Why the AI Capex Bill Looks Like a Telecom Bubble — and Why It Probably Isn't

The number that gets thrown around at every macro desk in London right now is $1.4 trillion. That is the cumulative announced AI infrastructure spend, across hyperscalers, sovereign clouds, and the tier of dedicated GPU buildouts that did not exist eighteen months ago, between 2024 and the end of 2027. Microsoft's $190 billion commitment. Meta's $145 billion run-rate. Amazon's $125 billion FY26 capex line. Google's $90 billion. Alibaba's $52 billion three-year pledge. Stargate's headline $500 billion that nobody can quite source-attribute but everybody quotes anyway.

The number is large enough that it has started to attract a specific historical comparison. In 1999 and 2000, telecom and equipment vendors collectively committed about $500 billion (around $900 billion in 2026 money) to fibre-optic and switching infrastructure on the back of a thesis that internet traffic was going to double every hundred days. Traffic did grow, eventually, but it took until roughly 2008 for fibre utilisation to catch up with the capacity laid down in 2000. WorldCom went bankrupt. Global Crossing went bankrupt. Nortel imploded. Cisco lost 90% of its market cap from peak and took fifteen years to get back.

If you have heard one comparison, you have heard them all. Hyperscaler capex looks like Lucent in 1999. The depreciation cliff is coming. The AI applications do not yet justify the spend. And so on.

The comparison is doing real intellectual work, and the people making it are not stupid. But it has at least three holes in it, and the holes are the interesting part.

Hole One: The Demand Curve Is Already There

The telecom bubble was built on projected demand. Companies committed capital to capacity that, in 1999, did not yet have customers. The famous Worldcom slide deck claiming internet traffic would 16x every year was, in retrospect, a fabrication — actual traffic was growing closer to 100% annually, which is fast but nothing like the assumptions baked into infrastructure financing.

The AI capex story is, in a structurally important way, different. The hyperscalers are not building speculative capacity. They are building to meet customer demand that has already overrun existing supply. Microsoft Azure has been GPU-constrained on its Stargate-pre-cursor instances since Q2 2024. AWS Trainium2 wait-lists have been multiple quarters long. Google's TPU v6 capacity has been allocated to internal use plus a handful of priority external customers, with no public-cloud SLA for the rest of 2026.

Demand is not the bottleneck on hyperscaler revenue right now. Supply is. That is precisely the inverse of the 1999 dynamic, and it changes the risk profile substantially. When the iron arrives, it gets bought. The question is not whether the capacity will be utilised; it is whether utilisation pricing holds up once the supply-demand imbalance unwinds.

Hole Two: The Useful-Life Question Is Different

Fibre laid in 1999 is still in service. The actual physical capacity built during the telecom buildout was structurally sound — it was the financing that broke, not the assets. WorldCom went under because it could not service the debt it had taken on to build the network, not because the network did not work.

GPUs depreciate on a much shorter horizon. NVIDIA H100s, which were the apex predator of training compute in early 2024, are already being repriced as inference workhorses against H200s, B100s, and B200s. The depreciation schedules hyperscalers are quietly extending — Microsoft moved from a six-year to a four-year useful life last quarter, which the CFO addressed in detail on the earnings call — are the canary. If the useful life of a $40,000 GPU keeps compressing, the unit economics of training infrastructure get genuinely difficult.

The counterargument from the hyperscaler CFOs is that older GPUs cascade down to inference, where the workload is more forgiving, the margins thinner, and the depreciated hardware genuinely competitive for years. That is true. It is also a much smaller revenue pool than training, and it is the pool the merchant GPU cloud providers — CoreWeave, Lambda, Crusoe — are also fighting over with similarly depreciating fleets.

Hole Three: The Revenue Models Are Both More Mature and More Fragile

Telecom equipment vendors in 1999 sold to carriers, and the carriers sold to enterprise and consumer customers under long-term contracts. The value chain was three layers deep and the contractual relationships were long.

The AI capex value chain is shorter. Hyperscalers sell compute directly to model labs (Anthropic, OpenAI, Mistral, the in-house Meta and Google teams) and to enterprise AI customers. The contractual horizons are months to a couple of years, not the seven-to-ten-year carrier deals that underpinned 1999 telecom financing. That cuts both ways. The revenue is harder to model out beyond eighteen months, but it is also less load-bearing on the financing structure: the hyperscalers are spending out of operating cash flow, not debt-financed at the WorldCom multiple.

The fragility risk is concentration. If three or four model labs account for the majority of training-tier GPU consumption, and any one of them has a Stability-AI-style implosion or pivots to a wildly more efficient architecture, the demand-supply equation re-tilts overnight. The 2025 DeepSeek moment, when a Chinese lab published a frontier-quality model trained on a fraction of the conventional compute, is the template for what a true AI capex correction would look like. It would not be a slow-rolling debt unwind. It would be a single architectural breakthrough that reset the cost curve and stranded a billion-dollar order book.

What This Actually Looks Like When It Cracks

If the supercycle does crack, the failure mode will look more like 1986 semiconductors than 2000 telecom. Memory chips in the mid-1980s went through a brutal oversupply correction when capacity additions, designed for one demand trajectory, met a demand trajectory that turned out to be slightly different. The companies that survived were the ones with the lowest unit costs and the most diversified end-customer bases.

Hyperscalers have both. The merchant GPU clouds — the CoreWeaves and the Lambdas — have neither. If you wanted to bet against the AI capex thesis without taking on the systemic risk of shorting Microsoft, the pure-play infrastructure financiers are where the asymmetry is. They have built large fleets on the assumption that hyperscaler GPU demand will overflow into the merchant tier indefinitely. The CoreWeave Q1 2026 results gave a hint of what happens when that overflow slows: revenue beat, utilisation softened, and the stock dropped 18% on the day.

The Honest Read

The AI capex supercycle is not the telecom bubble. The demand is real, the customers are paying today, and the cash flows funding the buildout are operating, not debt-financed. The asset depreciation profile is worse than fibre, but the revenue ramp is faster.

What it could be — and this is the comparison that should keep CFOs awake — is the 1995-to-2000 dot-com infrastructure layer. Most of that capacity got used. Most of the companies funding it did not survive. The technology mattered enormously. The equity returns, for those who picked individual names, were a coin flip.

If you are an LP wondering whether to underwrite another $200 million AI infrastructure fund, the question is not "will AI matter." It will. The question is "will the specific physical assets being built today still be economically productive in 2030." And on that question, the honest answer is: probably most of them. Probably not all of them. The difference between the two will determine whether the supercycle gets remembered as the railway boom that built modern Britain or the canal mania that broke a generation of speculators.

Both are real historical analogies. Both involved technology that absolutely mattered. The investors in the first did very well. The investors in the second mostly did not.