Abstract

Artificial intelligence is driving an infrastructure investment cycle of exceptional scale. Whether that investment earns adequate returns will depend not only on growth in AI demand, but also on market structure, utilization, contractual commitments, counterparty credit, technological change and pricing. This article examines which historical infrastructure business models provide the most useful precedents for evaluating AI Infrastructure investment prospects.

The fiber-optic communications boom demonstrates that enormous demand growth does not guarantee capital recovery. Capacity became excessive relative to demand at prices capable of remunerating the original capital base, as numerous suppliers competed to fill largely undifferentiated networks whose construction costs were already sunk. By contrast, the U.S. tower-colocation sector has produced resilient cash flows through locally scarce assets, anchor tenants, long and escalating leases, high renewal propensity and attractive incremental economics from additional tenants. Public cloud offers a different successful model: hyperscalers aggregate diversified customer workloads, benefit from statistical multiplexing, control capacity expansion and add software-platform value above raw infrastructure.

AI Infrastructure currently exhibits elements of all three models. Long leases, contracted power capacity, prepayments and guarantees make some projects tower-like. Neoclouds and hyperscalers may achieve cloud-like workload aggregation. Yet fragmented build plans, reciprocal financial relationships, uncertain future pricing and a potentially large number of suppliers also create fiber-like risks.

The article distinguishes four utilization measures relevant to AI Infrastructure investment: serving utilization, economic utilization, physical utilization and industry utilization. Separating these measures helps identify which party bears utilization risk at different layers of the AI ecosystem and provides a more useful basis for capital-recovery analysis than reliance on a single utilization percentage. The central investment test is not whether AI demand will grow and create substantial value. It is whether contracted and realized demand will grow quickly enough, at sufficiently high prices, for first owners to recover committed capital before contracts expire, equipment is repriced, counterparties fail or competing supply erodes renewal economics.

AI use declaration. Microsoft Copilot was used as a research and writing aid in preparing this article, including for literature searches, source identification, drafting assistance and editorial review. Its use was directed by the author, and all resulting material was reviewed, revised and verified by the author. The article’s subject matter, analysis, judgments, conclusions and opinions are the author’s own. The author accepts full responsibility for the final content.

Keywords: AI infrastructure, capital recovery, utilization, fiber optics, tower colocation, public cloud, neoclouds, infrastructure investment, take-or-pay contracts, artificial intelligence, data center

Three key takeaways

• AI infrastructure returns increasingly depend on time to revenue. Power availability, utility deposits, construction schedules and permitting delays determine how long investors must commit capital before data centers begin generating income.
• Powered land and construction costs are raising the investment hurdle. DigitalBridge’s Brent Mayo said land premiums have reached $1 million–$2 million per megawatt, while First Citizens Bank’s Jeremy Wolfe cited turnkey construction costs of $12 million–$14 million per megawatt.
• Lenders are placing greater emphasis on tenant creditworthiness and execution risk. Hyperscalers, neoclouds and AI laboratories present different financing profiles, while refinancing exposure and community opposition add uncertainty to projected returns.

AI Infrastructure ROI Starts With Power, Permitting and Capital Commitments

AI infrastructure developers have plenty of prospective customers. The harder question is how much capital they must commit, and for how long, before those customers begin paying for capacity.

That distinction was central to a panel at Structure Research’s infra/STRUCTURE 2026 conference featuring Greg Miller and Remington Yee of Citizens, Brent Mayo of DigitalBridge and Jeremy Wolfe of First Citizens Bank.

Wolfe described utilities requiring larger upfront commitments to protect against stranded infrastructure investment. Developers may need to reserve power and fund preliminary infrastructure work before securing a signed lease.

Those expenditures increase capital at risk and extend the period before revenue begins. Delays can also increase financing costs and reduce the return on invested capital.

The constraint is consistent with RCRTech’s reporting on power availability and the projected $3 trillion data center capital expenditure cycle

For investors, time to power is becoming a financial variable, not simply a construction milestone.

Powered Land Premiums Raise the Cost of Reaching First Revenue

Mayo offered a measure of how much developers are willing to pay for speed to market.

He said land premiums that previously ranged from $300,000 to $500,000 per megawatt are now being quoted at $1 million to $2 million per megawatt.

These are Mayo’s observations of market pricing, not a standardized national land-price index.

Construction costs have also increased. Wolfe said turnkey data center development costs have risen from approximately $8 million–$10 million per megawatt in an earlier development cycle to $12 million–$14 million per megawatt today.

Together, these increases make speculative construction more expensive.

Panelists distinguished between horizontal development — securing land, power and permits — and vertical development, including the building and technical infrastructure.

Large AI training facilities present particular speculative risk because GPU architectures, cooling systems and rack densities continue to change. Developers risk completing facilities that no longer meet prospective tenants’ requirements.

Inference may support a different approach. Mayo described developers converting existing properties into smaller facilities ranging from 5 to 50 MW, sometimes ahead of committed demand.

The investment calculation depends on whether earlier availability produces enough additional revenue to justify higher acquisition and development costs.

RCRTech has examined the relationship between power access and changing data center development locations⁠

AI Infrastructure Financing: Matching Capital to Revenue Timelines

The capital required to build AI infrastructure extends beyond the physical data center.

Panelists described an expanding financing market that includes high-yield project finance, asset-backed securities (ABS), commercial mortgage-backed securities (CMBS) and private credit.

GPU financing adds another substantial requirement. Depending on the deployment, compute equipment expenditures can approach the capital cost of the data center itself.

Financing structures therefore need to account for different assets, borrowers and revenue schedules.

Investment-grade hyperscalers can generally obtain financing on terms that differ from those available to emerging neoclouds and AI laboratories. Less-established borrowers may require additional guarantees, private credit support or other financial backstops.

An April 2026 transaction illustrates the scale of project financing. First Citizens Bank announced a $525 million facility for T5 Data Centers⁠ supporting construction and expansion in Chicago and Charlotte.

The transaction also illustrates the importance of financing identifiable infrastructure projects with defined development requirements.

For AI infrastructure investors, the objective is to align debt service, construction expenditures and capital commitments with the timing and durability of expected revenue.

Neocloud Credit Risk and Refinancing Pressure Threaten Investment Returns

Mayo identified emerging neocloud operators as a potentially attractive but challenging investment segment.

Some are well capitalized and securing substantial customer commitments. Their operating histories, revenue concentration and business models, however, differ from those of established hyperscalers.

Mayo suggested that certain neoclouds could eventually achieve investment-grade credit profiles. Investors able to assess that transition may find opportunities, but the outcome remains uncertain.

The panel also identified refinancing as a growing consideration.

Infrastructure debt originated during the current development cycle will eventually mature. Borrowers may encounter different interest rates, credit conditions and asset valuations when those obligations need to be refinanced.

That creates a second financial test beyond completing construction: whether operating cash flow and asset values can support future financing requirements.

For neoclouds, the analysis also extends to GPU utilization, customer retention and the economics of operating compute capacity.

RCRTech’s analysis of neoclouds and the race to cut Time to Token⁠ examines the relationship between deploying compute infrastructure and beginning revenue-generating operations.

Community Opposition Extends Time to Revenue for Data Center Developers

Local opposition has become another source of financial uncertainty.

Panelists described development delays, canceled transactions and litigation involving projects that had already received approvals.

Yee emphasized transparent engagement and investment in community infrastructure, including schools and public facilities, rather than relying primarily on tax incentives.

The panel also discussed concerns about electricity rates and water consumption. Some participants cited utility claims that data center investment had benefited local electricity customers, although such outcomes require project-specific evidence.

The underlying financial exposure is more readily measurable.

A project delayed by permitting disputes or litigation may continue incurring financing costs while land, utility deposits and development expenditures remain committed.

For developers and lenders, community engagement increasingly belongs in the initial project-risk assessment rather than being treated as a separate public relations issue.

Conclusion: From AI Infrastructure Capex to Revenue and Return on Investment

The Structure Research discussion described an investment market increasingly divided among powered-land developers, power-generation providers, data center operators and compute infrastructure businesses.

Each assumes different financial risks and operates on a different revenue timeline.

Those distinctions matter when evaluating the economics of AI infrastructure.

Time to power measures how quickly a project can obtain usable electricity. Time to operational capacity measures when the facility and its equipment can support customers. Time to revenue measures when those assets begin producing income.

None of these milestones, individually, establishes profitability.

For neocloud operators, sustainable returns also depend on GPU utilization, operating margins, financing costs and customer demand. For data center developers, lease economics, construction costs and capital structure remain central.

The broader investment question is whether accelerating development and revenue generation produces returns sufficient to justify the additional capital and risk.

For AI infrastructure investors, the defining metric is shifting from how many megawatts or GPUs can be deployed to how quickly that capital produces sustainable revenue and an acceptable risk-adjusted return.

Editorial disclosure: This article is based on the panel “Capital Allocation in the Age of AI — Scale, Constraints and Execution Complexity” at Structure Research’s infra/STRUCTURE 2026 conference. Participants included Greg Miller and Remington Yee of Citizens, Brent Mayo of DigitalBridge and Jeremy Wolfe of First Citizens Bank. Otter.ai was used to transcribe the session. AI-assisted tools were used to organize source material and prepare an initial draft. Financial estimates attributed to panelists reflect their observations during the session and are not independently verified industry benchmarks. 

Three key takeaways

• Circular financing does not establish an AI infrastructure bubble. Structure Research argues that GPU utilization, contract pricing, customer commitments and underlying revenue provide more useful indicators of market health.
• Older GPUs may remain commercially productive long after newer chips replace them in frontier training. Their ability to support inference and batch processing could extend revenue-generating lives and change infrastructure financing assumptions.
• Inference is reshaping data center investment. Workloads will increasingly span metropolitan, regional and remote facilities, placing greater importance on network connectivity, compute utilization and revenue generated per megawatt.

Is AI infrastructure repeating the telecom bubble?

The AI infrastructure financing cycle has an uncomfortable resemblance to the telecom investment boom of the late 1990s.

Chip suppliers invest in AI companies. Those companies contract for compute from hyperscalers and neoclouds. The infrastructure providers purchase GPUs, often supported by financing tied to customer commitments.

The money moves among a relatively small group of companies.

But circular financing does not necessarily mean artificial demand.

That was the argument presented by Jabez Tan, CTO and head of research at Structure Research, during infra/STRUCTURE 2026.

Tan compared the current buildout with the telecom industry’s dark-fiber expansion. During the dot-com boom, operators installed substantial fiber capacity without sufficient paying customers.

Today’s AI infrastructure market presents a different picture, according to Structure Research. Demand for compute continues to challenge available supply, including capacity using older GPU generations.

The distinction matters for investors. The question is not simply whether suppliers finance customers, but whether those customers can generate sufficient revenue to support the capital invested.

Older NVIDIA GPUs challenge depreciation assumptions

One of Tan’s more revealing observations concerned the continuing demand for NVIDIA A100 and H100 accelerators.

Structure Research cited GPU rental pricing, contract renewals and extended customer commitments as evidence that older accelerators remain commercially valuable.

That challenges the assumption that GPUs become economically obsolete within three years.

New architectures may displace existing accelerators from frontier-model training, but older chips can continue supporting inference, fine-tuning and batch processing.

Structure Research described a progression in which GPUs move from frontier training into premium inference and subsequently into less time-sensitive workloads.

For lenders financing GPU infrastructure, the distinction between technological and economic obsolescence is important.

Equipment value depends on its ability to generate revenue above operating costs, not simply the arrival of a newer chip.

Longer equipment lives could support different depreciation and financing assumptions. However, utilization, electricity costs and declining inference prices remain important risks.

Frontier AI revenue growth drives compute investment

Structure Research estimates that aggregate annualized revenue across frontier AI labs could increase from approximately $46 billion in 2025 to $275 billion by the end of 2026.

The firm also estimates that OpenAI generates approximately $10 billion in annualized revenue per gigawatt of deployed compute.

These are Structure Research estimates, not independently reported company financial results.

The projections help explain why frontier AI developers are securing compute capacity years before deployment.

Competition among OpenAI, Anthropic, Google and open-weight model providers could increase overall token consumption while placing downward pressure on inference pricing.

That raises an important financial question: How much revenue generated by AI applications ultimately reaches the companies financing the infrastructure?

As examined in my earlier RCRTech analysis of neocloud time to token securing GPUs is only the beginning. Infrastructure must be powered, connected, commissioned and running customer workloads before it generates revenue.

Inference changes the data center map

Training and inference have different infrastructure requirements.

Large training clusters benefit from concentrated compute and high-performance interconnects. Interactive inference requires responsive systems, often closer to users and enterprise data.

Batch inference and some model-refinement workloads can operate farther from major metropolitan markets.

Tan described a future in which facilities accommodate different workloads throughout the day, serving interactive applications during business hours and shifting available capacity toward batch processing or fine-tuning at other times.

Structure Research calls attention to the growing importance of fungible compute: infrastructure capable of supporting different workloads as demand changes.

The concept introduces operational challenges. Moving workloads between facilities requires compatible hardware, orchestration, sufficient network capacity and compliance with customer security and data requirements.

RCRTech Principal Analyst Sean Kinney explored related architecture questions in his analysis of test-time inference scaling and edge AI⁠.

For data center developers, the question becomes which combinations of workloads a facility can economically support over its operating life.

Connectivity becomes part of compute economics

Distributed inference also changes the importance of networking.

Moving workloads among metro facilities, regional campuses and remote AI clusters requires high-capacity connections and predictable latency.

A remote facility with inexpensive power may still be unsuitable for certain inference applications if network performance or connectivity costs undermine its economics.

Kinney and industry executives examined these requirements in RCRTech’s Scale-Up, Scale-Out, Scale-Across webinar⁠.

The same relationship between power, geography and infrastructure investment was examined in my recent article, AI Data Center Capacity Could Approach 100 GW in the U.S. by 2031 — but Power Is Deciding What Gets Built⁠.

Power determines where capacity can be developed. Networking and workload requirements help determine how productively that capacity can be used.

Neoclouds transfer time and technology risk

Structure Research characterizes neoclouds as infrastructure intermediaries providing access to power, accelerated deployment and GPU ownership arrangements that hyperscalers and AI labs may prefer not to undertake directly.

According to Structure Research’s conference estimates, Microsoft and Anthropic account for approximately 56% of signed neocloud compute contracts in the firm’s tracked dataset.

That concentration creates counterparty exposure for infrastructure lenders and operators.

Structure Research also reported increasing interest in smaller 20- to 50-megawatt compute deployments alongside much larger AI campuses.

These smaller blocks can offer more manageable financing and deployment requirements, particularly for inference workloads.

The broader migration toward secondary data center markets has also been examined by RCR Wireless News journalist James Blackman in his reporting on the generational AI build-cycle⁠.

Conclusion: From installed megawatts to productive compute

The AI infrastructure industry may eventually confront overinvestment. Customer concentration, GPU financing and rapidly changing hardware generations all create legitimate risks.

But Structure Research’s analysis suggests that circular financing alone is an insufficient basis for declaring a bubble.

The stronger indicators will be GPU utilization, contract renewals, compute pricing, customer cancellations and returns on deployed capital.

For neoclouds and infrastructure investors, the operating measures worth tracking include time to first token revenue, revenue per GPU, revenue per megawatt and gross profit generated from installed capacity.

The next phase of AI infrastructure investment will not be measured solely by how many gigawatts get built.

It will be measured by how much productive, profitable compute those gigawatts deliver.

Three key takeaways

• U.S. colocation capacity is entering a massive expansion cycle. Structure Research estimates operational U.S. colocation capacity at roughly 23 GW in 2026 and projects it could approach 100 GW by 2031, representing a 33.6% five-year growth rate.
• The headline pipeline overstates what will actually get built. The increasingly important distinction is between operating capacity, projects under construction and enormous land banks that may still lack committed customers, financing or deliverable power.
• Time to power is becoming time to revenue. With GPUs, capital and customers waiting on capacity, securing and delivering power increasingly determines how quickly AI infrastructure can move from a development plan to productive compute and revenue.

U.S. data center capacity is moving from megawatts to gigawatts

The scale of the U.S. data center pipeline is getting difficult to comprehend.

During the Americas market update at infra/STRUCTURE 2026, Structure Research analyst Vivian Kong put the current U.S. pipeline at roughly 40–42 GW of operating data center capacity, another 25 GW under development and approximately 150 GW in land-banked capacity.

The more revealing number may be the growth trajectory for commercial colocation.

Structure Research estimates approximately 23 GW of operational U.S. colocation capacity in 2026, growing to nearly 100 GW by 2031. That represents a five-year growth rate of 33.6%.

Those numbers need context. A land bank is not a data center, and an announced gigawatt is not necessarily a future operating gigawatt.

That distinction is becoming increasingly important as established data center developers, real estate owners, former cryptocurrency infrastructure operators and new entrants compete for land and power.

Structure Research estimates speculative entrants account for approximately 17% of capacity currently under development and 27% of land-banked capacity.

The next question for the market, then, is no longer simply: How many gigawatts have been announced?

It is: How many gigawatts can actually be energized?

That distinction is consistent with other RCRTech reporting. In February, RCRTech examined the increasingly difficult search for what JLL calls “digital dirt” — land that can actually support power-intensive data center development. More recently, RCRTech reported that power availability is actively reshaping where AI infrastructure and the networks connecting it are being deployed.

Power is redrawing the U.S. data center map

Northern Virginia remains the largest individual U.S. colocation market in Structure Research’s dataset, with approximately 16.4 GW across operating and pipeline capacity.

But growth is moving outward.

Texas has emerged as the largest state-level colocation market and pipeline, with Structure Research tracking approximately 42.4 GW of total capacity. West Texas, in particular, is developing as what Kong described as a “pure AI” market — infrastructure being created specifically around large AI campuses rather than expanding from a traditional enterprise data center base.

Similar movement is occurring into Ohio, Illinois, Nevada, Arizona and less traditional data center markets.

The driver is increasingly straightforward: power.

ERCOT’s March 2026 data showed 43.5 GW in its large-load interconnection queue for 2026, rising to 150.6 GW for 2027. Those figures include projects at different stages of development and should not be interpreted as committed future load.

Texas regulators have also moved toward a batch-review process for large loads of 75 MW or greater as the grid responds to unprecedented requests for large-load interconnections.

RCRTech has been following the same constraint from the infrastructure side. Recent reporting found that power access is influencing not only where AI data centers get built but the fiber and optical networks required to connect increasingly distributed facilities. RCRTech has also examined how shortages of transformers, switchgear and other electrical infrastructure can delay data center construction.

For AI infrastructure economics, that changes the meaning of Time to Token.

Before a GPU can generate a token — or a dollar of revenue — the underlying facility has to be financed, constructed, connected and energized.

Time to power is becoming one of the first components of time to revenue.

Former crypto infrastructure is becoming AI infrastructure

Another important shift is occurring among companies that already control large blocks of power.

Structure Research estimates GPU-oriented providers account for 40% of U.S. colocation capacity currently under development.

Several of these operators built their original infrastructure around cryptocurrency mining and now control something AI developers badly need: sites with access to large blocks of power.

The transition is visible in actual contracts.

Cipher Digital, formerly Cipher Mining, signed a 15-year agreement with AWS covering approximately 300 MW of gross capacity at its Black Pearl facility near Wink, Texas. According to Cipher’s SEC filings, bitcoin mining at Black Pearl was decommissioned in February 2026 and approximately 85% of the site’s existing data center infrastructure is expected to be repurposed for the AWS lease. Cipher puts the contracted revenue associated with Black Pearl at approximately $5.5 billion.

Applied Digital provides another example. According to the company’s 2026 SEC filing, CoreWeave has contracted for 400 MW of critical IT load at Polaris Forge 1 in North Dakota under long-term leases.

The broader numbers show how quickly the model is scaling. Applied Digital reported 1,410 MW of contracted critical IT load across five campuses as of May 31, representing approximately $36.2 billion of contracted revenue over the initial lease terms.

The economics help explain why this part of the market is moving so quickly.

A developer that already controls a viable site and power position may be able to shorten the distance between an AI customer’s capacity requirement and revenue-producing compute.

Hyperscaler self-build is not replacing colocation

Structure Research also highlighted an important distinction in the relationship between hyperscaler-owned infrastructure and commercial colocation.

The firm’s forecast shows self-build capacity growing at a slower rate than colocation through 2031. But Kong cautioned against interpreting that trend as evidence that hyperscalers are abandoning self-build.

Long-dated projects simply have less visibility.

At the same time, hyperscalers remain major customers of colocation providers. Their own campuses and leased capacity can therefore expand simultaneously.

That dynamic is also pushing infrastructure into markets outside the traditional data center hubs.

Structure Research estimates that roughly 48% of approximately 60.9 GW of hyperscale self-build capacity sits in what it classifies as the “rest of U.S.” market — locations outside the established major data center hubs.

States including Wyoming, Iowa and Missouri are attracting large hyperscale projects. In some cases, hyperscaler investment can establish the power, connectivity and infrastructure ecosystem that subsequently attracts colocation development.

Canada’s AI data center growth shifts west

Canada shows a similar relationship between power availability and geography.

Toronto and Montreal remain major markets, but Structure Research expects slower growth as power and land constraints limit expansion.

Western Canada is moving in the opposite direction.

Structure Research projects a 22.5% five-year growth rate around Calgary and the broader western Canadian market, where developers have access to land and more opportunities to pursue large-scale power strategies.

Meta provided a significant proof point in July, announcing a 1 GW AI-optimized data center in Sturgeon County, Alberta. Meta says the project represents more than C$13 billion of investment and will be the company’s first data center in Canada.

The shift resembles what is happening in Texas: large AI campuses are increasingly going where power, land and permitting can support gigawatt-scale development.

Regulation becomes another variable in time to power

Power availability is only part of the constraint.

Community opposition and regulation are beginning to affect project schedules as states wrestle with electricity costs, transmission requirements, water use and the local economic impact of enormous AI facilities.

On July 14, New York Gov. Kathy Hochul signed a one-year statewide moratorium on new hyperscale data centers while the state develops a new regulatory framework intended to address ratepayer, environmental and community impacts.

The issue is becoming an increasingly important part of data center development. RCRTech has also reported on the growing role of community opposition in delaying or blocking data center projects.

For developers and AI infrastructure operators, permitting and regulatory risk now join power procurement, financing, construction and hardware availability in determining when a project can begin generating revenue.

From gigawatts of capacity to productive AI compute

The Americas data center market does not appear short of proposed capacity. It has hundreds of gigawatts of it.

The constraint is deliverable capacity.

That creates an important distinction between several stages of the AI infrastructure development cycle:

Announced capacity → credible capacity → powered capacity → operational compute → revenue-generating compute

Structure Research’s data documents the enormous amount of capacity moving through the first stages of that chain.

From an RCRTech AI infrastructure economics perspective, the next step is measuring how efficiently operators move through the rest of it.

Power alone does not make AI infrastructure productive.

Once the facility is energized, GPUs need to be interconnected, fed with data and kept highly utilized. As RCRTech Principal Analyst Sean Kinney has examined in his work on distributed inference, the shift from centralized AI training toward inference introduces another infrastructure question: where should compute run, and how should those compute resources be connected?

That extends the economics beyond the data center itself.

Kinney’s work on scale-up, scale-out and scale-across AI network fabrics examines the networking required to connect GPUs inside racks, clusters and data centers and, increasingly, across geographically distributed compute environments.

For RCRTech, those infrastructure layers fit into a broader operating sequence:

Time to power → time to production → time to first token → GPU utilization → revenue per GPU and revenue per MW

The nearly 100 GW U.S. colocation market Structure Research projects for 2031 is therefore about more than real estate.

It represents a test of execution.

Demand for AI infrastructure may be enormous. The harder question is how much of the announced infrastructure can actually be powered, built, connected and converted into productive AI capacity — and how quickly that capacity begins generating revenue.

Disclosure: This article is based on the “USA Market Pipeline & Americas” presentation at infra/STRUCTURE 2026 in Las Vegas. RCRTech reviewed the session transcript and independently checked selected company, regulatory and infrastructure claims against company filings, government sources and other primary materials. Otter.ai was used to transcribe the session, and AI-assisted tools were used in preparing the initial draft. The article was reviewed and edited by the author before publication.

Three key takeaways

• The AI infrastructure constraint is moving from acquiring GPUs to putting them into production. Structure Research sees demand continuing to run ahead of available supply as hyperscalers, AI labs and neoclouds compete for power and data center capacity.
• Neoclouds are selling speed as much as compute. Structure Research estimates roughly 14 GW of neocloud capacity could be live by 2028, much of it serving hyperscalers and frontier AI labs that need compute before other capacity is ready.
• Time to power is becoming part of time to revenue. The economics depend on how quickly capital, GPUs, power, networking and data center capacity can be converted into running workloads — and how productively those assets are used once online.

AI infrastructure demand is running into a deployment constraint

The AI infrastructure race is no longer only about getting GPUs.

Structure Research CTO and Head of Research Jabez Tan put numbers behind the next constraint during his presentation, “AI Infrastructure: Who is buying AI capacity, how it is sourced and financed, and what drives rent,” at infra/STRUCTURE 2026.

Tan examined how AI capacity moves from chip suppliers and data center developers through hyperscalers and neoclouds to frontier AI labs — and how that infrastructure gets financed.

“It’s never really been a demand issue,” Tan said. “It’s more how can we bring online credible supply to power these workloads.”

That distinction fits RCRTech’s Time to Token framework: measuring how quickly capital committed to AI infrastructure can be converted into productive compute and ultimately revenue.

Structure Research sees neocloud capacity reaching 14 GW

Structure Research estimates approximately 14 GW of neocloud capacity could be live by 2028, subject to execution, construction and power availability.

The customers behind that capacity are concentrated.

Tan said Anthropic, Microsoft, OpenAI, Meta, xAI and NVIDIA represent about 65% of the live neocloud capacity being purchased among the buyers Structure tracks. Microsoft and Anthropic account for 56% of signed contracts in its dataset.

These are Structure Research estimates presented at the conference, rather than industry-wide reported totals. But they highlight an important characteristic of the market: neocloud demand cannot always be separated from hyperscaler and frontier-lab demand.

Microsoft, for example, contracts with multiple neocloud providers. Frontier labs are sourcing compute through hyperscalers, neoclouds and direct infrastructure arrangements.

Structure’s public research also finds frontier labs driving neocloud leasing. Structure Research analysis of frontier labs and neocloud leasing

Neoclouds are selling time to power

Tan offered a concise description of the neocloud value proposition.

“What the neoclouds provide is time,” he said, specifically citing “time to market” and “time to power.”

For an AI lab waiting for a self-built campus or hyperscale capacity, that time has economic value.

A neocloud can secure GPUs, lease powered data center capacity, arrange financing and make compute available as a service. The customer gets capacity sooner. The neocloud assumes much of the financing, deployment, utilization and GPU-generation risk.

That maps directly to RCRTech’s Time to Token framework.

The clock does not stop when a GPU is purchased. It runs through installation, power, networking, cluster commissioning and customer deployment before the infrastructure starts producing tokens and revenue.

Financing is part of time to token

Getting there quickly takes capital.

Tan said neoclouds had raised more than $120 billion in committed equity and debt during 2026 through the time of his presentation, including undrawn facilities and excluding letters of intent.

CoreWeave provides a public example of how financing connects directly to deployment.

In August, a CoreWeave subsidiary entered into a $2.6 billion delayed-draw term loan facility primarily to fund GPUs and associated infrastructure required to support customer contracts. CoreWeave SEC filing

The financing is therefore part of the production timeline. Access to capital affects how quickly GPUs, power and data center infrastructure can be assembled into revenue-producing capacity.

AI data center rent prices credit and technology risk

The same economics extend to the data center developer.

Tan said Structure Research is paying closer attention to how tenant credit and cost of capital affect AI data center rents.

An investment-grade hyperscaler does not present the same underwriting profile as a young neocloud making a large capacity commitment. Developers and lenders have to account for tenant credit, contract duration and the cost of financing the facility.

They also have to price technology risk.

GPU generations can move considerably faster than data center financing and lease terms. Tan identified chip obsolescence as one reason AI labs have historically been cautious about leasing large amounts of infrastructure directly.

Neoclouds can take some of that risk onto their own balance sheets. The economics of the contract have to compensate them for doing it.

Time to power is only the first AI infrastructure metric

Getting a GPU cluster energized solves the deployment problem. It does not guarantee attractive economics.

Utilization comes next.

For an AI infrastructure operator, the sequence is straightforward:

capital committed → GPUs acquired → power secured → cluster deployed → customer workload running → tokens produced → revenue generated

A delay anywhere in that sequence pushes revenue farther away from the original investment. Low utilization after deployment creates a different problem: infrastructure has been financed and built but is not producing its expected economic output.

That is why RCRTech’s AI Infrastructure Economics framework extends beyond Time to Token to four operating measures: revenue per GPU, revenue per MW, gross profit per GPU and gross profit per MW.

RCRTech has examined the same issue through GPU utilization: expensive infrastructure creates economic value only when customers are using it. RCRTech analysis of AI infrastructure utilization

Conclusion: Measure the time between capital and revenue

Tan’s presentation helps explain why neoclouds have become an important part of the AI infrastructure market.

They are not simply another place to rent GPUs.

They connect capital, chips, data centers and customers, taking on risk to make AI compute available sooner. Tan’s description — “time to market” and “time to power” — captures the first part of that value.

RCRTech’s Time to Token framework extends the measurement to the economic outcome.

How long does it take from committing capital to putting GPUs into production and generating revenue?

Once the infrastructure is running, the measurement shifts to productivity: revenue per GPU, revenue per MW, gross profit per GPU and gross profit per MW.

For neoclouds and their capital providers, both sides of the equation matter: get GPUs into production faster, then generate more economic output from every GPU and every megawatt.

Editorial disclosure: This article is based on an infra/STRUCTURE 2026 presentation attended by RCRTech. Otter was used to transcribe the session, and AI tools assisted in preparing the initial draft. The article was reviewed, fact-checked and edited by RCRTech before publication.