What once (and only recently) seemed like a capital-light software boom has morphed into one of the largest infrastructure developments in modern corporate history.
The chart below, shows hyperscaler capex surging toward 25 percent of revenue, levels once associated with ‘old-world’ telecommunication networks.
The comparison is not dissimilar. Just as the dot-com era drove an arms race to lay fibre optic cable in anticipation of a digital future, today’s AI leaders are racing to construct vast compute networks.
More revenue demands more computational power, which demands more hardware, which demands more capital expenditure…
The worry with this analogy is that the unrealistic revenue expectations and ambitious overinvestment of the dot-com era removed discipline and pushed firms to chase scale ahead of returns that ultimately never materialised.
This escalating cost structure is also reshaping strategic alliances. The recent triangular investment pact between Anthropic, Nvidia and Microsoft is telling. AI has become so expensive that even the giants are banding together.
Under the arrangement, Anthropic will scale its large language models on Microsoft’s cloud network using Nvidia’s chips. In return, Microsoft and Nvidia will each commit billions to back Anthropic. Microsoft CEO Satya Natella’s comment that they will increasingly be customers of each other shows a growing circularity in the ecosystem, a bit like a powerboard plugged into itself with no real external source of energy.
When competitors start taking equity positions in one another simply to guarantee compute and chip supply, it shows that scarcity of resources at the top-end is shaping the race.
