Meta(META.US) terminates self-developed high-end chips, expands collaboration with NVIDIA and AMD

According to people familiar with the matter, Meta Platforms (META.US) faced challenges while developing AI chips internally. The company has abandoned its most advanced chip方案 and instead moved to a simplified version. Last week, due to technical difficulties in the chip design phase, the company officially terminated its project to develop the most advanced AI model training chip that was under development. The report also noted that Meta, last week, shared the latest update to this technical roadmap with employees in its artificial intelligence infrastructure department.

Meta’s decision to abandon its in-house chip design reveals the common predicament companies face when trying to build AI chips capable of taking on market leader Nvidia (NVDA.US).

Meta’s adjustment to its chip roadmap comes shortly after it reached new partnerships with AMD (AMD.US), Nvidia, and Alphabet’s Google (GOOGL.US). Reports say the company has signed multi-billion-dollar agreements to lease AI chips from Google.

Earlier this week, AMD announced a partnership with Meta to deploy up to 6 gigawatts of AMD Instinct chips, providing compute power to support its next-generation AI infrastructure. In addition, earlier this month Meta also reached a “cross-generational” strategic partnership with Nvidia, committing to large-scale deployment of Nvidia chips across its data centers.

Meta’s AI chips developed in-house fall under its Meta Training and Inference Accelerator (MTIA) project framework. The core goal of this initiative is to reduce long-term operating costs by vertically integrating chip design capabilities, and to strengthen its independent control and self-governance over data center infrastructure.

In response, a Meta spokesperson said clearly: “We continue to invest in building a diversified set of chip supply options to match business needs, and advancing the MTIA product line is an important strategic direction. This year, we will disclose more about the product line’s R&D progress and go-live plans.”

According to reports, Meta has discontinued its second-generation training chip version code-named Iris, and then also ended the more advanced training chip project Olympus that it had launched afterward.

An insider involved in Meta’s chip development said that within the company, there is widespread caution about plans to build chips whose performance can match that of Nvidia’s. The main concern is the risk of delays or needing to redesign the project. The source noted that related reports indicate that developing chips like these requires assembling a large engineering team to handle key areas such as chip design, debugging, and power consumption control. If power consumption issues cannot be effectively addressed, these in-house chips may not offer enough value when benchmarked against Nvidia’s mature products.

The Iris training chip uses a single-instruction, multiple-data (SIMD) computing architecture. While this architecture has a lower design barrier for hardware engineers, software engineers face significant programming challenges when training AI models. According to the report, Olympus uses a single-instruction, multi-threading (SIMT) architecture that is derived from Nvidia’s AI chip lineage—an approach that may make programming easier for software engineers, but imposes higher technical requirements on hardware design.

The SIMT architecture promoted by Nvidia, with its greater flexibility, is better suited to the needs of training modern AI models and has been favored by multiple technology companies. Meta originally planned to complete the Olympus chip design as early as the fourth quarter of 2026. However, the report added that from initial development to mass production, a new chip typically still takes nine months or even longer, meaning the actual mass-production timeline could be further delayed.

The core component for AI computing in Olympus—the graphics processing unit (GPU)—was originally planned to use the design from chip startup Rivos, which Meta acquired last year. According to reports, Rivos had claimed that its GPU could efficiently run Nvidia’s CUDA software code—code that is currently the mainstream software framework for training and running AI models.

The report also pointed out that Meta originally planned to use Olympus to build large server clusters, but company executives believed that, during a critical period of racing to compete with OpenAI and Google, this move could pose potential risks to training new models.

Specifically, the training software paired with these chips is already unable to match the stability of Nvidia’s products. In addition, Olympus’s complex design could further hinder its large-scale mass production. Therefore, Meta’s current plan is to continue using training chips produced by other vendors, because their paired software is more mature and reliable, enabling it to better support AI model training requirements.