Deepseek’s new open source AI prostitution R1 has caused the sale of NVIDIA shares, and consumer apps have risen to the top of the app store.
Last month, DeepSeek stated that it had trained a model using about 2,000 data centers of NVIDIA H800 GPU at about $ 5.5 million in about two months. Last week, we published a dissertation that shows that the latest model performance matches the world’s most advanced inference model. These models are trained in data centers that spend billions of dollars on NVIDIA’s faster and very expensive AI chips.
The reaction in the high -tech industry as a whole to the high -performance and low -cost models of Deepseek was wild. For example, PAT GELSINGER photographed X with Glee and posted, “Thank you for the DeepSeek team.”
Of course, GELSINGER is the current chairman of GLOO, an IPO bounce startup, a recent CEO of Intel, a hardware engineer, and a message for churches and engagement platforms. He left Intel four years later and tried to chase NVIDIA with Gaudi 3 AI, an alternative AI GPU for Intel.
GELSINGER wrote that DeepSeek should remind you of the three most important lessons in the high -tech industry. Under the constraints, ingenuity flourishes. He wrote, “Open victory. DeepSeek will help you reset more and more closed worlds in basic AI model work.” Openai and humanity are both closed sources.
Gelsinger told TechCrunch that R1 was very impressive, so GLOO has already decided not to pay Openaii. GLOO has built an AI service called KALLM. It provides chatbots and other services.
“My adhesive engineer runs R1 today,” he said. “They may have been able to execute O1. Well, they can only access O1 via API.”
Instead, GLOO expects Kallm to rebuild Kallm from zero in our own basic model. “It’s exciting.”
He said DeepSeek thinks that AI will be a very affordable price. Good AI is everywhere. “I want a better AI with my welding. I want a better AI in hearing aids. I want more AI on the phone. Like the EV voice recognition, AI is better for embedded devices. I need it. ”
GELSINGER’s happy reaction was probably confronted with other people who were not excited to have a much more affordable challenger in the basics of reasoning. AI is growing more expensive.
Others implicitly responded that Deep Sheque had somehow destroyed the number, and training must have been more expensive. Some people thought that it was not used to use high -end chips for restricting AI chips to China. Others have driven a hole in the performance and found a spot where other models have improved. Furthermore, O3, the next model of Openai, thinks that it will overtake R1 when released and restore the current situation.
GELSINGER scoops all of them. “Given that most of the work was done in China, there is no complete transparency,” he said. “Still, all evidence is 10 to 50 times cheaper than O1.”
DeepSeek says, “I prove that AI can be advanced by the creativity of engineering. This is not throwing more hardware power and cannot calculate the resource in problem. It is thrilling.” I said.
This is, as in privacy and censorship concerns, Gelsinger shakes a pHOR -like head for all Chinese developers who suggest everything.
“It’s embarrassing for our community to remember the power of an open ecosystem for Chinese people,” he said.
TechCrunch has a newsletter focusing on AI! Sign up here and get it on the reception tray every Wednesday.
