The Trillion-Dollar AI Capex Gap: Why Too Much Hardware Could Be Nvidia’s Trap and Microsoft’s Opportunity

The rapid expansion of artificial intelligence infrastructure spending is outpacing the revenue streams that can sustain it, according to estimates from Goldman Sachs. While OpenAI and Anthropic have generated over $105 billion in annualized revenue by August 2026, this remains a modest fraction of the $7.6 trillion in cumulative AI capital spending anticipated between 2026 and 2031.

To prevent infrastructure write-downs on short-lived hardware, the industry must scale its annual recurring revenue to surpass $1 trillion by 2030. Critics argue that the industry is constructing hardware capacity faster than customers can utilize it profitably. If enterprise demand does not meet the new capacity, hyperscalers may reduce purchases, negotiate lower prices, and migrate routine workloads to in-house chips.

Nvidia Corporation (NASDAQ:NVDA) faces the most significant exposure due to its 74.9% gross margin, which is contingent on customers competing for scarce, high-end GPUs. Google's Tensor Processing Units (TPUs), Amazon's Trainium, and Microsoft's Maia chips can handle predictable inference workloads, decreasing Nvidia's purchases and enhancing the bargaining power of hyperscalers.

Despite expectations that Nvidia will maintain its market advantage, shares traded at 25.64 times forward earnings as of August 17, and 285 elite hedge funds in Insider Monkey's second-quarter database held long positions. Slower orders coupled with weaker pricing would negatively impact both earnings expectations and the prevalent trading pattern.

Conversely, the cost of AI inference has plummeted by a factor of 50, from around $20 per million tokens to approximately $0.40. Deloitte estimates that inference will account for roughly two-thirds of AI compute by 2026. Although falling unit costs do not automatically imply reduced demand, developers can leverage cheaper inference calls more frequently as costs decline.

Microsoft (NASDAQ:MSFT) exemplifies how this cost reduction can generate enduring revenue. Its AI operations achieved a $37 billion annual run rate in Q3 2021, a 123% increase year-over-year, while Microsoft 365 Copilot surpassed 30 million paid seats by June. Microsoft benefits not only from the model but also from its position within software companies that already utilize its products.

However, concerns about rising AI capex, depreciation, and weaker free-cash-flow conversion could outpace Copilot and Azure revenue, potentially leading to a staged infrastructure digestion phase. The adjustment might gradually shift value from hardware scarcity towards software monetization. As inference becomes more affordable and custom silicon assumes routine workloads, Nvidia could encounter slower growth and more erratic margins, even with its platform leadership.

Microsoft could capitalize on the upside by integrating cheap inference into products its customers already pay for, converting the infrastructure buildout into recurring revenue after absorbing short-term depreciation and cash-flow pressure. While the authors acknowledge the potential of both MSFT and NVDA as investments, they believe certain AI stocks present greater upside potential with less downside risk.

For an extremely undervalued AI stock that stands to benefit from Trump-era tariffs and the onshoring trend, the authors recommend consulting a free report on the best short-term AI stock.