NVIDIA Is Still Skynet. Here Is What I Got Wrong.

In August 2018 I published a piece on LinkedIn called “NVIDIA Is Skynet.” The thesis was simple: analysts did not have the tools or the perspective to value the company, it was quietly underpinning self driving cars, VR, gaming and server side AI, and it was run by a founder with a tattoo of the company logo on his arm. I was not an analyst then and I am not one now. I was a guy running a digital agency who had just discovered that CUDA rendering turned an 8h-our render into a 4-minute one and that saved my editors about 30 hours of downtime a week, per PC.

That got my attention. Eight years later, running an AI startup, ChatBar AI, it is worth asking what I got right, what I got wrong, and, more importantly for the people reading this newsletter, what does it all mean for a business that will never buy a GPU?

The Scoreboard

When I wrote that article NVIDIA’s share price was $256.95 and a P/E of 42 was considered rich. After a 4 for 1 split in 2021 and a 10 for 1 split in 2024, that is the equivalent of about $6.40 a share today. The stock currently trades in the low $200s. Call it a 30x return for anyone who bought the “overpriced” shares that week and did nothing clever at all.

The numbers I quoted then feel quaint now. Q1 FY2019 record revenue was $3.21 billion. Data center revenue, the part I argued nobody understood, was $701 million. In the most recent quarter NVIDIA reported $96.2 billion in revenue, with $89 billion of it coming from data centers, roughly 92 percent of the company. Data center alone grew about 127x. Guidance for the current quarter is $108 billion. The company is worth somewhere north of $5 trillion depending on the day, and it is still run by Jensen Huang, who still has the tattoo.

What I Got Right

1. The “picks and shovels” framing held. I argued the value was in the hardware and software layer underneath AI, not in any single application. That turned out to be the whole story.
2. Simulated training for robots. I described VR environments plus physics engines plus deep learning letting a robot run thousands of hours of boot camp in a few seconds. That is now a mainstream product line, not a demo video.
3. Losing Tesla did not matter. Elon had just announced Tesla was dropping NVIDIA for its own chip. I said Wall Street would shrug. It shrugged.
4. Leadership was the most important factor. Founder led, small officer team, a CEO who apologised publicly for the GTX 970 spec issue. Try to imagine the CEO of Apple doing that. Eight years of execution later, I would weight that point even higher.

What I Got Wrong

1. I was thinking about the wrong customers. I pictured only self-driving cars, robotics labs and crypto miners. I did not picture a handful of cloud providers and AI labs buying GPUs by the stadium. The demand came from fewer buyers, with far bigger cheques, than I imagined.
2. I underestimated it. Badly. I called NVIDIA a “technology juggernaut the likes of which we’ve never seen before” and I still had it wrong by an order of magnitude. If that is what I got wrong when I was bullish, imagine what the bears got wrong.
3. I thought the hard part was building the brain. The hard part turned out to be the electricity, the real estate, the networking and the capital. NVIDIA sells into all of those constraints and it still cannot ship fast enough.
4. I used the words “deep learning” because “AI” was being debated endlessly. Today nobody debates whether the technology works. They debate whether their business is using it.

My Company Will Never Buy A GPU – Why Should I Care?

Here is the part I did not understand in 2018 and that I think most SME owners still do not see.

In 2018 the AI arms race was being fought by maybe a dozen companies that could afford the chips. Today the chips are built, the models are trained and the compute is rented by the minute – by the token. The capital intensive part of the race has already been run by people with trillions of dollars. What is left is the cheap part: putting that intelligence in front of your customers, your staff and your training content in a way that adds value.

That is a very different race, and it is one an SME business can actually win.

I see this every week at ChatBar AI. A logistics company with offices across Asia uses it to answer thousands of repetitive client queries every day, and it repurposed the same engine to collect stories from golfers at its annual tournament. A business school publisher put a conversational case search tool in front of the faculty who buy its cases, so they can find the right one by describing the class they are teaching instead of scrolling a catalogue. A European optical brand practises selling to different customer types through role play that talks back, with actionable coaching feedback after every round.

None of these companies are buying hardware. They are renting the output of NVIDIA’s data centers for a fraction of a cent per conversation, and the differentiator is no longer who has the compute. It is who has their content, their process and their customer data organised well enough to use it.

“I’ll be back… with the contract.”