Raytheon awarded $71 million in Navy contracts for missile systems
Investing.com -- In a note to clients Thursday, Rosenblatt analysts offered further insights into DeepSeek-R1, the AI model that has drawn industry-wide attention for its low training costs and competitive performance.
While the model was reportedly trained for less than $6 million in just two months, Rosenblatt says the actual cost was likely much higher.
“All-in costs to train R1 are likely well above $6 million as it appears it was a refined version of the older R1-Zero,” the firm wrote.
A key concern raised by Rosenblatt is that DeepSeek-R1 appears to be ‘"distilled’ from larger, more capable model outputs such as OpenAI’s, and is in violation of service agreements.”
This is said to raise ethical and legal questions about its data sourcing and model development process.
Additionally, Rosenblatt highlights transparency issues, noting that “DeepSeek/China have related issues on data collection, answers that favor China ‘views’, and lack of transparency.”
One of the most notable aspects of DeepSeek-R1 is its use of Test-Time scaling, a relatively new AI scaling approach that shifts compute resources away from training to the inference stage.
“Test-Time scaling is believed to be the next wave of AI scaling as pre-training and post-training scaling are seeing diminishing returns,” Rosenblatt noted, adding that this method was “pushed by Nvidia (NASDAQ:NVDA) openly at CES recently.”
Rosenblatt expects Test-Time scaling to become a mainstream AI development strategy, stating that “most AI models will adapt to incorporate Test-Time scaling as a normal evolution on how to best balance where compute is located.”
The firm also suggests that this shift will particularly benefit edge AI plays, as inference workloads gain more importance.