Before any of this will make sense you have to understand that our current approach to AI does not work and never will. I’m not saying AI is impossible and that we’ll never build it, but the GPT/ML approach was a known dead end for decades in academic research. We didn’t need modern technology to figure that out, you can do it with a slide rule and many did. All we’re seeing now is what came of a bunch of idiot VCs deciding to try anyway, some of it is genuinely impressive but for the $400 billion that’s went into it it’s still catastrophically disappointing.
The business case for AI is dead simple: “Hey there Manager, hiring staff is an expensive pain in the butt right? What if we made a computer that did their job for a tiny fraction of the price and always said yes?”. Think about the mindset of the C-suite and upper management of any major company. They didn’t work their way up, they don’t have professional expertise in whatever it is their company actually does. They’re the accountants and lawyers who say things like “just make it work” and “do more with less”. AI as promised is their dream, they can fire all their staff and never have to think about physical reality again, they think they’ll be able to tell GPT 7.0 “just make it work” and it’ll be so gosh dang smart that it can.
It’s neat that OpenAI and a few other companies have built vaguely serviceable chatbots, it’s neat that Midjourney can spit out cool pictures, but it’s embarrassingly short of what investors were promised. Ford’s executive team doesn’t want to replace 40% of it’s level 1 customer support team with ChatGPT, it wants to be able to end a board meeting by saying “FordGPT, please design and start manufacturing next year’s F150 model, make sure it’s available in the usual colours plus moss green” and it’ll just do it. They don’t want a search assistant, they want J.A.R.V.I.S.
And GPT/ML can’t do that, no matter how many GPUs you throw at it.
Why now?
A ground rule for machine learning is “exponential investment for linear gains”. That means exponential increases in training time, in compute, in training data, in storage, in power. Neither the world nor the economy is infinite and many of these are bumping into both of those.
- They’re running out of training data. Llama 3 and GPT 4o’s input training data is a 2-digit percentage of “all words ever written by humans”.
- They’re running out of power.
- They’re running out of money.
- ML-focused GPUs aren’t seeing that much generation-to-generation improvement in terms of cost/power efficiency.
But most of all the investment levels are getting so high that even major financial institutions are starting to feel uneasy. Remember they don’t want a search assistant, they want Jarvis, and as we enter year 9 of “the next version will finally be it, trust me bro” with stagnating improvement and not a single financially viable product to show for it, they want to see more results before signing off on the next trillion dollars in capex.
Why Nvidia?
Nvidia’s importance within the AI ecosystem is grossly overstated. Their work on CUDA opening up GPUs for general-purpose computing tasks deserves a lot of credit for starting the broader machine learning gold rush, but that was a decade ago now. Today none of the software to train or run AI models is beholden to Nvidia’s hardware. All of Microsoft’s recent datacentre expansion has used AMD Instinct cards, not Nvidia. Microsoft are already developing their own chips in-house, as are Meta, and both Google and Amazon have had their own deployed for years. Nvidia is merely a chip designer, not a manufacturer. They can design and market all the chips they want but the sector has grown large enough that it’s a no-brainer for the real tech giants to vertically integrate.
The expectation is Blackwell selling faster than they can manufacture them for the next 2 years, the way Hopper has for the previous 2 years. Even if Blackwell is as big of an improvement over Hopper as promised (which it probably isn’t), and even if the delays aren’t indicative of serious manufacturing problems with TSMC’s 3nm process (which it probably is), it’s hard to see a winning formula.
Nvidia is a sound business long term, they’ve got a well-deserved dominant position in several markets, but their current valuation is based on future revenue estimates that not even the most bullish AI future could possibly deliver. With the confluence of so many factors, I think this is the quarter where the penny finally drops.
Are we in for a broader tech crash?
Personally, I think not. The tech bubbles that most of you will be familiar with are the financial ones comprised of crapy unprofitable startups that IPO and inevitably explode some time later. AI is more of a tech bubble where its just the new hot thing tech bros are obsessed with. Rather, it’s 40% of the new product development budget at the blue-chip tech giants. It’s a waste of money that will no doubt result in corrections at some point, but that money mostly came from the revenue of existing profitable products. It was wasted on some stupid bullcrap before, and it’ll be wasted on some new stupid bullcrap after.
The real hit in the medium term will be to “shovel sellers”. Nvidia, AMD, TSMC, Equinix, Dell, HP, Supermicro, etc. Any company for whom those capex numbers are a considerable source of revenue.