Making Compute Underwritable
Huang expanded the argument a day later in an NVIDIA blog describing AI factory compute as an emerging investable asset class.
NVIDIA’s case begins with a definition.
The company does not describe its compute platform simply as a GPU. It includes accelerated computing, networking, systems software, AI frameworks and the CUDA software ecosystem surrounding the hardware.
That wider platform matters to the financing thesis because NVIDIA argues it increases the number of potential users for an installed AI system.
An NVIDIA DSX AI factory could support language models, vision, speech, biological computing, robotics, physical AI and other workloads. The same infrastructure could potentially move among customers, clouds or operators as demand changes.
In financial terms, NVIDIA is arguing for fungibility.
That could become particularly important to lenders and infrastructure investors trying to determine what happens if an original customer disappears, a contract expires or the economics of a particular workload change.
A GPU cluster tied economically to one speculative tenant is one thing.
Compute that can be redeployed across a large global market of clouds, enterprises, AI developers and model providers is a different risk proposition.
NVIDIA contends that this breadth of potential offtakers helps protect residual value.
Whether institutional markets ultimately price that risk the way NVIDIA hopes remains to be seen. But the company is now explicitly trying to establish a financial framework around that premise.
Challenging the Traditional Depreciation Curve
NVIDIA’s second argument is that software can extend the economic life of installed hardware.
CUDA is central to that case.
The company maintains that successive software improvements can increase the performance and efficiency of systems that have already been deployed, allowing the same hardware to produce more useful work at lower cost over time.
That does not eliminate hardware obsolescence. New GPU generations continue to arrive at a rapid cadence.
But NVIDIA is arguing that the useful economic life of an AI accelerator may be longer than a conventional IT depreciation schedule suggests.
The company points to its Ampere-generation A100 as evidence.
Introduced in 2020, A100 systems remain in commercial use six years later across AI training, fine-tuning, inference and high-performance computing. NVIDIA says customers continue to commit A100 capacity to multiyear deployments, pushing the potential economic life of some systems toward a decade.
Rental pricing provides another piece of NVIDIA’s argument.
The company said one-year H100 rental pricing increased from approximately $1.70 per GPU-hour in October 2025 to about $2.35 in March 2026.
Across providers, NVIDIA cited median on-demand pricing that rose from roughly $2 per GPU-hour in October 2025 to $2.70 by June 2026.
Newer Blackwell B200 capacity commands considerably more, with NVIDIA citing reported cloud pricing between approximately $5.30 and $7.05 per GPU-hour.
Those figures do not guarantee future residual values, utilization rates or investment returns.
They do, however, illustrate what NVIDIA wants financing institutions to evaluate: not simply the acquisition cost of GPUs, but the productive life and revenue-generating potential of the installed compute.
Independent Underwriting — With an NVIDIA Backstop
NVIDIA also directly addressed one of the obvious questions surrounding the financing initiative: whether the arrangement risks creating a circular market in which capital is effectively being supplied to customers so they can purchase more NVIDIA equipment.
The company’s answer centers on independent underwriting.
Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR would evaluate individual opportunities based on customer credit, demand, expected utilization, cash flow and residual equipment value.
NVIDIA supplies the AI factory platform. The financial institutions decide where their capital goes.
That separation will be important if the financing model is to establish credibility beyond the current AI investment cycle.
But NVIDIA also disclosed that it could provide a limited residual-value support mechanism in some transactions.
The company said that support could amount to as much as 25% of an opportunity, evaluated on a project-by-project basis.
That detail deserves particular attention.
NVIDIA portrays the mechanism as a limited backstop designed to complement independent underwriting rather than substitute for it. The company argues that it can assume some residual-value exposure because its compute is widely adopted, software-upgradable and redeployable across a broad customer ecosystem.
Even so, the arrangement underscores how early this prospective asset class remains.
A mature market for AI compute finance will ultimately depend on real utilization, durable customer demand, recoverable residual values and observable secondary markets — not merely on expectations for continued AI growth.
Goldman Sachs CEO David Solomon pointed directly toward that next step, describing an opportunity to create a market for credit backed by NVIDIA compute.
If such a market develops at scale, it could begin to move GPU infrastructure into financial territory previously occupied primarily by the buildings and utility infrastructure around it.
Financing the Factory Is Only Half the Problem
The capital story arrived alongside another NVIDIA argument: that the physical architecture of the AI factory must change if the industry expects compute performance to continue scaling.
The issue is no longer simply obtaining enough megawatts.
It is getting those megawatts efficiently from the grid to the accelerator.
Traditional data center power distribution brings AC electricity into the facility and moves it through multiple conversion stages before it reaches the compute.
At conventional rack densities, the losses and equipment associated with those conversions have been manageable.
At the power levels envisioned for next-generation accelerated computing, NVIDIA argues that they become increasingly consequential.
The company’s answer is 800 VDC distribution.
Higher-voltage DC allows power to move through fewer conversion stages between the electrical source and the GPU. NVIDIA says the result is a more efficient and scalable path for delivering very high levels of power to dense AI systems.
The architecture is becoming a core component of NVIDIA’s DSX reference designs, which increasingly function as a system-level blueprint connecting the compute rack to the electrical and mechanical infrastructure around it.
A Migration Path, Not Just a Greenfield Design
One of the more consequential elements of NVIDIA’s 800 VDC strategy is that the company is not limiting it to future greenfield AI campuses.
Most existing data centers were built around AC distribution.
NVIDIA’s near-term answer is an MGX-compatible 800 VDC power rack, scheduled to arrive in the second half of 2026.
The system is intended to sit inside existing AC facilities and convert power locally for delivery to 800 VDC compute racks within the row.
According to NVIDIA, that means operators could introduce next-generation rack-scale systems without first redesigning the electrical architecture of the entire building.
For an industry struggling simultaneously with long utility queues, scarce powered land and rapidly changing equipment requirements, that distinction could be substantial.
The hybrid model effectively attempts to preserve the value of existing land, interconnection rights and building infrastructure while creating an on-ramp to significantly higher compute density.
In other words, NVIDIA is trying to prevent the transition to higher-voltage DC from becoming a greenfield-only proposition.
From the Rack to 2 MW Per Row
The roadmap then moves outward.
For larger deployments, NVIDIA plans a row power center that centralizes power conversion for an entire rack row and distributes 800 VDC through overhead busway.
The architecture is expected in 2027 and is designed to support up to 2 megawatts per row.
That number illustrates the scale of the density transition underway.
Rack power has already moved well beyond the conventional enterprise data center range, and infrastructure designers are now preparing for hundreds of kilowatts per rack and potentially much higher configurations around future AI systems.
At those levels, row-level electrical architecture becomes a first-order design consideration rather than simply a supporting subsystem.
Beyond the row power center, NVIDIA’s roadmap extends to what it calls the DC power block.
Designed for new facilities, the architecture would convert medium-voltage grid power directly to 800 VDC at facility scale in a single step.
That represents the long-term version of the idea: not an AC data center adapted to accommodate DC-powered compute, but a facility whose electrical architecture is designed around high-density AI systems from the outset.
The progression is deliberate.
Existing AC facility.
Hybrid 800 VDC power rack.
800 VDC row distribution.
Native facility-scale DC architecture.
It gives operators different entry points depending on where they are in the development cycle.
Building an Ecosystem Around 800 VDC
NVIDIA is also working to make sure the architecture does not depend on a proprietary supply chain.
The company has been developing the 800 VDC approach with Google and Microsoft through the Open Compute Project.
The companies published a joint low-voltage DC white paper in March, followed in July by version 0.3 of an OCP solid-state transformer specification.
NVIDIA says more than 80 equipment manufacturers and infrastructure companies are now developing products around the specifications.
That ecosystem could prove as important as the electrical theory behind 800 VDC.
A facility-scale transition in power architecture requires far more than GPUs. Power conversion equipment, protection systems, busway, connectors, controls, cooling integration and other infrastructure components all have to be available at scale.
Common interfaces give multiple vendors a target around which to design products, potentially reducing the risk that operators are forced into a single-source electrical architecture.
It is another example of NVIDIA pushing farther beyond the accelerator itself.
The company increasingly needs the surrounding infrastructure industry to move in synchronization with its compute roadmap.
