Steam engines 2.0: The world will find a way to meet insatiable demand for GPU’s

In the early 1990s three engineers met in a Denny’s diner in San Jose (the same Denny’s where Jensen previously worked as a dishwasher). Video games were exploding in popularity and complexity. Characters were no longer flat sprites — they were three-dimensional worlds made of millions of pixels, all needing to be calculated, shaded, and rendered in real time.

The founders realised something fundamental...

Background to Revenue and Earnings

This narrative focuses on the data-centre aspect of Nvidia, which is the fastest growing segment and the one that the valuation of the company is primarily based on.

We can represent this as:

Data Centre Revenue ≈ Number of GPUs shipped × (Average GPU + Networking price)

To put some numbers against this:

• We use a total data centre revenue of 4x the last reported quarter (4x $51 billion = $204 billion). This accounts for the fact that the quarter ends in October and not December.
• We assume 5.2 million Blackwell GPUs were sold in 2025 (based on estimates from JP Morgan/ Morgan Stanley)
• We know from CNBC interviews with Jensen that that the cost of a Blackwell-class data-centre GPUs is around $30-40k.
• We know from that ratio of Compute/ Networking sales above that 85% of data centre is Compute and 15% is networking - and assuming that in general cloud providers are purchasing both CPUs and Networking equipment simultaneously

Therefore for 2025:

$204 billion = 5.2 million GPUs x $39,230 (average total price)

In terms of earnings, the net profit margin has been stable around 50% since 2024, and previously sat closer to more typical fabless semiconductor producer (think AMD, Qualcomm or Broadcom),which is typically closer to 25-30%.

Forecasting Revenue

We forecast $570 billion in revenue by 2031. Using our formula this could be achieved by:

$570 billion = 10.4 million GPUs x $55,000 (average total price)

$570 billion = 12.7 million GPUs x $45,000 (average total price)

This would require 25% YoY growth (currently forecast at 23% and previously >50%)

Sense checking this (recent LTM annual revenue):

• Amazon: $716.9bn
• Walmart: $681.0bn
• Apple: $416.2bn
• Alphabet: $310bn
• Saudi Aramco: $600bn

$570bn is beyond currency “big tech” revenue, it’s comparable to “global infrastructure” revenue (which AI could become.

Forecasting Earnings

We assume the net profit margin remains the same as today of 50%.

The company recently provided FY‑2027 guidance / commentary: management said they are working to hold gross margins in the mid‑70s next year, driven by mix, cycle‑time improvements and input cost work; they noted cost improvement, mix and cycle time as the levers.

Forecasting PE Multiple

We assume the PE ratio in 2031 to be the same as today of 45x. This is lower than the 5 year average that Nvidia has traded at.

Areas for Additional Research

Demand and committed spend

We already know hyperscalers are planning AI infrastructure years in advance. NVIDIA’s management has talked about hundreds of billions of dollars in long-term AI demand visibility, and cloud providers are guiding to sustained, elevated capex.

Depending on sources Amazon, Meta, Google, and Microsoft are expected to spend around ~$700B in capex in 2026, up ~60% vs 2025

Supply

NVIDIA doesn’t make its own chips. It relies heavily on TSMC, advanced packaging, memory, and a complex global supply chain.

To reach hundreds of billions in revenue, NVIDIA would need sustained access to leading-edge capacity — not just wafers, but packaging, HBM memory, and networking components.

TSMC own capex - which is needed years ahead manufacturing demand - has actually been flat or even in decline - and is only recently started to pick up.

Physical data-centre capacity

AI GPUs don’t live in isolation. They need buildings, cooling, racks, networking, and land.

A $570 billion revenue world implies an enormous expansion of AI-specific data centres — far beyond traditional cloud growth - and that is essentially what we are seeing with projects like Stargate in the US.

Power

AI data centres are extremely power-hungry. In many regions, electricity availability is already becoming the binding constraint. In fact the White House even has a draft bill which specifically looks to prevent data centres increasing the price of household electricity.

If power generation does become a bottleneck, it will be the hardest one to solve — it takes a long time to bring on new power generation facilities and why would cloud providers continue ordering racks of GPU’s if there is no power available to run them.