Capacity for
GPU clouds.

What is a neocloud? Cloud computing built for AI workloads

A neocloud is a cloud provider built specifically around GPU compute for AI workloads and high performance computing. Where traditional cloud providers such as AWS, Microsoft Azure and Google Cloud offer hundreds of services, neocloud providers concentrate on GPU clusters, high performance networking and the software to run model training, fine tuning and inference workloads. Customers buy GPU as a service: reserved clusters, on demand access or bare metal access to NVIDIA GPUs. Model builders, AI labs and enterprises use that dedicated AI infrastructure to train and serve AI models without buying hardware themselves, and bare metal tenancy avoids the noisy neighbor problem of shared public cloud instances.

That focus makes physical infrastructure the constraint. A neocloud company can order graphics processing units faster than it can find data centers with the power capacity, power density and liquid cooling those GPUs require. GPU cloud data center capacity, the powered space where the hardware runs, is what decides how fast it can scale AI services for its customers.

Hyperscalers usually build their own campuses at very large scale. Neoclouds mix approaches, and the trade off between speed, cost and control drives each decision: leasing from former bitcoin miners, colocation facilities, build-to-suit projects and, increasingly, sites they own.

How neoclouds source AI data center capacity and colocation facilities today

Public announcements show several patterns. None of the companies below is a QIT customer; they are examples of how the neocloud market buys capacity.

Leasing from bitcoin miners

Miners already hold energized sites, substations and power contracts. Several have converted them into AI data centers under long leases.

• Nscale and Ionic Digital. Nscale announced a Microsoft contract for about 200,000 GB300 GPUs, with about 104,000 of them in a roughly 240 MW Texas campus leased from Ionic Digital, a bitcoin miner, scaling toward 1.2 GW.
• Fluidstack and TeraWulf. Fluidstack signed 10-year HPC leases for more than 200 MW at TeraWulf’s Lake Mariner site in New York, with Google backstopping $1.8 billion of Fluidstack’s lease obligations. The companies later formed a joint venture at Abernathy, Texas, for 168 MW of critical IT with a $1.3 billion Google backstop.
• Fluidstack and Cipher. Cipher signed a 10-year Fluidstack lease for 168 MW of critical IT at Barber Lake in Colorado City, Texas, with a $1.4 billion Google backstop, and later expanded it to cover the full 300 MW site.
• CoreWeave and Core Scientific. CoreWeave leased about 590 MW of high-density colocation capacity across six Core Scientific sites, including a $1.2 billion expansion at Denton, Texas. CoreWeave agreed to acquire Core Scientific in July 2025, but Core Scientific’s stockholders rejected the deal and it was terminated on October 30, 2025.

Colocation and conversions

• Lambda uses colocation and build-to-suit space, including more than 30 MW with EdgeConneX in Chicago and Atlanta using hybrid liquid-to-chip and air cooling, and a partnership with Prime Data Centers in Vernon, California. It has also announced the conversion of an unoccupied, 2009-built facility in Kansas City into an AI factory for more than 10,000 NVIDIA GPUs.
• Nebius describes a mix of its own data center in Finland, colocation in several countries and a build-to-suit greenfield project in New Jersey of up to 300 MW.

Owned and modular campuses

• Firmus describes Project Southgate in Tasmania as a campus of modular, 100% liquid-cooled AI factories, with a 90 MW first stage. In September 2026 the company reported more than 900 MW of contracted capacity.
• Nscale lists a mix of owned and partner-run sites, and says its Ward County, Texas, campus went from an empty steel shell to a live AI data center in about nine months.
• CoreWeave reported 51 active data centers, more than 1.5 GW of active power and about 4.2 GW of contracted power in October 2026.

Sources: Nscale, October 15, 2025; Nscale, October 1, 2026; Nscale, April 14, 2026; TeraWulf, August 14, 2025; TeraWulf, October 28, 2025; Cipher, Barber Lake lease; Cipher, Barber Lake expansion; Core Scientific, February 26, 2025; Core Scientific Q2 2026 10-Q; CoreWeave, October 27, 2026; Lambda and EdgeConneX; Lambda and Prime Data Centers; Lambda, Kansas City; Nebius, March 5, 2025; Firmus, Project Southgate; Firmus, September 8, 2026. Company statements; reviewed November 2, 2026.

Why GPU cloud data center capacity is the constraint

GPU availability is only half of the problem. The other half is available capacity in data centers ready for high-density racks. A GPU sitting in a warehouse earns nothing; GPU capacity only becomes revenue once it is installed, cooled, networked and sold. For a neocloud company, the time between a hardware delivery and a customer going live is a direct cost.

The deals above share a pattern: large, long-term commitments, often hundreds of megawatts over ten years or more, sometimes with a hyperscaler or AI lab standing behind the lease. They suit anchor tenants and big training run commitments. They are less suited to what many neocloud providers also need:

• Bridge capacity while a large campus is under construction.
• Smaller blocks for a new customer, a regional inference deployment or a sovereign cloud infrastructure requirement, without long term commitments sized for a gigawatt campus.
• High-density racks for the latest NVIDIA platforms, which require liquid cooling that many existing colocation facilities were not designed for.
• Locations near customers to reduce latency for inference workloads and AI applications with latency requirements.

Building traditional data centers for these needs takes time and capital investment that a fast-growing cloud platform may prefer to spend on GPU hardware. Older air-cooled halls in major markets may have space but not the power density for current systems, which limits AI performance or forces lower rack counts.

Where modular 1 MW blocks fit in AI infrastructure

QIT develops liquid-cooled modular GPU data centers planned around roughly 1 MW of IT load per block, phased into 5–10 MW sites, with a focus on Texas and a preference for existing powered properties. For a neocloud, that format is useful in three places.

1. Bridge capacity

A few megawatts on an existing powered site can host customer clusters while a larger campus is built. QIT containers install on a prepared site in 15 hours; utility work and permits, not construction, set the critical path.

2. Phasing with demand

Adding capacity one block at a time lets GPU capacity follow signed customer demand rather than a speculative build. Each phase can be matched to a specific GPU order and customer contract.

3. Liquid-cooled readiness

Current rack-scale NVIDIA systems are liquid cooled. NVIDIA’s reference architecture for GB300 NVL72 describes a full rack requiring up to 142 kW. A liquid-cooled GPU container designs coolant distribution into the module from the start instead of retrofitting an air-cooled hall, so the AI compute can run at its rated power rather than throttled below peak performance by heat.

The comparison is general, not a quote. Actual block size, rack count and density depend on the GPU platform and the site.

Source: NVIDIA Enterprise Reference Architecture, GB300 NVL72. Reviewed November 2, 2026.

What QIT needs from a GPU cloud capacity buyer

A GPU colocation or capacity discussion moves faster with these inputs:

• Capacity and timing: how many MW of IT load, by when, and how the ramp should be phased.
• Hardware selection: GPU platform and generation, rack configuration, power density per rack and network design.
• Workload: model training, fine tuning or inference, and the uptime and latency each needs.
• Location: preferred markets or regions, latency targets and any data residency or sovereign cloud infrastructure constraints. QIT’s current focus is Texas.
• Connectivity: carrier preferences, bandwidth to your other regions and how the site joins your backbone or SD-WAN.
• Commercial model: who owns the GPUs and modules, contract term, and which operational teams handle hardware, network and facility operations.
• Acceptance: the tests and service levels that define a block as ready for customers.

The total cost of capacity includes more than the module: power, cooling, network and operations all belong in the comparison. See the modular data center cost guide.

Common questions about neocloud data centers

What is the difference between hyperscalers and neoclouds?

Hyperscalers run broad public cloud platforms with many services and usually build their own campuses. Neoclouds focus on GPU infrastructure for AI and often lease, colocate or partner for data center capacity.

What is a neocloud data center?

It is a data center, or part of one, that hosts a neocloud’s GPU clusters. It needs far more power per rack than general cloud infrastructure, usually liquid cooling, and high performance networking between racks. It may be owned by the neocloud, leased from a developer or provided through GPU colocation.

How do neoclouds make money?

Mostly by renting GPU capacity: reserved contracts for training and inference, on demand GPU hours and managed services such as inference endpoints or managed databases for machine learning pipelines. Utilization of installed GPUs drives the economics, which is why fast, predictable capacity matters.

Is QIT a GPU cloud?

QIT’s focus here is physical infrastructure: liquid-cooled modular data center blocks and the sites to put them on. Neocloud providers are potential customers, not competitors.

Can modular blocks support large-scale training?

Large training clusters need high performance networking between racks and enough contiguous capacity. Multiple blocks on one site can host a cluster if the network and power are designed for it from the first phase; very large training runs are usually better suited to large campuses.

Talk to QIT about capacity

If you run a GPU cloud and need liquid-cooled capacity in Texas, send QIT your target MW, timing, GPU platform and preferred locations. We will discuss a phased modular plan around a specific site.

Related resources: GPU shipping containers, behind-the-meter data centers, bitcoin miners moving to AI and HPC, Texas site selection and the deployment plan. Then contact QIT.