIA 450 and MTIA 500 —codenamed Arke and Astrid— are geared towards large-scale advanced inference tasks and will be more widely adopted starting in 2027. Astrid, the more powerful of the two, increases memory capacity by approximately 80% and access speed by 50% compared to its predecessor, positioning it for generative AI scenarios and large language models.
Yee Jiun Song, vice president of engineering at Meta, noted that the four products are being built in parallel and will be released at intervals of approximately six months . He explained that the rapid pace of change in AI workloads necessitates constant revisions to the roadmap and adjustments to development based on emerging needs.
Multimillion-dollar investment and long-term strategy
The shift towards its own silicon comes with an unprecedented spending plan. Meta plans to allocate between $115.000 billion and $135.000 billion in capital investment for its entire technology infrastructure by 2025 , a figure that is expected to remain very high in subsequent years as the new data centers become established.
Designing and producing chips is not exactly cheap: the transition from design to manufacturing typically costs billions , and it only becomes profitable with large-scale use and high utilization rates. This is where Meta hopes to gain an advantage, leveraging the massive volume of traffic and requests on its platforms.
To bolster its internal capabilities, the company has made significant corporate moves. After a failed attempt to acquire South Korea's FuriosaAI , Meta ultimately purchased Santa Clara-based semiconductor firm Rivos , along with several hundred of its employees. The goal is to accelerate the design of custom silicon for AI and to be able to tackle multiple chip projects simultaneously.
In parallel, Meta collaborates with Broadcom on specific design aspects and maintains high-volume agreements with Nvidia and AMD. In February, it signed contracts worth tens of billions to ensure a continuous supply of GPUs that will remain key for both training and some inference processing in the coming years.
Balance between in-house chips and external GPUs
The company's management insists that the investment in MTIA does not imply breaking with Nvidia or AMD , but rather complementing their products. Currently, Meta's in-house chips are primarily used in content recommendation and inference systems , while training large language models still relies on hardware from external vendors.
Susan Li, the company's chief financial officer, explained that the strategy involves combining different types of processors depending on the task . For certain highly specialized workloads, their own custom-designed chips can provide added efficiency and lower costs; for others, such as training next-generation models, high-performance GPUs from Nvidia or AMD remain essential.
Li herself has acknowledged that Meta's ambition is to develop processors capable of training its most advanced AI models in the future , although she admits that reaching the current level of major third-party manufacturers will take time. For now, the priority is expanding the use of MTIA in classification, recommendations, and generative inference.
Alongside its hardware development, Meta is strengthening its internal AI structure with a new engineering organization led by Maher Saba , which will work alongside the company's Superintelligence Lab. It has also secured content access agreements, such as the one signed with News Corp. , valued at up to $50 million annually, to train its models with journalistic material and archives.
Competition with Google, Amazon, and Microsoft in the silicon race
Meta's move fits into a broader trend in the tech sector. Google has been developing its TPUs for years to train and run AI models in its cloud, while Amazon has its own processors like Trainium and Inferentia. Microsoft, for its part, has begun deploying internally designed chips for specific functions, combined with the third-party hardware that powers much of Azure.
With MTIA, Meta fully joins the race for direct control of the most critical layers of the infrastructure . However, the company is aware that Nvidia maintains a dominant position , with over 90% of the AI ​​chip market and more than a decade's technological advantage in specialized GPUs.
Meta's approach isn't to dethrone Nvidia overnight, but rather to build a hardware layer tailored to its own services , allowing it to adjust performance, power consumption, and costs to the actual needs of its network of social and AI applications. In a context of soaring demand for computing power, any unit savings are multiplied when applied to tens of thousands of servers.
The scale of Meta's infrastructure helps explain this strategy. The company is expanding data centers in several US states and strengthening its global capacity to support both its social networks and new AI products, such as app-based assistants and product search features with image and price-based recommendations.
This entire plan for proprietary chips places Meta at a key stage in its technological evolution: if it can enable MTIA to reduce costs while maintaining the performance necessary for its services, the company will gain flexibility in the face of the fluctuations in the GPU market and will be able to continue expanding its large-scale AI projects without relying so heavily on third parties. The challenge lies in executing a complex and very expensive roadmap on time in a sector where technology is advancing at a speed that, as the company itself acknowledges, is surprising even those at the forefront.
