The latest datapoint regarding how the hyperscalers are feeling about AI’s potential came this week at AWS re:Invent 2025. AWS CEO Matt Garman put a stake in the ground that we are still early in AI.

The big question going into re:Invent was whether he would signal any cooling in AI demand or a pause in spending. Matt Garman answered that at the top of his keynote:

“Demand keeps skyrocketing. We’ve added 3.8 gigawatts of new data center capacity in the last 12 months alone. And we’re not slowing anything down. We’re only speeding that up.”

That sets the tone for the broader AI infrastructure spending outlook. AWS is spending on both Nvidia and its own silicon. Garman highlighted his long running message that AWS is the best place to run Nvidia GPUs.

Alongside Nvidia, AWS is pushing hard on its own Trainium chip. The latest version, Trainium 3 UltraServers, has more than 4x the compute compared to the prior generation, with their internal Project Rainier already using hundreds of thousands of Trainium chips to run Anthropic’s Claude models. AWS also teased Trainium 4, which underscores the company is determined to quickly rev the product.

For some context, Amazon’s Trainium chips are AWS’s in house alternative to Nvidia, designed specifically for AI training and inference in its data centers. In practice, most large AI customers use Nvidia GPUs when they want maximum performance and compatibility with the broader AI ecosystem, and consider using Trainium when they want to lower costs.

On top of raw compute, AWS announced AI Factories. These let large customers deploy dedicated AWS AI racks inside their own data centers, effectively creating a private mini region that still runs AWS hardware and software. For investors and enterprises, the message is simple, AWS sees AI infrastructure as a long runway and is building as if demand will stay strong.

Garman’s outlook was that AI has moved from basic chatbots to something more powerful. In his words, the true value of AI has not yet been unlocked, and the inflection point is the rise of AI agents that can take actions, not just answer questions. He argued that this shift will have as much impact on business as the Internet or the cloud, and predicted billions of agents inside of every company across every industry.

The main platform for the agent story is Amazon Bedrock and AgentCore. Bedrock continues to expand as a model marketplace, adding new models from Google and Mistral. AgentCore sits on top of that as the platform that allows customers to create their own agents. The basic idea is that Amazon is shifting its focus to give customers a suite of tools to build and maintain agents rather than starting from scratch.

Over the two trading days since the AWS keynote ended, shares of AMZN are down 3.5% versus the Nasdaq which is flat. While shares of AMZN were up 2.5 % in the week prior to the keynote, compared to the Nasdaq up 1.2%, in anticipation of good news from Garman’s keynote, the fact that the stock gave back more than its gains over the past week is the latest evidence that AI investors seem exhausted with good news.

Additionally, the market is overweighting negative data points. For example, the recent Information report that Microsoft salespeople are missing targets triggered a 3% sell off in MSFT. The company later refuted the report, resulting in shares only gaining back about 0.5%.

The bottom line: It seems I am not getting the joke. The market is looking past optimistic comments that suggest we are still early in AI. This raises the question of what good news, if any, could turn investor sentiment positive again. My view is that December earnings and full-year 2026 capex guidance should do the trick.

Disclaimer