{"slug": "can-neoclouds-corner-ai-compute", "title": "Can neoclouds corner AI compute?", "summary": "Neocloud revenue exceeded $25 billion in 2025, and Gartner predicts neoclouds will capture 20% of the $267 billion AI cloud market by 2030, according to Synergy Research Group and Gartner. These purpose-built GPU platforms like CoreWeave, Lambda, Nebius, RunPod, and Vultr offer NVIDIA-native infrastructure with InfiniBand networking and lower costs, positioning them to take business from hyperscalers Amazon Web Services, Google Cloud Platform, and Microsoft Azure. However, McKinsey & Company reports that only 10 to 15 of over 100 neoclouds operate at meaningful scale in the U.S., and enterprise adoption remains limited.", "body_md": "Neoclouds have staked their claim to AI compute infrastructure with low-cost GPU-based alternatives and purpose-built features for AI training and [inference](https://www.infoworld.com/article/4117620/edge-ai-the-future-of-ai-inference-is-smarter-local-compute.html). But can this new class of [alternative clouds](https://www.infoworld.com/article/4040239/what-alternative-clouds-are-good-for.html) actually take business from the hyperscalers?\n\nSome argue that neoclouds *are* poised to capture a growing share of AI-native workloads. Neocloud revenue exceeded $25 billion in 2025, reports [Synergy Research Group](https://www.srgresearch.com/articles/neocloud-market-forecast-to-approach-400b-by-2031-driven-by-surging-ai-infrastructure-demand). And [Gartner predicts](https://www.gartner.com/en/newsroom/press-releases/2026-06-23-gartner-predicts-neocloud-providers-will-capture-20-percent-of-the-267-billion-dollar-ai-cloud-market-by-2030) neoclouds could take 20% of the $267 billion AI cloud market by the year 2030.\n\n“Neoclouds offer purpose-built, NVIDIA-native infrastructure optimized for AI, with InfiniBand networking, bare-metal access, and lower costs,” says [Hardeep Singh](https://www.linkedin.com/in/hardeepsingh31/), senior principal analyst at [Gartner](https://www.gartner.com). “They focus on high-performance GPU clusters and flexible contracts, unlike hyperscalers’ general-purpose, legacy-heavy platforms.”\n\nThere are over 100 neoclouds in the market today, yet only 10 or 15 are operating at a “meaningful scale” in the United States, reports global consultancy [McKinsey & Company](https://www.mckinsey.com/capabilities/tech-and-ai/our-insights/the-evolution-of-neoclouds-and-their-next-moves#/). But while the neocloud category is growing, there is little evidence that a large percentage of enterprises have standardized upon neoclouds to date.\n\nNeoclouds are a [new type of public cloud](https://www.networkworld.com/article/4011187/neoclouds-roll-in-challenge-hyperscalers-for-ai-workloads.html) that has emerged as an alternative to hyperscalers Amazon Web Services, Google Cloud Platform, and Microsoft Azure. Neoclouds include usage-based [GPU](https://www.networkworld.com/article/3966130/what-are-gpus-inside-the-processing-power-behind-ai.html) computing platforms like [CoreWeave](https://coreweave.com/), [Lambda](https://lambda.ai/), [Nebius](https://nebius.com/), [RunPod](https://www.runpod.io/), and [Vultr](https://www.vultr.com/), among others. Given their specialization in AI computing, neoclouds have demonstrated a growing utility within modern AI projects.\n\n“Neoclouds have become key components of the AI infrastructure ecosystem,” says [Nii Osae](https://www.linkedin.com/in/starkosae/), CEO and founder of [Mindbeam AI](https://mindbeam.ai/), an AI services company. “They provide vertically integrated GPUs along with other complementary hardware, including high-bandwidth storage and networking, to AI research labs and enterprises to enable reliable training and inference of AI models,” he adds.\n\nNeoclouds are designed with a more modern computing paradigm in mind. As [Rohan Gupta](https://www.linkedin.com/in/rohanhariomgupta/), vice president of cloud, security, and devops at [R Systems](https://www.rsystems.com/), a digital product engineering and technology services company, notes, “Hyperscalers were designed for the internet of 2010: millions of small, multi-tenant virtual machines (VMs) over oversubscribed Ethernet. Neoclouds were designed for a workload that didn’t exist back then: a single tenant, thousands of GPUs, all talking to each other at the same time.”\n\nNeoclouds could support the unique requirements of today’s burgeoning [agentic systems](https://www.infoworld.com/article/4154570/best-practices-for-building-agentic-systems.html). As [Kevin Cochrane](https://www.linkedin.com/in/kevinvcochrane), CMO at [Vultr](https://www.vultr.com/), says, neoclouds can “extend traditional cloud native compute with all the specialized AI services and infrastructure needed to embed agentic experiences into existing application portfolios.”\n\nA primary selling point for neoclouds is lower costs. “Neoclouds often run at roughly a third of the cost of hyperscalers for the same GPU capacity,” says [Lauri Kien Kotcher](https://www.linkedin.com/in/lauri-kien-kotcher-836306a/), CEO and co-founder of [Different Day](https://differentday.ai/), an AI application development company. “They also provision in days, not the months hyperscalers quote for high-density AI infrastructure.”\n\nAccording to [ComputerWeekly](https://www.computerweekly.com/opinion/Gartner-Why-neoclouds-are-the-future-of-GPU-as-a-Service), neoclouds’ GPU-based instances can provide cost-savings of 60% to 70% compared to traditional hyperscaler GPU instances. These savings are made possible due to a closeness to chip manufacturers, making neoclouds alluring during [GPU shortages](https://cloudwars.com/cloud-wars-minute/microsoft-anthropic-and-nvidia-forge-ai-super-alliance-poised-to-shape-the-next-era-of-innovation/).\n\n“Many neoclouds have cultivated strong [relationships with NVIDIA](https://www.nvidia.com/en-us/data-center/gpu-cloud-computing/partners/), and the economics favor them on raw GPU-hour pricing,” adds [Michael Byrne](https://www.linkedin.com/in/michael-byrne-9b0b051/), vice president of data center solution architecture at [Presidio](https://www.presidio.com/), a digital services and infrastructure solutions provider.\n\nBeyond cost-savings, neoclouds are differentiating with AI-optimized compute and networking. “Bare-metal access eliminates hypervisor overhead, delivering the full performance of the underlying GPU,” says [John Q. Martin](https://www.red-gate.com/blog/author/johnqmartin/), technology partner and alliances manager at [Redgate Software](https://www.red-gate.com/), a database devops software company. Their optimized infrastructure for distributed training and high-bandwidth interconnections helps at scale, he adds.\n\nIf we look at storage, neoclouds also offer high-bandwidth [data retrieval](https://www.infoworld.com/article/4060211/do-vector-native-databases-beat-add-ons-for-ai-applications.html) options, enabling them to feed massive GPU clusters more efficiently. As R Systems’ Gupta explains, “Neoclouds standardized early on parallel filesystem architectures from vendors like [VAST](https://www.vastdata.com/), [WEKA](https://www.weka.io/), and [DDN](https://www.ddn.com/) because large-scale AI training is fundamentally a parallel-filesystem problem, not an object-storage one.”\n\nThe result? Improved *goodput*, not just throughput, says [David Johnson](https://www.linkedin.com/in/djohnsonjr/), director of product marketing at [Backblaze](https://www.backblaze.com/), a cloud storage provider. “Goodput is the fraction of AI infrastructure capacity that converts into useful training or inference work after failures, restarts, and I/O stalls,” Johnson says. Backblaze recently announced a five-year [$335 million deal with CoreWeave](https://www.backblaze.com/blog/coreweave-announcement-2026/).\n\nStill, there are plenty of other layers to consider. The neoclouds with more holistic offerings will likely win out in the long-term.\n\n“They need high-speed storage, low-latency networking, secure access, and a way to govern and meter usage across teams,” says [Haseeb Budhani](https://www.linkedin.com/in/budhani/), CEO and co-founder of [Rafay Systems](https://rafay.co/), provider of a cloud and AI infrastructure management platform. “This is especially critical for inference, where the bulk of enterprise AI spending will land.”\n\nGiven their lower costs, direct GPU access, and faster provisioning, experts say neoclouds make sense for tasks like large-scale AI/ML training, generative AI platforms, and high-performance compute. Even considering the sorts of companies that run those workloads, it may surprise you *who* the buyer ends up being.\n\n“The most striking thing about the neocloud customer list isn’t who’s on it — it’s who’s at the top of it,” says Gupta. “The largest neocloud customers right now are the hyperscalers themselves.”\n\nMicrosoft [reportedly accounted for 67%](https://www.forbes.com/sites/phoebeliu/2026/04/29/coreweave-depends-on-openai-what-if-the-ai-chatgpt-maker-cant-pay-up/) of CoreWeave’s 2025 revenue, and OpenAI remains one of CoreWeave’s top clients as well. As CoreWeave disclosed in its [2025 annual SEC filing](https://www.sec.gov/Archives/edgar/data/1769628/000176962826000104/crwv-20251231.htm), “We recognized an aggregate of approximately 77% of our revenue from our top two customers for the year ended December 31, 2024.”\n\n“The hyperscalers appear to have recognized some of the risks of the interconnected dependencies leading back to NVIDIA, OpenAI, and some of the major neoclouds such as CoreWeave,” IT consultancy Futuriom writes in its report, [AI, GPU Clouds, and Neoclouds in the Age of Inference](https://www.futuriom.com/signup?report=Futuriom-AI-Clouds-GPU-Clouds-1.5-final.pdf). “By investing and outsourcing some of their GPU clouds to other players, the hyperscalers are reducing the risks to their own balance sheet.”\n\nCoreWeave’s [lingering debts](https://tech-insider.org/coreweave-30-billion-capex-ai-cloud-2026/) and high customer concentration have raised questions about its long-term sustainability. And for some, the fact that hyperscalers are both a user and competitor complicates the longevity of neoclouds.\n\n“Building a competitive AI-native software stack requires enormous capital, engineering talent, and time,” says Martin. “And doing it while your biggest customers are the same hyperscalers you’re now competing with is a genuine strategic tension.”\n\nDespite the allure of more cost-effective GPU infrastructure, not everyone is ready to jump to a neocloud. “We went with a hyperscaler for our AI back end,” says Redgate Software’s Martin. “At our current stage, the managed services, ecosystem integrations, and ability to scale incrementally made more sense than committing to dedicated GPU infrastructure.”\n\nRedgate’s [2026 State of the Database Landscape](https://www.red-gate.com/solutions/state-of-database-landscape/2026/) found that 43% of respondents operate hybrid database estates. And as Martin explains, the operational fragmentation incurred from using a neocloud for AI compute while keeping data on a hyperscaler (or multiple clouds) is tough to stomach. “It introduces real complexity, data egress costs, fragmented security and identity management, and split operational overhead,” Martin says.\n\nWhile neoclouds may provide raw bare-metal GPU performance, you sacrifice some managed service conveniences in return. “You get bare-metal GPU access but lose the guardrails around auto-scaling, observability, and enterprise SLAs that hyperscalers provide out of the box,” adds Martin.\n\nDifferent Day opted to stick with the hyperscalers too, citing continuity with existing tech. “Owning your cloud account means you inherit the security certifications, the identity controls, the audit logs, and the procurement contracts that come with a major provider,” says Different Day’s Kotcher. “Neoclouds are not there yet on most of that.”\n\nOthers agree. “The biggest trade-off is enterprise maturity,” Rafay Systems’ Budhani says. “Hyperscalers spent years building governance, security, identity, auditing, billing, and operational controls that enterprises expect.”\n\nContinuity with the surrounding architecture often matters more than GPU performance. “Hyperscalers retain significant advantages in breadth and integration,” adds [Scott Sanders](https://www.linkedin.com/in/ssanders/), corporate vice president of engineering at [Sonar](https://www.sonarsource.com/), a code quality and security company. “For organizations whose AI workloads sit inside a larger enterprise, hyperscalers offer continuity that a specialized provider can’t easily substitute.”\n\n[Resilience](https://www.cio.com/article/3615770/how-resilient-cios-future-proof-to-mitigate-risks.html) is another area to consider. “While performance and cost for AI may improve, risks include portability challenges, operational overhead, and reliance on niche providers,” says [Nigel Gibbons](https://www.linkedin.com/in/vunrg/), director and senior advisor at [NCC Group](https://www.nccgroup.com/), a cybersecurity consulting firm. “This means enterprises must balance optimization against resilience, support depth, compliance coverage, and long-term vendor viability, particularly for mission-critical or regulated workloads.”\n\nLastly, while neoclouds offer competitive prices now, they are highly reliant on the fluctuating GPU hardware market, placing their top unique differentiator on shaky ground. “Unlike railroads or even fiber optics — capital intensive booms of the past that spurred massive investment — GPUs don’t retain their value for decades,” Futuriom notes. “In fact, they depreciate fast — sometimes as fast as four to five years.”\n\nTo sum up, the neoclouds have both clear advantages and drawbacks compared to their hyperscale competitors. They also face other challenges.\n\nA10’s [The State of AI Infrastructure Report 2025](https://www.a10networks.com/wp-content/uploads/A10-EB-The-State-of-AI-Infrastructure-Report.pdf) found that 35% of respondents primarily host their AI workloads in public-cloud environments, compared with 42% that favor a balanced hybrid approach. This means neoclouds are competing not only with hyperscalers, but also with on-premises and hybrid infrastructure strategies.\n\nWhat’s more, emerging hyperscaler-bred custom silicon could also upset future neocloud advances. “AWS’s Trainium, Google’s Ironwood, and Microsoft’s Maia give hyperscalers cost and scale advantages that neoclouds structurally can’t match because they are, by design, NVIDIA distributors,” says Gupta.\n\nStill, hype around neoclouds continues to follow excitement in the greater AI services market. [Writing for Forbes](https://www.forbes.com/sites/rscottraynovich/2026/07/09/behind-the-ai-gold-rush-at-raise-summit-paris/), Futuriom’s founder and chief analyst [Scott Raynovich](https://www.linkedin.com/in/scott-raynovich-9a40784/) describes a path to maturity that could mitigate some of the aforementioned trade-offs: “These companies will evolve out of being more than GPU rental services to provide global infrastructure, applications services, and data privacy and security.”\n\nIn the end, will neoclouds succeed in differentiating themselves from hyperscalers by providing AI-native software stacks? “Neoclouds win a real market,” says Kotcher. “They do not win the broader one. The honest answer is that this is a bounded victory, not a takeover.”\n\n“The leading neoclouds aren’t trying to replace AWS or Azure across the portfolio — they’re building a complementary layer for the AI workloads where hyperscaler economics break down,” says Backblaze’s Johnson.\n\nTherefore, the most likely outcome is coexistence. “Neoclouds keep frontier training and NVIDIA-tied inference,” Gupta predicts, “while hyperscalers keep the data plane, regulated workloads, custom-silicon inference, and enterprise commercial relationship.”", "url": "https://wpnews.pro/news/can-neoclouds-corner-ai-compute", "canonical_source": "https://www.infoworld.com/article/4211105/can-neoclouds-corner-ai-compute.html", "published_at": "2026-08-24 09:00:00+00:00", "updated_at": "2026-08-24 09:12:57.416541+00:00", "lang": "en", "topics": ["ai-infrastructure", "artificial-intelligence"], "entities": ["Synergy Research Group", "Gartner", "CoreWeave", "Lambda", "Nebius", "RunPod", "Vultr", "McKinsey & Company"], "alternates": {"html": "https://wpnews.pro/news/can-neoclouds-corner-ai-compute", "markdown": "https://wpnews.pro/news/can-neoclouds-corner-ai-compute.md", "text": "https://wpnews.pro/news/can-neoclouds-corner-ai-compute.txt", "jsonld": "https://wpnews.pro/news/can-neoclouds-corner-ai-compute.jsonld"}}