Nvidia beats estimates: revenues at $57 billion. Stock up in after hours

The Gaming segment, for its part, recorded a 30% increase in revenue compared to a year ago, supported by the continued demand for Blackwell products. On a sequential basis, Gaming sales decreased by 1%, as channel inventories reached more normal levels in the run-up to the Christmas season.

Finally, revenue in the Professional Visualisation segment in Q3 grew 56% year-on-year and 26% sequentially, thanks to the launch of the new DGX Spark™ and growth in sales related to the Blackwell line.

In short: in the end, the expectations that had arisen in the wake of the signals - launched some time ago - about the existence of robust demand were not disappointed. Last month, for instance, the CEO of Tsmc - Nvidia's main fabless manufacturing partner - described demand for Ai as 'very strong', higher than previously expected. Jensen Huang - Nvidia's CEO - had then remarked on 'exceptional' interest in the new generation of Blackwell chips, adding that the company would have visibility over more than USD 500 billion in cumulative revenues until 2026, including the future Rubin line.

However, investors are increasingly looking beyond the accounts, seeking confirmation of the structural durability of demand. In this context, an initial focus is on partnerships with major players in the sector. In September, as is well known, Nvidia and OpenAI signed a letter of intent for a collaboration that could envisage investments of up to 100 billion to build capacity in data centres for Artificial Intelligence. This is a situation that has given rise to much controversy. Several experts point out how, through circular investments with e.g. Oracle, a sort of self-induced demand for Ai has been created, where the one who finances OpenAI is the same one who provides the hardware to make artificial intelligence 'work'. In such a context, Ubs states - quite rightly - that a lot will be played on the materialisation of demand on the part of businesses. If companies start 'wanting' Ai in the right quantities - and above all - in the right timeframe (2026-2027) then everything will be 'OK'. Otherwise, there could be problems. In particular, with respect to the huge amount of capitalised investment that hyperscalers have planned.

Already the mega Capex of big tech. Historically, large platforms have been able to sustain capital expenditures by leveraging the robust cash flows generated from operations. However, the current outlays required by Artificial Intelligence's leap in scale are rapidly eroding this capacity. According to Bank of America's most recent analysis, hyperscalers have used, on average, 72% of their operating cash flow to fund outlays. And the trajectory looks set to rise. Projections for 2026 speak of over USD 500 billion in digital infrastructure spending, with market consensus indicating that cash flow absorption could exceed 90 per cent in the coming years. No wonder, then, that concerns are arising and the market is treading lightly.

But it is not only a question of too high Capex and circular investments. Another issue that is catching the attention of practitioners is the effective life of GPUs and how it is accounted for. Some observers, including Michael Burry, have expressed concerns about the six-year depreciation schedules used by big tech for their large capitalised infrastructure investments. According to Burry, the real useful life of chips - a central component of the infrastructure itself - would be shorter. As a result, the big US technology companies are in fact artificially inflating profitability. Not everyone, on closer inspection, shares this view. Stacy Rasgon of Bernstein reported that the A100 GPUs, despite being five years old, are still generating high margins. The analyst's audits of industry players indicate that GPUs can remain operational for six or seven years.