Bursting Your Own Bubble
- Ryan Bunn
- Oct 18
- 7 min read
Updated: Oct 19
Disruptive innovation is always unforeseen — These disruptions invalidate economic forecasts — Unforeseen technological advances will burst the AI bubble.
BURSTING YOUR OWN BUBBLE
Disruptive innovation is always unforeseen. These breakthroughs take the world by storm and alter life as we know it. The investment world is quick to support disruptive technology and rewards innovators with abundant capital and soaring share prices.
In eras of technological euphoria, investors extrapolate linearly or exponentially, resulting in incredible (if not preposterous) forecasts. Bubbles form when exciting discoveries combine with these forecasts to suggest massive economic potential.
But investors often forget that unforeseen breakthroughs can disrupt the same technologies they create. When bubbles burst, it is further technological progress that shatters economic models and the valuations they support.
The Fiber Infrastructure Bubble
The AI infrastructure buildout is often compared, disparagingly, to the US fiber optic network rollout from the tech bubble. At the time, the newfangled internet was a truly disruptive invention. Soaring projections for internet adoption and bandwidth needs underpinned massive investments in the US fiber optic network.
Overly optimistic projections were partly to blame. A PwC study in late 1998 forecasted internet usage doubling every 100 days.1 Although growth was exponential, global internet traffic doubled only once a year, far below the PwC forecast.2 This forecast led to telecom providers spending over $90 billion laying 39 million miles of fiber across the US.
With internet usage doubling every year, 80% of this capacity should have been utilized by 2006.3 Instead, utilization remained low. An unforeseen breakthrough was to blame for this miscalculation.
Exponential Growth Everywhere
Even if investors had correctly forecasted internet traffic, the fiber optic rollout would have still been a disaster. Investors and telecom providers both failed to consider future technological breakthroughs.
Not only was internet traffic growing exponentially, but so was fiber optic cable capacity. Wavelength-division multiplexing, a new technology, allowed for multiple signals to be sent across a single fiber. This innovation exponentially grew the capacity of every existing fiber optic cable.4
Investors failed to appreciate this trend, despite fiber optic capacity doubling annually since 1975.5 Forecasts of fiber optic cable requirements did not account for this phenomenon resulting in significant unutilized fiber optic capacity even today.
Investors are at risk when following forecasts that do not account for future innovation.
The Oil & Gas Shale Boom
Unforeseen disruption does not only occur in the tech space. Although hydraulic fracturing has existed since the 1940s, the ability to drill horizontally made the US’s vast shale reserves an economically viable resource.
The “shale boom” commenced in the early 2010s and, supported by massive amounts of capital from Wall Street, resulted in the US becoming the world’s largest oil producer.
Until the “shale bust” in 2015, drilling and hydraulic fracturing service companies saw an explosion of demand for their equipment. With oil trading at $100 per barrel there was money aplenty to pay top dollar for the drill rigs and horsepower that turned shale into gold.
This resulted in a massive investment in the space. In 2014 alone, over $200 billion was spent on acquiring, developing and drilling shale wells.6 Horizontal rig count rose from less than 400 in 2011 to over 1,000 in early 2015, just prior to the crash.7
Today there are 612 horizontal drill rigs in operation in the US. It is unlikely that over 1,000 rigs will ever be needed again, even though the US is producing more oil than ever.
Why did businesses spend billions of dollars on drill rigs that would hardly be used?
Never Ending Innovation
Drill rig efficiency, measured by the length a rig can drill, has improved by 5x since before 2010. This was a surprise to market forecasters and invalidated rosy projections for drill rig demand.
Similar advancements were made in other parts of the oil & gas extraction process. Hydraulic fracturing horsepower grew 10x between 2004 and 2019. Proppant (sand) mass used to support shale formations grew 40x over this same period.8
Today, with improvements in valve automation technology, service providers can drill and frac 12 wellbores at the same site (“zipper fracing”), offering another 10x efficiency increase over single well locations.
These technological advancements, which initially enabled shale oil extraction, served to pop the bubble they created. Notably, the potential for these unforeseen efficiencies was not captured in economic forecasts.
The Risk of Picks & Shovels
AI, like the internet and shale oil, is expected to revolutionize the world. Generative AI token demand is forecasted to grow by 115x through 2030, a CAGR of 120%, eerily similar to global internet traffic growth from the early 2000s.9
We do not doubt this forecast. Despite being value investors and bubble avoiders, we are also tech enthusiasts. AI is clearly useful and growing in value every day.
Like the fiber and drill rig rollouts though, we are wary of investing in the beloved “picks and shovels” that will underpin AI’s growth.
$750 billion is expected to be spent on data center GPUs through 2030.10 Data center power demand is expected to grow 2.5x to 2030.11 Overall global AI spend will top $2 trillion in 2026.12 These investments are being made based on AI usage projections but likely ignore, as they must, unforeseeable future advances.
Bottlenecks, Brute Force, and Disruption
Wall Street loves certainty. When industries experience bottlenecks, such as bandwidth constraints, drill rig availability, or data center capacity, investors are ready to provide capital to address these impediments head-on.
The brute force approach is preferred because of its certainty. Doing more of the same is certain to solve the problem: providing capital to lay another fiber optic cable directly addressed the bandwidth bottleneck; another drill rig was certain to drill another well.
AI is currently bottlenecked by computing power, data center capacity, and energy. The immediate market response is to demand more GPUs, data center facilities, and power plants. Companies that address these bottlenecks have already benefitted.
But this is the myopic approach taken during all bubbles. These projections ignore the innovation that is creating the opportunity in the first place. Ultimately, the brute force approach will expose investors to the risk of future technological disruption.
Unforeseen Technological Disruption: Moore’s Law
While investors will claim disruption is unforeseen, in reality, like fiber optic capacity and rig laterals, future tech advancement is with us today. Although no disruption can be predicted with certainty, there are many breakthroughs that could result in overcapacity in the AI market.
We are all familiar with Moore’s Law, the principle that computational power tends to double every two years. This principle is alive and well with Nvidia’s GPUs. The next generation chip will offer 15x the compute at only 2x the power. Forecasters are incorporating this trend into their projections, but what if the trend accelerates?13
Unforeseen Technological Disruption: Hardware Breakthroughs
In addition to Moore’s law, innovation continues outside of the GPU. A new type of memory called a “gain cell” can address cache memory bottlenecks. This enhancement can increase speed and reduce energy consumption by 2-5x.14 Whether or not this advancement can be widely deployed, it highlights the potential for hardware efficiency gains beyond Moore’s Law.
Or how about skipping the GPU altogether? MIT has developed a framework that allows your CPU to locally run an inference model at human reading speed.15 A convenient side effect is a 55% reduction in energy consumption and, by running locally, no payment or subscription required.
Unforeseen Technological Disruption: Software Breakthroughs
While AI investment is currently focused on enabling infrastructure, innovation can come from anywhere. The software running inside data centers holds promise for unforeseen breakthroughs.
The release of DeepSeek-R1 in January 2025 highlighted the potential from software enhancements. While details are sketchy, it is estimated that this model required only $1.6 billion in hardware costs to train.16 DeepSeek prompted the market, briefly, to question the hundreds of billions of capex being spent to train models that were cheaply replicated by DeepSeek.
Similarly, while AI leaders spend frantically to achieve artificial general intelligence (“AGI”), which promises a single, massive model to solve problems across domains, advances are being made in domain-specific models.
Retrieval-augmented generation (“RAG”) allows users to customize a model with their own data (I’ve tried it). This more focused model can often outperform general models such as ChatGPT. RAG can be used to enhance the output of free, locally-run, open-sourced models, eliminating the need to pay for tokens.
Unforeseen Technological Disruption: Power Efficiency
Finally, the power efficiency of chips and data centers tends to improve over time. If historical trends continue, there is risk of overbuilding power generation.
In their power demand forecast to 2030, Goldman Sachs assumes gains of 2-3% per year in data center power efficiency. But this is a sharp deceleration from the >10% gains generated from 2015—2020. An unforeseen breakthrough could allow for continued double-digit improvements.
While forecasters will brush off the inaccuracy of their predictions, blaming “unforeseen circumstances,” prudent investors, by learning from past bubbles, can protect their capital from this knowable risk.
Bursting Your Own Bubble
We are optimistic about the future of AI. We are even more optimistic about technology innovation across chips, training models, and power generation. But it is exactly this optimism that makes us skeptical of most economic forecasts regarding artificial intelligence.
AI should soon be smart enough to address its own bottlenecks with innovation. After all, even the human mind came up with enough innovation to end the fiber optic and shale booms. If AI is as powerful as the market claims it to be, surely it will burst its own bubble.
Sources:
2 Cisco
9 Forbes
12 CRN
15 BitNet
