Latest news with #H100s
Time of India
2 days ago
- Business
- Time of India
What Nvidia CEO Jensen Huang does differently with AI (and why you should too)
Jensen Huang doesn't use AI like a magic box. He treats it like a debate partner. When he wants to understand something new, he tells the chatbot to start from scratch 'Explain it like I'm 12' then build up to postgrad level. He'll ask the same question across multiple models, compare their responses, then make them critique one another. 'It's no different than getting three doctors' opinions,' he said at a Milken Institute panel. 'Then I ask them to compare notes and give me the best answer.' Explore courses from Top Institutes in Select a Course Category Technology CXO Finance Design Thinking Artificial Intelligence MBA Public Policy Degree Others Product Management Digital Marketing Data Science Project Management Data Science PGDM Management Healthcare others Data Analytics healthcare Operations Management Leadership Cybersecurity MCA Skills you'll gain: Duration: 12 Weeks MIT xPRO CERT-MIT XPRO Building AI Prod India Starts on undefined Get Details This isn't about outsourcing thinking. It's the opposite. Huang argues that good AI use demands more mental effort, not less. by Taboola by Taboola Sponsored Links Sponsored Links Promoted Links Promoted Links You May Like Indonesia: New Container Houses (Prices May Surprise You) Container House | Search ads Search Now Undo 'You have to be analytical. You have to reason. You have to formulate questions,' he said. 'When I'm interacting with AI, it's not passive. It's active interrogation.' Which is why he's dismissive of a recent MIT study suggesting generative AI weakens cognitive effort. If people are using AI in a way that makes them think less, Huang says, they're doing it wrong. Live Events Huawei's approach: Full spectrum control While Huang fine-tunes how individuals think with AI, Huawei is building a system designed to scale across sectors—and across borders. Once a telecom hardware supplier, Huawei is now positioning itself as China's national AI champion. It's not just developing models. It's creating the chips, the software frameworks, the training infrastructure, and the real-world deployments—everything Nvidia currently dominates in the West. And it's doing it under pressure. US sanctions, imposed in 2019, cut Huawei off from advanced chips and key American tech. What followed wasn't retreat—it was reinvention. The company redirected R&D into building its own semiconductors, working with blacklisted Chinese chipmaker SMIC to produce 5G-capable chips and AI processors. Today, Huawei's answer to Nvidia's H100s are its Ascend 910B and 910C chips. These are paired with its CANN software framework, meant to replace Nvidia's CUDA. Together, they power Huawei's AI data centres, now operating across China. In April, Huawei unveiled CloudMatrix 384, a high-performance cluster using 384 Ascend 910C chips. Analysts say it rivals Nvidia's GB200 NVL72 in raw throughput, and outperforms it in some areas like power efficiency and memory bandwidth. So yes, Huawei is sanctioned. But it's still shipping product and now it's shipping scale. AI that builds industry, not chatbots Huawei's models aren't designed for broad public use like ChatGPT or Gemini. Its flagship Pangu series is built for specific sectors: mining, healthcare, finance, logistics, and government. This isn't theoretical. Pangu models are already deployed across more than 20 sectors. In mining, electric trucks operate autonomously to transport coal. In healthcare, Pangu supports diagnostic tools and hospital infrastructure. In government, it helps with everything from city planning to crisis response. Jack Chen, VP of Huawei's mining business, says this isn't just national. The company plans to export its AI infrastructure globally, especially to Belt and Road countries across Central Asia, Africa, Southeast Asia, and Latin America. To grease adoption, Huawei has open-sourced parts of its Pangu model stack, a strategic move to win over partners where Western firms are either absent or constrained. Nvidia's Warning Shot Jensen Huang sees what's happening. He's called Huawei 'one of the most formidable technology companies in the world.' He's also warned that if export controls tighten further, Huawei won't just compete—it'll replace Nvidia inside China. This matters. Nvidia currently dominates global AI training infrastructure. Its chips and CUDA platform are the backbone for most large language models and research labs. But Nvidia can't sell into China's biggest AI buyers—not without US government sign-off. And as restrictions expand, its grip on the Chinese market weakens. Huawei is stepping into that vacuum, not with imitation, but with a parallel system built from top to bottom. Two AI worlds, two philosophies So what does all this add up to? Two dominant players, building two separate versions of AI infrastructure. But it's more than that. What's really unfolding is a divergence in how AI is used, built, and governed. Nvidia's model is decentralised, open-ended, developer-driven. It gives tools to people, researchers, startups—whoever can build with them. The logic is: if you make AI accessible and flexible, progress will follow. Huawei's model is centralised, vertical, and tightly integrated with the state. Its AI is built for strategic resilience, industrial application, and geopolitical alignment. The logic is: control the stack, scale the stack, and use it to drive national advantage. One sees AI as a platform. The other sees it as infrastructure. One bets on users. The other bets on systems. And both are scaling fast. This isn't just a commercial rivalry. It's a slow bifurcation of global AI—two tech stacks, two ecosystems, two sets of rules. It mirrors what's already happened in areas like 5G, semiconductors, and cloud infrastructure. But AI cuts deeper. It's not just about how fast models train or how efficient chips are. It's about how knowledge gets processed, who has access to it, and what decisions it drives. That's why this matters. Not because of which company sells more hardware next quarter—but because the future of intelligence itself is being shaped by a power struggle most people barely see. And whether you're using AI to answer questions or to run cities, that split is coming for you too.
Hans India
16-07-2025
- Business
- Hans India
Zuckerberg Says AI Talent Wants More GPUs, Less Bureaucracy — Not Just Big Money
As Meta continues its bold push into artificial intelligence, CEO Mark Zuckerberg has offered insights into what truly drives top AI researchers to join his company — and it's not just the massive paychecks making headlines. In a recent conversation on The Information's TITV, Zuckerberg addressed reports claiming that Meta has offered salaries as high as ₹1,600 crore ($200 million) to AI specialists. While he acknowledged that compensation plays a role, he emphasized that elite researchers are more interested in autonomy and cutting-edge tools than in managing large teams or collecting oversized paychecks. "Historically, when I was recruiting people to different parts of the company, people are like, 'Okay, what's my scope going to be?'" Zuckerberg shared. "Here, people say, 'I want the fewest number of people reporting to me and the most GPUs.'" GPUs — or graphical processing units — are essential for training advanced AI models. Meta, like other major tech firms, relies heavily on Nvidia's top-tier chips, such as the H100s, to power its large language models and AGI initiatives. According to Zuckerberg, ensuring that each researcher has access to maximum computing power is a major strategy in attracting world-class talent. This approach seems to be working. Meta's current AI hiring spree is among the most aggressive Silicon Valley has seen. The company has successfully drawn in top minds from rivals including OpenAI, Google DeepMind, Apple, and Anthropic. Among the most notable hires are Trapit Bansal, previously with OpenAI, and Ruoming Pang from Apple. Both are now part of Meta's ambitious new Superintelligence Lab. Led by former GitHub CEO Nat Friedman and Scale AI founder Alexandr Wang, the lab is quickly becoming a heavyweight in the race toward Artificial General Intelligence (AGI) — AI systems capable of reasoning and learning like humans. The team now includes at least 11 prominent names from Google and OpenAI, including Lucas Beyer, Xiaohua Zhai, Jack Rae, Johan Schalkwyk, Ji Lin, Shengjia Zhao, and Jiahui Yu. Backing this talent pool is an equally massive infrastructure buildout. Meta recently unveiled plans for a next-generation AI supercluster named Prometheus, expected to go live by 2026. It will be supported by sprawling data centres — including the 5-gigawatt Hyperion — making it one of the largest AI infrastructure efforts globally. In a move reminiscent of Elon Musk's 'tent factories' at Tesla, Meta has also started constructing temporary tent-based data centres. These prefabricated facilities allow the company to begin AI model training earlier, even before the full infrastructure is in place. The urgency behind these moves' stems in part from mixed reactions to Meta's Llama 4 model earlier this year. Since then, the company has revised its roadmap, committing $14 billion to Scale AI for improved training data and dramatically stepping up its AI recruitment game. With its unique mix of raw computing power, operational freedom, and a robust infrastructure plan, Meta is positioning itself as a serious contender in the global race for AGI — and not merely by throwing money at the problem.
Economic Times
11-07-2025
- Business
- Economic Times
Intel facing a BlackBerry moment as thousands laid off - CEO says too late to catch up with AI as firm falls out of global top 10
Synopsis Intel CEO Lip-Bu Tan made a shocking admission, saying it's now 'too late' for Intel to catch up in the AI race, as the company falls out of the top 10 semiconductor companies. Once a leader in chipmaking, Intel now faces massive layoffs, a $16 billion loss, and increasing reliance on TSMC for chip production. The company has lost ground to rivals like Nvidia, AMD, and Apple, especially in AI and data centers. With a renewed focus on edge AI and agentic systems, Tan promises change, but Intel's future remains uncertain. Read how the tech giant plans to reinvent itself. Reuters Intel CEO Lip-Bu Tan says it's too late to catch up with Nvidia in AI, as Intel drops out of the top 10 semiconductor companies and faces global layoffs, heavy losses, and a complete shift toward edge AI and chip outsourcing. Intel CEO admits it's 'too late' to catch up in AI race as chip giant slips out of top 10 semiconductor firms- Intel, once the undisputed leader of the semiconductor world, now finds itself at a critical crossroads. In a leaked internal discussion, Intel CEO Lip-Bu Tan made a brutally honest admission—he believes it's already 'too late' for Intel to catch up in the AI competition. The statement, shared during a global employee Q&A, reveals just how far the tech giant has fallen, even slipping out of the top 10 semiconductor companies by Tan's own words. This sharp self-assessment highlights the company's struggle to stay relevant amid fierce competition from AMD, Nvidia, Apple, TSMC, and Samsung. While Intel still holds legacy clout, that alone may not be enough to power through the rapidly evolving AI era. And with layoffs underway and massive losses stacking up, the pressure to turn things around has never been greater. Despite having the resources and infrastructure once deemed untouchable, Intel has fallen behind in the AI hardware race. Lip-Bu Tan's comment—"On training, I think it is too late for us"—makes it clear that Nvidia's runaway success in data center GPUs has created a gap that may now be impossible to close. AI development today heavily depends on powerful training hardware, most of which runs on Nvidia's CUDA-based ecosystem. While Intel tried to enter this space with its Habana Labs acquisition and Gaudi AI chips, it never gained the market traction needed to compete with Nvidia's H100s or AMD's MI300X chips. The rise of large language models like OpenAI's ChatGPT only widened the gap, further cementing Nvidia's lead. Instead of pushing further in data center AI, Intel plans to pivot towards edge AI, focusing on bringing artificial intelligence to personal devices like laptops, desktops, and embedded systems—where it still sees growth potential. In what's perhaps the most jarring part of Tan's talk, he reportedly said: 'Twenty, 30 years ago, we were really the leader. Now I think the world has changed. We are not in the top 10 semiconductor companies.' This admission shocked many across the tech industry. While Intel is still a recognized name globally, competitors like TSMC, Nvidia, Samsung, Apple, and AMD have outpaced it in terms of innovation, revenue, and market relevance. Even relatively smaller firms like Broadcom, MediaTek, Micron, and SK Hynix are making waves in specialized markets. According to recent financial data, Intel reported a $16 billion loss in Q3 of 2024, and it's struggling to reverse the trend. The company that once nearly acquired Nvidia for $20 billion is now watching from the sidelines as Nvidia crosses a staggering $4 trillion market cap. Intel's decision to exit the AI training chip space comes as the AI chip market explodes in value. Key stats: Global AI chip market was worth $23.2 billion in 2023 Projected to reach $117.5 billion by 2029 at a CAGR of ~31% Nvidia holds an estimated 90% share in AI training chips Meanwhile, AMD is quickly gaining ground with its MI300 series, and TSMC is dominating as the primary chip manufacturer for Nvidia, Apple, and AMD. While conceding the AI training space, Intel is attempting a pivot: Investing $20 billion in a new Ohio AI chip plant Backed by the CHIPS Act, receiving billions in grants Eligible for a 35% tax credit on investment in U.S. fabs This move aims to revitalize Intel's role as a domestic foundry powerhouse, producing edge and agentic AI chips rather than competing directly with Nvidia's data center dominance. Company AI Focus Area Market Position Key Advantage Nvidia AI training & inference ~90% market share Dominant CUDA software ecosystem, H100/Blackwell chips AMD Data center AI GPUs Rising rapidly Competitive MI300X chips with increasing adoption TSMC Manufacturing/foundry Backbone of AI industry Manufactures chips for Nvidia, AMD, Apple Intel Edge AI (future focus) Minor AI share Investing in U.S. fabs, but far behind in AI chips Intel's once-vibrant CPU leadership has failed to translate into GPU or AI-specific success. Analysts note that even if Intel builds capacity, it lacks the software stack, developer loyalty, and ecosystem that power Nvidia's moat. Intel now hopes to find success in: Edge AI: Chips powering autonomous devices, cars, and smart sensors Agentic AI: AI chips focused on decision-making in real-time systems Foundry services: Becoming a U.S.-based manufacturer for others, not just itself However, these bets are long-term, with profitability and success far from guaranteed. A big part of Intel's decline can be traced to delays in its own chipmaking technology. While AMD partnered with TSMC to produce cutting-edge 5nm and 3nm chips, Intel stuck with its internal foundries. Unfortunately, those fabs fell behind schedule. Intel's own hybrid architecture, similar to ARM's design, didn't take off the way it had hoped. Its Arrow Lake and Meteor Lake CPUs failed to gain significant ground on AMD's Ryzen and EPYC series, especially in high-performance computing. AMD now powers everything from handhelds like the Steam Deck and Rog Ally X, to gaming consoles like the PlayStation 5 and Xbox Series X/S. Meanwhile, Intel's attempts at entering the discrete GPU market with its Arc lineup were too little, too late. The GPUs suffered from driver issues, performance gaps, and poor market timing. By the time Intel showed up, Nvidia and AMD had already cornered the market. There's growing speculation that Intel may split into two entities—one focused on designing chips (like AMD and Apple) and the other running its foundry operations as a separate business. While nothing has been confirmed officially, the strategy could relieve some pressure and allow Intel to act more flexibly. As of 2025, Intel already outsources about 30% of its chip production to TSMC, a move that would have been unthinkable years ago. TSMC is now producing major parts of Intel's upcoming Lunar Lake and Meteor Lake chips, including the GPU and compute tiles. Intel's long-delayed 18A node—the supposed game-changer—isn't expected to be ready until late 2026. This shift to a fabless model, or something close to it, could be Intel's path to survival. Both AMD and Apple have succeeded by focusing entirely on chip design and leaving manufacturing to TSMC. Nvidia has always followed this model, too. To cut costs, Intel has been laying off thousands of employees globally. These layoffs come as part of a larger cost-cutting initiative after heavy R&D spending and failed product launches. According to OregonTech, the company is in "a fight for survival." CEO Lip-Bu Tan, who replaced former chief Pat Gelsinger in late 2024, has signaled a major cultural reset. He emphasized that Intel's comeback would be a 'marathon', not a sprint. The new approach? Fewer distractions and a laser-sharp focus on areas where Intel can still compete—namely edge AI, low-power computing, and eventually reclaiming performance leadership in CPU markets. Tan is also betting big on agentic AI, a fast-growing field where AI systems can operate independently without constant human input. He teased that more executive-level hires are coming to help accelerate the transformation, saying, 'Stay tuned. A few more people are coming on board.' The honest reality is, Intel has already missed the first AI wave. Nvidia owns the training market. AMD is now winning in data centers. TSMC continues to dominate manufacturing. Even Apple's M-series chips are setting new standards in efficiency and performance. Still, Intel isn't dead. It's wounded—yes—but not out. With the right leadership, sharper product focus, and a little humility, the company could still stage a comeback. The road ahead won't be easy, and it won't be fast. But if there's one thing we've seen from tech turnarounds, it's that big brands can rise again—if they're willing to let go of the past. Intel's survival now depends on its ability to adapt—not just to AI, but to a world where speed, specialization, and scale matter more than legacy. The next 18 months will likely determine whether the company can climb back into relevance, or fade deeper into the background. Intel CEO Lip-Bu Tan admits it's "too late" for Intel to catch up in AI training. The company reportedly no longer ranks among the top 10 global semiconductor firms. Intel posted a $16 billion loss in Q3 2024 and is laying off thousands worldwide. The company now outsources 30% of chip production to TSMC. Intel's future focus includes edge AI, agentic AI, and potentially a fabless business model. Q1: Why did Intel CEO say it's too late for AI? Intel CEO Lip-Bu Tan said Intel missed the AI training wave led by Nvidia and can't catch up now. Q2: Is Intel still a top semiconductor company? No, according to Tan, Intel is no longer among the top 10 semiconductor firms globally.
Yahoo
21-04-2025
- Business
- Yahoo
Huawei Just Crushed Nvidia -- $230 Billion Gone in a Flash
Nvidia (NASDAQ:NVDA) shares sank Monday after Reuters reported Huawei is about to launch its new 910C AI chipsdesigned to rival Nvidia's powerful H100s. That's a big problem. These chips arrive just days after the U.S. government banned Nvidia's China-specific H20 GPU line, designed to comply with trade restrictions. JPMorgan now projects Nvidia could lose up to $16 billion in FY25 alone. The hit was immediate: Nvidia stock slid around nearly 5.5% at 10.26am, extending last week's $230 billion market cap wipeout. AMD (NASDAQ:AMD), Qualcomm (NASDAQ:QCOM), and Broadcom (NASDAQ:AVGO) followed, each down roughly 2% premarket. Warning! GuruFocus has detected 3 Warning Signs with NVDA. It's not just a one-off product ban. It's a patternand a growing one. Between the AI Diffusion Rule from the Biden administration, capping overseas chip exports, and Trump's new semiconductor tariff investigation, Nvidia's facing a geopolitical pincer move. Even as CEO Jensen Huang touched down in Beijing last week, pledging $500 billion to scale the U.S. AI supply chain, the company is losing ground in its second-largest market. And despite Nvidia's best efforts to build China-compliant GPUs, the U.S. keeps tightening the screws. Now the political pressure is spilling into public markets. The House Select Committee on China is demanding answers from Nvidia, while Wall Street is asking a different question: are U.S. policies pushing China to build better, faster? Bernstein's Stacy Rasgon didn't mince words: banning the H20 chips is like handing the Chinese AI market to Huawei. And based on Monday's selloff, investors seem to agree. The real risk now? America's trying to defend its leadwhile accidentally accelerating the very race it fears. This article first appeared on GuruFocus.
Yahoo
13-03-2025
- Business
- Yahoo
When the AI hype dies, I hope Nvidia pivots back to gaming
Nvidia doesn't seem to care about gaming much anymore. From its lacklustre Blackwell RTX 50 launch, to its stock price booming as its H100s were used in all manner or AI training, to hyperbolic marketing that suggests we should just take our fake frames and be happy about it. Nvidia's clearly more of an AI company these days. But I hope that doesn't last. This is the company that brought us iconic gaming graphics cards, like the 1080 Ti and the ludicrous 4090. It popularized dynamic upscaling with DLSS, and helped make raytracing kind-of viable. But it's been very clear for a number of years that gaming is not Nvidia's major focus, and indeed this past one, makes it feel like an afterthought. I really hope if and when the AI hype dies down, or the bubble bursts, that Nvidia returns to its roots and makes some great gaming hardware again. If gaming became a bigger focus for Nvidia again, it would have take it more seriously. That would mean coming to terms with some of the negative habits it's cultivated in recent years. Instead of taking the gaming market for granted, it would need to be more aggressive with its pricing, making more of its higher-end graphics cards more affordable. That might even mean it stopped undercutting its board partners with overbuilt Founders Editions, so they can be more creative with their designs without sky-rocketing the price. Maybe Nvidia could sweet talk EVGA to come back too? How fun would it be see KINGPIN cards make a comeback? It would have to start putting more than 8GB of VRAM on its most affordable graphics cards, too. If there's one thing I wish it would do, though, it's be more honest about its products. Jensen Huang confidently stating that the RTX 5070 would offer RTX 4090 performance ('with AI') is as close to an outright lie you can get without actually doing so — it's not even faster than the 4070 Super all the time. It could stop using misleadingly labelled graphs and show real frames per second against its own cards, and the competition. It could highlight native performance, like AMD did during its recent RX 9070 XT debut. I am a staunch critic of Nvidia's predatory, monopolistic practices. Its price increases in recent generations have been laughable, and the RTX 50 launch has been downright insulting. But I'd be a liar if I said I didn't love the ridiculousness of its Titan-esque halo cards. The RTX 1080 Ti, the RTX 3090 Ti, the RTX 4090. Less so the RTX 5090, but still, it's bonkers. Look how tiny the actual PCB is. Look at how the power connectors are melting again because these graphics cards are so ridiculous. They're the fastest cards there's even been at the time of their release. The lizard part of my brain that just likes to see the numbers go up loves what the hardware can do, and the jaded tech journalist side of me adores the unique ways they pushed the boundaries of what's possible. I'd love it if graphics cards were more competitive, and I'd love to see AMD take a shot at the top spot again even if it doesn't make much financial sense. But I also want to see Nvidia do crazy Nvidia stuff. Just make a 1,000W GPU that needs its own power supply already. Why not? If Nvidia took gaming more seriously, it could make some of its best features more applicable to more people. As it stands, some of its flagship gaming features, like DLSS and raytracing, are effectively paywall locked behind the highest-end graphics cards. Even almost seven years on from the debut of these new features with the RTX 20 series, and no one I know uses raytracing outside of those with the most high-end graphics cards. And how many people do you know who have an XX90 class card? I work in this industry and I only know a couple. DLSS at the very least should be just as usable on an XX60 card as it is on an XX90. The whole point of the high-end card is that it has the local rendering power to do what the lower-end cards can't. Giving it more Tensor cores so that it can DLSS even harder than the cards that actually need the help is backwards. The same goes for raytracing. The 5090 has all the power it needs to drive high frame rates in most games. Giving it almost four times the RT cores of the 5070 just seems unfair. The 5070 is the card that needs that extra help tracing those paths. Sure, make the future 6090 faster in all senses than every other card, but couldn't we have it so that turning on raytracing on low-end cards doesn't mean making major sacrifices? Nvidia really botched the RTX 50 series, and the RTX 40-series wasn't that impressive either. But its tanking reputation doesn't have to continue to slide — if it starts treating gamers as more than an afterthought. The AI boom won't last forever, and though Nvidia's datacenter business will certainly remain a more profitable part of it than consumer hardware, the gaming market will always be an important component in its system. If Nvidia wants to see off new competition from AMD and Intel, and the ever-greater-encroachment from ARM, it needs to rethink its approach: Fairer pricing, more competitive mid-range options with democratised features, would be a great start. That doesn't mean ditching the crazy designs and over the top GPUs that push the boundaries of what's possible — I want to see more of that, not less — but that can be folded into a more cohesive strategy that understands what gamers actually want and need: Affordable GPUs that let them enjoy a more customized, performative experience than consoles and handhelds. If Nvidia wants to keep being the best, it needs to act like it still cares.
