Artificial Intelligence - Page 18

Chinese tech companies are reportedly preparing for an AI future without NVIDIA as their stockpile of AI GPUs will run out in early 2026, with accelerated testing of homegrown AI chips, and a small window for the major decision (and the full switch) from NVIDIA to Huawei is expected to kick start a 3-month disruption.

With more and more US export restrictions hitting China, it is getting harder and harder for the country to acquire the bleeding-edge in AI silicon. Chinese tech giants including Alibaba, Tencent, and Baidu are all beginning to test alternative semiconductors to meet the ever-growing AI-related internal and consumer demands, according to industry executives, reports the Financial Times.

These companies have been forced to bring forward their contingency planning as the continued tensions between US and China in the new Trump administration causing headaches in China. Last week we saw the Trump administration stopping the NVIDIA H20 AI GPU from being sold in China, which was already a cut-down version of the AI GPU, and now it's been banned.

AMD is expected to adjust their downgraded chip designs to form a new family of AI GPUs that will comply with the constantly evolving US export restrictions starting in July 2025, with a new rumored AI GPU that will battle NVIDIA and Huawei in China.

In a post by insider @Jukanlosreve we're hearing that AMD will soon have an AI GPU answer for the Chinese market that complies with US export restrictions, with DigiTimes reporting that AMD is expected to offer an RDNA 4-based Radeon AI PRO R9700 workstation GPU for the Chinese market.

AMD would be cutting down its specs to meet the US export restrictions, but will have just GDDR6 memory on its China-specific Radeon AI PRO R9700 GPU, versus NVIDIA with its rumored B20 shifting from HBM to GDDR memory, and GDDR7 at that (with far more memory bandwidth, which is an important part of AI workload performance).

NVIDIA has said that Huawei's new AI developments are far more capable than people though, so much so that they're more powerful than what NVIDIA is allowed to get into China, effectively forcing them out of the market.

Considering NVIDIA kick-started the AI revolution without competition from anyone (including AMD and Intel) but over the last few years multiple competitors have entered the ring, and one of them is from China: Huawei. Huawei has been hard at work developing new AI hardware to compete with NVIDIA, which has been facing ever-increasing US export regulations that have stopped the company from pumping China full of its AI chips.

In a recent interview with Bloomberg, NVIDIA CEO Jensen Huang said: "Huawei's technology, based on our best understanding at the moment, is probably comparable to an H200. They've been moving quite fast. They've also offered this new system called Cloud Matrix, which scales up to even a larger system than our latest generation, Grace Blackwell. Huawei, as you know, is a formidable technology company. And they're not sitting still. They look for ways to compete, and they're quite formidable".

NVIDIA supply chain partners are accelerating the production of its flagship AI data center racks, after technical issues had delayed shipments, but now these issues have been resolved and supply of GB300 AI server racks will now flow.

In a new report from The Financial Times, we're hearing that NVIDIA supply chain partners including Foxconn, Inventec, Dell, and Wistron, have "made a series of breakthroughs" that have allowed them to start shipments of NVIDIA's high Blackwell AI servers. NVIDIA experienced technical problems that started at the end of 2024 that disrupted production, and their annual sales targets.

This new advancement is also seeing NVIDIA laying the groundwork of its next-gen GB300 NVL72 AI server racks, featuring 288GB of HBM3E memory per GPU and offering even more compute power than GB200 NVL72 AI servers, which will handle even more complex reasoning models including OpenAI O1 and DeepSeek R1 models. NVIDIA CEO Jensen Huang said last week that GB300 will launch in Q3 2025.

NVIDIA recently announced that it would be adopting a new 800V high-voltage DC power distribution system, which will power the waves of next-generation data centers.

In a new post on X by insider @Jukanlosreve we're hearing that 3 semiconductor companies have been officially named as NVIDIA partners in developing this new 800V high-voltage system, referred to as the "three major power IC players" with Infineon, Texas Instruments (TI), and Navitas.

NVIDIA's new 800V high-voltage DC power distribution system has an extremely complex design, with companies forced to offer robust and diverse power solutions to help NVIDIA achieve its lofty goals. NVIDIA is most likely preparing to introduce more power semiconductors in the future, but in the early stages of this new ultra-high-end battleground, these three companies will hold a very important competitive edge.

Pavel Durov, the founder and chief executive officer (CEO) of Telegram, has announced that Elon Musk's xAI is paying the messaging platform $300 million for it to adopt Grok, the AI-powered chatbot created by xAI.

Durov explained in a recent X post the details of the deal, which when reading appear to be extremely in the favor of Telegram, as the $300 million the social messaging platform is getting paid is only for a year-long partnership, but it also includes other benefits such as 50% of revenue from xAI subscriptions sold via Telegram. This means that if any users sign up for Grok Premium or Premium+ subscriptions through Telegram, Durov's social messaging platform will take 50% of that sale, presumably for the subscription duration, or until the partnership is ended.

Telegram is home to more than 1 billion monthly users, which is about double the size of X. Perhaps X thinks of this deal as marketing expenditure to get Grok into many more hands, potentially attracting them to the subscription service. However, there could be another motive. 1 billion users = a lot of user data, and Grok is currently training on data on X. It remains to be seen if the deal between the xAI and Telegram includes Grok being able to train on Telegram user data.

OpenAI CEO Sam Altman and former Apple chief design officer Jony Ive are currently developing a new hardware product that will ship under the ChatGPT creators company, and when it's ready, there will be 100 million units going out to consumers.

OpenAI is preparing to broaden its horizons beyond ground-breaking AI-powered software and step into the realm of hardware. In a recent interview with The Wall Street Journal, OpenAI CEO Sam Altman revealed he is developing this new secret device alongside Ive, which neither Altman nor Ives specified a description for. 

Presumably, it would be a device that enables ChatGPT to be taken on the go by users but somehow provides more value to consumers than the currently available software downloadable on smartphone devices via an app. If it's an AI companion device, possibly wearable, it would likely be running an AI model locally rather than requesting processing power from OpenAI's servers for an answer to a user's question.

Samsung will be bursting onto the next-gen HBM4 memory business later this year, with its new 1c DRAM driving its upcoming HBM4 memory, which the South Korean firm will be sampling to tech giants including AI GPU leader NVIDIA, in Q3 2025 ready for next-gen AI chips arriving later this year, and into 2026.

In a new post from @Jukenreslove on X and reported on by South Korean media outlet The Bell, we're hearing that Vice Chairman Jeon Young-hyun -- who was freshly minted the "firefighter" of Samsung Electronics' semiconductor business -- looking closely over the DRAM business as the new boss as the head of the Device Solutions (DS) division.

The chairman has been personally overseeing and instructing Samsung's huge redesign of 10nm 6th generation DRAM (1c DRAM) which is being prepared as the next-generation HBM4 core die, with Jeon "taking care" of major product development issues. Jeon was appointed the head of the DS division back in May 2024, and promoted to the head of the memory division and president of Samsung Advanced Institute of Technology (SAIT) in November.

NVIDIA showcased some impressive AI technologies at Computex 2025, with the company leading the charge of AI integration into the gaming space. One of those technologies is Project G-Assist, and I got a behind-the-scenes demo of what is new with the feature. For those who don't know, it is a local AI agent designed to enable a user to control their PC through voice/text chat.

G-Assist has received some considerable improvements, with NVIDIA demonstrating the AI agent's ability to launch titles through a user activating the microphone and simply saying, "Open Counter-Strike 2." Moreover, G-Assist, currently available to download within the NVIDIA App, can also optimize a game's settings without user intervention. 

A user will interact with G-Assist either through text chat or via a microphone and simply ask it to optimize the settings for a game. The AI agent will go ahead and optimize the in-game settings based on the PC configuration, improving average framerate.

NVIDIA's new special edition Blackwell AI GPU for China is expected to roll out later this year, and could be named 6000D or B40 according to the latest reports.

In a new post on X by insider @Jukanlosreve, we're hearing that the new NVIDIA 6000D or B40 -- whatever it ends up being called -- will be based on Blackwell and not Hopper, with NVIDIA CEO Jensen Huang saying last week that NVIDIA's new AI GPU for China wouldn't be based on Hopper as H20 "can't be modified anymore".

NVIDIA's new 6000D or B40 will reportedly shift to GDDR7 instead of HBM, also lining up with previous rumors, with @Jukanlosreve saying that the GDDR7 memory bandwidth will be around 1.7TB/sec, and that the NVLink speed is estimated to be around 550GB/sec per direction, and that it will support CUDA.

Streamlabs and NVIDIA have showcased a significant upgrade to the suite of features available within the livestreaming software, with the head of Streamlabs, Ashray Urs, giving us a live demo of the new AI-powered software working in action.

A partnership between Streamlabs, Inworld AI, and NVIDIA has created what is called the Streamlabs Intelligent Streaming Agent, a new digital human that is capable of assisting a livestreamer by carrying out commands in real-time. Ashray showed three sets of functionality with the new AI Agent: It being a "sidekick" to the streamer, which includes how a streamer can interact with the agent and leverage it being aware of what is happening in-game, the agent being used a stream producer, which includes it leveling up the overall production quality, and how the agent can be used as technical support source.

The above video shows Ashray asking the AI agent, "Where should I drop?" while in a simulated Fortnite game lobby. The AI responds with a location. Ashray then proceeds to showcase how a streamer can use the AI agent to poll the stream chat for an answer, which it will then scrape and provide as an audio answer. Another way streamers can use the AI agent as a sidekick is by having the agent ask the streamer questions when the streamer is quiet. The purpose of that feature is to help the streamer be entertaining/engaging when people are watching the stream.

NVIDIA announced it is accelerating its trillion-dollar IT infrastructure transition to enterprise AI factories with NVIDIA RTX PRO Servers at Computex 2025, and a new NVIDIA Enterprise AI Factory validation design for building data centers that provide universal acceleration for AI, design, engineering, and business applications.

Inside, the new RTX PRO Servers are powered by NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs, RTX PRO Servers extend the leading performance and energy efficiency of the NVIDIA Blackwell architecture to data centers that can run virtually every enterprise workload -- driving the shift from CPU-based system to efficient GPU-accelerated infrastructure.

NVIDIA Enterprise AI Factory validated design, partners are now building a new class of on-premises infrastructure, featuring RTX PRO Servers, NVIDIA Spectrum-X Ethernet networking, NVIDIA BlueField DPUs, NVIDIA-Certified Storage systems, and NVIDIA AI Enterprise, all to "accelerate product design and engineering simulation applications, as well as a quickly growing catalog of AI-enabled business systems and teams of digital AI agents".

NVIDIA's beasty new GB300 NVL72 AI server racks are on show at Computex 2025, where an early walk around the halls of Nangang seeing me instantly attracted me... nerd vibes were at 11/10. Check it out:

ASRock had the new NVIDIA GB300 NVL72 AI server rack at their booth, with the fully liquid-cooled, rack-scale system combining 72 x GB300 "Blackwell Ultra" AI GPUs with an upgraded 288GB of HBM3E memory, 1.4kW of power, and 50% more performance than GB200 NVL72 AI servers.

NVIDIA is expected to have its new GB300 NVL72 AI servers enter mass production in Q3 2025, according to recent supply chain sources. Apple reportedly placed an order for around $1 billion worth of new NVIDIA GB300 AI chips, with each GB300 NVL72 AI server costing between $3.7 and $4 million a pop... sadly, it wouldn't fit in my backpack so I left it on the trade show floor.

AMD has said that TSMC's new 2nm process node (N2) is the superior to any other semiconductor foundry, highlighting that it has the first product on TSMC N2 NanoSheet technology with its upcoming EPYC "Venice" CPUs.

IMG

TSMC's new bleeding-edge 2nm process node is far out and ahead of any other semiconductor foundry, with Samsung lagging far behind, and Intel coming up fast with billions of dollars of ASML's new High-NA EUV lithography machines purchased and being built and installed right now.

NVIDIA founder and CEO Jensen Huang is here on the ground in Taipei for Computex 2025, where he met with the biggest business partners in Taiwan ahead of the show where he's provided some updates on its new China-exclusive AI GPUs.

Huang said that NVIDIA's next AI GPU after its H20 chip won't be Hopper, saying: "it's not Hopper because it's not possible to modify Hopper anymore". We've previously heard rumors that the new H30 (and now possibly B30) was dropping HBM (High Bandwidth Memory) for consumer-class GDDR memory (probably GDDR7).

Jensen provided the update during a livestream posted by Taiwan's Formosa TV News network, so it looks like NVIDIA will be jumping from the Hopper AI GPU architecture to Blackwell for China. We've seen NVIDIA's new RTX PRO 6000 Blackwell workstation GPUs packing a huge 96GB of GDDR7 memory, so the rumors of its new China-destined AI GPUs shifting to GDDR memory is now making more and more sense.

InWin, a PC and workstation chassis pioneer, is partnering with Accordance System, a leader in integrated system solutions, to showcase the 'Most Powerful Edge Solution' at Computex 2025. The system features InWin's new cubePRO edge computing chassis, alongside Accordance's ARAID M500 and ARAID 5000 disk array solutions. Combine the two and get incredible performance, reliability, and data security.

This marks a strategic partnership between the two companies, focusing on delivering a one-stop AI and AIoT hardware and ecosystem solution for customers. The new cubePRO edge computing chassis includes four full-height PCIe expansion slots that can support the latest high-end GPUs, support for 'high-wattage' power supplies, and Micro-ATX motherboards.

With its perforated design and efficient fan configuration, the cubePRO is designed for AI workloads, whether working with large language models (LLMs), inference, or lightweight training, while ensuring optimal thermal performance. The cubePRO also includes three front 5.25-inch bays for modular expansion, which is where Accordance's ARAID solutions enter the picture.

NVIDIA CEO Jensen Huang was in Saudi Arabia today, announcing it would be supplying AI semiconductors to Saudi Arabia AI company Humain, for its new 500-megawatt AI data centers.

The announcement was made on-stage at the Saudi-US Investment Forum in Riyadh, where President Trump, SpaceX and Tesla CEO Elon Musk, OpenAI CEO Sam Altman, BlackRock CEO Stephen Schwartzman, and many other high-profile US players attended. NVIDIA CEO Jensen Huang announced the news, adding that AI needs a lot of power, with Saudi Arabia being an energy-rich country that will be able to use NVIDIA AI GPUs to unlock new capabilities in the field.

Humain is owned by the Saudi kingdom's Public Investment Fund, which was unveiled on Monday, with the task of developing data centers, AI infrastructure, and Arabic large language models (LLMs). Humain CEO Tareq Amin said his company will build 1.9-gigawatts of AI data centers by 2030.

Samsung Electronics has its Samsung Foundry semiconductor division conducting process evaluations with NVIDIA and Qualcomm to secure orders for its next-gen 2nm process node.

We had rumors from early January 2025 that said NVIDIA and Qualcomm were considering Samsung Foundry and its new 2nm process, but in a new report from Korean media outlet Chosun, we're hearing Samsung Foundry is expected to enter the final stage of 2nm process performance evaluation for NVIDIA GPUs and Qualcomm mobile processors "soon".

Chosen spoke with a semiconductor industry insider, who said: "with geopolitical risks in the Taiwan Strait intensifying by the day, global big tech companies cannot rely solely on TSMC for production. While yields were lower than expected initially, recent evaluations have produced positive results". Samsung Foundry has been struggling for years now, with TSMC absolutely dominating the semiconductor space and now Intel with billions of dollars worth of ASML's bleeding-edge High-NA EUV lithography machines, Samsung needs all the clients (and big ones like NVIDIA and Qualcomm) as it can get.

NVIDIA's new BG200 NVL72 AI server cabinet rack output increased throughout April 2025, with 1500 units shipped in April alone, compared to 1000 units shipped during the entire Q1 2025 period from January to March 2025.

New reports suggest that industry-wide NVIDIA GB200 NVL72 server racks reached 1500 units in April versus 1000 units in Q1 2025, with Quanta shipping 300-400 GB200 racks in April, Wistron shipping 150-150 GB200 racks, and Hon Hai shipping 1000 GB200 racks, all in April alone.

This report focuses on the Taiwanese industry, not counting what Dell, HPP, and Supermicro are all pumping out in terms of NVIDIA GB200 NVL72 AI server racks, which would increase these numbers again. Insider @Jukanlosreve reported the numbers, stating that they're from a US brokerage firm, so there is no link to this report other than what is on the X post.

NVIDIA is reportedly preparing an even more downgraded H20 AI GPU for China in the next two months, following further tightening US export restrictions on the original H20.

The news is coming from the usual "people familiar with the matter" reports Reuters, with NVIDIA notifying major Chinese retailers -- including large cloud computing providers -- with a new modified H20 AI GPU in July, according to "two of the sources".

We have had recent rumors that NVIDIA was preparing a next-generation H30 AI GPU that would drop from using HBM to GDDR memory, but it seems NVIDIA is tweaking the GPU again. The original H20 chip was the most powerful AI chip that NVIDIA could legally sell in China, but the US government informed the company last month that H20 would now require an export license, effectively cutting NVIDIA off from the Chinese market.