Is Google playing with fire?

We may have drastically underestimated the complexity of this 2026 AI race.

While most observers fixate on overt indicators such as chip manufacturing processes, model parameters and market share, a genuine strategic shift is quietly taking place in the notes to a quarterly report.

Alphabet disclosed that the nominal exposure of its credit derivatives has surged from $16.9 billion to $43.8 billion in six months, with another $24.1 billion of guarantees expected to be put in place in the future. These figures do not appear in prominent positions on traditional balance sheets, yet the scale of commitments they represent is already close to the annual GDP of some medium-sized countries.

We are accustomed to regarding technological progress as the primary driving force of industrial transformation, a premise that has proven effective time and again over the past few decades. The party that owns more advanced technology can define industry standards and reap excess profits.

Intel dominated the PC era with the x86 architecture, while NVIDIA locked in the AI training market with its CUDA ecosystem.

The logic of the technological moat is simple, clear and repeatedly verified.

But what Alphabet is doing now is challenging this deep-rooted perception.

It has deployed a weapon completely unrelated to technology, such as its AAA credit rating, in an attempt to carve out a moat along the TPU technology roadmap that NVIDIA cannot replicate.

Equitization of Credit, the Misunderstood "Guarantee" Business

What is so special about this new approach? Let's first look at this specific transaction.

According to public documents, Google provided data center lease guarantees for a developer named TeraWulf.

Many people may be familiar with the name TeraWulf. Its previous main business was cryptocurrency mining, and it is now transitioning to AI infrastructure. When a company like this applies for a bank loan to build a data center, the bank will only worry: if the project fails, who can they sell the collateral in their hands to?

This question is relatively easy to answer for data centers using NVIDIA GPUs.

Countless enterprises, research institutions and startups around the world need NVIDIA's computing power. Although the second-hand GPU market has discounted prices, liquidity still exists. Banks can roughly estimate a residual value and set loan terms based on that.

But for data centers packed with Google TPUs, this problem is almost unsolvable.

TPUs are dedicated chips designed by Google from scratch for its own software stack, deeply integrated with Google's TensorFlow framework, network architecture and storage system. Without Google's ecosystem, these expensive silicon chips are essentially no different from a pile of sand.

Any rational credit assessment agency will draw a harsh conclusion: the residual value is close to zero. This means that according to normal business logic, no one will be willing to lend money to a TPU data center unless the financing cost is high enough to cover the full risk of loss.

Google's intervention method is quite clever. It promises the bank: if the tenant defaults, it will take over the lease responsibility or pay the termination fee.

The economic consequence of this commitment is that banks no longer need to assess the residual value of TPUs, nor do they need to dig deep into the business prospects of TeraWulf, this former mining company.

They only need to assess one thing: will Alphabet go bankrupt? The answer is obviously "no".

As a result, the loan pricing benchmark has shifted from "a highly speculative data center project" to "unsecured debt of one of the companies with the highest credit ratings in the world".

The financing cost instantly drops from the unfeasible range to the feasible range.

It does not end here. Google does not provide guarantees for free. TeraWulf has issued deep in-the-money warrants to Google for this purpose, with an exercise price of $0.01 per share, involving tens of millions of shares.

This means that when TeraWulf's project is completed with the obtained financing and its valuation rises, Google can share this equity appreciation at a cost close to zero.

Essentially, this structure is "equity consideration for credit intermediary services".

Put simply, Google's core resource is its AAA credit. The traditional way for it is to use this credit to issue bonds and then invest the raised funds. But now it has found a new usage: to sell credit as a service, and the compensation it receives is not interest, but equity call options of the counterparty company.

It separates the value of its own credit from the bond market and injects it into the equity market of artificial intelligence infrastructure.

As Google promotes this template across about ten projects with a total power scale of 2.4 GW, it is actually building a unique industrial entry barrier for itself.

NVIDIA can make faster chips, but it cannot conjure up a AAA credit rating out of thin air to provide guarantees for customer projects around the world.

The CUDA ecosystem locks in software developers, while Google's financial ecosystem is locking in infrastructure developers and end computing power tenants.

The firmness of these two types of "lock-in" will show significant differentiation in future industrial competition.

If Risks Suddenly Converge in the Same Direction

There is a well-known curse in the financial field called "false independence of risk models".

Suppose you run an umbrella factory. To hedge risks, you buy flood insurance and invest in the stock of a raincoat company. Your logic is: if rainfall is insufficient, umbrellas will be unsalable, but the raincoat company's stock price may rise due to more outdoor activities; if a flood occurs, the umbrella factory will be damaged, but the flood insurance will pay compensation.

You submit this plan to the board of directors, using an exquisite correlation matrix to prove that various assets can hedge against each other.

But you have overlooked one possibility, such as a hurricane.

The strong wind it brings will destroy the factory building, the heavy rain will cause floods, and the paralyzed logistics will also prevent the raincoat company from delivering goods.

The assets you marked as "low correlation" or even "negative correlation" in the model will collapse completely synchronously in the face of the same hurricane.

The entire guarantee network that Alphabet is building now is such a structure that exposes all risks to the same "hurricane", for example, "the commercialization of cutting-edge AI collapses".

We need to break down every key point in this network. First is the end demand: those AI labs that rent TPU computing power, such as Anthropic, their ability to pay is entirely based on being able to continuously obtain the next round of financing or generate considerable commercial returns.

If the market doubts the profitability of AI models and venture capital shrinks, the default rate of this group of tenants will rise sharply.

Second is Google's own undertaking obligation: once a tenant defaults, Google needs to take over the lease contracts of these data centers according to the guarantee terms. On the surface, this is not a loss, because Google itself also needs a large amount of computing power.

But there is a problem here: the location of the data center, power agreement, cooling system and network wiring are optimized for the specific needs of a specific tenant at the beginning of construction. When Google is forced to take over, these facilities may not meet its technical requirements, and the transformation cost and time delay will erode all the book value.

Third is inventory: within six months, Google's TPU inventory has expanded from $2.4 billion to $10 billion. These chips should have been deployed as the guarantee projects progress.

If market demand shrinks, there is hardly a second-hand market for these dedicated chips, and the write-down pressure will directly impact the income statement.

Fourth are the more than $850 billion of various procurement commitments, energy contracts and unstarted data center leases. When these contracts were signed, the expectation was a future of supply shortage.

Once the slope of the demand curve changes, huge fixed expenditures will become pure financial burdens.

Do you see the key to the problem? All these risks, such as tenant default, guarantee trigger, inventory write-down and committed expenditure, have highly overlapping trigger conditions. When AI commercialization cools down, they will occur at the same time, instead of rising and falling alternately as assumed in risk management textbooks.

In the traditional industrial chain, fluctuations in market prices can be dispersed to different entities upstream and downstream, forming a buffer mechanism.

However, Google has deeply bound chip design, infrastructure financing, computing power leasing and cloud service sales to itself through financial means, eliminating the originally existing market buffer zone.

In the rising cycle, this kind of binding manifests as unparalleled synergy efficiency. But in the downward cycle, it will become a rigid structure with no room for force dissipation.

In the history of financial development, the lessons learned are that the most dangerous moment is often not the moment with the highest leverage, but the moment when everyone suddenly finds that all the risks marked as "independent" are actually on the same boat. Google's current growth rate of guarantee exposure, with a 158% increase in six months, means that the market may be much closer to this cognitive tipping point than imagined.

Are Enterprises Starting to Perform the Functions of Banks?

Finally, it is necessary for us to view this phenomenon from a more macro perspective, because it highlights profound institutional changes. In the emerging strategic industry of artificial intelligence, super-large enterprises are evolving from commercial entities to operators of quasi-market infrastructure.

It should be noted that the sign of a mature industry is the existence of developed financial intermediaries and market pricing mechanisms.

Behind the oil industry, there is a series of tools such as crude oil futures, project financing and reserve securitization. The shipping industry has the Baltic Index and the ship financial leasing market.

These financial infrastructures make asset pricing transparent and allow risks to be transferred to entities that are willing and suitable to bear them.

But what stage is the AI computing power industry in now? It is no exaggeration to say that it is almost a financial wilderness.

How to assess the value of a TPU? How to price a data center lease contract to be delivered seven years later? How much is a revenue-sharing agreement based on the revenue of future AI models worth?

Wall Street does not have standard answers to these questions yet.

It is precisely this vacuum zone that gives Alphabet a stage to display its capabilities.

Due to the lack of a market, Google simply uses its own balance sheet and credit rating to act as this "market".

It creates liquidity for TPU computing power, making lenders willing to accept these assets as collateral through guarantees; it sets risk pricing, using its own credit rating to lower the financing cost of the entire ecosystem; it also acts as the lender of last resort, promising to step in and take over when the market has problems.

This is an amazing innovation. It bypasses the bottleneck of the financial system's slow response and insufficient understanding of new industries, and leverages hundreds of billions of dollars in infrastructure investment with the decision-making efficiency of a single company.

But it also brings systematic information asymmetry and risk concentration. When the core financing pricing of an industry is not generated by the game of a large number of buyers and sellers in the open market, but anchored by the credit rating of a single company, this pricing is fragile.

It does not reflect the market's true judgment on the prospects of the AI industry, but the market's judgment on Alphabet's solvency.

There is a dangerous information gap between the two. Alphabet's credit status may still be very healthy, but those underlying projects it guarantees may have already fallen into trouble due to technological route changes, soaring energy costs, or delayed realization of demand.

This is exactly what the Bank for International Settlements and the Bank of England are concerned about as they begin to pay attention to the role of private credit in AI financing.

No one can currently quantify the impact of a hundreds of billions of dollars shadow debt network built by guarantees from super-large enterprises on overall financial stability. Existing accounting standards classify these guarantees as regular commercial commitments, and the nominal exposure of $43.8 billion corresponds to a fair value liability record of only $815 million on the balance sheet.

This treatment complies with the current rules, but when the rules were formulated, legislators could not have imagined that a technology company would use its own credit to directly leverage the capital formation of an entire industry.

Regulators eventually require enhanced disclosures, including counterparty concentration and the sensitivity of fair value to key assumptions. This is probably not out of harshness, but simply because the market is actually no longer able to correctly price this huge asset class anchored by a single company's credit.

When investors buy Alphabet's stocks, they think they hold a technology company whose main businesses are advertising and cloud services.

The truth is that they have unknowingly taken on the risk exposure of a super guarantee portfolio deeply embedded in the global AI infrastructure cycle.

Historically, every wave of technology has spawned matching financial innovations. The canal era had government-guaranteed bonds, the railway era had land grants and stock issuance, the telecommunications era had supplier financing, and now the AI era seems to be creating its own financial instruments and institutional arrangements.

The credit distribution machine forged by Google with its balance sheet is very likely to be the landmark invention of this era.

Its ultimate fate does not depend on the level of financial skills, nor on the strictness of regulatory attitudes, but on whether the commercialization process of artificial intelligence can outrun the self-reinforcing expansion speed of this guarantee network.

If it outruns, this system will be a visionary strategic initiative.

If it fails to outrun, it will constitute the most expensive correlation risk management lesson in business history.

And the final grade of this lesson will be borne by all investors, creditors and users who participate in this AI craze.

This article is from the WeChat public account "Dongzhen Strategy", Author: Dongzhen Strategy, Published with authorization from 36Kr.