Nvidia's Slide Deepens as Broadcom's CEO Calls AI a "Casino" and Microsoft Rolls Out a Rival Chip

Nvidia (NVDA) shares slipped 0.8% to $224.45 in Tuesday's premarket trading, extending a five-session decline to 5.4%. Two headlines hit at once: Broadcom's chief executive questioned whether investors really understand what they are paying for in AI, and Microsoft unveiled a new in-house chip it plans to run most of its AI inference on. Together they revive two of the market's biggest worries about Nvidia, whether AI spending can keep growing and whether its largest customers will keep buying its chips.

Premarket moves

Prices are from Tuesday premarket trading on October 13, not closing figures.

Broadcom's CEO sounds a warning

Broadcom (AVGO) chief executive Hock Tan said many AI investors lack a clear understanding of what they are investing in, likening the environment to a casino. He also cautioned that the supply of capital could eventually limit the trillions of dollars projected for AI infrastructure.

The remarks carry weight because of who made them. Broadcom is one of the biggest winners of the AI build-out, designing custom AI chips for several of the largest cloud companies. When a supplier that benefits from the spending questions how long it can last, investors listen, and Broadcom's own shares fell 1.1% premarket.

Microsoft's Maia 300 targets inference

Microsoft (MSFT) unveiled Maia 300, the latest generation of its custom AI processor, and said it plans to use it as the primary system for AI inference, the work of running trained models to answer user requests. That is the fastest-growing part of AI computing, and a market Nvidia is counting on as training demand matures.

Microsoft expects Maia 300 to process 40% more AI tokens per megawatt of power than leading GPU-based systems. With power now one of the main limits on how fast data centers can grow, getting more output from each megawatt matters as much as raw speed. Maia 300 follows Maia 200, which Microsoft introduced in January as an inference chip aimed at cutting the cost of serving AI models.

What it means for Nvidia

Microsoft is one of Nvidia's largest customers, so any sign it is shifting work to its own chips raises the concern that the big cloud providers will need fewer Nvidia GPUs over time. Google, Amazon and Meta are pursuing the same strategy with their own processors.

Microsoft was careful to say it will keep offering multiple chip options to its cloud customers, so custom silicon is an addition to Nvidia, not a replacement. In practice, that likely means Maia handles more of Microsoft's own inference work while Nvidia GPUs remain available to Azure customers who want them.

The bigger shift may be in where Nvidia finds its next wave of demand. As hyperscalers build their own chips, Nvidia's push to build an ecosystem of smaller AI companies, including neocloud providers and AI startups, could become increasingly important to keeping its order book full.

What to watch

• Hyperscaler capex guidance. Microsoft, Alphabet, Amazon and Meta report later this month. Their spending plans, and how much goes to custom chips, will matter more than any single comment.
• Maia 300 deployment. How quickly Microsoft moves inference onto Maia, and whether independent benchmarks back up the 40% efficiency claim.
• Nvidia's support level. Whether buyers step in after a 5.4% slide, or the decline extends as AI spending doubts spread.
• Neocloud demand. Orders from smaller AI cloud providers, which are increasingly important to Nvidia's growth outside the hyperscalers.