Amazon's annual cloud computing conference opens, CEO to appear tonight—can the new AI chip "catch up with Google's TPU"?

Amazon's annual cloud computing event—AWS re:Invent—opened in Las Vegas on December 1 (Monday).

Although a series of AI applications and collaborations were announced on the first day, the highlight of the conference is the upcoming CEO keynote speech. The market is holding its breath for Amazon’s soon-to-be-released next-generation Nova AI model, as well as the latest updates on its self-developed chip—Trainium 3. The market is closely watching whether these self-developed products are powerful enough to withstand fierce competition from rivals like Google and Microsoft, and to shake up the current market landscape.

These announcements are critical to Amazon. According to a report from JPMorgan, despite strong recent revenue growth, AWS’s cloud market share has seen a slight decline, dropping from 49.7% to 45.1%. The company's stock price has recently pulled back, partly due to investor concerns about its AI strategy and the timetable for new chip releases. The performance of Trainium 3 and Nova will be key to assessing whether Amazon can stay ahead in the race for top AI infrastructure.

New Chip Trainium 3: 40% Performance Boost, But Facing Fierce Competition

Whether Trainium 3 can serve as AWS’s trump card for maintaining competitiveness in the AI era depends on its performance, cost, and market acceptance. The market has high expectations for Trainium 3, but remains cautious.

JPMorgan predicts that Trainium 3 will deliver about a 40% improvement in cost-performance ratio compared to its second-generation product, Trainium 2. The report also notes doubts over whether Amazon’s self-developed chips can catch up with competitors. Currently, Amazon’s Trainium chips have only reached the second or third generation, while Google’s TPUs are at the seventh generation and Nvidia’s GPUs are roughly at the tenth generation.

Additionally, the report cites media sources stating that some customers have encountered technical challenges in using Trainium, and even engineers at Anthropic—an important AI partner for Amazon—prefer to use Google’s TPU. Therefore, the market expects Amazon to provide more specific performance data for Trainium 3 at re:Invent and clarify its “go-to-market” strategy for expanding the customer base.

On release timing, JPMorgan expects Trainium 3 to launch a preview version by the end of 2025, with larger-scale deployment in early 2026. However, there are market concerns that it could be delayed to the second quarter of 2026. This conference is expected to provide a clearer timetable.

Despite these doubts, the Trainium business has shown strong growth momentum. Amazon CEO Andy Jassy revealed in last month’s earnings call that Trainium is now a multi-billion dollar business, with last quarter’s revenue alone growing by 150%, mainly thanks to usage by Anthropic.

New Model Nova: An “All-in-One” Challenger to Gemini?

Aside from chips, Amazon’s self-developed AI foundation model Nova is also a focus of the conference. According to The Information, citing sources familiar with the matter, AWS is expected to release a brand-new Nova model, which can process text, speech, images, and video, and generate text and images—a multimodal model.

The report says AWS is positioning this new model as an “all-in-one” product, aiming to directly compete with top models like Google’s Gemini. This is seen as a crucial step for Amazon to catch up in the domain of cutting-edge models. An AWS salesperson told The Information that the current Nova model is not yet a “mainstream choice” for complex reasoning tasks like OpenAI or Anthropic models.

However, Amazon officially refuted the claim that its AI models are lagging. A company spokesperson stated that Nova is the second most popular model series on its AI service platform Bedrock (Q2 2025 data), with over 10,000 customers including Siemens and Coinbase.

Expanding Partnerships with Anthropic and OpenAI

While developing in-house technology, Amazon is also actively managing its complex relationships with external AI giants.

Anthropic is one of Amazon’s most important AI allies, with its Claude model leading sales on AWS's Bedrock platform. To support this partnership, the jointly built AI supercomputer “Project Rainier” is rapidly expanding, with the number of Trainium 2 chips expected to more than double from about 500,000 to over 1 million by year-end. JPMorgan estimates the project could bring AWS annualized revenue of about $9 billion by 2026.

However, Anthropic is also actively seeking diversified computing support and has entered into large-scale agreements with AWS’s competitors. It has agreed to rent cloud servers worth tens of billions of dollars powered by Google chips and announced last month it will rent servers powered by Nvidia chips from Microsoft Azure. JPMorgan believes this reflects Anthropic’s massive demand for computing power and current supply chain limits, rather than a departure from AWS.

To hedge risks and seize market opportunities, Amazon has also reached out to another AI giant, OpenAI, signing a 7-year, $38 billion agreement to run and scale its workloads on AWS infrastructure, including use of tens of thousands of Nvidia GPUs obtained via AWS.

According to JPMorgan’s report, OpenAI has committed $38 billion over the next seven years to use AWS infrastructure to scale its workloads, which is undoubtedly a major win for Amazon in the AI computing power market.

Market Dynamics: The Fight to Defend Market Share?

From investors’ perspective, Amazon’s AI narrative is full of growth opportunities and uncertainties. JPMorgan’s report notes that although AWS posted its fastest growth in 11 quarters in Q3, Amazon’s stock price retreated to pre-earnings levels, partly due to market concerns over Trainium 3’s launch timing and Anthropic’s partnerships with other cloud providers.

According to IDC data, AWS’s global cloud infrastructure market share has declined from about 49.7% in 2022 to 45.1% in the first half of 2025.

Nevertheless, Wall Street analysts are still optimistic about its prospects. JPMorgan believes that clearer signals on AI strategy at re:Invent should help alleviate the gloom over Amazon’s stock price. Its analysts point out that AWS’s demand trends remain healthy, with Q3 backlog up 22% year-over-year to $200 billion, and new backlog in just October exceeding the entire Q3. The bank forecasts that AWS growth will accelerate in 2026.

Intensive First-Day Announcements, AI Ecosystem Takes Shape

Before the CEO keynote, a slew of announcements on re:Invent’s opening day already showcased AWS’s determination to infuse its AI capabilities across industries.

According to TechRepublic, these announcements include:

• Multicloud Networking: AWS and Google Cloud jointly launched a service called “AWS Interconnect – multicloud,” allowing customers to establish private, high-bandwidth connections between different cloud platforms.
• Financial Services: BlackRock confirmed its Aladdin investment tech platform will run on AWS infrastructure for US clients starting in the second half of 2026. Visa also announced a collaboration with AWS to enable AI agents to securely complete multi-step transactions.
• Industry Applications: Lyft is working with AWS and Anthropic to use the Claude model to provide AI assistance to drivers, shortening problem-solving time by 87%. Nissan is accelerating development of its software-defined vehicle platform on AWS.

These intensive announcements indicate that AWS is endeavoring to build a vast, deeply embedded AI ecosystem to cement its leadership in the cloud computing market and translate it into tangible business growth.

Risk Warning and DisclaimerThe market involves risk and investment needs caution. This article does not constitute personal investment advice, nor does it take into account individual users’ specific investment objectives, financial situations, or needs. Users should consider whether any opinions, viewpoints, or conclusions in this article fit their particular circumstances. Investment decisions based on the article are at their own risk.