Nvidia Asset-Backed "Securitization" Moves

Nvidia is working with six private equity and financial entities to create a financing mechanism for servers that essentially aims to turn hardware capex  into infrastructure. 

Nvidia signed memorandums of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR. 

The partners will assemble capital pools at rates Nvidia characterized as attractive, with intended beneficiaries spanning frontier AI labs, enterprises, and cloud providers. 

The commercial aim is to enable compute buyers to obtain capacity without showing that “capex” on balance sheets, much as an airline leases aircraft rather than buying planes.

In other words, the mechanism shifts server and compute hardware depreciation schedules to longer-lived categories similar to commercial real estate or toll roads. 

Essentially, the effort aims to effectively securitize compute, functionally if not in a textbook form. 

“Securitization” means a special purpose vehicle pools financial assets such as loans, leases, or receivables, then issues notes whose repayment comes principally from those pooled cash flows. 

There are other similar forms that accomplish the same ends, if not using precisely the same means. 

Structure	What investors finance	Is it securitization?
Equipment-secured loan	GPU servers and sometimes customer contracts	No; it is secured/private credit
Direct infrastructure or project loan	Data center, power, cooling, and network assets	No; it is project finance
SPV ownership plus lease	SPV owns GPUs/data center and leases compute capacity to an operator	Not automatically
Sale-leaseback	A sponsor sells assets to an SPV and leases them back	Not automatically
Asset-Backed Securities (ABS) and Commercial Mortgage-Backed Securities (CMBS)	Pool of GPU leases, compute receivables, data-center loans, or tenant leases	Yes, or close economic analogy

The model resembles aircraft-lease or equipment forms of asset-backed securities, where: a bankruptcy-remote vehicle owns equipment and receives contractual lease or service payments. 

Equipment ABS also have been used to finance shipping, and rail assets as well. 

That, in turn, will help customers access scarce compute at scale by moving such compute capabilities off the balance sheet, in principle alleviating investor concern about the timing of AI capital expenditure and near-term financial returns.

How Big a Problem is Buyer Concentration in the AI Value Chain?

If OpenAI and Anthropic drive 48 percent of Google Cloud revenue in 2027, is that a problem?

• Some might say “yes,” to the degree that customer concentration is generally viewed as a problem. 

• Others might say “maybe,” if customer concentration or solvency danger is not ameliorated.

• But the history of computing might suggest “no” is a possible answer, as there are parallels in mainframes, supercomputers, military computing, and even early semiconductors.

The closest historical analogy is probably the mainframe/supercomputer market of the 1960s to 1980s. 

The current AI infrastructure boom therefore looks less like the PC industry and more like an earlier era when a computer system could be a multimillion-dollar capital project.

The interesting question is not simply "Are there few buyers?" There are. It is why there are few buyers and what dimensions of concentration matter.

In the 1950s and 1960s, computers were so expensive that the customer base was intrinsically concentrated.

IBM's 1401, introduced in 1959, in the mid-1960s, represented more than half of the world's computers. The IBM System/360 arguably deepened the pattern, as a relatively small number of customers represented a very large percentage of industry revenue.

Early on, government was an extraordinarily important buyer of such computers, for example. 

There also are parallels in the semiconductor industry, where  military and aerospace customers were key. 

NASA's Apollo program, for example, purchased roughly 60 percent of U.S. integrated-circuit output in the early 1960s.

And the Minuteman missile program subsequently became an even larger individual consumer of ICs.

Of course, skeptics will argue that was different as the U.S. federal government was in no danger of defaulting, where OpenAI or Anthropic are not immune from that outcome.

In the context of demand for high-performance computing services, though, optimists might argue we must separate the small number of direct buyers from the much-larger end-user demand.

The number of organizations using HPC can be much larger than the number actually buying the infrastructure, for example, even if a few buyers stand out. 

So today's AI compute services market resembles other industries where direct buyers are few, such as:

• aircraft;

• electric utilities;

• telecommunications infrastructure;

• semiconductor fabs;

• power generation.

The closest non-computing analogy may actually be aircraft, where large commercial aircraft are sold to a relatively small number of airlines.

In that sense, the small number of passenger airline providers represents an aggregation of relatively large and dispersed  demand. 

That arguably resembles Nvidia selling to a few hyperscalers whose customers are highly concentrated at the moment, but also representing lots of dispersed enterprise demand. 

There might also be similarities to telecommunications or power generation, where a concentrated buyer base is not necessarily evidence of a small market.

In essence, there is a two-level demand structure, with a few key buyers (frontier language model suppliers) supporting a relatively small number of large HPC suppliers, which in turn support end-user demand that is highly distributed.

And, of course, optimists say the market will broaden over time. 

Era	Principal buyers	Buyer concentration
1940s–50s	Military, government, universities	Extreme
1960s mainframes	Government + very large corporations	Very high
1960s supercomputers	Government, defense, scientific institutions	Extreme
1970s–80s minicomputers	Corporations, universities, government	High → falling
1980s–90s PCs	Millions of businesses/consumers	Low
1990s–2000s servers	Businesses, Internet companies	Moderate
2010s cloud	Large enterprises + hyperscalers	Increasing
2020s AI/HPC	Hyperscalers + frontier AI + governments	Very high at infrastructure layer
Ultimate AI consumption	Potentially billions of people and millions of organizations	Potentially very low

The point is that customer concentration in computing is not unusual, especially at early stages of deployment.

Aside from the resemblance to mainframe, minicomputer, integrated circuit precedents, HPC might also suggest parallels to railroads, electric utilities and telephone service industries, where a small number of infrastructure buyers also has been key. 

Are there risks? Yes. But are the risks also manageable over time and structurally consistent with other capital-intensive industries? Perhaps also yes.

Who Benefits Most from AI: Highly-Skilled or Less-Skilled Workers?

In early language model studies, generative AI seemed to produce  larger  gains for less-skilled or less-experienced workers.

Later studies now show different outcomes. Apparently, studies now also suggest that workers with greater competence using AI seem able to capture disproportionately large benefits.

That seems quite contradictory. One potential answer might be that two different processes are at work. 

The first-generation effect of LLMs may be basic skill compression. The least-competent workers become more productive as AI lifts basic performance.

A second-generation effect might involve a new level of skill polarization around AI competence. As often happens, the most-productive workers then become even more productive, even as basic skill competence of the least-skilled also rises. 

The process might be akin to the “strength begets strength” phenomenon we often see with firms in general, which is that the best-managed firms also tend to be the organizations able to reap the most benefit from any new technology. 

Study	Setting / sample	Finding about skill differences	Implication
Noy & Zhang (2023), Science	453 college-educated professionals doing writing tasks	ChatGPT cut time by 40% and increased quality by 18%. The gains were largest among weaker writers, reducing inequality between workers. (DOI)	Strong evidence against the idea that higher-skilled workers benefit more—at least for standardized writing.
Brynjolfsson, Li & Raymond (2025), Quarterly Journal of Economics	5,172 customer-support agents	Productivity increased about 15%. Less-skilled and less-experienced workers gained substantially more; the lowest-skilled group gained roughly 30–35%, while the highest-skilled saw little benefit and sometimes small quality declines. (OUP Academic)	AI can transfer the practices of high performers to low performers, effectively compressing skill differences.
Dell'Acqua et al., BCG/Harvard "Jagged Frontier" experiment	758 consultants performing realistic consulting tasks	AI increased performance substantially on tasks within its capability frontier, with particularly large benefits for consultants who initially performed less well. But AI could also hurt performance on tasks outside its frontier.	AI can be an equalizer, but only if workers know where the technology is reliable.
Humlum & Vestergaard (2024), Denmark	Survey of 100,000 workers in 11 occupations	ChatGPT adoption was higher among younger, less-experienced and higher-achieving workers, particularly men. (IZA)	Here the inequality is partly an adoption/selection effect: workers with greater capability may be more likely to use AI.
Cruces et al. (2026), NBER	Randomized experiment, 1,174 adults, workplace-style problem solving	AI improved everybody's performance, but lower-education participants gained substantially more. The initial performance gap of 0.548 SD fell to 0.139 SD with AI—roughly three-quarters of the gap disappeared. Higher-education participants nevertheless used AI somewhat more effectively. (National Bureau of Economic Research)	Very strong recent evidence that GenAI can reduce education-based productivity inequality, while leaving an important advantage in AI utilization to better educated users.
Idan & Anand (2026)	Randomized experiment with early-career knowledge-worker analogs learning a technical domain	Average performance rose with an LLM, but gains were highly uneven. GPA and prior knowledge did not predict gains; AI Interaction Competence (AIC) did. High-AIC users obtained outsized gains while low-AIC users sometimes received little or negative marginal benefit. (SSRN)	This is probably the strongest evidence for the proposition in your question—but it changes "skill" into skill at working with AI.
Freund & Mann (2026)	Model of occupational/task-specific skills and GenAI exposure	Moderate AI exposure can benefit workers, while high exposure can hurt; effects vary greatly within occupations. Their model projects larger gains for lower earners and higher returns to social skills. (IZA)	Again, the aggregate prediction is not simply "AI favors the highly skilled." It depends on which skills remain complementary to AI.
Bloom et al. (2024), NBER/IZA	Theoretical model of AI and skill premia	Unlike empirical workplace experiments, the model assumes AI is particularly useful for high-skill tasks. Depending on substitution relationships, AI can nevertheless reduce the skill premium. (National Bureau of Economic Research)	Theoretical models don't necessarily predict widening inequality even when AI works disproportionately on high-skill tasks.
Humlum & Vestergaard (2025/26), Denmark	Administrative labor-market records linked to ChatGPT adoption	Despite widespread adoption and reported productivity improvements, they find no detectable earnings effect, with effects larger than 2% ruled out two years after ChatGPT's launch. AI is changing tasks more than aggregate earnings so far. (National Bureau of Economic Research)	Productivity gains do not automatically translate into higher wages for high-skilled workers—or anybody else.

The research arguably makes more sense  if "skill" is divided into three different things:

• Underlying job skill: how good you were at the task before AI

• Domain knowledge: how well you understand the subject being worked on

• AI-use skill: how well you can prompt, interrogate, evaluate, correct and integrate the model's output.

The first category has generally not favored high-skilled workers in the experiments, but the third category increasingly appears to do so.

So it seems AI can reduce the productivity gap between workers while simultaneously creating a new gap between good and bad AI users.

The 2026 Idan-Anand experiment is especially interesting because it explicitly finds that prior knowledge and GPA were not good predictors of who gained most from the LLM. 

What mattered was AI interaction competence; the ability to elicit, filter and verify the model's output.

That might be quite a different matter than the ability of a new employee to master basic skills.

Consider customer support roles:

• Top worker without AI: 100 units

• Bottom worker without AI: 50 units

• AI increases bottom worker to 70

• AI increases top worker to 105

In such instances, we can argue that productivity inequality is lessened. The least-skilled workers perform better. But the performance of the best workers also increases. 

But such instances also are of job categories where the “best” level of performance might be effectively limited. 

Other job roles involving higher cognitive input and output do not have the same degree of limits on “best” performance. In other words, there is no obvious ceiling on output quality. 

In those instances, AI is not going to help a person with cognitive, creative or other limits to approach the possible performance of workers with more innate capabilities. 

The upside from AI might  depend less on whether a job is "high skill" than on whether the job has a high ceiling on the quality of its output.

The key distinction is between tasks with an externally imposed performance ceiling and tasks where better reasoning, creativity, judgment, or synthesis can keep improving the result.

Consider a customer-service representative processing a standard billing inquiry.

There may be a fairly objective definition of a good answer:

• correct information;

• appropriate procedure;

• polite interaction;

• resolution of the customer's problem.

Once the AI gets the worker from, say, 80 percent to 95 percent effectiveness, there isn't necessarily much additional value in getting to 98 percent.

The value of being 10 times better than average is limited because the task itself doesn't have much room for exceptional performance.

Non-routine knowledge work arguably is quite different. Consider a:

• scientist developing a new drug;

• engineer designing a new semiconductor;

• entrepreneur developing a new business;

• investment analyst finding an overlooked opportunity;

• architect designing a building;

• software engineer designing a new system;

• lawyer developing an unusual legal strategy;

• management consultant developing a corporate strategy;

• researcher developing a new theory.

There isn't necessarily a point at which the answer becomes simply "correct” or “good.”

Exceptional or revolutionary value creation can happen, with enormous implications for value.

These are jobs where:

• the problem isn't completely specified

• there are many possible approaches

• success requires combining information from multiple domains

• there isn't an obvious algorithm for solving the problem

1. better answers can have enormous economic value.

In those scenarios, where upside is unlimited or bounded, AI might make a huge difference.

For jobs where a human can produce 0 to 100 units of value, and AI raises the average worker from 50 to 75, that’s helpful.

But in other instances, where a human can produce 0 to 100 units of value today, but potentially 1,000 or 10,000 if they discover something exceptional, the order of magnitude upside is where AI can produce exponentially-better outcomes.

Creative work has a peculiar economic property: “quality” is unrestricted or uneven. 

• A brilliant advertisement can transform a brand

• An exceptional bit of software can create an entirely new category

• An exceptional financial bet can produce enormous returns

• A breakthrough scientific paper can create an entirely new field. 

“Adequate” is one thing. “Transformative” is something else.

The conventional automation story is that machines take the routine jobs while humans retain the non-routine jobs.

But the generative-AI story might also be that AI takes the routine portions of non-routine jobs, leaving humans with increasingly open-ended problems:

• A consultant focuses on business options

• A programmer asks “What should we build?”

• A scientist asks “What should we investigate next?”

• A CEO continues to make decisions about what the company makes

• An entrepreneur decides which possibilities are worth pursuing.

So whether AI helps lower-skilled or highly-skilled workers more, the greatest long-term economic upside of AI might occur where AI can dramatically expand the number and quality of solutions that a capable human can explore, while the value of an exceptional solution has no obvious upper bound.

So open-ended, non-routine work may ultimately be a much more important AI opportunity than routine automation.