--- title: "Meta's self-developed chip project encounters significant obstacles" type: "News" locale: "zh-CN" url: "https://longbridge.com/zh-CN/news/277175266.md" description: "Meta's self-developed AI chip project has encountered significant obstacles, canceling the high-end AI training chip Olympus due to design difficulties and shifting towards developing a simpler version. This move reflects the challenges of competing against Nvidia's dominance. Meanwhile, Meta has signed new chip supply agreements with AMD and Nvidia, planning to purchase a large number of data center chips to reduce reliance on external chip manufacturers. Meta expects capital expenditures to reach $115 billion to $135 billion in 2026, primarily for chips and servers" datetime: "2026-02-27T09:41:40.000Z" locales: - [zh-CN](https://longbridge.com/zh-CN/news/277175266.md) - [en](https://longbridge.com/en/news/277175266.md) - [zh-HK](https://longbridge.com/zh-HK/news/277175266.md) --- > 支持的语言: [English](https://longbridge.com/en/news/277175266.md) | [繁體中文](https://longbridge.com/zh-HK/news/277175266.md) # Meta's self-developed chip project encounters significant obstacles **Authors**: Jyoti Mann, Wayne Ma, Liu Qianer According to six people with direct knowledge, at the same time that Meta signed new chip supply agreements with AMD and NVIDIA, its **self-developed AI chip project has encountered serious problems**. Insiders said that due to significant design difficulties, Meta **canceled the development of its high-end AI training chip** last week and shifted its focus to a simpler version. The company informed employees in the AI infrastructure department of this adjustment last week. This decision highlights the immense difficulty for tech giants to design AI chips that can compete with market leader NVIDIA. **Key Points** - Meta has scrapped its self-developed high-end AI training chip **Olympus** due to design challenges. - This decision underscores the difficulty of challenging NVIDIA's dominance. - Meta has reached data center chip supply agreements with AMD and NVIDIA. Before adjusting its chip roadmap, Meta had established new collaborations with Advanced Micro Devices (AMD) and NVIDIA in recent weeks: - On Tuesday, Meta announced it would procure AMD chips with **a power of up to 6 gigawatts**, which can roughly meet the needs of multiple large data centers. - Earlier this month, Meta reached a **multi-generational long-term cooperation** agreement with NVIDIA, committing to deploy millions of existing and next-generation NVIDIA chips in data centers. Meta's self-developed AI chips are part of the MTIA (Meta Training and Inference Accelerator) project, an important initiative for the company to develop AI hardware independently and reduce reliance on external chip manufacturers like NVIDIA, aiming to lower costs and strengthen control over data center infrastructure. For example, Meta expects its capital expenditures to reach **$115 billion to $135 billion** in 2026, most of which will be directed towards chips and servers. A Meta spokesperson stated: “We remain committed to investing in a diverse chip portfolio to meet our needs, including advancing the MTIA series products, with more information to be announced this year.” Other tech companies, including Microsoft, have also encountered similar issues when developing their own AI chips. Last year, NVIDIA CEO Jensen Huang publicly stated that most tech giants would ultimately abandon plans to develop competing chips and claimed that these chips would continue to lag behind NVIDIA products in performance. **Multiple Self-Developed Chips from Meta Encounter Setbacks** Several self-developed chips from Meta have encountered problems: - The company has scrapped one version of its second-generation training chip, internally codenamed **Iris** - Subsequently, a more advanced training chip project, codenamed **Olympus**, was launched, but it has now been canceled. A person involved in the Meta chip project stated that there is **skepticism** within the company about whether self-developed chips can catch up to NVIDIA's capabilities, as the project faces risks of delays and redesigns. This individual mentioned that this work requires a large engineering team for design and debugging, and to ensure that power consumption does not become too high; otherwise, it would be of no cost-effectiveness compared to NVIDIA chips. - **Iris** adopts a **SIMD (Single Instruction, Multiple Data)** architecture, which has a relatively simple hardware design, but the software programming difficulty is higher when training AI models. - **Olympus** uses a **SIMT (Single Instruction, Multiple Threads)** architecture similar to NVIDIA's AI chips, which is more user-friendly for software programming, but the hardware design is extremely challenging. Many tech companies favor the SIMT architecture popularized by NVIDIA because it is more flexible and better suited for training modern AI models. According to four sources, Meta originally planned to complete the design of Olympus by the fourth quarter of 2026, and new chips typically require nine months or longer from initial design to mass production. Olympus is responsible for the core GPU part of AI computing, and it plans to use designs from **Rivos**, a chip startup acquired by Meta last year. Rivos claimed that its GPU can efficiently run NVIDIA's proprietary **CUDA** software code, which is the mainstream software ecosystem for training and running machine learning models. According to one source, Meta initially planned to use Olympus to build a large server cluster, but executives ultimately determined that this would pose a **significant risk** for training new models in the context of fierce competition with established rivals like OpenAI and Google. Multiple sources indicated that the training software stability of Olympus is not as good as NVIDIA's, and the complex design may lead to difficulties in large-scale production. Therefore, Meta has currently chosen to **continue using training chips from third-party vendors**, whose software ecosystems are already mature ### 相关股票 - [Meta Platforms (META.US)](https://longbridge.com/zh-CN/quote/META.US.md) - [Roundhill META WeeklyPay ETF (METW.US)](https://longbridge.com/zh-CN/quote/METW.US.md) - [iShares Semiconductor ETF (SOXX.US)](https://longbridge.com/zh-CN/quote/SOXX.US.md) - [Invesco AI and Next Gen Software ETF (IGPT.US)](https://longbridge.com/zh-CN/quote/IGPT.US.md) - [GraniteShares 2x Long NVDA Daily ETF (NVDL.US)](https://longbridge.com/zh-CN/quote/NVDL.US.md) - [VG Info Tech (VGT.US)](https://longbridge.com/zh-CN/quote/VGT.US.md) - [iShares Expanded Tech Software Sector ETF (IGV.US)](https://longbridge.com/zh-CN/quote/IGV.US.md) - [Direxion Semicon Bull 3X (SOXL.US)](https://longbridge.com/zh-CN/quote/SOXL.US.md) - [Direxion Daily Meta Bull 2x Shares (METU.US)](https://longbridge.com/zh-CN/quote/METU.US.md) - [Direxion Daily AMD Bull 2X Shares (AMUU.US)](https://longbridge.com/zh-CN/quote/AMUU.US.md) - [AMD (AMD.US)](https://longbridge.com/zh-CN/quote/AMD.US.md) - [SPDR S&P Semicon (XSD.US)](https://longbridge.com/zh-CN/quote/XSD.US.md) - [Invesco PHLX Semiconductor ETF (SOXQ.US)](https://longbridge.com/zh-CN/quote/SOXQ.US.md) - [VanEck Semiconductor ETF (SMH.US)](https://longbridge.com/zh-CN/quote/SMH.US.md) - [Invesco S&P 500 Equal Weight Tech ETF (RSPT.US)](https://longbridge.com/zh-CN/quote/RSPT.US.md) - [GraniteShares 2x Long META Daily ETF (FBL.US)](https://longbridge.com/zh-CN/quote/FBL.US.md) - [GraniteShares 2x Long AMD Daily ETF (AMDL.US)](https://longbridge.com/zh-CN/quote/AMDL.US.md) ## 相关资讯与研究 - [Meta Opens Manhattan Store As Capital Shifts From Metaverse To AI](https://longbridge.com/zh-CN/news/279834632.md) - [Meta is having trouble with rogue AI agents](https://longbridge.com/zh-CN/news/279688226.md) - [Meta Platforms Plans Mass Layoff To Help Offset A.I. Spending](https://longbridge.com/zh-CN/news/279265642.md) - [Meta actually is keeping its VR metaverse running, for now](https://longbridge.com/zh-CN/news/279825430.md) - [Meta Expands Retail Presence with New York City Flagship Location](https://longbridge.com/zh-CN/news/279610884.md)