推理时代下的AI、GPU云和新云市场趋势报告

Sponsor ed by: Key Findings: •AI datacenter spending is enormous and shows no signs of slowing.It’s likely thatmore than $1 trillion has been committed to AI datacenter buildouts. •Neoclouds—clouds purpose-built for AI—are grabbing headlines in the datacenterindustry.Most prominent in this group is CoreWeave, which went public in 2025 •NVIDIA is providing fuel by funding neoclouds, which are also NVIDIA customers.It’s legal, and it’s in NVIDIA’s best interest, but the circular nature of the NVIDIA AIeconomy is raising uncomfortable bubble questions. •Alternative clouds have a promising niche in enterprise AI.These clouds, whichtout lower prices than the likes of AWS, provide both GPU clusters and enterprisecloud computing services. Vultr in particular is finding traction with this •The need for power is driving AI cloud expansion.The heated competition to buildmore AI capacity is primarily a race to lock down power guarantees. •This new market is driving the evolution of a new infrastructure stack with needsin networking, storage, and security.Each layer of the infrastructure has new •Debt is starting to outweigh equity in neocloud funding.The scale is tipped bycompanies like CoreWeave. The debt usually goes toward buying GPUs, but forthose neoclouds that are building datacenters themselves, the debt also finances •Hyperscalers are implementing diversification strategies to offset risks.Themarket abhors a monopoly, which in this case includes not only NVIDIA GPUs butalso the rest of the NVIDIA software and hardware stack. Chip and networking •Companies mentioned in this report:Accton, AMD, Arista, AWS, Cerebras, Cisco,CoreWeave, Crusoe, DDN, Dell, DigitalOcean, DriveNets, Eclypsium, Fluidstack,Fortanix, Google, Groq, Hedgehog, HPE, Juniper, Lambda Labs, Nebius, Nscale,NVIDIA, OpenAI, Runpod, Shakti Cloud, Supermicro, TensorWave, Together AI, VAST Smart Networking forSmarter AI CentersSmart Networking forSmarter AI Centers arista.comarista.com800 GbE for AI Clusters800 GbE for AI Clusters Evolve on yourown terms. Not a hyperscaler. Not a neocloud. Vultr is the only AI-first cloud that combines GPUpower with security, compliance, and worldwide scale. A successful AI strategy starts with a foundationbuilt to scale. Where neoclouds stop at raw compute,Vultr delivers the full stack – globally distributed, From scaling LLMs to running real-time inference,Vultr gives you everything you need to Evolve. Elevate. Excel.Your cloud, your way. Table of Contents 1.Introduction: A Seismic Shift for Datacenters About Our Methodology6 2. What’s in a Neocloud? The Rise of the GPU Cloud The Neoclouds that Really Matter8Frontier AI Models Pay the Bills, but Small Jobs Are Welcome Too9It Really Is All About Power10Massive Funding, Massive Debt10The Circle of Life (Is OpenAI Too Big to Fail?)11 3. Hyperscaler and GPU Cloud Strategies 13 Hyperscalers and Other GPU Clouds13Alt Clouds Are In on GPUs, Too14 4. GPU Cloud Infrastructure: Vendor Landscape 15 GPUs: NVIDIA Rules, but Customers Want Alternatives15Networking: Applying Lessons from Hyperscalers16What About Security?18Storage Needs to Be More than Just “Storage”19 5. Conclusion: Potential Risks and Rewards of the GPU Cloud Market 22 The Depreciation Question22Have the Hyperscalers Outsourced Risk?23 Appendix: Selected AI and GPU Cloud Ecosystem Players to Watch Proactive Security for AIHardware and Infrastructure Supply Chain Security Trustedby Top AI Data Centers 1.Introduction: A Seismic Shift for Datacenters The AI boom has generated big dreams in the tech and finance worlds, and that's most blatantlyon display when it comes to clouds and datacenters. What we have on our hands is a worldwidearms race to build out AI clouds of all stripes, an endeavor which requires lots of resources—land, This is the domain of GPU clouds as well as neoclouds, which are GPU-laden clouds dedicated to AItraining and inference. They’re not the only clouds deploying GPUs, but they've grabbed thespotlight due to their audacious scale. The splashiest neocloud projects are campuses costing And yet, it’s still not enough, according to AI company executives. Neocloud activity intensified in2025, with bigger datacenter plans, enormous funding rounds, and cavernous debt financing. Aswe noted in a recent Cloud Tracker Pro report, in total more than $1 trillion has been pledged for Some of this money will be internal investment (Apple, xAI), but most of it, including likely all $500billion of the Stargate figure, is fueling the AI cloud boom, and the pace is accelerating. Nine-figuredeals seem to get announced nearly every day, meaning by the time this report goes to market, AI Clouds, GPU Clouds, and Neoclouds| The surge has also propelled neocloud-related Initial Public Offerings (IPOs) by CoreWeave andWhiteFiber, providing public capital to pick up the pace of buildouts. CoreWeave projects it willspend $20 billion to $23 billion in capex this year, up roughl