NextFin News - The global race for artificial intelligence has hit a structural bottleneck that mirrors the most frantic periods of the 1990s internet boom, as leading labs like Anthropic and OpenAI struggle to secure enough compute power to satisfy a surging mainstream user base. Despite a projected $700 billion in capital expenditure from hyperscalers this year, the supply of high-end semiconductors remains so constrained that major AI providers are being forced to implement aggressive usage limits and "peak hour" scheduling for model training.
Michael Parekh, a veteran market analyst and former Internet Equities Analyst at Goldman Sachs, argues that the current "AI Compute shortage" is a direct rhyme of the 1990s dial-up modem capacity crisis. Parekh, who has spent decades tracking tech cycles and helped finance the early build-out of TCP-IP networks, suggests that the industry has entered a phase where capital allocation is becoming more critical than model architecture. His perspective, while rooted in historical precedent, highlights a growing frustration among enterprise customers who find that paying for premium subscriptions no longer guarantees unfettered access to the most advanced models.
The data reflects a tightening grip on the hardware market. Lead times for Nvidia’s Blackwell and H200 series GPUs have extended to between three and seven months, with pricing for the flagship Blackwell architecture climbing as much as 23% in early 2026. Meta Platforms recently executed a multi-billion dollar acquisition of these chips to solidify its generative AI lead, a move that has effectively crowded out smaller players and forced them into what some analysts call "compute-aware" software development—designing systems specifically to run on less efficient, older hardware because the premium supply is locked up by the "Magnificent Seven" tech giants.
This scarcity has created a precarious balancing act for AI lab CEOs. Anthropic’s Dario Amodei recently noted that there is "no hedge on earth" against overbuying compute; purchasing too much capacity could bankrupt a firm if demand fluctuates, while buying too little risks ceding the market to competitors. This dilemma is manifesting in the secondary markets, where demand for OpenAI shares has reportedly cooled relative to Anthropic, as investors weigh OpenAI’s ferocious spending against Anthropic’s more disciplined, albeit capacity-constrained, approach.
However, the narrative of a permanent shortage is not without its detractors. Some industry observers point to the "Jevons Paradox," where increased efficiency in chip design and software actually drives up total demand, making the shortage feel perpetual even as supply increases. Furthermore, Nvidia has begun easing some constraints by maintaining dual production lines for both the H200 and Blackwell architectures through 2026. This strategy is designed to alleviate the backlog, though it has yet to significantly lower the barrier to entry for firms without massive balance sheets.
The frustration is now trickling down to the end-user. As AI labs prioritize internal training for next-generation models during off-peak hours, retail and enterprise users are facing variable cost bills and intermittent service outages. The transition from $20-a-month "pro" tiers to $200-a-month enterprise subscriptions has not yet solved the underlying physical reality: there are simply not enough high-end tokens to go around. For the time being, the AI revolution remains tethered to the slow, physical grind of semiconductor fabrication and data center power grids.
Explore more exclusive insights at nextfin.ai.
