A year ago, I became intrigued by the parallels between the current AI infrastructure build-out and the dot-com period. I asked, "Is the AI bubble bursting?" Much has happened since then, and I thought it appropriate to revisit the subject. I addressed this topic while recently speaking to a large group of public-sector retirement-system sponsors and service providers. What follows builds on my comments to that audience.
In summary, I highlighted the importance of balancing the two main elements of their fiduciary duty: driving returns and managing portfolio risk with an eye on the long-term health of the plan. With that in mind, one cannot miss the opportunity to participate in the growth driven by AI. However, there are signs of stress in the system.
AI is already a major driver of growth in the U.S. economy. In the first half of 2025, AI-related capital expenditures contributed 1.1 percentage points to U.S. GDP growth, outpacing the consumer as the main economic growth driver. The investment in AI has only increased since then. This concentrated growth can, at the same time, be injecting fragility into the economy. In the meantime, we are seeing the signs of a profound transformation of labor markets.
Large sums of capital are going into building infrastructure for the industry, from data center construction to semiconductors to power generation. There is the promise of productivity increases, as seen during previous technology-driven change. AI can raise firms' output while increasing their margins. The sharp increase in tech firms' valuations and their positive impact on the stock market affects household wealth. This is concentrated in a small number of players, and a repricing would damage investor confidence.
The credit and financing side of the boom may be its Achilles' heel. The Wall Street Journal identified about $600 billion in on-balance-sheet infrastructure commitments by Alphabet, Amazon, Meta and Microsoft as reported in their most recent quarterly filings. The same report listed $2.4 trillion in off-balance-sheet purchase commitments and not-started leases. The $600 billion in traditional capital expenditures (CapEx) is the tip of the iceberg. In parallel, complex structures are being built using private equity and credit, lacking the transparency of public markets.
While investments grow and adoption climbs, the impact on labor markets is not as clear. Earlier general-purpose technologies, such as electricity or the internet, have been associated with net new job creation. This time, there's a lot of discussion about the impact on entry-level and white-collar jobs. The jury is still out, and one can see academic papers and studies arguing both sides. It may be too early to tell, but the change is happening faster than before, when society had time to adapt institutions and policies, a luxury we may not be able to afford this time.
The surge in AI investments can, at the same time, strengthen near-term growth while increasing financial fragility.
The Dutch tulip mania of the 1630s is one of the first recorded speculative financial bubbles, and there are repeating patterns in this long history. When driven by a new technology, the cycle starts when the new technology appears, displacing what existed before. This shock changes expectations of future returns and attracts capital faster than the ability to prove returns. The boom leads to euphoria when narratives replace evidence. Over time, signs of distress appear, smart money starts to take profits, financing tightens and the weakest links break. This could lead to a panic. Eventually, prices can reset sharply to below the original levels.
The cycle of boom-and-bust repeats itself: a plausible new vision is articulated, abundant capital obfuscates weak business models and, after a reset, useful infrastructure remains. Most of the time the value of the underlying technology persists. One can look at the dot-com era for reference. Telecommunications companies made massive investments to build the fiber-optic infrastructure to "wire" the economy. This led to overcapacity which, coupled with financial fraud and questionable business models, brought down several large players. Using 1995 as a base of 100, the Nasdaq reached 505 in March 2000 and then fell to 111 by October 2002. Despite the turmoil, that infrastructure is still being used today, as it enabled the world of e-commerce. But it took more than a decade for the returns on the original investment to become visible.
There is an open question as to whether today's market will have the patience to wait for AI infrastructure returns and which market players have the financial wherewithal to weather volatility and sustain prolonged losses. Technology adoption and value can be real while the market drives overly aggressive valuations. The financing of the cycle can provide clues to where the fragility and weakness reside in the system. Look for where vendor financing or circular deals can be masking weak revenue streams.
Those with the financial wherewithal to survive the period can define the next phase of economic growth, as it happened with Amazon during the early 2000s. Founded in 1994, Amazon prioritized growth and sustained losses for years before reporting its first quarterly profit in late 2001, a strategy many questioned at the time. Today, Amazon dominates not only e-commerce but also cloud computing and is making significant inroads into AI.
Turning back to today, one can see four sources of financial risk in the current AI cycle. The first is concentration. The Magnificent Seven account for more than one-third of the S&P 500's value; at the height of the dot-com boom, the leading technology companies represented approximately 15%.
Second, CapEx investments are being made ahead of monetization. The math doesn't close with trillions going into data centers. For comparison, Anthropic reported an annualized revenue run rate of $65 billion in July of 2026. While sustaining incredible growth, the gap between investment and profit is material. The impact is also seen in free cash flow statements, with Alphabet (Google's parent company) reporting its first quarter of negative free cash flow since going public.
A third source of financial risk is the opacity in how deals are done in private markets and the circular nature of some of these investments. A web of circular investments can be traced back to two main participants: Nvidia and OpenAI. A hiccup in the structures affecting one of these two can trigger contagion with the rest of the industry.
Finally, there is timing pressure to exit private markets. Notably, the frontier AI labs, Anthropic and OpenAI, will need the depth of public markets to continue to fuel their growth ambitions. The first of what can become a wave of "giga-IPOs" happened when SpaceX, which had earlier merged with Elon Musk's AI venture (xAI), went public in June. Public markets tend to be less forgiving of unprofitable business models. During 2026, private investors committed nearly a quarter-trillion dollars to OpenAI, Anthropic and xAI/SpaceX. As they become public, their valuations, revenues, cash burn and contractual obligations will be subject to greater scrutiny.
To keep the investment and growth flywheel turning, one needs demand for the technology. So far, this is a good news story: AI is diffusing and adoption is fast. According to the Pew Research Center, 49% of American adults have used AI chatbots. At work, Gallup estimates that 30% of American employees use AI several times a week and the Office of Management and Budget inventoried 3,611 AI use cases across 56 federal agencies, doubling the number from the year before. Adoption is happening at scale and faster than with previous technologies. A National Bureau of Economic Research study found faster adoption of generative AI than what happened during the early introduction of either personal computers or the internet.
These early trends support the need for data centers to satisfy the growing demand. There is logic to the investment thesis, and one cannot dismiss the optionality created by a strong consumer market, a maturing enterprise demand and the consumption backstop generated by government demand.
At the same time, there are three limits that we need to pay attention to: an economic limit, a physical limit and an institutional limit. I previously addressed the economic limit when talking about the tension around business models. A lot of red flags should go up when one sees investments in the trillions and revenues in the billions coupled with negative cash flows. Investors need to pay attention and monitor the window of tolerance of public markets.
There is also a physical limit. For as much as we talk about AI in the abstract, it is delivered by data centers, which require land, energy and physical labor during their construction. There is a physical capacity associated with that that limits how fast AI can grow. As an example, the International Energy Agency expects data centers to account for nearly half of the growth in U.S. electricity usage through 2030. If continued demand growth is needed to yield a return for capital investments, and if one cannot build fast enough to satisfy the demand, then the equation does not close.
Finally, there is an institutional or moral limit associated with society's perceptions of AI. During 2026, the pressure on large AI companies and policymakers has increased. Until recently, when one talked about AI risks, one referred primarily to existential risks, such as whether AIs could create a bioweapon that would exterminate humankind. These are esoteric and easy to discount as very remote. Many accept trading personal privacy for convenience or do not appreciate the extent to which personal information is being used. Recently, the discussion shifted to more concrete issues: "Is AI going to take my job? Are my kids going to graduate without a job? How will a data center project impact my community? Is my electricity bill going up?"
People vote on kitchen-table issues, which are harder for policymakers and elected officials to ignore. This may finally trigger regulatory action in the U.S. It could be bad news if it slows down progress. It could be good news if it establishes an environment of trust in which progress can continue.
It is like when riding a bicycle. If you stop pedaling, you are going to fall. On the consumer side, we have massive adoption, but little monetization. OpenAI, the leader in consumer AI, has more than 90% of its users on the free tier. Its ability to convert free users into paid users or create new revenue streams, such as advertising, or enter new businesses like hardware will determine success in this space.
Enterprise demand continues to accelerate. The segment drives higher-revenue transactions, integration into business workflows creates high switching costs, and enterprise consumption can be more predictable, making it easier for the AI supplier to manage infrastructure costs. However, enterprise implementations–other than the most demanding clients–can be satisfied with "good enough" technology, currently implemented by lower-cost open-source models. CFOs are more conscious about return on investment and disciplined when considering infrastructure investments.
Enterprises want predictability and an environment in which they can operate with an understanding of liabilities they may be exposed to. They welcome some level of governance as a platform for trust to sustain growth.
Investors want good returns, and AI-related public stocks have appreciated sharply, as did private AI company valuations. As of the market close on August 4, 2026, the Nasdaq was up 38% from June 2025, Nvidia 54% and Alphabet 123%. Meta was an outlier in the tech trade, declining 12% during this period. Meta is considered a laggard in AI and is a defendant in increasing litigation relative to its social platform businesses.
Private market valuations have soared, and we could have two or three multitrillion-dollar companies entering public markets. SpaceX IPO was priced at $1.77 trillion in June, while markets speculate that Anthropic could seek a record $2 trillion valuation for an IPO later in the year. OpenAI's March 2026 funding round valued it at $852 billion.
These are larger than any IPO valuations in history. One should create a measured level of exposure to AI to participate in the market gains while keeping a close eye on real risks that are increasingly visible.
Speaking at the Public Pension Funding Forum of the National Conference on Public Employee Retirement Systems (NCPERS), I felt the weight of the more than 650 public-sector retirement systems represented by the organization. Collectively, they serve more than 20 million teachers, firefighters, police officers, municipal workers and public servants. I used averages that don't perfectly reflect their portfolios, but I endeavored to show how they are already exposed to AI. State and local defined-benefit public pension systems held $6.5 trillion in assets in 2025 with an average 42% equity allocation. The S&P 500 can be used as a rough proxy for the market, and, as of July 2026, 37% of its weight was in information technology stocks. Using four illustrative stocks and their shares of the S&P (Nvidia 7.6%, Alphabet 5.8%, Meta 2.2% and Micron 1.6%), if a hypothetical plan invested its entire 42% public-equity allocation in the S&P 500, those four stocks would represent approximately 7% of total plan assets. The conclusion is that you are already exposed to AI even if you are not trying to seek alpha with a particular AI allocation strategy.
Similar, and hopefully more detailed, estimates can be calculated by individual investors to measure exposure in their portfolios.
Investors are already participating in the upside created by the AI boom. On the flip side, they are also exposed to downside, even if they have not proactively invested in AI. To responsibly balance investment returns with risk mitigation, I recommend:
Measure your level of exposure. Instead of using averages, assess your portfolio mix, calculate your real level of exposure and be intentional about it. There's nothing wrong with an aggressive or a conservative strategy if you execute it with clarity.
Monitor capital flows in the industry and whether they translate into revenue and profits. Pay special attention to data center utilization, but do not limit your analysis to this. Follow the energy sector and the semiconductor industry.
Monitor public-market liquidity and whether markets can absorb large AI IPOs while maintaining their tolerance for unproven business models.
Identify sources of circular financing and flag them as risks. Understand who is connected to whom and identify the weakest links. Pay special attention to private credit markets and off-balance sheet debt.
Keep an eye on developments on the policy front. Good policy can create an environment of trust and propel growth while mitigating risks. Alternatively, poorly designed policy could constrain supply or demand, accelerating a market correction.
Be intentional about capturing market upside while looking for signs of a potential bubble bursting. I believe in the value of this technology, and it is sufficiently embedded in our economy to be considered "too big to fail." The AI bubble may not burst, but it may deflate and, in the process, sort winners from losers.
This article was originally published on Forbes.com