Latest news with #compute
Yahoo
23-07-2025
- Business
- Yahoo
OpenAI's skyrocketing spending could see billions of dollars in silicon headed down the AI mines in the next few years, including 2 million Nvidia chips headed to Texas Stargate facility
When you buy through links on our articles, Future and its syndication partners may earn a commission. OpenAI consumes compute capacity like few have ever done before it. A recent report expects it to gorge itself on datacentre capacity and research between 2025 and 2030—burning cash at a rate of swimming pools per minute by some estimations. The Information reports that OpenAI is chasing fresh investment to allow it to expand its compute capability—buying new graphics cards, accelerators, and processors to jumpstart new AI models. The company is said to be spending around $13 billion on Microsoft-owned datacentres this year, which could rise to around $28 billion in 2028. But the love affair with Microsoft is not set to last. OpenAI says one key investor, Japan's SoftBank, could be providing $30 billion of a hopeful $40 billion it hopes to raise in coming months, with a large amount of that cash headed toward Stargate. Stargate isn't as cool as it sounds and has nothing to do with space-age Egyptian folk. It's a plan between OpenAI, Softbank, and Oracle to build out AI infrastructure in the US. As much as $500 billion worth over four years. The first site for development is in Abilene, Texas. It's called Stargate 1, and the first Nvidia GB200 racks are being installed and already running 'early' workloads at the facility. Just today, OpenAI and Oracle inked a deal to develop over 5 gigawatts of capacity at the site, which is nearly five-fold its initial expected capacity and will incorporate… 2 million chips. OpenAI isn't footing the bill for that joint venture, which has attracted investment from the company's partners, though still needs to raise more cash. All told, The Information projects OpenAI will end up burning through as much as $20 billion in cash flow in 2027, up from $2 billion in 2024. Its fees for researching and developing new models could raise up to as much as $40 billion starting in 2028. Overall, The Information projects the company will spend something like $320 billion between 2025 and 2030. Further to all of this wild internal spending, OpenAI also has the risk of further unplanned spending in the case of, well, court cases. AI companies are under constant and historical scrutiny for their use of copyright materials in training data. The UK government has waved away many complaints by artists and passed a bill that would allow some degree of usage for copyrighted materials in training AI. Not cool. Though OpenAI is facing a slew of cases in the US by authors, and The New York Times has sued the company for use of its articles in training data, too. Similar cases are ongoing against other AI providers, such as Anthropic, and Meta has already won an early case fighting over similar grounds Whether there'll be hell to pay, that's up to the judge in each case—that's just one judge for the copyright cases put forward by US authors, as the cases are now being consolidated. OpenAI has stated in response to the authors' cases that it believes its "models are trained on publicly available data, grounded in fair use, and supportive of innovation." Though admittedly these cases and any repercussions are unlikely to matter to OpenAI's bottom line either way. It's projected to earn up to $12.7 billion this year, according to The Information, and it's already roughly around the $10 billion mark, reports Reuters. That isn''t anywhere near its expenses but, hey, it's not entirely footing the bill itself. You'd think there'd be some cash spare to pay some of those rights holders too, but alas…
Yahoo
22-07-2025
- Business
- Yahoo
Sam Altman says OpenAI could own 100 million GPUs by the end of the year, estimated at $3 trillion worth of silicon — ChatGPT maker to cross 'well over 1 million' AI GPUs by end of year
When you buy through links on our articles, Future and its syndication partners may earn a commission. Credit: Getty / Bloomberg OpenAI CEO Sam Altman isn't exactly known for thinking small, but his latest comments push the boundaries of even his usual brand of audacious tech talk. In a new post on X, Altman revealed that OpenAI is on track to bring 'well over 1 million GPUs online' by the end of this year. That alone is an astonishing number—consider that Elon Musk's xAI, which made waves earlier this year with its Grok 4 model, runs on about 200,000 Nvidia H100 GPUs. OpenAI will have five times that power, and it's still not enough for Altman going into the future. 'Very proud of the team...' he wrote, 'but now they better get to work figuring out how to 100x that lol.' The 'lol' might make it sound like he's joking, but Altman's track record suggests otherwise. Back in February, he admitted that OpenAI had to slow the rollout of GPT‑4.5 because they were literally 'out of GPUs.' That wasn't just a minor hiccup; it was a wake-up call considering Nvidia is also sold out till next year for its premier AI hardware. Altman has since made compute scaling a top priority, pursuing partnerships and infrastructure projects that look more like national-scale operations than corporate IT upgrades. When OpenAI hits its 1 million GPU milestone later this year, it won't just be a social media flex—it'll be cementing itself as the single largest consumer of AI compute on the planet. Anyhow, let's talk about that 100x goal, because it's exactly as wild as it sounds. At current market prices, 100 million GPUs would cost around $3 trillion—almost the GDP of the UK—and that's before factoring in the power requirements or the data centers needed to house them. There's no way Nvidia could even produce that many chips in the near term, let alone handle the energy requirements to power them all. Yet, that's the kind of moonshot thinking that drives Altman. It's less about a literal target and more about laying down the foundation for AGI (Artificial General Intelligence), whether that means custom silicon, exotic new architectures, or something we haven't even seen yet. OpenAI clearly wants to find out. The biggest living proof of this is OpenAI's Texas data center, now the world's largest single facility, which consumes around 300 MW—enough to power a mid-sized city—and is set to hit 1 gigawatt by mid-2026. Such massive and unpredictable energy demands are already drawing scrutiny from Texas grid operators, who warn that stabilizing voltage and frequency for a site of this scale requires costly, rapid infrastructure upgrades that even state utilities struggle to match. Regardless, innovation must prevail, and the bubble shouldn't burst.
Bloomberg
14-07-2025
- Business
- Bloomberg
Meta to Spend Hundreds of Billions on AI, Says Zuckerberg
Bloomberg's Kurt Wagner discusses Meta's plans to invest in compute power to build out AI superintelligence. He joins Jackie Davalos on 'Bloomberg Tech.' (Source: Bloomberg)
Gizmodo
14-07-2025
- Business
- Gizmodo
Meta Is Building Massive Gigawatt Data Centers as It Pours Billions Into AI
Meta's AI spending spree continues. After reportedly offering tens (or in some cases, hundreds) of millions of dollars to lure away top researchers from competing artificial intelligence operations, CEO Mark Zuckerberg announced via Threads that the company plans to build several multi-gigawatt superclusters to ramp up its compute power. Zuckerberg said the first of the superclusters, called Prometheus, will come online sometime in 2026, with 'multiple more titan clusters' to follow. According to Zuck, 'Just one of these covers a significant part of the footprint of Manhattan.' According to a report from SemiAnalysis, an AI research firm, Prometheus is being built in Ohio. Another one of its clusters, reportedly named Hyperion, is currently being built in Louisiana and is expected to go online in 2027. A supercluster is an interconnected network of compute resources, including graphics processing units (GPUs) or tensor processing units (TPUs), designed to do the kind of large-scale data processing needed to train AI models. Most active clusters built to support AI efforts have hundreds of megawatts of electrical power capacity. Meta's gigawatt approach would theoretically give it a leg up in compute capabilities, and Zuckerberg claims that Meta is on pace to be the first to bring a supercluster with gigawatt capacity online. Elon Musk might dispute that last achievement. Earlier this year, he claimed that xAI, his AI project behind Grok, was working on its next-generation data center and that it would be the 'first gigawatt AI training supercluster.' So I guess we've got ourselves an old-fashioned showdown to see who actually gets their data center online first. Musk's project in Memphis, Tennessee, and Zuck's in Ohio have something in common other than the goal of achieving gigawatt capacity for a supercluster data center: they both have some major environmental concerns. As Musk has ramped up construction in Memphis, the company reportedly brought in 35 portable methane gas turbines without air permits to power the project. Those turbines, capable of providing power to a neighborhood of 50,000 homes, could also emit up to 130 tons of harmful nitrogen oxides per year, per NPR. Meta's project in Ohio doesn't appear to be much cleaner. According to SemiAnalysis, Meta is building two separate 200MW on-site natural gas plants to help meet the energy demands of its data center. While natural gas plants are cleaner than alternatives like coal (which the Trump administration has given AI companies the go-ahead to use if they want), they still produce a considerable amount of pollutants, including nitrogen oxides linked to heightened cancer risks for exposed communities. Earlier this year, Zuckerberg pledged to drop up to $72 billion on AI projects to compete with the OpenAIs and xAIs of the world. It might be nice if he spent a fraction of that on making sure those projects aren't significantly harming the quality of life of people who have the misfortune of living near his data centers.
Forbes
11-07-2025
- Business
- Forbes
HPE Discover Showcases Networking And Hybrid Cloud For Enterprise AI
At HPE Discover 2025, CEO Antonio Neri made the company's case for helping enterprises embrace AI ... More with networking and compute solutions built for hybrid environments. It turned out that the HPE Discover 2025 keynote — held for the second year running at The Sphere in Las Vegas — took place only a couple of days before the news broke that the U.S. government would allow HPE's acquisition of Juniper Networks to go ahead. I attended the conference in person, plus had the chance to visit CEO Antonio Neri at an HPE headquarters outside of Houston. I talked with him in depth about the conclusion of the Juniper deal and how he sees HPE helping enterprises embrace AI, bringing networking as well as compute to data-intensive hybrid settings — where HPE was early to market. HPE faces stiff competition from Dell, Lenovo, Cisco and others as all of these vendors race to build out AI capacity for their customers. Each of these companies has its own areas of emphasis; I believe that HPE has positioned itself strongly to compete, not only with its $14 billion purchase of Juniper, but in particular with its acquisition of Morpheus Data nearly a year ago and its launch just now of GreenLake Intelligence, backed up by its new ProLiant 12 servers, HPE OpsRamp and AI-specific offerings from HPE Aruba Networking. Neri and the rest of HPE's leadership have put a lot of thought into a full-stack offering to enable AI, and I expect it to make a mark in the industry. (Note: HPE is an advisory client of my firm, Moor Insights & Strategy.) Why Enterprise AI Needs So Much Networking I was a little surprised when Neri opened his Discover keynote by talking about networking. I would have put that topic a little lower in the batting order, but maybe he knew something the audience didn't about the imminent Juniper news. I also recognize that Aruba is a juggernaut in edge networking, but was still surprised. Networking is so important today in enterprise technology because bigger and more complex AI models require more expansive network capabilities to run — and to be fully exploited by the companies that use them. That applies to everyone, but in HPE's case there's the important added dimension that the company wants to excel in both AI-for-networking, where it is already very active, and in networking-for-AI, through which it wants to boost customers' ability to use AI on their data. It's no accident that Neri talked early in his keynote about the common problem faced by businesses of all sizes making the shift to AI: legacy IT architecture is hard to modernize. And of course it's no accident that HPE wants to layer its (now significantly expanded) networking capabilities on top of its expertise in hybrid cloud to capitalize on the AI boom, for its customers and for itself. My colleague Will Townsend, who's an expert on networking and security, also attended Discover and joined me for the interview with Neri at the HPE headquarters in Texas. You can watch an edited version of that interview here. For more on the nuts and bolts of HPE's announcements at Discover in Will's specialty areas, I encourage you to read his analysis published last week on Forbes. What The Juniper Deal Does For HPE Juniper brings 10,000-plus new employees into HPE, and during our interview, Neri told us that in the wake of the acquisition the combined company now has nearly 7,000 engineers just for networking. I like to think that Bill Hewlett and David Packard, both of them engineers down to their bones, would smile at incorporating an engineering-heavy culture like Juniper's into HPE. Aside from the strong cultural fit, Juniper rounds out the connectivity functions HPE needs for a complete enterprise offering. In the AI world, networking is required for both scale-up and scale-out applications in the datacenter, plus connecting the datacenter to the outside world, including the AI-infused edge. As Neri put it, when you stand up a new datacenter, one of the first things you have to do is bring in the pipe that connects it to the outside world. Juniper has a great routing business to address exactly that, not to mention its Mist networking platform. The Juniper portfolio — and client base — complements what HPE was already doing, for instance with Aruba's edge networking capabilities (and related security functions). So now HPE has an even broader range of cloud-native, AI-driven connectivity solutions to address campus, branch and datacenter functions. It's clear that Neri regards the Juniper acquisition as a strategic way to add value to HPE's offerings, not some consolidation play. Juniper's engineering chops extend to custom silicon and really smart software, and I feel sure there will be chances to fold that expertise into HPE's product lines over time. In short, I believe this acquisition is going to improve HPE's ability to deliver efficient, cost-effective, scalable connectivity both into the datacenter and within the datacenter, which will only get more important as AI (especially agentic) continues to grow. Time will tell what this means in the context of Cisco, particularly for core switches. Enabling Enterprise AI With Everything As-A-Service Neri sees the network as the first step for enterprise AI — 'how you connect this vast pool of data that lives everywhere.' And he emphasizes that AI workloads are hybrid and data-intensive in their essence, which further plays to HPE's strengths. The company also has an existing delivery mechanism to make it easy for customers to capture these strengths; for years now, the GreenLake platform has given HPE the ability to offer everything as-a-service, including private or hybrid cloud infrastructure. At Discover, Neri announced the next iteration on this — GreenLake Intelligence. This is essentially an integrated way to deliver 'autonomous IT' that spans observability, GreenOps (eco-friendly operations), provisioning, resiliency, FinOps, storage, support, security, networking and orchestration. AI agents within GreenLake Intelligence act across the entire hybrid cloud estate to help with optimization, management and so on — which addresses the emerging need for IT organizations to reevaluate how they deploy workloads in this era of AI. For me, GreenLake Intelligence was the biggest announcement of the show, because it puts every element of hybrid cloud operations into one stack. GreenLake Intelligence is just one of the offerings that show how HPE has been building toward truly integrated, full-stack hybrid operations for years. The HPE CloudOps software suite, for example, allows your IT team to observe, provision and manage everything from one place — which, among other benefits, should lower your operating costs. CloudOps builds in part on the very clever purchase of Morpheus with its hybrid-cloud management capabilities, which I think is the best acquisition HPE has ever made. (That's really saying something, considering the past decade of contributions from Aruba and the promise embodied in the Juniper deal.) All of this is about enabling the hybrid application stack and helping it connect to whatever else is in an enterprise's unique IT architecture — now including connectors to the public cloud, which is something I criticized HPE for lacking in the past. When we interviewed him, Neri told us that HPE aims to enable tailoring the hybrid environment for each customer — in terms of both experience and cost — then help them scale the whole thing, including silicon IP, compute, storage, software, security and services. I have to give him and his team credit for taking the steps to make that vision a reality. Implementing A Vision For Hybrid Cloud Kudos to HPE for being the first big vendor — years ago — to buy in so strongly on hybrid cloud. These days, hybrid cloud is the company's bigger differentiator, and it's clear that HPE continues to build in that direction. From my perspective, it has made one smart move after another to enable enterprise AI, for example with the networking both in the datacenter and at the edge that allows customers to put the compute near the data where it belongs. Another example that I didn't have room for here — but which my colleague Matt Kimball has analyzed in depth — is the new line of ProLiant 12 servers, full of innovations to make datacenter computing more secure and performant for AI workloads. HPE is so dedicated to engineering that the often-overused term 'innovation' really does apply. That said, of course I also have some questions. HPE Discover 2025 was full of discussion about implementing AI, but last year's version of the event made a very big splash about the company's AI partnership with Nvidia. Yet I'm still uncertain whether HPE has signed up many customers for that private AI platform. The company says it has more than 250 AI use cases in the works, with more than 50 already in production. Could we have some examples with real live companies? I would also love to see HPE make the value proposition for GreenLake Intelligence more tangible. Please don't get me wrong: I think the value prop is strong. But what I said last year about HPE's homegrown software applies to GreenLake Intelligence, too: trot out customers that are using it, delighted with it and optimizing TCO because of it. We don't need all the details, just a few metrics that show progress in pipeline, sales, usage. So that's another thing I'm super interested in learning more about in the future. This year marks the 10th anniversary since HPE became an independent company. It has established itself as a powerhouse in hybrid cloud and networking — one that should only get stronger with the Juniper acquisition. As I look ahead to the company's next 10 years, I love the aggressive approach to the market and the focus on customers. I'm looking forward to seeing how those customers ultimately integrate HPE capabilities across compute, storage, networking and infrastructure software. The market positioning and the technology are compelling. At this point, execution is everything.
