AI, Biometrics, Robotics etc. 02 (Jun 23 - Dec 25)

[winston]

DeepSeek Model Development Uses GPUs at 70-90% Cost of Most Advanced Versions: JP Media

The Chinese AI startup DeepSeek's development of the DeepSeek model used GPUs at a unit cost of about 70% of the most advanced versions, the Nikkei reported.

According to the report, the DeepSeek model used NVIDIA (NVDA.US) H800 GPUs, which are a scaled-down version of the H100 GPUs used by American companies.

The report cited semiconductor industry sources as saying that the unit price of the H800 GPUs was between RMB190,000-240,000, about 70-90% of the H100 GPUs' RMB260,000.

Based on the development of the DeepSeek-V3 model using 2,000-3,000 H800 GPUs, the GPU cost was RMB385-721 million.

Source: AAStocks Financial News

http://www.aastocks.com/en/stocks/news/ ... -news/AAFN

[behappyalways]

Guard against U.S.-China dominance!
Mark Hong and Modi hold AI summit 3.6 trillion investment law industry

防美中獨霸！馬克洪偕莫迪辦AI峰會　3.6兆投資法產業｜秦綾謙｜FOCUS全球新聞20250211 ‪@tvbsfocus‬

https://m.youtube.com/watch?v=i_kmzCfLUgM&t=194s

[behappyalways]

The world's fastest land-based robot dog "Black Panther", which can run 10 meters in 1 second, is unveiled

1秒跑10公尺 陸製全球最快機器狗「黑豹」亮相｜TVBS新聞 ‪@TVBSNEWS02‬

https://m.youtube.com/watch?v=55nTl4jaRrg

[behappyalways]

"Swindler": Musk And Altman Throw Punches After Musk-Led Group Leads $97.4 Billion Bid For Control Of OpenAI
https://www.zerohedge.com/political/mus ... rol-openai


"I Don't Think He's A Happy Guy": Sam Altman Talks Trash About Musk After OpenAI Buyout Offer
https://www.zerohedge.com/technology/i- ... yout-offer

[winston]

The fears over DeepSeek are vastly overblown

by Eric Fry

It appears that the AI that DeepSeek uses, may simply optimize various packets of data, which wouldn’t use a lot of computing power.

However, other kinds of machine learning utilize a lot more computing power – and that’s why more data centers and an enhanced electrical grid are still necessary.

With more folks starting to come to this same conclusion, AI-related stocks are firming back up, especially in the wake of positive quarterly results from market leaders.

Source: Smart Money

[winston]

The fears over DeepSeek are vastly overblown

by Eric Fry

It appears that the AI that DeepSeek uses, may simply optimize various packets of data, which wouldn’t use a lot of computing power.

However, other kinds of machine learning utilize a lot more computing power – and that’s why more data centers and an enhanced electrical grid are still necessary.

With more folks starting to come to this same conclusion, AI-related stocks are firming back up, especially in the wake of positive quarterly results from market leaders.

Source: Smart Money

[winston]

How did DeepSeek build its AI with less money?

AI companies typically train their chatbots using supercomputers packed with 16,000 specialised chips or more. But DeepSeek said it needed only about 2,000.

Typically, chips multiply numbers that fit into 16 bits of memory. But DeepSeek squeezed each number into only 8 bits of memory – half the space. In essence, it lopped several decimals from each number.

This meant that each calculation was less accurate. But that didn’t matter. The calculations were accurate enough to produce a really powerful neural network.

Source: NY Times

https://www.thestar.com.my/tech/tech-ne ... e=smartech

[winston]

<Research>M Stanley: Quicker AI Adoption Benefits Baidu/ Alibaba/ Tencent; Top Pick Tencent

Since the launch of DeepSeek on January 24, the HSTECH has grossed surges about 20%, fueled by a number of AI-related developments, including ByteDance's and BABA-W (09988.HK) 's launch of cost-effective large language models (LLMs),

DeepSeek's integration into the Chinese and U.S. platforms, and reports of a possible partnership between Apple (AAPL.US) and Alibaba, as well as reports that Chinese President Xi Jinping may host a forum with business leaders, Morgan Stanley’s report said.

The broker saw lower training or reasoning costs to drive faster, favoring BIDU-SW (09888.HK), Alibaba, TENCENT (00700.HK) (collectively known as “BAT”). As the results period is approaching this week, the broker considered the market will focus more on management's comments on AI (e.g., applications, monetization and capex) than on core business performance.

Among them, Tencent remains the broker's top pick as it is already pampered by AI-driven ad tech upgrades on the monetization front, whereas the rollout of AI features in WeChat (e.g., access to DeepSeek for gray testing) is expected to be a catalyst.

Secondly, the broker preferred Alibaba and maintained its Equalweight rating on Alibaba (BABA.US), believing that the stock's recent gains have reflected the upside from its cloud business.

Under the base case, the broker foresaw Baidu's cloud revenue to grow at an average annual rate of 18% to RMB163 billion in 2026 and 2027, with EBITA margins stabilizing at 15%.

Under the bullish case, the broker estimated cloud revenue growth to accelerate to 20% and 25% in 2026 and 2027 to RMB179 billion, with EBITA margins of around 20%.

As for Baidu (BIDU.US) , the broker kept an Equalweight rating on the company. Although the company has a first-mover advantage in its all-AI strategy, its core search business faces the risk of being affected by generative AI, and the intensified competition from Tencent and ByteDance makes monetization uncertain.

Related News: UBS Elevates TENCENT (00700.HK) TP to $593 on 3 Potential Surprises in 2025

Source: AAstocks.com

[behappyalways]

AI competition increases!
Musk strongly promotes Grok 3 DeepSeek, but South Korea pulls it off the shelves

AI競爭加劇! 馬斯克力推Grok 3 DeepSeek南韓遭下架｜方念華｜FOCUS全球新聞 20250219 ‪@tvbsfocus‬

https://m.youtube.com/watch?v=UpUpNn_P8zw

[behappyalways]

Asked in Chinese about the origin of "Xinqi", DeepSeek answered "China"
中文提問「辛奇」原產地 DeepSeek答「中國」｜FOCUS午間新聞 20250220 ‪@tvbsfocus‬
https://m.youtube.com/watch?v=AoFj0o3lQq4


Meet the Protoclone
https://x.com/BrianRoemmele/status/1892316274539319493