What's the variant view on AI compute demand and semis capex versus consensus?

The consensus view is that the AI industry is defined by a structural supply deficit and a massive capital expenditure arms race, not a speculative bubble [13, 21]. Hyperscalers and cloud providers are projected to spend between **$660 billion and $690 billion** on AI compute and infrastructure in 2026 alone, a figure that has already been revised upward by nearly $120 billion since the beginning of the year [6, 8, 17, 18]. This spending is driven by an insatiable and unpredictable demand for compute that is growing exponentially, creating a multi-billion dollar backlog for hardware [12, 21, 22]. The scale of this buildout is seen as a competitive necessity, with firms racing to avoid being out-scaled by rivals [7, 20]. This capital influx is not limited to chips but extends to the entire physical infrastructure layer, including data centers and power components, causing stocks for companies like Vertiv and Eaton to trade like semiconductor stocks [16, 19].

While the consensus points to current supply chain chokepoints in high-bandwidth memory (HBM), advanced packaging, and TSMC's wafer capacity as the primary constraints on growth [3, 7, 13, 23], a variant view suggests the critical bottleneck is shifting. The emerging constraint is more foundational: the availability of **power generation, distribution, and data center components** like transformers and switchgears . Looking further ahead, another analysis predicts that by 2028-2029, the ultimate cap on the entire AI buildout will be ASML's production capacity for EUV lithography tools, the machines required to manufacture advanced chips [7, 24]. This suggests a multi-stage evolution of bottlenecks, moving from specific chip components today to fundamental industrial and manufacturing capacity in the future. Some companies, like Cerebras, are pursuing a variant strategy by designing architectures that specifically avoid current HBM and packaging bottlenecks, offering a potential path around near-term constraints .

A significant tension exists between the current narrative of scarcity-driven high costs and a variant view predicting a dramatic long-term price decline. While present market dynamics show soaring memory prices and high GPU rental costs forcing AI model vendors to increase prices [7, 9], one forecast anticipates a **1000x reduction in AI compute cost** over the next five years . This decline would be driven by a confluence of 10x improvements from silicon advances, 10x from new chip designs, and 10x from software and memory optimizations . This potential for rapid cost reduction aligns with an observed shift in customer behavior, where enterprises prototype on expensive frontier models but increasingly scale on more cost-effective, tunable open-source models to optimize performance and manage their cost of goods sold [1, 14]. This suggests a future where demand becomes highly price-elastic and the market diversifies away from a few dominant, high-cost model providers .

Beyond the hyperscaler arms race, the landscape of compute demand and supply is becoming more complex. The next major wave of infrastructure demand may be driven by the adoption of "reasoning" models, which are significantly more expensive per query than current models . Concurrently, broader corporate AI adoption is expected to increase demand for traditional servers, benefiting established players like Dell and HPE . The competitive landscape could also be altered by new entrants like SpaceX and significant, market-directing government investments, such as the reported **$9B CIA/NSA compute cluster** [26, 27]. This indicates that while the current buildout is massive, its future trajectory will be shaped by evolving model architectures, enterprise cost-optimization strategies, and the entry of non-traditional infrastructure players.