AI’s Great Infrastructure Boom: Bullwhip or Building the Future?

I tend to avoid writing about topics that others are already discussing. So I usually avoid the “AI is going to kill us”/”The AI is stealing our jobs”/ “The AI is not as good as I thought” articles.

However, when the topic of AI intersects with supply chain management and the Bullwhip effect, I have to write about it.

I am not making up the rules here.

One thing is clear: The artificial intelligence boom has sparked one of the most costly technology buildouts in history.

Hyperscale cloud companies and chip manufacturers are pouring capital into data centers, GPUs, and power infrastructure at unprecedented rates. By 2028, total investment in AI-centric chips, servers, and facilities could approach $3 trillion, a spree likened to the most significant industrial booms of the past. Microsoft, Google, Amazon, and Meta alone spent on the order of $100–$200 billion in just six months on AI hardware and data centers, a record pace of capital expenditure.

This massive scale-up is giving tech giants even more centralized control over AI development, prompting observers to ask whether we are witnessing a classic supply-chain bullwhip effect or simply laying the groundwork for a new era of computing.

The question is whether the rush to build AI infrastructure is a transient overreaction that will overshoot actual demand, or a rational response to a secular shift with decades of growth ahead?

Below, I delve into evidence for both interpretations, drawing on research, and examine what each scenario could mean for prices, capacity, investors, and public infrastructure.

The Bullwhip Effect and the AI Supply Chain

The bullwhip effect describes how small demand shocks amplify upstream due to long lead times, poor information sharing, and behavioral over-ordering. It’s the classic pattern where orders become more volatile than demand, often illustrated by the beer game.

Applied to AI, the bullwhip framework provides a lens for understanding today’s expansion of chip production and data centers. In late 2022, OpenAI’s ChatGPT burst onto the scene, demonstrating the power of large language models to millions of users. ChatGPT’s overnight popularity created an abrupt and unprecedented demand spike, a textbook example of the bullwhip effect.

To understand the AI supply chain clearly, it helps to separate it into three bottleneck layers:

1. Chips (centralized, capex-heavy, slow to expand): Cutting-edge GPUs depend on TSMC, EUV tools, and advanced packaging, all with >6–12 month cycle times. Orders placed in early 2023 were not delivered until 2024 or 2025.

2. Data Centers (deploying hardware into usable compute): Even when GPUs are available, data hall buildout, cooling, networking, and server integration lag. The construction cycle for hyperscale facilities typically ranges from 12 to 24 months.

3. Power (local, slow-moving, and often binding): Unlike chips and server racks, grid capacity is local, regulated, and can take 3–5+ years to expand. Several U.S. data center hubs (Northern Virginia, Georgia, and Ohio) are already hitting grid constraints, delaying deployment.

These three layers rarely scale in sync, and whichever has the longest lead time becomes the real bottleneck. This mismatch amplifies any bullwhip effect because each layer responds to the same demand spike on different time scales.

Whispers of Whiplash: Signs of an AI Bullwhip Cycle

Several signs suggest the current AI buildout is following a bullwhip-like boom–bust pattern. First is the initial demand surge.

Fransoo et al. (2025) make the case for the Bullwhip.

Following ChatGPT’s breakthrough, orders for AI hardware surged. By early 2023, data-center operators and cloud platforms were scrambling to secure Nvidia’s latest GPUs – effectively buying every high-end chip they could get. Anecdotal reports indicate “extreme increases in orders for Nvidia [and competitors’] products” far above business-as-usual levels, driven by widespread shortages and allocation limits.

This is a classic case of over-ordering in response to perceived scarcity. Big players with deep pockets placed bulk orders, in some cases to build technological moats that keep rivals out. At the same time, smaller AI startups found it hard to find top-tier chips, often having to rent cloud GPU time instead of purchasing hardware. Such shortage-induced hoarding amplifies the upstream demand signal beyond the true end-need – a hallmark of bullwhip behavior.

Secondly, long lead times in semiconductor fabrication and data center construction virtually guaranteed a lagged supply response. Manufacturing cutting-edge AI chips is not a quick process: even under ideal conditions, GPU production lead times exceed six months, and critical equipment, such as extreme ultraviolet (EUV) lithography machines or advanced chip packaging tools, can take well over a year to acquire and deploy. Building out massive data centers and connecting them to the grid is still a slow process – in many regions, obtaining a new high-voltage grid connection can take four years or more. These delays mean the huge orders placed in 2023 only translated into available chips and installed servers in 2024, with some deliveries pushed into 2025.

In the interim, customers queued up on waitlists, and backlogs grew, reinforcing the perception of scarcity and prompting even more aggressive orders. It’s the classic bullwhip feedback loop: demand spike → shortages → over-ordering → delayed supply → catch-up shipments.

By mid-2024, the supply chain’s response finally began arriving. Fabs ramped output, shipments of AI accelerators surged, and new data centers started coming online. The modeling by Fransoo et al. (2025) suggests that this sequence produced a pronounced bullwhip oscillation. In their simulation (depicted in the figure below), orders for AI chips (solid line) and actual chip deliveries (dashed line) swing dramatically out of phase, mirroring the boom-bust dynamics seen in other industries.

Speculative overbuilding is another classic symptom. History is replete with examples of industries overshooting capacity during hype cycles (think of fiber-optic cable in the dot-com boom, or more recently, semiconductor fabs during the cryptocurrency mining frenzy). There are indications that AI may be entering a similar phase.

Quoted in the articles above, analysts at Morgan Stanley warn that tech firms have taken on substantial debt to fund their expansion of AI, and they flag the risk of a bubble. The sheer dollar figures are eye-popping – at nearly $3 trillion of projected investment by 2028, the AI infrastructure boom would rival the railroad build-out of the 1800s or the space race-era highway system in scale. Timeline estimates suggest that mid-decade (2025–2026) may be a potential inflection point, when supply catches up to the inflated order rate and potentially surpasses actual consumption.

Fransoo et al. note that by 2025, periods of overcapacity and unstable pricing are likely to emerge as the bullwhip unwinds. In other words, the industry could experience a sudden surge in demand for AI chips and data center space, leading to a rapid decline in prices and utilization.

One critical constraint driving this cycle is the electrical grid. Unlike chips (which can be produced centrally and shipped globally), power infrastructure is inherently local and slower to scale. Major U.S. AI hubs are already hitting grid constraints, with some regions projecting severe strain.

These bottlenecks have real consequences: if a data hall is built but must wait months or years for a sufficient grid hookup, operators may delay or cancel further server and chip orders, creating ripple effects upstream.

Jan Fransoo and co-authors’ analysis explicitly warns that mismatches between compute deployment and energy availability can trigger “unexpected dynamics”. For example, a shortage of grid capacity could force customers to defer turning on new GPU racks, which in turn might lead chip suppliers to face sudden cancellations of orders. This coupling of two out-of-sync supply chains (technology vs. infrastructure) can amplify volatility. Over-investment in one area (say, too many AI servers for a constrained grid) can produce logjams that then reverberate back as lulls or cutbacks in the other regions.

All of this aligns with bullwhip logic: long lead times and multi-stage dependencies result in cyclical over- and undershoot.

Not Just Hype: Why It May Not Be a Bullwhip

On the other hand, there is a compelling argument that the AI infrastructure boom, while cyclical in the short term, is driven by genuinely transformative, long-term demand – more akin to the beginning of a multi-decade platform shift than a mere inventory overreaction. Proponents of this view argue that even if we see some temporary gluts, they will be absorbed by continuous growth in AI adoption. Several points bolster this optimistic case.

Fundamental demand growth for AI compute appears robust and nowhere near saturation.

The world’s appetite for intelligent applications, from chatbots and image generators to AI copilots in software, medicine, and science, is only increasing. Today’s models still run at relatively limited scales for most users; tomorrow’s are expected to handle far more complex tasks (real-time video generation, advanced personal assistants, etc.), which will require orders of magnitude more computation.

As the New Yorker reported, each user query to an AI model expends significant energy: writing a short essay with an AI can consume a few minutes of microwave-level electricity per query. Multiply such use cases across billions of daily tasks, AI becoming as ubiquitous as electricity or the internet itself, and it’s easy to see demand outstripping today’s capacity many times over. One Microsoft data-center executive went so far as to say he’s more worried that “we are underbuilding rather than overbuilding” for future AI needs. In other words, from this perspective, the current build-out may feel frenzied, but could actually prove insufficient a few years down the line as AI becomes a mainstream utility.

Forecasts for data center growth reflect this long-term trajectory. A 2024 McKinsey analysis (cited by Fransoo et al.) projected that supporting widespread AI adoption could require an additional 130–240 gigawatts of data center capacity by 2030. That represents a roughly 5-fold increase in global compute power from 2023 levels (around 55 GW) to the high-end scenario of ~293 GW by 2030. Such exponential growth curves imply that even if there are periods of overshoot, the overall trendline points sharply upward. Any excess GPUs or servers purchased now might find use a year or two later as new AI applications come online. This is analogous to early internet infrastructure: investors laid vast fiber networks in the 1990s that went underutilized for several years after the dot-com bust. However, a decade later, much of that capacity became essential for handling streaming, social media, and cloud computing. From a macro perspective, AI appears to be a similar, general-purpose technology with a long runway of adoption, rather than a fad that peaks and fades.

Another factor mitigating the bullwhip is the nature of the investors and firms driving the boom. Unlike many past bubbles, which were fueled by speculative startups or highly leveraged new entrants, the AI infrastructure surge is primarily funded by cash-rich tech incumbents. The likes of Google, Microsoft, Amazon, and Meta are among the wealthiest companies in history, with enormous free cash flows from their current businesses. They are deploying these resources to secure leadership in the next era of computing.

This means two things: (1) They can afford a few years of low returns on these investments if needed – they won’t go bust just because some data halls sit half-full for a quarter or two. (2) Their spending is at least partly strategic, not purely reactive.

These giants are positioning themselves as the indispensable landlords and toll collectors of the AI age, much as they already are in cloud computing. Even if demand growth pauses, they will continue to invest to widen their lead. For example, hyperscalers have been known to build out data centers in anticipation of future regions gaining connectivity or demand. This strategic long view tempers the bullwhip effect: the build-out isn’t entirely based on myopic forecast overshooting; it’s also based on a rational expectation that AI will become ubiquitous. The “Magnificent 7” tech firms’ record capex can thus be seen as building durable assets for a new computing platform, not simply piling up excess inventory.

Additionally, high barriers to entry in AI infrastructure may prevent the kind of uncontrolled capacity glut that is often seen in more fragmented industries. Constructing leading-edge chip fabs or massive AI training centers is extraordinarily expensive and complex, effectively limiting it to a handful of players. This oligopolistic structure can impose a form of self-discipline. Suppose only a few companies (who all know each other’s capabilities) are adding capacity. In that case, there is a chance for some coordination or at least mutual awareness that can avoid the most extreme overshoot. It’s not a perfect guarantee – indeed, the big competitors are all racing simultaneously – but it’s a far cry from, say, dozens of telecom firms laying overlapping fiber networks in the 1990s.

Moreover, the fact that state-of-the-art AI hardware has remained mainly in the hands of the top firms means demand is being aggregated into those firms’ cloud platforms. They can dynamically allocate and share that capacity among many customers, achieving higher utilization rates than if each company built its own siloed AI cluster. In essence, the cloud model can reduce waste: excess capacity can be rented out or repurposed for others on the platform, thereby cushioning the impact of any one user overestimating their needs. Smaller organizations no longer need to over-invest in their own infrastructure (they can scale to the cloud), which could reduce the overall bullwhip effect in the industry.

Like fiber-optic overbuild in the 1990s, early AI infrastructure may look excessive today but become essential later. Even within the next 5–10 years, there’s reason to believe the supply-demand gaps will equilibrate. The research by Fransoo et al. anticipates a “bumpy road” for 3–4 years due to bullwhips and pork cycles, but then a sustained growth trend after 2030 as the ecosystem matures.

By that time, improvements in chip efficiency and new GPU generations will have arrived, and data center operators may have adopted more flexible strategies (the authors even muse that cloud providers might evolve to resemble contract manufacturers like Foxconn, redeploying hardware to where it’s needed most).

Implications: Will the Boom End in a Whiplash Bust?

If the current AI infrastructure boom is indeed a classic bullwhip overshoot, there are several implications to watch for in the coming years.

Prices for AI hardware and cloud computing services are likely to undergo a rollercoaster ride. In the short term, intense demand has driven up costs – for instance, cutting-edge Nvidia GPU prices skyrocketed in 2023 amid shortages. But once the backlog of orders gets fulfilled and new supply floods the market, we could see a glut of chips leading to sharp price declines.

This would mirror what happens in memory chips or disk drives after boom periods: prices crash, and suppliers’ margins evaporate until the excess inventory is worked off. The research chapter by Fransoo et al. explicitly warns of “periods of overcapacity and unstable pricing for multiple years” due to bullwhip effects. Cloud compute costs (for AI workloads) could similarly fluctuate – today, some AI models are costly to run due to limited compute resources. Still, tomorrow we might see cloud providers slashing prices or offering steep discounts on idle GPU clusters after overbuilding.

Periods of overcapacity could leave expensive GPU clusters underutilized, pressuring margins until demand catches up. If a bust follows the boom, we may see some investors realize that projected AI revenues (or cost savings) won’t materialize as quickly as hoped. Stock market enthusiasm could cool, and companies that overextended (especially smaller AI firms that took on debt for hardware) might face financial distress. On the other hand, a period of oversupply could benefit AI practitioners and research labs by offering lower costs – a buyer’s market for compute, where startups can acquire top-notch chips at a lower price (or acquire distressed data center assets). In fact, an overshoot in hardware could help democratize access to AI capabilities that were previously the exclusive domain of tech giants, leveling the playing field somewhat once the dust settles.

For the public infrastructure sector, a bullwhip outcome carries its own lessons.

Power utilities and grid planners that rushed to accommodate the AI data center boom might suddenly find growth stalling. Regions that invested in new power plants or transmission upgrades to serve data centers could end up with overcapacity in the electrical grid if some of those data centers don’t fully materialize or operate at peak load. This isn’t entirely negative – a respite could give grids time to stabilize and incorporate more renewable sources (many AI operators contract renewable energy, so overshoot might temporarily reduce strain on fossil fuel plants).

However, utilities recovering costs on large projects might raise rates for other customers if the anticipated AI demand doesn’t fully cover the investments. Grid planners may find themselves with temporary overcapacity if expected demand doesn’t materialize.

On the other hand, if the boom decelerates gracefully, excess capacity in the grid and data centers could be redirected. For example, unused data center space might be leased for general cloud computing or made available to new AI researchers who previously couldn’t access compute time.

In a bullwhip scenario, we could also see industry consolidation. Just as the dot-com crash weeded out weaker players, an AI infrastructure bust might leave only the strongest standing. The largest firms (with diversified businesses) would weather the storm, while some smaller, pure-play AI cloud providers or hardware startups could fold or be acquired at bargain prices. Ironically, this would further concentrate the power in companies like Google, Microsoft, and Amazon, the very outcome some feared at the start. Indeed, the Wall Street Journal noted that this AI boom could make the big tech firms “even more powerful,” as smaller competitors lean on their platforms or get squeezed out in the frenzy. A rough ride in the market could validate that: the giants survive and scoop up assets, emerging even more dominant in AI.

Of course, there’s an alternative ending. If the optimistic view holds and this boom is more about platform buildout than a bubble, the current investments will eventually be justified by robust demand. In that case, prices might stabilize at profitable levels, capacity will gradually be absorbed without a deep bust, and investor returns will come not from immediate AI product revenues but from longer-term shifts (e.g., AI-as-a-Service becoming a core utility that businesses pay for steadily).

Public infrastructure enhancements – including increased grid capacity, renewable energy projects, and edge data centers in new regions – will find ample use as our digital and physical economies integrate AI into every sector. In reality, the outcome may lie somewhere in between: a mild bullwhip where a period of oversupply in the late 2020s corrects the most extreme exuberance, followed by a continued upward trajectory in the 2030s. Strong demand growth may increase the amplitude of the cycle, but it doesn’t much change its timing, meaning we might be in for a wave motion, not a one-time crash.

Bottom line

The current AI infrastructure boom is walking a knife-edge between overenthusiasm and genuine necessity.

The bullwhip effect serves as a helpful warning, reminding us that the supply chain can and does overshoot, particularly with long lead times and uncertain forecasts. All the ingredients for a classic bullwhip are present in AI’s rise.

Yet, unlike a transient fad, AI has the makings of a general-purpose technology that will underpin the economy for decades – more akin to electricity or the internet than to a one-off gadget craze.

Thus, even if we receive a correction, it may be a shakeout that ultimately benefits the sector by eliminating inefficiencies and driving the next phase of stable growth. Policymakers and investors would be wise to prepare for some volatility, as price swings, capacity gluts, and localized bottlenecks are likely to occur in the near term.

The real takeaway is this: the AI infrastructure boom is both a bullwhip cycle and the early buildout of a general-purpose technology.

We should expect periods of overshoot, falling GPU prices, idled data halls, grid mismatches, because long, multi-layered supply chains always create waves. But we should also expect those waves to rise over time as AI becomes a foundational utility across industries.

In other words, this is not a bubble; it is a bullwhip riding a secular platform shift. The cycles will be violent, but the direction is unmistakably upward.

The challenge for policymakers, investors, and operators is to navigate the oscillations without losing sight of the long arc. Today’s surplus will likely be tomorrow’s backbone. The question is not whether we are overbuilding, but whether we are building fast enough in the places the grid will allow.