The performance gains from Nvidia's Hopper to Blackwell GPUs come from increased size and power, not efficiency. This signals a potential scaling limit, creating an opportunity for radically new hardware primitives and neural network architectures beyond today's matrix-multiplication-centric models.

The current AI spending spree by tech giants is historically reminiscent of the railroad and fiber-optic bubbles. These eras saw massive, redundant capital investment based on technological promise, which ultimately led to a crash when it became clear customers weren't willing to pay for the resulting products.

When power (watts) is the primary constraint for data centers, the total cost of compute becomes secondary. The crucial metric is performance-per-watt. This gives a massive pricing advantage to the most efficient chipmakers, as customers will pay anything for hardware that maximizes output from their limited power budget.

The current AI infrastructure build-out is structurally safer than the late-90s telecom boom. Today's spending is driven by highly-rated, cash-rich hyperscalers, whereas the telecom boom was fueled by highly leveraged, barely investment-grade companies, creating a wider and safer distribution of risk today.

Hyperscalers are extending depreciation schedules for AI hardware. While this may look like "cooking the books" to inflate earnings, it's justified by the reality that even 7-8 year old TPUs and GPUs are still running at 100% utilization for less complex AI tasks, making them valuable for longer and validating the accounting change.

The massive capital rush into AI infrastructure mirrors past tech cycles where excess capacity was built, leading to unprofitable projects. While large tech firms can absorb losses, the standalone projects and their supplier ecosystems (power, materials) are at risk if anticipated demand doesn't materialize.

Current AI spending appears bubble-like, but it's not propping up unprofitable operations. Inference is already profitable. The immense cash burn is a deliberate, forward-looking investment in developing future, more powerful models, not a sign of a failing business model. This re-frames the financial risk.

Arvind Krishna forecasts a 1000x drop in AI compute costs over five years. This won't just come from better chips (a 10x gain). It will be compounded by new processor architectures (another 10x) and major software optimizations like model compression and quantization (a final 10x).

Accusations that hyperscalers "cook the books" by extending GPU depreciation misunderstand hardware lifecycles. Older chips remain at full utilization for less demanding tasks. High operational costs (power, cooling) provide a natural economic incentive to retire genuinely unprofitable hardware, invalidating claims of artificial earnings boosts.