“Fear and greed” notoriously are drivers of equity market sentiment and that is clear in the yo-yo behavior surrounding artificial intelligence equities recently. The fear is that AI investment levels are a bubble, overinvestment that will ultimately not pay off.
The greed flows from the belief that AI is a transformative new technology that will create new winners and losers in the broader economy.
A likely third position is that AI is not a bubble on the order of the do-com mania, but will produce excess investment that has to be rationalized, eventually, as all great new technology waves have done so.
Optimists might agree with Mara Holdings CEO Frederick Thiel that “ electrons are the new oil,” referring to the idea that computational resources underlie the ability to wring value from AI, while the data centers that provide the computation now are dependent on access to large and affordable amounts of electricity.
Mara believes future winners will be high-performance compute providers who have the lowest costs to produce insight per token; insight per kilowatt of power consumed, especially for enterprise private compute operations.
As Thiel puts it, his firm, which originally ran bitcoin mining operations, now provides a high-performance computing infrastructure that converts energy into both value (bitcoins) and intelligence (AI computing).
The broader vision for the company, as is true for many other former bitcoin miners, is "transforming energy into intelligence.” In other words, consuming electricity to power AI models and the inferences to be drawn from using those models.
The analogy is not unlike that sometimes made to the export of alfalfa from the U.S. great plains to the Middle East. The production of the alfalfa consumes water, which becomes livestock food, which essentially also represents the value of the water consumed to grow the produce. So exporting alfalfa also is akin to exporting the water used to grow it.
“We believe energy, not compute, really becomes the primary constraint on AI growth,” says Thiel.
Pursuant to that belief, Mara has a venture with MPLX, formed by Marathon Petroleum Corporation, the largest petroleum refinery operator in the United States, to develop and operate multiple integrated power generation facilities and state-of-the-art data center campuses in West Texas.
MPLX will provide long-term access to lower-cost natural gas at scale, while Mara will develop and operate on-site power generation and compute infrastructure.
The initial capacity is expected to reach 400 megawatts with the option to expand to up to 1.5 gigawatts across three plant sites.
But Mara also is basing its business on “inference” rather than model training, as that allows it to use application specific integrated circuits (ASICs) rather than graphics processor units (GPUs), thus lowering its capital investment.
That approach also enables use of smaller data centers and air cooling rather than the more-expensive liquid cooling. The strategy is not especially new, as others in the data center and connectivity spaces have chosen to become specialists in smaller markets (either in terms of geography or types of customers).
But all that happens within the context of a market that is volatile.
A positive development such as a new chip announcement, a major partnership like the AWS/OpenAI compute services deal, or strong earnings from an AI leader pushes the market into "extreme greed" territory, driving up prices quickly.
But then reports of high AI capital expenditure, delayed profitability for end-users, or a general sentiment survey warning of a "bubble" causes profit-taking and selling, plunging the market into "fear" sentiment, leading to sharp, temporary pullbacks.
Month | Major Event | Sentiment | Notable Impact |
2025-01 | DeepSeek Launch | Fear | Sharp drop, infrastructure risk flagged |
2025-04 | Trump Tariffs Threat | Fear | Market volatility spiked, quick rebound after walkback |
2025-09 | NVIDIA-OpenAI Chip Deal, Fed Rate Cut | Greed | Strong surge, positive sentiment returned |
2025-10 | Bubble Talk Surge | Fear | Renewed caution, market exhaustion warnings |
The cycle resets because the fundamental belief in AI's future remains generally strong. Investors who sold out of fear often rush back in for fear of missing the next leg up (greed), making the dips short-lived and creating the current high-volatility, upward-trending cycle.
But skepticism and hope continue to coexist and oscillate.
Beyond the volatility, we might argue that “high-performance computing capability” has become a strategic commodity.
High-performance compute capacity arguably has become the single most critical, scarce, and expensive strategic resource in the AI industry.
If so, long-term, multi-billion-dollar compute contracts are now a competitive necessity, resembling procurement models for essential commodities like energy or raw materials. But volatility will persist until some future time when there is much more predictability about AI investments and revenue gains.
So nobody knows yet whether the investment boom in artificial intelligence we now see is a bubble, or not. Much conventional wisdom seems to suggest AI is a bubble, but there is disagreement.
And if some argue it is a bubble, there remains an argument that there is a significant difference between a dot-com style bubble and an “ordinary” investment bubble associated with introduction of any major new technology.
To be sure, for some of us, there are hints to parallels of excesses akin to the excessive dot-com investment at the turn of the century. As I was writing one startup business plan, I was told “there’s plenty of money, make it bigger.”
As it turned out, “this time is different” and admonitions that some of us “did not get it” were wrong. Economics was not different and normal business logic was not suspended.
But some might note that there are important differences between AI investment and dot-com startup investment. Back then, many bets were placed on small firms with no actual revenue.
Today, it is the cash flow rich, profitable hyperscalers that dominate much of the activity. Investment burdens are real, but so are immense cash flows and profits to support that investment.
And by some financial metrics, valuations do not seem as stretched as they were in the dot-com era, though everyone agrees equity market valuations are high, at the moment.
We also can’t tell yet what impact artificial intelligence might have on productivity and economic growth, much less future revenues for industries and firms.
And that might be crucial to the argument that there actually is not an investment bubble; that there are real financial and economic upsides to be reaped; new products and industries to be created.
There is some thinking by economists that AI impact could be greater than electricity and at least as important and positive as information technology in general.
General-Purpose Technology | Primary Timeframe of Peak Impact | Estimated Annual Productivity Boost (Peak Rate) | Macro-Level Impact Metric |
Steam Engine | Mid-19th Century (Decades after invention) | 0.2% - 0.3% | Contribution to annual TFP* or Labor Productivity Growth |
Electrification | 1920s - 1940s (30+ years after initial adoption) | ~0.4% - 0.5% | Contribution to annual TFP or Labor Productivity Growth |
Information Technology (IT) / Computers | Mid-1990s - Early 2000s | ~1.0% - 1.5% | Acceleration in annual Labor Productivity Growth (U.S.) |
Artificial Intelligence (AI) (Current Forecasts) | Early 2030s (7–15 years after GenAI breakthrough) | 1.0% - 1.5% | Projected increase in annual Labor Productivity Growth over 10 years |
Study/Source | Projection Focus | Estimated Gain (Over Baseline) | Caveats |
Goldman Sachs (2023) | Macroeconomic Forecast (Global/U.S.) | 7% increase in Global GDP over 10 years; 1.5 ppt annual U.S. labor productivity growth | Highly optimistic, assuming rapid adoption and task automation. |
McKinsey Global Institute (2023) | Economic Potential of Generative AI | $2.6 to $4.4 Trillion added annually to the global economy. | Based on value from 63 specific use cases across business functions. |
Acemoglu (MIT) | Conservative Macroeconomic Model | 0.7% increase in TFP over 10 years (U.S. economy). | More modest, based on historical adoption rates and cost-benefit analysis of task automation. |
Brynjolfsson et al. (Micro Studies) | Firm/Task-Level Productivity | 10% - 40% increase in productivity for tasks like coding, customer service, and professional writing. | These are early, firm-level gains, which historically take time to translate into aggregate macro statistics. |
Each of us has to make a call: bubble or not; big bubble or only “normal” overinvestment?
