Eclipse Ventures founder Lior Susan shares a quote from Sam Altman that flips a long-held venture assumption on its head. The massive compute and talent costs for foundational AI models mean that software—specifically AI—has become more capital-intensive than traditional hardware businesses, altering investment theses.

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 debate over AI chip depreciation highlights a flaw in traditional accounting. GAAP was designed for physical assets with predictable lifecycles, not for digital infrastructure like GPUs whose value creation is dynamic. This mismatch leads to accusations of financial manipulation where firms are simply following outdated rules.

Some tech companies have doubled the depreciable life of their AI hardware (e.g., from 3 to 6 years) for accounting purposes. This inflates reported earnings, but it contradicts the economic reality that rapid innovation is shortening the chips' actual useful life, creating a significant red flag for earnings quality.

Traditional valuation multiples are increasingly misleading because GAAP rules expense intangible investments (R&D, brand building) rather than capitalizing them. For a company like Microsoft, properly capitalizing these investments can drop its P/E ratio from 35 to 30, revealing a more attractive valuation.

While the current AI phase is all about capital spending, a future catalyst for a downturn will emerge when the depreciation and amortization schedules for this hardware kick in. Unlike long-lasting infrastructure like railroads, short-term tech assets will create a significant financial drag in a few years.

Adjusting financial statements to capitalize R&D provides a more accurate book value for tech firms. However, this input-based approach is limited, as the value of an intangible asset, like a successful drug patent, is non-linear and disconnected from its historical cost.

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.

Perplexity achieves profitability on its paid subscribers, countering the narrative of unsustainable AI compute costs. Critically, the cost of servicing free users is categorized as a research and development expense, as their queries are used to train and improve the system. This accounting strategy presents a clearer path to sustainable unit economics for AI services.