GPU’s new rival has arrived! Amazon officially announces Trainium3: over 50% lower cost compared to GPUs.

Another major tech company has entered Nvidia’s territory.

On Tuesday, Amazon Web Services (AWS) officially launched its third-generation custom AI chip, Trainium3, directly targeting the GPU market dominated by Nvidia.

The cloud computing giant claims that the new chip offers four times the performance of the previous generation, and compared to equivalent GPU systems, can reduce the cost of AI model training and operation by up to 50%.

In recent months, more and more AI companies have been seeking supplier diversification. According to reports, Meta Platforms is negotiating with Google to purchase tens of billions of dollars’ worth of TPU chips, while OpenAI has partnered with Nvidia’s rival AMD and custom chip designer Broadcom.

Trainium3 has already been adopted by several clients including Anthropic and Decart.

Dean Leitersdorf, co-founder and CEO of AI video startup Decart, said that after testing various competing chips including Nvidia’s processors, Trainium3 helped them achieve a technological breakthrough, with video generation frame rates reaching four times that of other chips.

Nevertheless, the progress of new entrants does not mean that AI companies are abandoning Nvidia chips. Many of AWS’s largest customers are also buyers of Nvidia, but this trend shows that Nvidia's near-monopoly may not last forever.

Custom Chips Challenge GPU Dominance

AWS, through its Annapurna Labs custom chip design business, launched Trainium3 with a differentiated strategy focused on cost-effectiveness.

AWS Vice President and Trainium Chief Architect Ron Diamant stated:

Ultimately, the main advantage is cost-effectiveness.

Diamant emphasized:

We don’t see ourselves as trying to replace Nvidia.

However, the client roster AWS revealed on Tuesday demonstrates its increasing market penetration, including companies such as Anthropic, Karakuri, Metagenomi, Neto.ai, Ricoh, and Splash Music.

As early as 2015, AWS acquired Israeli startup Annapurna Labs, beginning to design chips powering data center servers, including network security chips, central processors, and later its AI processor series Inferentia and Trainium.

Startup’s Technological Breakthrough

Decart tested AWS chips against top processors including Nvidia and found Trainium3 was able to continuously generate video at a frame rate four times higher than other chips.

This San Francisco-based AI video startup has spent months trying to train its flagship product, Lucy—an AI-powered video generation app—to achieve seamless real-time rendering.

Reportedly, after a two-week coding marathon at a rented house in Silicon Valley, the breakthrough occurred. Leitersdorf recalled:

At the moment it successfully ran, I saw four people jump up with joy.

He said:

Before this, we could only run for 1.5 to 2 seconds before noise would appear on the screen. The model would become too chaotic, and everything would blur together... Trainium really allowed us to run a bigger, smarter model that wouldn’t crash quickly.

Founded two years ago by 27-year-old Dean Leitersdorf with his brother Orian and their friend Moshe Shalev, the company has grown from 14 employees to over 80 and in August raised $100 million from Sequoia Capital, Benchmark, Zeev Ventures, etc., with a valuation of $3.1 billion.

Shifting Market Landscape, but Not Disrupted

Although custom chips are drawing attention, Nvidia’s market position remains hard to shake in the short term. Many major buyers of AWS chips are also Nvidia customers.

At the end of October, AWS stated that Anthropic was using over one million Trainium2 chips to build and deploy its Claude AI model. A month later, Nvidia announced a $1 billion investment in Anthropic as part of a large-scale computing chip sales deal.

Following news of the Meta and Google negotiations, Nvidia’s stock price dropped noticeably.

In response, Nvidia said on X platform that it was “happy for Google’s success,” and stressed that Nvidia was “a generation ahead of the industry—the only platform capable of running all AI models and anywhere compute happens.”

The company said its chips offer “performance, versatility, and interchangeability” superior to custom chips produced by Google and AWS.

The trend of AI companies seeking supplier diversification is becoming more obvious, but this is mainly to spread risk and optimize costs rather than to completely replace Nvidia’s technical solutions. Analysts believe the market is evolving from a single supplier-led model to a diversified competitive landscape.

Risk Warning and DisclaimerThe market involves risk; investment requires caution. This article does not constitute personal investment advice, nor does it take into account individual users’ specific investment goals, financial situation, or needs. Users should consider whether any opinions, views, or conclusions in this article are appropriate for their particular situation. If you invest based on this, you do so at your own risk.