© The Indian Express Pvt Ltd
Tags:
Journalism of Courage
In late January this year, DeepSeek R1 triggered a massive sell-off in AI stocks as investors were concerned about an open-weight AI model performing as well as models developed by frontrunners OpenAI and Google while using lesser resources. Nvidia shares plummeted by 17 percent in a single session and close to $600 billion was wiped off from the chipmaking giant’s market cap, making it the largest-ever single day drop for a US company.
Months later, markets appear to be rattled again. As much as $250 billion was wiped off Nvidia’s market cap Tuesday, November 25, with shares falling by 3 per cent.
Gemini 3 is Google’s new series of large language models (LLMs) comprising Gemini 3 Pro, Gemini 3 Pro Image, and Gemini 3 Deep Think reasoning mode.
Gemini 3 Pro is a multimodal reasoning model that is able to provide better answers to more complex questions as it can understand text, images, audio, and spatial cues with support for multiple languages. It offers a one million-input token context window, enabling users to ask longer and more nuanced questions.
According to Google, the model is designed to grasp the intent and context behind prompts so that users get what they need with less prompting. Gemini 3 Pro also uses a technique known as sparse mixture-of-experts (MoE), making it more compute and cost efficient.
Behaviourally, the model is said to have been tested to curb sycophancy, meaning it is less likely to give flattering or overly-agreeable responses. Hallucinations, though, could still be an issue. Google’s progress with Gemini 3 is reflected in the model’s performance across several benchmark tests. It has topped the LM Arena leaderboard and earned top marks on Humanity’s Last Exam and GPQA Diamond.
Early reviews of Gemini 3 have been positive. Marc Benioff, the founder and CEO of Salesforce, said he “is not going back” to using ChatGPT after trying out Gemini 3. Analysts at investment firms DA Davidson and Bank of America Securities were effusive about the model, describing it as the “current state of the art” and “another positive step” for Google.
Major hyperscalers such as Google, Microsoft, Amazon, Meta, and Oracle have come to rely heavily on Nvidia’s graphics processing units (GPUs) to train and develop their own AI models as well as rent them out to AI startups like OpenAI and Anthropic. However, these tech giants have also been looking to develop in-house, custom-built AI chips to reduce their dependence on chipmakers and cut costs — at least in the long run since custom chips have a steep upfront cost that can run into tens of millions of dollars.
While these efforts have been years in the making, Gemini 3 is likely perceived as a meaningful milestone since it runs entirely on Google’s tensor processing units (TPUs). GPUs are said to work well for training AI models, but as the technology matures and competition increases, companies are looking to optimise AI model inferencing, where the model is made to generate outputs on previously unseen data.
This is where TPUs come in. They fall in the category of custom application-specific integrated circuits (ASICs). Other types of custom ASICs include Amazon’s Trainium, Microsoft’s Maia 100, and Tesla’s AI5 chips.
First launched in 2015, Google’s TPU is said to have contributed to the invention of the underlying transformer architecture of LLMs. Earlier this month, Google launched its seventh generation of TPU chips (Ironwood), one million of which are expected to be used by Anthropic to run its Claude models amid growing customer demand.
“TPUs are specifically designed to handle the massive computations involved in training LLMs and can speed up training considerably compared to CPUs. TPUs often come with large amounts of high-bandwidth memory, allowing for the handling of large models and batch sizes during training, which can lead to better model quality,” as per Gemini 3 Pro’s model card.
To be sure, Google does not currently sell its TPUs to other companies and instead, makes their compute and processing power available for customers through Google Cloud. The tech giant has been pitching companies such as Meta on using its specialised AI chips, according to a report by The Information.
The possibility of Meta using Google’s chips to develop its AI models not only triggered a stock market reaction but also prompted a not-so-thinly veiled response from Nvidia.
Stating that it was “delighted by Google’s success”, the chipmaker wrote in a post on X, “NVIDIA is a generation ahead of the industry — it’s the only platform that runs every AI model and does it everywhere computing is done.”
“NVIDIA offers greater performance, versatility, and fungibility than ASICs, which are designed for specific AI frameworks or functions,” it added. While Nvidia’s GPUs may be flexible enough for adoption by many AI companies, they can also be quite expensive, reaching up to $40,000 per unit.
In the past, the company has aggressively sought to entice buyers to stick with it using several strategies, including financing its customers’ chip purchases in circular deals that have fueled speculation of an AI market bubble — and an impending pop.
A few months ago, amid reports that OpenAI was eyeing Google’s chips, Nvidia announced it would invest up to $100 billion in the company as part of a deal in which the ChatGPT-maker would use Nvidia’s next-generation chips. Similarly, Nvidia has said it will invest $10 billion in Anthropic as part of an arrangement that reportedly ensures the startup behind Claude uses Nvidia’s new hardware, in addition to Google’s TPU and Amazon’s Trainium chips.
“This multi-platform approach ensures we can continue advancing Claude’s capabilities while maintaining strong partnerships across the industry,” Anthropic said in a blog post in October this year.
When OpenAI introduced ChatGPT three years ago, it was the starting pistol in the AI race that woke up Google from its slumber. Since then, the tech giant is said to have disrupted itself by breaking down internal silos, streamlining leadership, consolidating AI projects, and overhauling its AI development strategy. Sergey Brin, one of Google’s co-founders, came out of partial retirement and has taken on a day-to-day role at the company.
Google’s deployment of TPUs to take on Nvidia’s dominance is part of a two-front battle. On the other front, Google is leveraging $70 billion in free cash flow to apply pressure on OpenAI, considered to have first-mover advantage in the AI race, whose CEO privately warned that the ChatGPT-maker is facing “rough vibes” and “temporary economic headwinds” following Gemini 3’s successful roll-out, as per The Information.
While tech leaders such as Microsoft’s Satya Nadella have repeatedly argued that cloud computing and AI model development is unlikely to be winner-takes-all scenarios, others such as billionaire entrepreneur Mark Cuban believe that the AI race could end up like the search race in the 1990s.
