- New emerging AI models showing that they can work with just a handful of Nvidia chips.
- Google is a company that follows Deepseek in the manufacture of a more powerful AI that requires less calculation.
- It does not look like a deep level problem for Nvidia – but there are warnings.
A new generation of AI models leads to more power of fewer fleas. If they trigger another Panic Because Nvidia is another matter.
Google led this week with a collection of smaller models – Gemma 3 – which seem to pack a serious punch with an out -of -competition feat: they work smoothly with a single NVIDIA chip, known as GPU.
Revealing the models on Wednesday, the CEO of Google Sundar Pichai highlighted their efficiency in a X postWriting that “you would need at least 10 times the calculation to get similar performance from other models.”
AdhereA startup based in Toronto led by former Googler Aidan Gomez, also published Thursday a new model called Thursday Order A, which is described as a “cutting edge” model that works on only two GPUs. (Business Insider, alongside other publishers, has Coheru understanding on violation of copyright.)
One of the main Deepseek lessons conferred on the world when it published an AI model in January was the ability to do more with less. The Chinese startup said its R1 model was competitive with the O1 model of Openai While saying that he needed fewer fleas.
The complaint sparked the biggest trial in one day in American stock market history, in which Nvidia has lost nearly $ 600 billion in value. The market wondered if a more efficient AI would reduce the demand for NVIDIA fleas – a request that helped it reach a record turnover in the year in 2024 of $ 130.5 billion.
At first glance, this new wave of models of AI seems to constitute an even greater threat, because they claim to be at the cutting edge of technology while needing a handful of GPU to execute.
A Gemma’s performance table on the Chatbot Arena industry workbook, for example – shared by Pichai – has shown that the model has surpassed those of Deepseek, Openai and Meta while being executed on less GPU.
But just because more companies learn to repress more performance from their AI with less jet Nvidia.
On the one hand, while the Deepseek saga took place, technological CEOs quickly noted Jevons paradox. The economic principle suggests that, as technology becomes more effective, the consumption of this technology will increase – and not decrease.
He could help explain why Google himself said he was planning to increase his Capital expenditure linked to AI This year at 75 billion dollars, which generally includes GPUs housed in essential data centers for AI.
Google was one of the main buyers of the latest generation of GPU Nvidia – the Blackwell chips Presented last year – it is therefore plausible to expect that they are among those who are ready to spend on the new NVIDIA GPU should unveil during its GTC event next week.
Until now, the market does not seem to be concerned by the latest developments in chip efficiency – the course of NVIDIA action has increased by around 6% since Tuesday.
There is a small warning to this.
Although the new Google Gemma models can use a single NVIDIA GPU to operate, it seems that the training of new models took place with the own Google chips, called tensor treatment units, or TPU.
Technology and Google technology giants have spent years working on their own silicon as a means of reducing their dependence on Nvidia, so Gemma poses a curious situation in which Google produced a competitive AI model without using NVIDIA GPU for training.
However, it is unlikely that a company like Google will reduce its dependence on Nvidia GPU in a significant way – and for a simple reason.
The company’s pressure to produce more effective models occurs in parallel with its development of more powerful and large -scale gemini ai models which aim to push the limits of intelligence. The success of these models, for the moment, depends on access to as much calculation as possible.
This approach was recently validated by the release of Grok 3, the latest AI frontier model of Elon Musk’s startup. Its version came with a note that a future version of the model would be formed on a larger one, GPU cluster of 200,000.
The future of the development of AI therefore resembles one in which two different paths emerge in tandem: one in which smaller and more effective models emerge which work on less GPU, and another in which large -scale models continue to browse as much GPU as possible.