The numbers are hard to ignore. In 2026, the four largest hyperscalers, Microsoft, Alphabet, Amazon, and Meta, are projected to spend somewhere between $650 billion and $700 billion on capital expenditures, up more than 60% from 2025's already historic levels. BloombergNEF puts the broader figure, including other major data center firms, closer to $750 billion. Roughly three-quarters of that is going directly into AI infrastructure: GPUs, servers, networking, and the massive facilities to house them all. Then Elon Musk walked on stage at a weekend event in late March and announced Terafab, a $25 billion chip-making factory in Austin, Texas, backed by Tesla, SpaceX, and xAI. The stated goal: to build "more chips than all the chip manufacturers in the world combined can provide today." Bernstein analysts quickly estimated that the full vision could cost $5 trillion, more than 70% of the total yearly US government budget. Tesla generates $3 to $6 billion in free cash flow. The math, as many observers noted, doesn't work without external capital on a scale that doesn't yet exist. But Terafab isn't really the story. It's a symptom. The real question is whether the entire AI infrastructure boom, the hundreds of billions flowing into data centers and chips every quarter, is a rational response to genuine demand or the early stages of the most expensive overbuild in history.

The bull case: you can't have too much compute

The optimistic argument is straightforward. AI inference demand is growing faster than anyone predicted. Inference now accounts for roughly two-thirds of all AI compute cycles, and that ratio keeps tilting as deployment scales outpace training runs. Every major tech company is racing to embed AI into its products. Enterprise adoption, by most measures, is still in its infancy. This is where Jevons Paradox enters the conversation. The idea, first articulated by economist William Stanley Jevons in the 1860s, is that when efficiency improvements reduce the cost of using a resource, total consumption goes up, not down. DeepSeek's breakthroughs in early 2025 demonstrated this in real time: dramatically cheaper inference didn't reduce demand for compute. It expanded the set of things people were willing to build with AI. KKR's investment research team made a version of this argument in their "Beyond the Bubble" report, noting that AI-related capex now accounts for about 5% of US GDP and contributed materially to GDP growth in the first half of 2025. In the bull case, every application becomes an agent, every company needs its own models, and compute demand compounds for decades. If that's the future, today's spending is just a down payment.

The bear case: we've seen this movie before

But efficiency gains don't always overwhelm supply. Sometimes you just overbuild. The most cited historical parallel is the fiber optic bubble of the late 1990s. Telecom companies, intoxicated by projections of exponential internet traffic growth, laid tens of thousands of miles of fiber optic cable across the US and under the oceans. By 2001, an estimated 85% to 95% of installed fiber capacity sat unused, earning the nickname "dark fiber." Companies like WorldCom, Global Crossing, and Qwest collapsed. The internet did change the world, it just didn't happen on the timeline or at the scale those early investments assumed. The pattern is instructive. The technology was real. The demand eventually materialized. But "eventually" can take a decade, and capital doesn't wait that long. As Yale Insights noted, the fiber bubble wasn't caused by bad technology. It was caused by a "technological breakthrough that made each line exponentially more powerful, multiplying existing capacity," rendering much of the new infrastructure unnecessary for years. The AI version of this risk is already visible. Microsoft reported a $13 billion annual run rate from AI initiatives while planning $40 billion in additional capex. That's a ratio that requires extraordinary faith in future revenue. Forbes estimated the total AI bubble at $1.7 trillion in data center spending worldwide by 2030. And as NPR reported in late 2025, tech companies are increasingly relying on debt and risky financial tactics to fund their buildouts, a dynamic eerily reminiscent of the telecom era.

The middle case: concentration, not shortage

There's a third possibility that gets less attention. What if the compute gets absorbed, but by far fewer players than the current spending implies? The hyperscaler capex race is overwhelmingly driven by four or five companies. Amazon planned $200 billion in 2026 capex. Google is close behind at $175 to $185 billion. The startup share of total AI infrastructure spending is tiny. If AI value concentrates in a handful of platforms, the infrastructure gets used, but the competitive dynamics look nothing like the broad-based tech boom that many investors are pricing in. This concentration story extends to chips. NVIDIA currently holds approximately 80% of the AI accelerator market, but the custom silicon movement is accelerating. Google's seventh-generation TPU Ironwood, released in late 2025, is described by analysts as "arguably on par with NVIDIA Blackwell." AWS has deployed over 500,000 Trainium2 chips for Anthropic's model training, the largest non-NVIDIA AI cluster in production. Meta announced plans to develop four new generations of its MTIA custom chips within two years. The New York Times reported in January 2026 that Amazon and Google are actively eating into NVIDIA's chip supremacy. The numbers tell a striking story of divergence. Custom ASICs are growing at 44.6% CAGR compared to 16.1% for general-purpose GPUs, with the custom chip market projected to grow from $18 billion in 2024 to $165 billion by 2033. NVIDIA still sells the shovels to everyone, but the biggest gold miners are increasingly forging their own tools.

The energy wall

Every chip needs power, and power is the constraint nobody has solved. The International Energy Agency projects that data center electricity consumption could reach 945 terawatt-hours by 2030, roughly equivalent to Japan's entire electricity consumption. McKinsey estimates meeting this demand will require $7 trillion in capital investment. But the grid isn't ready. At the end of 2025, data centers requiring 241 gigawatts of electricity were in the development pipeline, a 159% increase from the beginning of the year. Only a third of those projects are under active development. In Northern Virginia, America's "Data Center Alley," grid connection delays now stretch to seven years. A survey by AlphaStruxure found that 92% of data center decision-makers see grid constraints as the primary obstacle to construction. The result is a strategic pivot: capital deployment is increasingly directed at regions offering reliable large-scale power rather than network connectivity. The old logic of building near population centers is giving way to building wherever the electrons are. Researchers estimate a 9 to 18 gigawatt AI energy shortage that will persist through at least 2027, with partial relief only beginning in 2028 as renewable farms, nuclear restarts, and small modular reactor deployments come online. Fortune reported in March 2026 that US data center buildouts are hitting concrete snags because power grid limits are forcing "a bend in the trajectory."

The Singapore question

Singapore offers a microcosm of the tensions in the global AI infrastructure race. The city-state sits one degree north of the equator, making it one of the worst places on Earth to cool a data center. It's constrained by land and energy. And yet it has one of the highest densities of data center infrastructure per capita on the planet, with more than 70 facilities and over 1.4 gigawatts of capacity. The government is betting heavily on AI as national infrastructure. Budget 2026 announced national AI missions across manufacturing, transport, finance, and healthcare, alongside a new National AI Council, a dedicated AI park at one-north, and the "Champions of AI" program. Google announced expanded AI investments in Singapore in February 2026. Operators like ST Telemedia are piloting high-voltage direct current testbeds because, as their Singapore country head put it, traditional AC systems "are hitting their physical and efficiency limits." But can a small state really compete in an infrastructure race defined by scale? Bridge Data Centres is exploring barge-based hydrogen power generation, an innovative but unproven approach. The South China Morning Post noted in March 2026 that Southeast Asia's AI data center gold rush is testing power grids and tropical cooling systems to their limits. Singapore's play isn't to out-build the hyperscalers. It can't. The bet is that strategic positioning, regulatory clarity, talent density, and regional connectivity matter more than raw capacity. It's a fundamentally different wager than the one being made in Texas or Virginia, and it might be the smarter one.

Analyzing the bet

The honest answer is that nobody knows whether we're overbuilding. The bull case and the bear case both rest on assumptions about future demand that are, by definition, unknowable. What we can do is analyze the structure of the bet itself. First, the spending is irreversible. Data centers, once built, represent sunk costs. If demand doesn't materialize on schedule, the assets depreciate while sitting underutilized, exactly like dark fiber in 2002. Second, the financing is getting riskier. Free cash flow for the four largest hyperscalers dropped from $237 billion in 2024 to $200 billion in 2025, even as capex surged. MUFG projects that hyperscalers will fund about half of data center spend through 2028, with the rest coming from private credit, corporate debt, and securitized instruments. That's a capital structure that assumes continued growth. Third, the competitive landscape is shifting underneath the investment. Custom silicon is eroding NVIDIA's dominance. Efficiency improvements in models and inference are reducing compute per task. If Jevons Paradox holds, that's bullish for total demand. If it doesn't, that's a lot of expensive GPUs with shrinking utilization rates. Fourth, energy is a hard physical constraint, not a software problem. You can't deploy models on chips that don't have power. The grid bottleneck introduces timing risk that no amount of capital can fully resolve. The $25 billion Terafab might get built, or it might not. But the broader question it embodies, whether the AI infrastructure boom is building the future or just building, will define the next decade of technology investment. The answer probably isn't a crash or a vindication. It's something messier: a partial overbuild, absorbed unevenly, with winners defined not by who spent the most but by who spent most wisely.

References