What do most people call a chatbot these days?
You guessed it. In the U.S., when someone mentions a chatbot, most of us nowadays think, “ChatGPT.”
Even though Google’s Gemini, Anthropic’s Claude, Meta’s Llama, and xAI’s Grok are on par with OpenAI’s technology, that’s the advantage of being first to a market.
Well, in France, they call it Le Chat…
At least that’s what Paris, France-based Mistral calls its frontier generative AI model: “Le Chat: Your AI assistant for life and work.”
Mistral is Europe’s best chance at competing with the U.S.-based firms in this race to artificial general intelligence (AGI).
Founded in 2023, less than 24 months ago, Mistral has been making great progress with its technology and it has now raised about $1.2 billion to date.
It’s a paltry amount compared to what the others have been spending, but that doesn’t mean it’s not an impressive AI. Le Chat can perform all the functions that the other frontier models can, to varying degrees of accuracy.
And as we’ve been witnessing over the last two years, every month brings a new release of someone’s software, jockeying for performance improvements and accuracy on various benchmarks.
For anyone interested in trying Le Chat for free, you can experiment right here.
What’s unique about Mistral’s Le Chat is that it is “true” open-source software. Mistral grants users access to the software code, data, and weights used to train the AI model. This is unique in the industry, as it allows users to customize the AI model for their specific needs.
An example might be that a company or other organization might want to customize the weights to perform better at certain tasks than others. Mistral allows this with its open-source model. This flexibility is something that’s not offered by any of the other major frontier AI models.
While there is nothing really special about Mistral’s performance on popular AI benchmarks, what is interesting about Mistral’s latest announcement with Le Chat, is the inference performance (i.e. how quickly Le Chat responds to inquiries).
Source: Cerebras
Mistral now holds the prize for the world’s fastest AI assistant. What’s the secret? A currently private semiconductor company that my readers will know well – Cerebras.
Cerebras is unique in the AI-specific semiconductor industry, in that it has designed its product specifically for inference and on a scale that no other semiconductor company is doing today. Below is an image of Cerebras’ third-generation Wafer Scale Engine (WSE-3).
Source: Cerebras
Having such a physically large chip enables a level of performance and integration not possible by stringing separate chips together. This gives Cerebras performance advantages in memory speed, bandwidth, and overall inference performance.
Think of inference as the “working” mode of an AI model, as opposed to the training mode. During training, a model learns patterns and connections from vast amounts of data – like how a student studies and learns concepts. However, during inference, the model applies what it learned to new situations – like how a student uses what they learned to answer questions on a test.
Inference is what makes an AI feel human. When we interact with each other, we take in data from various inputs (i.e. our full senses) during the interaction, and then we apply that data to our pre-learned parameters. From that, we generate appropriate responses.
What’s remarkable about Mistral is that when we interact with Le Chat – running on Cerebras semiconductors – we receive near-instantaneous results.
For comparison, below is a screenshot of three different AI models tasked with writing software code for the snake game in the programming language Python. Software coding is one of the most-used applications of these frontier AI models, as it enables a major productivity boost in terms of both “writing” code, as well as in debugging code.
Source: Cerebras
As we can see above, Le Chat running on Cerebras completes the task in 1.3 seconds, compared to Anthropic’s Claude in 19 seconds, and OpenAI’s GPT-4o in 46 seconds.
And the best part is that Le Chat isn’t a nation-state-backed data surveillance piece of malware like DeepSeek. Le Chat is also available on both iOS and Android operating systems.
I really like this technology example because it highlights the importance of both the software (the AI) and the hardware (the AI-specific semiconductors) to a generative AI model. It also emphasizes one of the biggest trends of 2025 regarding AI: inference.
Now that the current batch of frontier AI models has immense utility, the need for AI data centers to support AI inference – the running as opposed to the training of AI applications – is driving even higher levels of AI infrastructure spend.
This is precisely why hyperscale players like Amazon have been increasing their forecasts for AI infrastructure spend, despite the incorrect conclusions many have drawn from the false narrative about how cheap it was to develop DeepSeek.
Amazon knows what’s real and what’s not, and it knows that training costs are going to continue to be expensive, and inference workloads are going to increase exponentially for years to come. That’s why it increased its forecast for AI infrastructure to $100 billion this year, a decision that was made after the DeepSeek news came out. That’s $100 billion for a single company in a single year.
Ironically, Amazon’s share price fell 4% after the news, but Amazon’s CEO Andy Jassy was clear…
Jassy knows investors will eventually be happy that Amazon is making the investment in “the biggest technology shift in business since the internet.”
Another quote that I particularly liked from Jassy was this…
Sometimes people make the assumption that if you’re able to decrease the cost of any type of technology component… somehow it’s going to lead to less total spend on technology. We’ve never seen that to be the case.
This is precisely why AI spending won’t slow down. It will continue to increase. It’s also why we’ll continue to see more investment into AI-related companies at increasing valuations.
Just on Friday, we learned that OpenAI is closing an additional $40 billion of funding at a $300 billion valuation. It’s an astounding number, especially considering the company raised capital last October at $157 billion. This latest round is nearly double the valuation from less than four months ago.
I know that might seem crazy, but consider this…
In October last year, the OpenAI revenue forecast for 2025 was $11.7 billion. That in itself is incredible, considering its GPT products had only been around for less than two years. That forecast put its $157 billion valuation around 13.4 times forecasted 2025 sales. It’s “rich,” but it’s not unreasonable considering how fast revenue is growing for OpenAI.
And I’ve now seen a recent forecast for 2025 as high as $30 billion. That’s how much has changed in just four months. That would put its current funding round valuation of $300 billion at “just” 10 times forecasted sales. Not unusual at all.
In fact, it might just be a bargain.
Looking at popular AI stock Palantir (PLTR), which is now worth $256 billion on forecasted 2025 revenues of $3.7 billion. It’s trading at an ebullient 69 times forecasted 2025 sales.
Institutional capital is willing to pay these multiples in both private and public markets because the revenues are growing so quickly that it’s hard to update their forecasting models fast enough to keep up.
DeepSeek hasn’t given anyone pause. It’s the opposite. Capital formation and investment continue to pick up.
Jeff
The Bleeding Edge is the only free newsletter that delivers daily insights and information from the high-tech world as well as topics and trends relevant to investments.
The Bleeding Edge is the only free newsletter that delivers daily insights and information from the high-tech world as well as topics and trends relevant to investments.
