On-chain capital structure in the era of artificial intelligence

Author: Charlie Liu

Last Wednesday, at NVIDIA’s much-anticipated financial report, seeing the exciting performance, the stone hanging in the hearts of thousands of investors finally fell to the ground: revenue increased by more than 60% year-on-year, the data center business was sold out, and the performance guidance was raised again.

However, the capital market gave another reaction.Nvidia’s stock price briefly surged and then fell back, broader AI concept stocks fell collectively, and credit spreads of companies that aggressively expanded AI infrastructure widened.The public market even saw a 2.5% plunge in just over an hour.

In fact, talk of an “AI bubble” has been spreading recently: MIT said that 95% of enterprise AI pilot projects failed to produce measurable returns on investment, central bank governors warned that valuations have become as distorted as in the late 1990s, and the media began to dig into the recycling revenue among major AI companies.

In other words, despite the high revenue numbers, the market has begun to doubt whether the industry’s underlying base can support this valuation.

The real bottleneck of AI: electricity and capital

In a recent report on the energy and power industry, Goldman Sachs made an interesting analogy. The current moment echoes two infrastructure super cycles in history.

nineteenth centuryRailroad construction gave rise to modern investment banking and bonds as a popular asset class.

late twentieth centuryInternet construction has given birth to venture capital and sparked the rise of high-risk growth stocks and IPOs..

In the current AI era, traditional stocks and bonds cannot meet the demand brought about by the explosion of electricity and computing power. We need a new capital formation model, or even a new capital market.

And the fundamental constraint is whether we can provide enough AI-grade power and finance it without overwhelming the financial system.

Power dilemma

Over the past two decades, the U.S. power grid has grown at an average annual rate of less than 1% — manageable in the era of web servers and smartphones, but a disaster for AI factories.

Some analysis shows that in order to meet the comprehensive needs of new data centers, electric vehicles and industrial reshoring, the United States now needs to add approximately 80 gigawatts of power generation capacity every year.However, the actual annual growth is only 50-60 GW, creating a gap of about 20 GW each year – enough to support two or three cities the size of New York.

The first reaction to filling the gap is always the intuitive option: more natural gas power plants, accelerated deployment of wind and solar energy storage, and hope for the resurgence of nuclear energy.But none of them can meet demand within a reasonable time:

• New natural gas power plantIt was tempting on paper, but in practice it has become a project that takes an average of four years, with turbine supply bottlenecks pushing equipment delivery times to three to five years, not counting approval and grid connection queues.

• Onshore wind powerIncluding preliminary planning and grid connection studies, it usually takes three to four years, and may even drag on to nearly ten years, although the physical construction phase only takes six to 24 months.

• Utility Grade SolarIt is more modular and faster to build. The typical construction period is 12-18 months, and the average battery energy storage development period is shorter than two years. Therefore, “photovoltaic + energy storage” accounts for more than 80% of the expected new installed capacity in the United States in 2025.

• Nuclear energy, especially small modular reactors, may be the most compelling long-term answer to 24/7 AI-level power, but the target commercial operation time of the first round of SMR projects in North America is also around 2030-2035.

All of these options are essential, but they are only mid- to long-term solutions in a world where grid connection queues are often four to seven years away.

The only way to achieve a significant speedup is to reuse sites that already have land, high-capacity grid connections, and power infrastructure—especially large Bitcoin mines.In practice, upgrading an existing mine to an AI facility only requires a few months of modification work (liquid cooling, power distribution, GPU), rather than the four to seven-year long journey required to apply for a new grid connection from scratch.

This is why AI companies acquire or cooperate with mining companies: CoreWeave bid for CoreScientific, with the core purpose of converting its approximately 1.3 GW of mining infrastructure to AI.

Although the amazingness of Gemini 3 has made everyone wonder whether TPU will replace GPU in the future, so the demand for power has been reduced, the consensus gradually formed in the market is still the pattern of “GPU as the mainstay and TPU as the supplement”.Just like the previous doubts about GPU demand brought about by the emergence of DeepSeek, Nvidia’s GPU has once again withstood the pressure, and power demand expectations remain strong.

capital dilemma

Since ChatGPT detonated the AI craze at the end of 2022, the demand for AI data centers has soared, and the financing model has evolved through several stages.

• first stageSupported almost entirely by operating cash flow from very large businesses.When you’re generating tens of billions of dollars in free cash flow every year, you can quietly build a lot of data centers and lock up a lot of GPUs.But the scale of the current vision — a multi-trillion-dollar AI stack globally — is starting to weigh on those balance sheets.

• So we enteredPhase Two: Debt and Private Credit.There has been a surge in investment-grade borrowing to fund AI construction; high-yield issuers (Bitcoin miners transitioning to AI, developers of new data centers) have entered the junk bond market; and a rapidly growing private credit system has added custom loans, sale-leasebacks, and revenue-sharing facilities to this foundation.

• It is worth noting that many funds never appear on the balance sheet as simple “debt” but are off-balance sheetprivate credit: They exist in project joint ventures, structured leases and other off-balance sheet vehicles, converting capital expenditures into long-term obligations, making the entire stack more like shadow financing.If trillion-dollar AI capex forecasts are broadly accurate, banks and bondholders will not be able to support them; by 2028, private credit and these quasi-stealth structures are expected to provide a significant share—perhaps even a majority—of the capital behind AI data centers and power trading.

• Even this is not enough, so we seeEarly trends in phase three: securitization.Asset-backed securities for data center rentals and leases have quietly grown to approximately $80 billion in outstanding value and are expected to reach approximately $115 billion by 2026.In terms of equity, REIT-type instruments and joint ventures split the economic interests of “land + shell + electricity vs. GPU vs. AI application income”.

Public credit markets are already aware of these potential risks.Bloomberg’s criticism of Meta’s “creative financing” of its $27 billion off-balance sheet data center joint venture, as well as its comments on Oracle’s aggressive leasing and lending strategies, all point to the same point:Tech giants are unable to fully self-fund their AI efforts, and every new financing trick they employ makes bond investors more nervous.

So, is this an AI bubble?To some extent it is — but not in the way the headlines suggest.

On the equity side, valuations are truly eye-popping.AI-related companies account for a disproportionate share of market revenue, the S&P 500 trades at Internet-era valuation multiples, and Nvidia’s market capitalization briefly exceeded the GDP of almost every country except China and the United States.But equity investors at least think they know how to price growth and hype.

The more interesting — and dangerous — action lies behind thesecapital stack.The problem isn’t that AI has no practical uses, it’s that we’re trying to use it with a combination of risks that aren’t specific to this (Long-term physical risks: power plant and power grid upgrades; short-term technical risks: old GPUs may become obsolete within five years) to design tools and intermediaries to finance a generation of infrastructure.

Returning to the historical analogy mentioned earlier: Railroads were not financed solely by universal loans for crude oil, but the need to finance thousands of miles of track and rolling stock gave rise to modern investment banking and standardized railroad bonds; the Internet was not simply grafted onto the balance sheets of conglomerates; it gave birth to venture capital partnerships and norms around using equity to fund other loss-making companies in the portfolio, because of the extremely asymmetric distribution of returns and the potential for appreciation.

Therefore, the real question is: What should be a more effective capital formation mechanism in the AI era?What are its native financial instruments?

RWA: A financial instrument for a new era

On the surface, it looks like Wall Street has the answer.

“RWA” has become an annual hot word in earnings calls and regulatory speeches. It is the general name for tokenized treasury bonds, stocks, bank deposits and on-chain repurchase experiments, and is considered a new era of financial market infrastructure.

According to the SEC’s narrative, it seems as if it was born to be the financial infrastructure of the AI era, just like railroad bonds are to steel and startup equity is to the Internet.

However, in essence, tokenized RWA itself is not a new form of capital, it is just a new packaging of familiar financial products: behind it are still senior and mezzanine debt; common stocks and preferred stocks; income sharing agreements, etc.

In an energy or data center context, this could mean tokenized shares of a 20-year power purchase agreement; tokenized project equity with on-chain waterfall logic; tokenized REIT units; or short-term overcollateralized notes backed by contracted GPU revenue.

So, if RWA is nothing new, what real advantages does it bring over traditional financial instruments that transcend the noise and hype?Through the analysis of some early projects, we can see four major practical benefits:

1. fine separability: A $50 million project share can be divided into thousands of on-chain positions, allowing the position size to match a wider range of investment requirements.

2. Global reach: As long as securities rules are followed, the same instrument can be held by funds, family offices, DAOs or corporations in different jurisdictions without having to rewire the underlying pipeline each time.

3. Programmable cash flow distribution: Smart contracts can escrow stablecoins, enforce waterfalls and contracts, and automatically pay coupons or revenue shares based on verifiable performance data without relying on spreadsheets and intermediaries.

4. Fast settlement based on USD stablecoin: You can move principal and interest across time zones and weekends in minutes, although secondary market depth is still far thinner than in traditional bond markets.

All of this sounds like a financial upgrade, but it still feels like it leaves unanswered the deeper question of capital formation.

In the railroad age, bonds worked because there was a whole apparatus surrounding them that turned steel and land into standardized securities; in the Internet age, high-growth equity worked because venture capital partnerships turned chaotic startups into bankable pipelines.But tokenized RWA cannot miraculously create that flywheel out of thin air.

The real financial problem to be solved in the AI+ energy cycle is not how the leading AI companies can continue to “smartly” use financial engineering to borrow debt to build AI data centers and power plants, butHow to originate, aggregate and de-risk thousands of small distributed assets (solar rooftops, batteries, micro data centers, flexible loads) and express their cash flows in a way that global capital can truly hold with confidence.

This is exactly the gap that DePIN RWA is trying to fill, and why energy and computing networks are more important in this context than another general “RWA narrative.”

Energy DePIN: Long Tail Capital Formation

This is where DePIN—the idea of using tokens to coordinate the deployment of physical infrastructure—gets interesting.

Today, DePIN is still small.Messari’s 2024 report data shows that the entire sector’s approximately 350 tokens have a total market value of approximately US$50 billion, trading at approximately 100 times consolidated revenue.Specifically, there are about 65 projects in the energy DePIN subcategory, with a total market value of less than US$500 million.

If you are a traditional infrastructure investor, these numbers are laughably small in the face of trillion-dollar AI capex plans.However, the form of the optimal energy DePIN design almost perfectly fits the power bottleneck that the AI ​​stack is encountering.

Take Daylight as an example.

Its core logic is that distributed energy resources—rooftop solar, home batteries, electric vehicle chargers—can be orchestrated into a kind of software-defined power plant if they can detect and pay for “flexibility” rather than just raw generation.Its “proof of flexibility” mechanism pays in $GRID tokens when smart devices promise to adjust consumption or charging and discharging behavior in future high-voltage moments; energy companies burn $GRID to purchase access to that flexible capacity.

On this basis, $GRID serves as an energy-backed currency that touches every part of the stack: installation discounts for homeowners; payments for data and analytics; staking and derivatives for regional capacity mispricing; and insurance for off-chain capacity commitments.In its U.S.-only model, the total across physical and financial energy markets is about $1 trillion per year.

Daylight’s model is tightly coupled to the existing grid.If you believe that AI data centers will primarily be located within or near the current transmission grid, and utilities are willing to pay a premium for flexibility, this is an important selling point.There is also a risk if grid connection delays and regulations slow everything down.

In contrast is Arkreen.

If Daylight is “grid-native and US-centric,” then Arkreen is “grid-agnostic and globally oriented.”

It connects distributed renewable energy resources to a Web3-powered data and asset network.Participants install “miners” or connect via API; the network records verifiable green energy generation data and tokenizes it into renewable energy certificates and other green assets.

Arkreen has connected more than 200,000 renewable energy data nodes, issued more than 100 million kilowatts of tokenized RECs, and facilitated thousands of on-chain climate actions.

Its vision is explicitly global and long-tailed: a peer-to-peer energy asset trading network where households and small producers can connect their DERs to the DePIN system, earn tokens through “profit-for-impact” activities, and indirectly form virtual power plants or green AI offsets.

Taken individually, none of these projects will fund the next 1 GW data park for hyperscale enterprises.But they point out the possible form of “capital formation” in the AI era——What if we stopped thinking only in terms of nine-figure chunks of project financing and started thinking in terms of atomic “kilowatt-hours.”

This is exactly the centralization worry pointed out by the encryption + AI story and a16z’s latest encryption status report: laissez-faire AI tends to be centralized – big models, big clusters, big clouds.In contrast, blockchain excels at aggregating large numbers of small distributed contributions and giving them fluid global market access.

A cryptographic bridge connecting kilowatt hours and AI tokens

Currently, the value chain from marginal “kilowatt hours” to “AI tokens” is fragmented.

Power plants sign PPAs with utility companies; utility companies or developers sign contracts with data centers; data centers sign contracts with cloud providers and AI companies; AI companies sell API access or seats; and somewhere at the top of the stack, users pay a few dollars to run an inference.

Each segment is financed independently, with different investors, risk models and jurisdictional constraints.The opportunity, and the crypto-native version of “capital formation,” is to make this chain transparent and programmable.

• on the supply side, you can tokenize kilowatt-hour-related output to represent claims on specific renewable energy generation flows; tokenize RECs and carbon credits; and tokens that represent flexible capacity commitments from batteries, smart devices, and VPPs.Projects such as Arkreen show that this is technically and commercially feasible at a reasonable scale.

• in the middle reaches, you can express infrastructure as tokenized RWA: equity and debt in data centers, grid connection upgrades, behind-the-meter generation and storage, GPU clusters.Here, securitization in the traditional sense still happens, but on-chain rails make it more transparent: when investors buy tiered products, they know exactly which assets back them, and cash flows are settled in stablecoins that move in minutes rather than days.

• on the demand side, you can link energy and computing with AI-native tools: GPU-hour tokens, inference-second credits, and even application-layer “AI service” tokens.As agent AI systems mature, some of these tokens will be held and spent directly by software agents—programs capable of evaluating where to buy compute and power at the margin and dynamically arbitrage among providers.

From this, every marginal kilowatt-hour used by an AI model, from its origin (rooftop, solar farm, nuclear SMR) to its consumption in the GPU rack to its monetization in the AI application, isTraceable, priceable, and hedging.

This does not require that each link must be on the same chain or denominated in the same token.Rather, it means that the status of each link is machine-readable and can be stitched together by smart contracts and agents.

If you can do that, you essentially create a new form of capital: any investor, anywhere, can choose where to take risk in the chain—energy, grids, data centers, GPUs, AI applications—and purchase tokenized exposure of corresponding size and duration.

The balance sheet of hyperscalers is not going away, but it is no longer the only way to store that risk.

Conclusion

This story is not a foregone conclusion.

The combination of “big factories + big capital” is enough to accomplish all this independently.Hyperscalers may decide to own the energy stack directly through vertical integration and keep cash flow in-house.

Long-tail energy DePIN may never outgrow centralized projects.

But even if only a small portion of AI-related energy and computing ends up being financed and coordinated through DePIN and tokenized RWA, we have answered the open questions left by Goldman Sachs and a16z’s call for decentralization.

At this moment, computing power and electricity are intertwined in an unprecedented way, and the form of capital is also quietly reshaped.