In a move aimed at boosting its in-house semiconductor development and reducing reliance on Nvidia, Meta has acquired Rivos, a startup working on its own graphics processing units, signaling an acceleration in the AI chip arms race.
Article contents:
- Meta doubles down on the chip race
- Why acquire Rivos?
- Challenges facing Meta’s internal chip project (MTIA)
- Zuckerberg’s “superintelligence” ambition and massive spending
- The ultimate goal: reducing reliance on Nvidia
- A consistent strategy following a previous failed attempt
- A decisive step in the battle for technological sovereignty
- Frequently asked questions
Meta doubles down on the chip race
Meta Platforms is acquiring chip startup Rivos, according to a person familiar with the deal, as part of a broader effort to strengthen its in-house semiconductor development and gain tighter control over its AI business infrastructure. The acquisition comes as major tech companies race to build their own chip design capabilities to cut costs and reduce reliance on third-party suppliers.
Why acquire Rivos?
Rivos has been working on developing its own graphics processing unit (GPU), the chip that powers most artificial intelligence workloads. Although Meta already has an internal team building chips for AI development, known as the Meta Training and Inference Accelerator (MTIA), the company still spends billions of dollars each year on GPUs from outside partners, including market leader Nvidia. The terms of the deal were not immediately known, though Rivos had been seeking new funding at a $2 billion valuation in August.
Challenges facing Meta’s internal chip project (MTIA)
Meta’s internal chip development efforts have not been progressing as fast as CEO Mark Zuckerberg would like, according to people familiar with the matter. Meta leadership had been scouring the market hoping to bring in reinforcements to speed up the internal group’s work. A Meta spokesperson denied this characterization, saying: “Our custom silicon work is progressing rapidly and this will accelerate our efforts further.”
By acquiring Rivos, Meta gains not only the technology the startup is developing, but more importantly, a team of talented and experienced chip design engineers, which could significantly accelerate its roadmap.
Zuckerberg’s “superintelligence” ambition and massive spending
Zuckerberg has made artificial intelligence the company’s top priority, spending aggressively on AI talent and infrastructure to keep pace with competitors like OpenAI and Google. The company has pledged to spend up to $72 billion this year on capital expenditures, including costs related to AI infrastructure, and recently raised $29 billion to fund the development of a massive data center in Louisiana.
The ultimate goal: reducing reliance on Nvidia
Developing custom AI chips is vital for Meta for several reasons. First, custom-designed chips can be more efficient at running Meta’s AI models, lowering power and operational costs. Second, it reduces its reliance on Nvidia, which currently dominates the market and commands high prices. Third, it gives the company greater control over its supply chain and protects it against potential future shortages.
A consistent strategy following a previous failed attempt
This is not the first time Meta has tried to acquire a chip company. Earlier this year, Meta explored acquiring Korean chip startup FuriosaAI to help improve internal efforts to develop chips designed for training AI systems. However, FuriosaAI turned down the $800 million offer, choosing instead to grow its business as an independent company. Its success in acquiring Rivos demonstrates its determination to push forward with this strategy.
A decisive step in the battle for technological sovereignty
Meta’s acquisition of Rivos is more than just a business transaction; it is a strategic statement. It shows that tech giants are no longer content being mere customers to chip companies. They want to control their technological destiny from start to finish. The outcome of this race to design the best AI chips will be a decisive factor in determining who leads the next technological revolution.
Frequently asked questions
Q1: What is the difference between a training chip and an inference chip?
A1: Training chips are extremely powerful graphics processing units used to “teach” AI models by processing massive amounts of data. Inference chips, such as Meta’s MTIA project, are designed to efficiently run the trained model to answer real-world user queries.
Q2: Why is chip design so difficult and expensive?
A2: Advanced semiconductor design requires highly specialized technical expertise, expensive software tools, and years of research and development. Even after design, the manufacturing process is extremely complex and costs billions of dollars to build a single fabrication plant.
Q3: Will Meta stop buying Nvidia chips?
A3: No, not anytime soon. It will take years for Meta to develop chips that can rival Nvidia’s best products at scale. In the meantime, it will continue to be one of Nvidia’s largest customers, but the long-term goal is to gradually reduce that reliance.
