The big conclusion of it has become a big job and a bigger business

Thanks to innovations like Deepseek, training has become cheaper. However, the conclusion is becoming more demanding as we ask him to think more before answering our questions. Nvidia, GROK and Cerebras systems (Cambrian-Ai research clients) have released all accelerators and mass infrastructure to support this trend. I suspect we will see more from Nvidia about the conclusion next week than training, including clouds, robots and cars. Jensen Huang has said that this style of reasoning of the conclusion processing is 100 times more demanding. I have found in a recent experiment that reasoning can be 200 times more expensive, but much more intelligent and valuable!

Cerebras takes conclusion to a new level

Cerebras Systems, the Creator of Mass -size chips, frisbee size, has flipped a plan to build six new data centers from entering the “high value” sign business. The company claims to become the largest provider of such conclusions at the end of this year. Thousands will exceed 40 million Llama signs of 70b per second.

High -value signs carry more contextual information and are usually more important to understand the overall meaning of a text. They often represent key concepts, rare words or specialized terminology. High value signs consume more calculator sources and can cost more to process. This is because they usually require more attention from the model and contribute more significantly to the end result. Low value signs, which are more common and less dense computing, usually require fewer processing sources. Clearly, Cerebras is aiming for problems that are suitable for its Mass scale access to it.

This level of performance in high value signs is attracting clients of new enterprises that also need elastic services to meet their needs. For example, Alphasense, a major market intelligence platform, has been moved to the Cerebras conclusion, replacing a model provider with closed resources three first. The company has also landed confused, mistral, hugging the face and other high -value inferential users, giving the conclusion performance 10 to 20 times faster than the alternative.

The conclusion revolution is starting

Next week, we will hear more about “high value” signs from Nvidia to GTC, as the conclusion market exceeds global income training. Markets such as autonomous vehicles, robots and sovereign data centers all depend on the rapid conclusion, and Nvidia does not plan to allow that market to pass them. The high -value concept is new, and platforms like Cerebras and Nvidia LVL72 are ideal for its distribution.

Unveils: This article expresses the author’s opinions and should not be taken as advice to buy or invest in the companies mentioned. My firm, Cambrian-Ai Research, is lucky to have many semiconductor firms as our clients, including Baya Systems Brainchip, Cadence, Cerebras Systems, D-Matrix, Esperanto, Flex, Groq, IBM, Intel, Micron, Qualcomm, Graphcore, Sima, Tensorrent, Ventana, Ventana, and Scortors of Investors. I have no investment positions in any of the companies mentioned in this article. For more information, please visit our website at https://cambrian-i.com.