Executive Summary
● Information processing equipment (4% of GDP) generated 92% of U.S. GDP growth in H1 2025.
● The remaining 96% of the economy grew at just 0.1% annualised.
● U.S. growth is infrastructure-driven, not productivity-driven.
● Magnificent 7 plus Broadcom: 32% of U.S. stock market value.
● AI stocks drove 80% of all U.S. market gains in 2025.
● Foreign investors poured $290 billion into U.S. stocks in Q2 2025 alone.
● Is this a rational infrastructure buildout (Goldman Sachs view) or a speculative “bubble within a bubble” (Jeremy Grantham view)?
● The paradox: Market concentration is hyper-rational given where growth lives, yet fragile due to dependency.
The One-Pillar Economy
The U.S. economy is exhibiting a remarkable and deeply fragile concentration risk that demands the immediate attention of sophisticated investors. There is no escaping the numbers. A startling 92% of all U.S. economic growth in the first half of 2025 was driven by investments in information processing equipment, which accounted for just 4% of the country’s GDP.
That is the entire story.
The implication is profound and troubling.
The remaining 96% of the economy, which includes manufacturing, housing, services, and consumer discretionary spending, grew at an annualised rate of just 0.1%. Because it is using artificial intelligence to boost productivity throughout its economy, the United States is not currently expanding. The rapid expansion of the physical infrastructure required for future productivity drives growth.
It’s still unclear how much this investment will yield in the near future.
The U.S. equity markets are a clear example of this extreme economic concentration. Together with chipmaker Broadcom, the “Magnificent 7” technology companies now make up 32% of the value of the US stock market. In 2025, AI-related businesses accounted for 80% of the gains in the US stock market. This concentration is now the main attraction for growth in global capital seeking. In just the second quarter of 2025, foreign investors invested a record $290 billion in U.S. stocks. In essence, the U.S. market has turned into a single, heavily leveraged wager on AI.
The defining paradox of today’s markets is born out of this perfect alignment. As the wise Jeremy Grantham cautions, is this a speculative “bubble within a bubble” that is bound to implode like every other tech fad? Or is this, as Goldman Sachs analysts argue, a logical and long-lasting infrastructure expansion that is essentially distinct from the dot-com bust due to its financing by “immensely profitable, free-cash-flowing hyperscalers” as opposed to leveraged startups?
The key finding is that the market’s concentration is hyper-rational rather than irrational: if the construction of data center infrastructure generates 92% of new economic value, it stands to reason that the small number of companies engaged in that activity would account for 32% of market value and 80% of gains.
The fragility that concentration causes, rather than the concentration itself, is the real danger.
A significant slowdown in AI capital expenditure would have an immediate and significant impact on the equity market and dependent macroeconomy.
The $400 Billion Arms Race and the Prisoner’s Dilemma
An unprecedented and quickening arms race in capital expenditures is a measure of the AI infrastructure boom. By investing almost $400 billion in AI infrastructure this year alone, America’s leading tech companies are placing a huge wager. Data centers, GPUs, networking, and the ability to control the future all come at that cost. Even more concerning is the fact that this spending is increasing rather than decreasing. The top four hyperscalers’ combined capital expenditure estimates for 2026 have been raised from $370 billion to roughly $430 billion.
Company-Specific Capital Deployment
The specificity of this commitment is worth examining in detail, as each hyperscaler’s spending pattern reveals the logic of mutual escalation.
- In order to fulfill its exclusive infrastructure partnership with OpenAI, Microsoft is carrying out a $80 billion capital expenditure plan for FY2025, which is mainly focused on growing the Azure data center footprint globally. In order to publicly defend this extravagant expenditure, CEO Satya Nadella stated that he “expects to double its data center footprint over the next two years, reflecting the demand signals”. Since demand, not speculation, is driving investment, the framing is crucial. But the buildout’s size and speed point to something more aggressive.
- Alphabet presents an even clearer signal of competitive escalation. The company has revised its 2025 CapEx guidance upward no fewer than three times: from an initial $75 billion in February, to $85 billion in July, and finally to $91–$93 billion in October. This repeated upward revision, with each quarter yielding new demands, suggests not a measured response to proven demand but a defensive race against rivals. The CFO’s candid admission captured the dynamic perfectly, “We exited the year with more demand than we had available capacity”. This statement is key as it signifies not excess demand but constrained supply, forcing accelerated capital deployment to avoid ceding market share. CEO Sundar Pichai crystallised the moment by announcing Google’s first-ever $100 billion quarterly revenue driven by AI delivering “real business results”.
- Amazon has projected a striking $125 billion in cash CapEx for 2025, a substantial increase from prior guidance exceeding $100 billion. This capital is directed at expanding AWS infrastructure, developing proprietary Trainium and Inferentia accelerator chips to reduce dependence on Nvidia, and securing the future model pipeline via an $8 billion cumulative investment in AI startup Anthropic. CEO Andy Jassy’s statement on the earnings call reveals the logic clearly, “You’re going to see us continue to be very aggressive investing in capacity because we see the demand”. This is the Microsoft playbook. Both CEOs claim they are simply meeting market demand yet their spending spree is a massive bet on future growth.
- Meta has raised its 2025 CapEx guidance to $70–$72 billion and issued a stark warning that 2026 spending growth will be “notably larger”. The company is deploying this capital on massive new data center campuses and acquiring “more than one million GPUs” to support its open-source Llama model family and the newly announced “Meta Superintelligence Labs”. CEO Mark Zuckerberg explicitly described this as “aggressively front-loading infrastructure” to pursue a “massive latent opportunity”. This candid language signals capital expenditure in advance of proven returns.
The Prisoner’s Dilemma Mechanism
This pattern reveals a classic Prisoner’s Dilemma concerning capital allocation.
No CEO can risk being the first to slow down CapEx, for to do so would be to cede the “next great platform” to a more aggressive rival. This mutual fear forces all players into an escalating, collectively destructive cycle, regardless of near-term ROI visibility. Each company’s spending can be justified individually; collectively, this spending threatens to destroy the profit pool it purports to enhance.
The market’s reaction to these announcements reveals investor desperation to distinguish “rational” infrastructure spending from “speculative” bubble-building. Zuckerberg’s “front-loading” comment instantly cost Meta 13% of its market value, a clear signal that Wall Street is punishing capital deployment that outpaces genuine demand. Conversely, when Microsoft’s Nadella framed spend as following existing “demand signals”, the market rewarded the stock.
The distinction is artificial.
Both companies are spending at scales that assume the most optimistic future scenarios. The psychological comfort of attributing spending to “demand pull” versus “front-loading” cannot obscure the underlying reality that this $400 billion (rising to $430 billion) annual commitment is based on faith in an infrastructure requirement that has not yet been validated by proven use cases with clear ROI.
The “Tangled Web” of Circular Capital Flows
Forget the simplistic CapEx numbers. The real risk is the deep, interwoven financial structure of a complex knot of cross-equity, “strategic” alliances, and massive cloud credit commitments. These arrangements deliberately obscure where revenue ends and investment or quasi-debt begins. This “Tangled Web” is not just a concern for bubble theorists, it is now on the radar of serious regulators.
The Vendor-Financed Mechanism
The most intricate and profound AI partnership on the market, Microsoft-OpenAI, serves as the best example of the mechanism. Azure cloud credits played a major role in enabling Microsoft’s large OpenAI investment. This non-cash contribution made use of Microsoft’s asset with almost zero marginal cost. At a post-money valuation of $135 billion, Microsoft acquired a 27% stake in OpenAI as a result of this investment. In a renegotiated 2025 agreement, OpenAI agreed to spend an additional $250 billion on Microsoft’s Azure services.
The following is how the circular mechanism works: Microsoft uses non-cash cloud credits to “invest” in OpenAI. Now with “billions” of dollars in capital, OpenAI uses these credits to “buy” Azure services. In order to inflate Azure’s growth metrics and support its exorbitant valuation multiples, Microsoft books this “purchase” as revenue. At the same time, OpenAI, which is now “fueled” by billions of dollars in “capital”, raises a significantly higher amount of money in its next round of funding. In March 2025, it raised $40 billion at a valuation of $300 billion.
Using the same non-cash capital, this process inflates the hyperscaler’s revenue and the AI startup’s valuation. It is a closed-loop system that uses the same asset to increase the financial metrics of both businesses, giving the appearance of value creation when none has. According to one analysis, OpenAI was at or close to cost for $10 billion of Microsoft’s Azure revenue. The system creates a perpetual motion machine for valuations by acting as an advanced type of vendor financing.
Interlocking Investments and Equity Stakes
This central relationship is mirrored and extended through the ecosystem:
- Amazon–Anthropic: With its $8 billion investment in the AI startup Anthropic, Amazon has established AWS as its “primary cloud and training partner”. Crucially, the agreement calls for Anthropic to use Amazon’s exclusive Trainium and Inferentia chips, establishing a feedback loop for co-development that improves the financial metrics of both companies.
- Google–Anthropic: Google has spent more than $3 billion to acquire a 14% stake in Anthropic, highlighting the competition for compute access through a “dual cloud” strategy. Anthropic announced a historic extension of its Google Cloud agreement in October 2025, acquiring access to “up to one million TPUs” in a deal worth “tens of billions of dollars”.
- Nvidia’s Portfolio Role: Nvidia is an investor as well as a seller. The chipmaker owns stock in CoreWeave, a significant Nvidia customer and critical infrastructure firm. Additionally, Nvidia co-invests with well-known venture capital firms like Sequoia Capital, as demonstrated by its $305 million Series B investment in Together AI. The financial incentives offered by Nvidia are exactly in line. Nvidia’s profits soar along with the valuations of their clients. This produces a strong and lucrative flywheel effect.
The arrangement can be described as a “Nvidia-VC” axis, in which venture capital’s desire to inflate the valuations of AI startups coincides with the chip manufacturer’s financial incentives. As the main vendor to these startups, Nvidia benefits from both hardware sales and the equity growth of clients whose expansion is reliant on Nvidia’s hegemony.
The Depreciation Accounting Question
Michael Burry, an investor, is the most vocal opponent of this structure, accusing it of systematic accounting manipulation. Burry has a $1.1 billion short position against AI bellwethers, which includes $912 million in puts on Palantir and $187 million in put options on Nvidia. He specifically argues that hyperscalers are “understating depreciation by extending the useful life of assets” in order to inflate profits.
Accounting sleight-of-hand is the main problem.
Servers and GPUs, which typically have a two to three year lifespan, are being stretched to five, seven, or even ten years on the balance sheet. Profits in the present are inflated by this practice. According to Burry, this understates depreciation through 2028 by an astounding $176 billion. According to reports, Oracle and Meta are overstating their earnings by 20.8 and 26.9 percent, respectively, by that year.
As this hardware is replaced and accrued depreciation charges catch up, future earnings face a disastrous cliff, and current profits are illusory if Burry’s thesis is right. The high-stakes argument in the market is over. Goldman Sachs notices the cash flow right away, but Burry spots a fatal flaw. That money is being invested in obsolescence. Fake earnings today and a harsh market reckoning tomorrow are the unavoidable outcomes.
The Skeptics’ Case and the Telecom Bubble Parallel
Historically astute investors have issued strong warnings in response to the unprecedented scope and structure of AI spending, each with unique but complementary criticisms.
Jeremy Grantham’s “Gold Rush” Thesis
Jeremy Grantham, the legendary investor and co-founder of GMO, has been unequivocal in his assessment, labeling the AI frenzy a “bubble within a bubble”. A genuinely transformative technology guarantees a speculative bubble.
That is his profound core argument.
Massive overinvestment, technological disruption, and eventual collapse are guaranteed by the certainty of the brilliant underlying idea. His main example is Amazon, which became one of the most valuable companies in the world but crashed 92% after the dot-com bubble.
Despite its current $5 trillion valuation, Grantham characterises Nvidia as “the guy selling shovels at the peak” of a gold rush. The gold rush itself is a speculative excess, but the shovel industry is genuine and long-lasting. As with all manias, even the shovel sellers lose a great deal when the rush is over.
The Telecom Bubble CapEx Intensity Parallel
A fundamental change in the business model of large technology companies is the most concerning quantitative indicator in favor of the skeptics’ argument. These businesses are quickly changing from high-margin, asset-light software companies to capital-intensive, asset-heavy infrastructure firms. CapEx has historically accounted for around 12.5% of revenue for large technology companies. This ratio has averaged close to 22% in 2025. In particular, the CapEx intensity for Meta, Microsoft, and Alphabet currently varies between 21% and 35% of yearly revenue.
This is not merely elevated; it is historically alarming.
This capital intensity (21–35%) is higher than the average for global utilities and, more importantly, higher than AT&T’s at the height of the dot-com and telecom bubble in 2000. A quarter of a century later, the telecom index is still 60% below its peak in 2000. The analogy is frightening and accurate. Through accounting manipulation and increased competition, what starts out as a sensible infrastructure buildout can turn into a traditional overbuild.
The historical record is not promising.
Wall Street’s Muted But Growing Concern
Institutional skepticism is evident, even though the majority of sellside analysts are publicly bullish. 54% of institutional investors think AI stocks are currently in a bubble, according to a Bank of America survey conducted in October 2025. More importantly, in November 2025, the CEOs of Goldman Sachs and Morgan Stanley warned the public that there would be a “stock market correction,” citing overvalued stocks.
This skepticism is not relegated to minor figures.
It comes from the top-tier institutions’ leadership, people who privately have concerns they can’t express in their public research mandates.
The Power Bottleneck Emerges
The severe shortage of high-performance GPUs that characterised 2024 has significantly subsided. After peaking at about eight dollars per hour, rental rates for the H100 chip have steadied within a predictable range, dropping to roughly two dollars and eighty-five cents to three dollars and fifty cents per hour.
However, the direction of artificial intelligence infrastructure is currently determined by a far more powerful constraint. The primary barrier to the sector’s physical growth is now electric power rather than processing power.
The Data
The sheer magnitude of the power crisis is undeniable.
Data centers consumed roughly four percent of all United States electricity in 2024. Current projections indicate this demand will balloon to between seven and twelve percent of the entire national grid by 2030. This represents a staggering 175 to 300 percent increase in a mere five years. Globally, data center power requirements are forecast to jump 165 percent by 2030, necessitating an overwhelming $5.2 trillion investment specifically in AI data center infrastructure.
Regardless of financial resources or technological prowess, the power constraint places a hard cap on the expansion of AI infrastructure. No amount of engineering or money can create electricity that doesn’t exist.
This is a physical limitation rather than a financial one.
The Nadella Warning
This is more than speculation. Microsoft CEO Satya Nadella has publicly conceded that electrical power, not processor capacity, represents the fundamental constraint.
In a 2025 podcast interview, Nadella stated the core issue starkly stating, “The biggest issue…is the power and…the ability to get the builds done fast enough close to power…you may actually have a bunch of chips sitting in inventory that I can’t plug in”. This admission from the CEO of the world’s largest cloud computing company signals that the “unlimited growth” narrative of AI infrastructure buildout is hitting a hard physical constraint.
Strategic M&A outside of traditional technology is already being driven by this power crisis.
The need to secure power generation assets for data centers was a clear motivator for Constellation Energy’s $26.6 billion acquisition of Calpine, a power generation company, in 2025. Once pedestrian utilities, data center REITs such as Equinix and Digital Realty Trust have emerged as hubs for infrastructure investment capital looking to gain exposure to the energy needs of the AI buildout.
Regulatory Risk Emerging
This buildout’s opacity and concentration have caught the attention of regulators. The Microsoft-OpenAI, Amazon-Anthropic, and Google-Anthropic collaborations are being closely examined for possible anticompetitive ‘lock-in’ and preferential compute access under the FTC’s 6(b) market study, which was initiated in January 2024.
More importantly, the Financial Stability Oversight Council (FSOC) now has AI infrastructure risk on its agenda thanks to U.S. Treasury Secretary Janet Yellen. The “complexity and opacity of AI models” and the extreme “concentration” among a small number of cloud and data providers are “significant risks” that Yellen has openly warned of.
Although the expansion of AI infrastructure is sound economically, top economic policymakers are becoming increasingly concerned that it may be unintentionally creating serious risks to financial stability due to excessive industry concentration and intricate systemic interconnection. This is highlighted by the focused regulatory scrutiny.
Tesla’s Project Dojo and the Nvidia Moat
Perhaps the worst example of the unlucky course of Tesla’s Project Dojo is the structural resilience of Nvidia’s competitive position and the inherent limitations of vertical integration within this specialised sector.
The Ambition and the Failure
Tesla tried a daring vertical integration approach, leveraging decades of world-class silicon design experience. Bypassing Nvidia completely, the company created its own Full Self-Driving (FSD) chips for in-vehicle inference and introduced Project Dojo, a highly ambitious custom supercomputer built from the ground up to train its computer vision models.
After years and billions of dollars of development, the Project Dojo team was formally dissolved by Tesla in August 2025. The core engineering team’s mass departure to a startup called DensityAI, headed by silicon veteran Peter Bannon, was the cause of the failure. The project was unjustifiable due to its manufacturing complexity, technical challenges, and escalating costs.
The Strategic Pivot and Its Implications
Tesla announced a major strategic realignment after the Dojo project was decommissioned. The business will now adopt a “hybrid architecture” and significantly rely on outside technology partners, particularly AMD and Nvidia. In the end, the company that had the internal technical capability to seek vertical integration in semiconductor development decided the effort was unnecessary.
This failure is arguably the single most bullish data point for Nvidia’s long-term dominance.
It demonstrates that even top-tier engineering firms with extensive knowledge of silicon could not create a competitive alternative to the market leader. The reason is not that Nvidia has better hardware (although it may have), but rather that its decades-old, deeply ingrained CUDA software ecosystem is its real competitive advantage rather than its newest chip architecture.
The millions of developer hours and institutional knowledge that are ingrained in CUDA cannot be readily replaced, even if Google’s TPU, Amazon’s Trainium, or a future Tesla chip prove to be quicker or more effective. Even businesses actively working to lessen their reliance on Nvidia remain “captive” to it, as evidenced by Tesla’s failure and its subsequent forced switch to purchasing from Nvidia.
The $400 billion+ yearly capital expenditure aimed at Nvidia and its ecosystem is justified by this captivity.
The Three-Layer Risk Framework and Strategic Considerations
The AI investment cycle is not a monolithic “bubble” but a multi-layered phenomenon with distinct risk profiles at each layer.
1) The Application Layer (Startups)
Every sign of conventional speculative overvaluation is present in the new layer of startup businesses, particularly those using Vertical AI solutions to target specialised white-collar industries.
Venture capitalists eager to invest in the only “hot” industry value companies like Harvey (legal AI), OpenEvidence (medical AI), and dozens of others on the basis of an unproven return on investment. According to historical precedent, this layer will undergo dramatic consolidation and failure rates that surpass 80%.
2) The Infrastructure Layer (Hyperscalers and Nvidia)
The infrastructure layer is a real, extensive physical buildout rather than a speculative one. The hyperscaler compute divisions, Nvidia, and data center operators are seeing significant free cash flow and actual demand. The danger here is not that the infrastructure is unrealistic, but rather that it is a prime example of logical, game-changing technology that ensures enormous overinvestment.
This overbuild appears to already surpass the severity of the telecom bubble, according to the CapEx intensity metrics. Whether the return on this capital investment can support the deployment rate will determine whether this overbuild causes catastrophic losses. This reality is obscured by current accounting procedures.
3) The Model Layer (Foundational Models)
There is an immediate and serious risk of commoditisation for the foundational model providers, such as OpenAI and Anthropic. This difficulty arises from the spread of excellent open-source competitors like China’s DeepSeek and Meta’s Llama, as well as ongoing efficiency improvements that make smaller models more competitive. The significant investment in proprietary training infrastructure runs the risk of becoming unprofitable if the market decides that open-source AI is “good enough”.
The Strategic Implication
This three-layer distinction is crucial for investors who manage corporate capital and legacy wealth. The market is a complicated, three-dimensional problem that necessitates precise capital allocation rather than a simple “bubble or not” choice.
The application layer demands skepticism.
The infrastructure layer demands scrutiny of accounting practices and return metrics.
The model layer demands vigilance regarding open-source and efficiency risks.
At Bancara, our philosophy is grounded in providing precisely this clarity to sophisticated investors. We believe institutional-grade analysis and access to a multi-asset platform like BancaraX that provides precision tools for navigating complex market cycles is essential during periods of extreme concentration and structural opacity.
Determining the layer of your investment is the crucial difference between compounding portfolio success and suffering a material loss. It is necessary to ascertain whether the capital is supporting a legitimate but possibly saturated infrastructure buildout, pursuing purely speculative momentum, or supporting a platform that is at risk of commoditisation while simultaneously positioning itself for a winner-take-all situation.
Although the financial structure of this AI wave is extremely complex and carries substantial, multi-layered risks, it is truly revolutionary. The current state of the market is a hyper-rational focus on the one source of observable growth rather than a mistake in judgment. But when combined with self-reinforcing capital flows, opaque financial reporting, and expanding power infrastructure constraints, this intense concentration leads to systemic vulnerabilities that necessitate close monitoring and extremely critical analysis.
We encourage you to speak with your Bancara advisor to discuss how these dynamics may influence your strategic asset allocation and to assess your portfolio’s exposure to each of these three layers. A nuanced understanding of the divergence between infrastructure, platforms, and applications, and the inherent risks tied to each category, mandates a sophisticated, empirically-driven analysis. This meticulous approach defines the Bancara advantage.
