Altimeter Capital partner Clark Tang said in a recent post that Nvidia is moving through a major repositioning that much of the market has barely noticed. His argument is that the company now controls both the software operating layer and the financing layer that define a hyperscaler, creating what he called a “synthetic hyperscaler” that is starting to replace Amazon, Microsoft, and Google in the AI compute supply chain.
Why Tang says the old hyperscaler model is breaking down in AI
Tang first laid out what a hyperscaler is meant to be. At its core, he described it as a large infrastructure bundle spanning CPUs, networking, and storage, with a developer platform built on top.
He said the business model rests on two pillars. Financially, hyperscalers convert customer capital expenditure into operating expenditure, smoothing financial obligations for enterprise clients. Operationally, they abstract the underlying resources through software so developers can consume compute without dealing with hardware complexity.
That model has supported operating margins of 35% to 40% for AWS, Azure, and Google Cloud Platform, according to Tang, because large cloud providers can buy hardware at scale, lower their cost of capital, and push utilization higher through software.
His view is that the AI era changes the assumptions behind that system. Training large models requires large synchronized GPU clusters. Inference workloads, especially agent-style workloads, are highly sensitive to tokens per watt and time to first token. Tang said both requirements run against the cloud-era design logic of multi-tenancy and maximum virtual machine density.
In training clusters, a delay at a single node can slow the whole job. He said the virtualization and network overlay layers widely used by traditional hyperscalers have become a drag on GPU performance. Tang framed this as an innovator’s dilemma: hyperscalers do not want to give up 35% to 40% margins, and they want their own ASIC efforts to compress Nvidia’s 75% gross margin down to 25%, yet they continue to lag on execution speed and technical flexibility.
Neocloud providers found the opening, but balance sheets remain a constraint
Tang said neocloud providers saw the shift early and built an ecosystem around it. In his telling, they recognized that hyperscalers were wrapping Nvidia hardware with a thin software layer and taking rich profits, while neoclouds could compete at gross margins of 20% or lower.
He also said neoclouds worked closely with Nvidia to build software specifically for AI workloads. That stack includes hot standby, predictive maintenance, and storage software optimized for training with lower pricing.
According to Tang, AI labs have shown a clear preference for working with neoclouds, and their complaints about hyperscalers usually fall into three areas:
- They move too slowly.
- Cluster deployment is too rigid.
- The virtualization layer reduces GPU performance.
Still, Tang said neoclouds face a structural weakness that is hard to solve: thin balance sheets. Because hyperscalers operate at scale and carry stronger credit profiles, they can raise money and build capacity before demand is fully visible, then capture the upside when that demand arrives. Lenders to neoclouds, by contrast, are generally willing to finance only hardware backed by signed contracts, and early AI labs themselves have tended to carry speculative credit profiles, which raises funding costs.
That, in his view, is why neoclouds have not been able to make pre-demand capacity bets on the scale of Musk’s Colossus.
Nvidia’s strategy: replicate both the operating layer and the financing layer
Tang’s core claim is that Nvidia has quietly spent the past few years reproducing the two capabilities that made hyperscalers powerful, and that this happened through a deliberate sequence rather than by accident.
On the operating side, he pointed to DSX OS and Mission Control for managing GPU fleets, DSX reference designs and Omniverse digital twins for standardized data center construction, and the Dynamo framework for inference services. Taken together, he said, these tools amount to a full operating system for AI data centers, one that can be used by any team with the technical expertise and physical site capacity to run it.
On the financing side, Tang said Nvidia first standardized its hardware assets, then used reference designs to show infrastructure investors that compute assets were interchangeable and transferable across customers. That, he argued, allowed the assets to pass financial due diligence at unlevered yields of 7% to 8%.
He said Nvidia then brought in $500 billion in third-party capital. The partners he named were Apollo, BlackRock, Blackstone, Goldman Sachs, and KKR, and he said the company has built six separate financing platforms.
Tang also provided a timeline to support the idea that this was a long-planned strategy. Nvidia signed a master agreement with CoreWeave in 2023, formed an AI infrastructure partnership with BlackRock in 2024, helped drive a $100 billion Brookfield fund in 2025, and advanced work with KKR Helix in 2026. He added that Jensen Huang spent much of 2025 traveling across Europe, the Middle East, and Southeast Asia to push the ground-level buildout of these new compute sites.
Tang responds to questions about capital circularity and asset finance
Tang also addressed what he described as the main objections to this framework.
On concerns about circular capital flows, he said the six institutions, including Apollo and BlackRock, underwrite independently rather than through a linked structure. In his view, that means the money is genuine outside capital rather than an extension of Nvidia’s own balance sheet.
On asset life and financeability, he cited CoreWeave as an example. He wrote that contracts for A100 systems shipped six years ago have been extended to 2029, and that CoreWeave raised pricing across its fleet by 25% in July 2025. To Tang, that suggests GPU depreciation looks more like aircraft than smartphones.
His conclusion was that CoreWeave has already shown the earnings potential of a neocloud model supported by Nvidia-backed financing, while the balance-sheet advantage long held by hyperscalers is weakening. He pointed to Google raising $50 billion after reporting its first quarter of negative free cash flow, and to Microsoft carrying $329 billion in leases not yet commenced on its books, as signs that the old financial moat is starting to erode.
Tang’s bottom line was that Nvidia’s “synthetic hyperscaler” may already exist, even if much of the market has yet to recognize it.