- Google’s new Gemma 3 models, as reported by CNBC’s Deirdre Bosa, run efficiently on a single chip – Nvidia’s H100 or Google’s TPU – outpacing DeepSeek’s R1 (34 H100s) and Meta’s Llama 3 (16), spotlighting a cost-effective edge in the AI inference phase.
- The announcement underscores Google’s TPU advantage, challenging Nvidia’s dominance as hyperscalers like Amazon (Trainium) and Meta, plus startups like Cerebras, push in-house chips, while Nvidia’s stock, down 15% yearly, rose nearly 6% today.
- With AI shifting to inference, Google’s efficiency play – backed by Anthropic’s TPU use and massive data center spending like Stargate – raises questions about Nvidia’s margins, with its GTC next week pivotal to its counter-strategy.
Google is once again making waves in the AI world, as CNBC’s Deirdre Bosa reported, with its new Gemma 3 family of open-source models that can run efficiently on a single chip – whether a GPU, like Nvidia’s H100, or Google’s custom TPU – highlighting a cost-effective edge over rivals like DeepSeek’s R1, which needs 34 H100s, or Meta’s Llama 3, requiring 16. This move showcases Google’s knack for building powerful yet lean AI tools, a strength Bosa notes is well-known to insiders but often missed by the broader market, offering businesses a cheaper way to tap into advanced tech. With Nvidia’s (NVDA) stock up 7% today yet down 14% for the year, Google’s announcement doubles as a flex of its TPU advantage, signaling a shift in the AI race from training models to running them efficiently, a phase called inference where competition is heating up.
The tech landscape is buzzing as big players like Google (GOOG, GOOGL), Amazon (AMZN), and Meta (META) push their own AI chips to cut costs and challenge Nvidia’s grip on the market, where its chips have long ruled both training and inference. Bosa pointed out that Anthropic leans heavily on Google Cloud’s TPUs for inference, while Amazon’s Trainium chips and Meta’s new in-house design are gaining ground, and even startups like Cerebras are powering models for companies like Perplexity and Mistral. Nvidia’s CEO Jensen Huang insists its ecosystem remains top-tier for inference, but with hyperscalers pouring money into alternatives – think hundreds of billions for projects like Stargate – the economics are tilting, raising questions about Nvidia’s sky-high margins as more efficient, lower-cost options emerge.
It’s worth noting that Google’s latest release isn’t just about tech—it’s a strategic play in a crowded field where efficiency could redefine who leads AI’s next chapter. Bosa emphasized that next week’s Nvidia GTC event will be a big moment, with investors eager to see how Nvidia counters this inference-phase push amid its 14% yearly drop, despite today’s 7% bounce. Google’s ability to run Gemma 3 on a single chip versus DeepSeek’s 34 or Meta’s 16 isn’t just a stat—it’s a signal that custom chips like TPUs could shake up the game, especially as OpenAI and others race to build their own solutions. This clash of giants and upstarts shows AI isn’t just about power anymore—it’s about smart, affordable delivery, and Google’s latest step could nudge the market in a new direction.
WallStreetPit does not provide investment advice. All rights reserved.
Leave a Reply