Artificial Intelligence - Page 14

The AI industry is reportedly preparing to spend "trillions of dollars" securing AI hardware, with Elon Musk's xAI planning to acquire the compute power equivalent to 50 million NVIDIA H100 AI GPUs.

In a new post on X, Musk said: "230k GPUs, including 30k GB200s, are operational for training Grok @xAI in a single supercluster called Colossus 1 (inference is done by our cloud providers). At Colossus 2, the first batch of 550k GB200s & GB300s, also for training, start going online in a few weeks. As Jensen Huang has stated, @xAI is unmatched in speed. It's not even close".

xAI's massive AI supercomputer cluster is called the Colossus 2, and it'll be online in the coming weeks, powered by NVIDIA GB200 and GB300 AI servers, with a total count of 550,000 units. Just on these numbers alone, that means xAI has spent around $2 trillion getting Colossus online, which is a colossal (pun intended) amount of money.

NVIDIA AI hardware worth over $1 billion has been smuggled into China, even the huge GB200 AI servers that weigh up to two tons, in the middle of a chip ban.

NVIDIA GB200 AI servers and AI GPUs are always available on Chinese black markets, with a new report from the Financial Times claiming that over $1 billion of AI hardware has ended up on China's AI black markets since the US government imposed strict export controls, including NVIDIA GB200 AI servers.

The Financial Times has had eyes-on with multiple sales contracts and filings, revealing that China's AI black markets are most interested in the NVIDIA GB200 AI servers, and that they're available in local markets. The US government banned the H20 AI GPU and quickly after the H20 banned, distributors were still getting their hands on them through multiple means: trade loopholes or grey channels that haven't been fixed by the US government.

Elon Musk has announced xAI will be resurrecting the popular short-form video app Vine, but there will be a big catch - it will be in AI form.

Musk took to his personal X account to announce the news, with the Tesla and SpaceX CEO writing, "We're bringing back Vine, but in AI form." Currently, it remains unclear what Musk exactly means by "AI form" as no further context was provided.

For those who don't know, Vine was an extremely popular short-form video-sharing app that played a pivotal role in internet culture. Users would upload 6-second looping videos that were easily shareable. The platform gave rise to many celebrities who are prominent on social media platforms today, such as Logan Paul, Shawn Mendes, David Dobrik, and others.

The Velvet Sundown is a band that is gaining intense popularity with the release of the song "Dust On the Wind," which is a track from the band's debut album, Floating on Echoes, which was published on Spotify on June 5.

Floating on Echoes has since exploded in popularity, with Dust On the Wind being streamed more than two million times on the platform, and the other songs on the album being streamed hundreds of thousands of times.

Listeners were surprised to see that Velvet Sundown, a band that includes singer Gabe Farrow, guitarist Lennie West, keyboardist Milo Rains, and drummer Orion "Rio" Del Mar, was able to release a second album on June 20, called Dust And Silence. An unusually fast turnaround time for even an extremely adept band, let alone an up-and-comer.

The rise of artificial intelligence-powered chatbots such as ChatGPT, while technologically impressive and undoubtedly useful in various situations, is also creating problems that appear to be mounting.

It wasn't too long ago that a ChatGPT user proposed to the AI-powered software after they confessed their love to it. ChatGPT accepted, and the partner of the individual who proposed was shocked at the relationship dynamic between the individual and ChatGPT.

Now, ChatGPT has confessed to contributing to episodes of mania developed in a 30-year-old man on the autism spectrum that had no previous diagnoses of mental illness. Jacob Irwin was hospitalized twice in May for manic episodes after he asked ChatGPT to find flaws in his theory on faster-than-light travel.

One of the key parts in Samsung returning to competitiveness in the HBM market lies with its new 1c DRAM, which will be an integral part of its next-gen HBM4 memory ready to fight SK hynix and Micron in the AI battle of 2026 with HBM4.

In a new report from TheElec picked up by insider @Jukanrosleve on X, we're hearing that Samsung has delayed its 12-stack HBM4 memory based on its new 1c DRAM until 2026. Samsung had originally aimed for mass production in the second half of this year, but it is being more cautious with its HBM4 rollout, setting a new target for Production Readiness Approval (PRA) in Q4 2025.

Samsung redesigned its 1c DRAM process, with performance improvements and yields that have reportedly reached 65%, with a source familiar with the matter adding: "the company has internally set the goal for HBM4 12-stack PRA in Q4. This means mass production is now targeting next year, not this year. Based on small-scale sample tests conducted at the R&D fab, the 1c DRAM yield recorded 65%, which is a hopeful situation.

ChatGPT usage has jumped in a big way over the course of 2025, and according to a new report, the AI is now dealing with over 2.5 billion queries every day.

As you might expect, those are mostly from free users, and this is data from Axios, which contends that 330 million daily prompts are from people in the US.

To put these figures into perspective, as TechRadar (which flagged the report) observes, at the end of 2024, Sam Altman told us that ChatGPT was crunching through around a billion prompts on a daily basis.

OpenAI CEO Sam Altman took to X to confirm that the AI firm will have over 1 million GPUs online by the end of the year, which is an impressive statistic to visualize. However, as you try to picture what 1 million cutting-edge GPUs looks like, Sam Altman added that he'd much rather see 100 million GPUs go online.

After saying that he's "very proud of the team" for reaching the 1 million GPU milestone, he joked that they "better get to work figuring out how to 100x that." To put this figure into perspective, xAI's headline-grabbing Grok 4 model is powered by around 200,000 NVIDIA H100 GPUs, which suggests OpenAI is working with five times the GPU power as xAI.

The 100 million GPU figure is not currently feasible, as earlier this year, Sam Altman announced that OpenAI was delaying the release of its GPT 4.5 model because it was "out of GPUs." Which is a good problem to have if you're NVIDIA, as companies like OpenAI, xAI, Microsoft, and others are buying up GPUs as quickly as they can be produced.

At Computex 2024, AMD and Stability AI announced a partnership that introduced the world's first block FP16 stable diffusion model for image generation, offering improved quality and accuracy. This week, AMD is announcing another world's first, the FP16 SD 3.0 Medium model optimized for the AMD Ryzen AI 300 Series XDNA 2 NPUs.

Promising a "significant increase in image quality" for AI image generation using Stable Diffusion, those with compatible Ryzen AI-powered laptops can download Amuse 3.1 by Tensorstack to test it out today. As long as you've got 24GB of memory, Amuse 3.1's HQ mode will deliver 16-bit quality without the need for a dedicated GPU.

The results are impressive, with the ability to create Stable Diffusion 3.0 Medium model-powered images at a 4MP resolution (2048x2048) with XDNA Super Resolution. Best of all, it's free and doesn't require a subscription.

The International Mathematical Olympiad (IMO) has been recognized as the premier mathematics competition for some of the world's brightest young minds since 1959. With participants from a wide range of countries, each competitor is tasked with solving six "exceptionally difficult" problems spanning fields such as algebra, combinatorics, geometry, and number theory.

Only 8% of participants earn a gold medal, and now we can add Google's Gemini Deep Think AI to the list. This advanced version of Gemini solved five out of the six problems "perfectly" according to Google, which was enough for it to achieve a gold-medal performance. If you're wondering what the math problems were and the solutions, head here (PDF). But fair warning, it's a lot more advanced than 12 + 57.

"We can confirm that Google DeepMind has reached the much-desired milestone, earning 35 out of a possible 42 points - a gold medal score," IMO President, Prof. Dr. Gregor Dolinar said, adding that it was not only the solutions that were impressive, but how well laid out and easy to follow they were. "Their solutions were astonishing in many respects. IMO graders found them to be clear, precise, and most of them easy to follow."

Samsung Electronics is reportedly focusing on substance over speed when it comes to its bleeding-edge 2nm process node, with the South Korean company "preparing for a 3-4 year long-term competition with TSMC".

In a new report from DigiTimes picked up by insider @Jukanrosleve on X, we're hearing that the launch of Samsung's new 2nm process node is expected to launch in the second half of this year, behind TSMC in yield, but "steadily identifying areas for improvement".

US chip giants Apple, AMD, and NVIDIA are running to TSMC to have their next-gen 2nm chips made at the Taiwanese semiconductor fabs, but Samsung plans to leverage its improved yields and cost-effectiveness by 2026 to prepare for the long game.

Foxconn is pushing hard into its AI server business, after pumping out NVIDIA GB200 AI servers in Q2 2024, and beginning production of NVIDIA's new GB300 AI servers, it's also preparing for NVIDIA's next-gen Vera Rubin AI servers starting this month.

In a new report from UDN picked up by insider @Jukanrosleve, the Chinese manufacturer expects NVIDIA's next-gen Vera Rubin AI servers to be the main product to drive performance growth in 2026 to 2027. Foxconn boss Liu Yangwei has previously said that the company has always been a major customer's co-development partner when it comes to AI servers, and that "major customer" is NVIDIA.

Foxconn participates in the development of next-gen products with NVIDIA, with supply chain sources saying that according to Foxconn's shipping schedule, GB200 AI servers are being shipped this year, next-gen GB300 AI servers have already started small-scale production, and will become its main product for AI server shipments in the first half of 2026.

NVIDIA's upcoming B30 AI GPU will begin shipping in Q4 of this year, with performance expected to be around 10-20% slower than the H20, while being 30-40% cheaper.

In new reports from Taiwanese outlet Ctee, we're hearing that thanks to the US government allowing NVIDIA to resume exporting its H20 AI GPUs into China, the market expects to see shipments in the last 6 months of the year hitting 400,000 units, with the server cooling supply chain to benefit from this.

NVIDIA originally launched the downgraded H20 AI GPU in response to previous US export restrictions, but with those loosening on the tail end of 2024, it saw NVIDIA experience $4.5 billion in lost revenue from China earlier this year.

NVIDIA's next-gen GB300 AI servers have entered production, with the new GB300 "Blackwell Ultra" AI servers to begin shipping in September... on time, and ready to rock and roll.

In a new report from DigiTimes picked up by insider @Jukanrosleve on X, we're hearing that NVIDIA's new GB300 "Blackwell Ultra" AI servers have entered production according to supply chain sources. Industry sources add that they expect a smooth production trajectory into the second half of 2025, which is said to be from a strategic shift that's making it easier for AI server manufacturers.

NVIDIA decided to reuse the motherboard design from its current GB200 platform -- known as the Bianca board -- for its new GB300 platform. This move has significantly shortened the learning curve for suppliers, many of which were struggling to keep up with NVIDIA's incredibly fast product update cycle in the past. One ODM representative noted: "there are no major issues with the GB300 at this stage. Shipments should proceed smoothly in the second half".

Meta is pushing into the AI supercomputer space in a big way, with Mark Zuckerberg saying that the social media giant plans to add over 5GW of AI compute power in the years ahead.

Zuckerberg explained on his Threads post: "For our superintelligence effort, I'm focused on building the most elite and talent-dense team in the industry. We're also going to invest hundreds of billions of dollars into compute to build superintelligence. We have the capital from our business to do this".

The Meta CEO continued: "We're actually building several multi-GW clusters. We're calling the first one Prometheus and it's coming online in '26. We're also building Hyperion, which will be able to scale up to 5GW over several years. We're building multiple more titan clusters as well. Just one of these covers a significant part of the footprint of Manhattan. Meta Superintelligence Labs will have industry-leading levels of compute and by far the greatest compute per researcher. I'm looking forward to working with the top researchers to advance the frontier!".

NVIDIA CEO Jensen Huang is in China right now, with news that the company is preparing to launch its new RTX 6000D AI GPU with the card expected to ship 2 million units in 2025.

NVIDIA has confirmed its new RTX 6000D will launch in Q3 2025, manufactured on TSMC's 4nm process node, and a shipment target of around 2 million units before the end of 2025, filling a revenue gap of over $10 billion according to a new report from DigiTimes picked up by insider @Jukanrosleve on X.

The new RTX 6000D and the Blackwell AI GPU series have driven 4nm production capacity at TSMC to "unprecedented levels" which has significantly contributed to its revenue. The US government banned NVIDIA's Hopper H20 AI GPU earlier this year, causing the company to immediately recognize $5.5 billion in losses, but the H20 is now ready to ship to China again, as well as the company preparing the new RTX 6000D card for the country, too.

SK hynix started supplying its next-gen HBM4 memory in March 2025, but they are "early" versions of its HBM4, with qualification tests expected to take longer than the same tests for HBM3E.

The reason for HBM4 qualification tests taking longer is due to generational changes from HBM3E, where HBM3E memory samples were sent in a nearly complete state. SK hynix sent over HBM3E samples to its clients in August 2023, with mass production starting in March 2024, just 7 months between sample supply and mass production.

HBM4 in its early state will require modifications, with projections that it could take longer than 7 months, with increased technical difficulty with the generational changes cited as a factor that could extend the qualification testing period. The number of I/O terminals for HBM4 has doubled from HBM3E (2048 for HBM4, 1024 for HBM3E). Additionally, HBM4 -- for the first time in this generation -- has the logic die (base die) produced at a foundry (TSMC).

NVIDIA's new China-specific B30 AI GPU has performance of around 75% of the H20 AI GPU, while demand for the new B30 is "significant" according to the latest reports.

In a new post on X by insider @Jukanrosleve, we're hearing from China's major internet companies that estimates that the performance of NVIDIA's new B30 AI GPU is "approximately 75% that of the H20". Chinese tech companies have reportedly placed orders for hundreds of thousands of units -- orders of over $1 billion -- in late-June, with deliveries expected in August.

Another large Chinese tech company reportedly plans to increase its Q3 2025 capital expenditure and intends to order 300,000 orders of NVIDIA's new B30 AI GPU, with a delivery schedule for September.

SK hynix is reportedly changing its wafer cutting process for next-generation memory manufacturing, paving the way for its new HBM4 and 400-layer and higher NAND flash memory as they become increasingly thin, pushing existing processes to their absolute limits.

In a new story from ETnews picked up by insider @Jukanrosleve on X, industry sources have said that SK hynix plans to introduce femto-second grooving and full-cut processes for HBM4 wafer cutting. The news was confirmed by the South Korean memory manufacturer in discussing a Joint Evaluation Project (JEP) for new wafer cutting equipment with laser equipment partners.

It's reported that technology tests are already underway with some partners, with an industry official saying: "SK hynix is planning a major change to its existing wafer cutting methods and is discussing numerous technical solutions with partners".

NVIDIA's new B30 AI GPU won't be arriving until September, with Chinese customers needing to wait a couple of more months according to the latest report.

In a new report published by the Financial Times picked up by insider @Jukanrosleve on X, we're hearing that the NVIDIA B30 AI GPU not being sold before September is because NVIDIA asked for prior assurance from the Trump administration that it wouldn't be breaching the new US export regulations, and then seeing its new B30 banned after it is introduced.

NVIDIA's new B30 AI GPU specifications could change before now and September, depending on its discussions with the US government, and if the specifications change, Jukan says that it might primarily involve enabling NVLink. The new B30 AI GPU is rumored to have NVLink disabled, making it more of a modified RTX PRO 6000 workstation GPU (as it doesn't have HBM, and uses GDDR7 instead).