Taiwan AI Conference Slides, Nvidia Fraud Hoax, Huawei, ARM IPO, and Intel
A few quick takes before SEMICON. I'll have an ARM IPO piece coming.
Hello readers. I’m sorry it’s been a bit. I have been on vacation, and now I’m in Taiwan. Last Sunday, we had an incredibly successful AI and Semiconductor conference. There were great slides by Warren Lau and Dylan Patel on AI content, the death of American Lithography by Asianometry, and Franklin Lin talked about the future of Andes Technology. I presented about China’s EV and lagging edge push.
Here’s a photo of the turnout! Thanks to all the readers who came; meeting you in person was so nice.
I’m happy that ~160+ people showed up, especially with a last-minute heads-up. I am not free to share all the slides, but here is the portion of the slides I made. If there’s interest, Dylan and I will replicate this again in the US. Maybe a few CEOs and a few interest tracks? Who knows! My slides are below.
I think this story that I presented is still early. We are likely years away from real impact, but I will write about it soon. There’s a long and a short angle here. I’ll be writing about it soon.
Nvidia Fraud Hoax
The funniest meme on the internet is the incredulous “Nvidia is a fraud” meme. Given the recent rise in Nvidia’s stock, some brains are breaking on the internet. The thread/train of thought seems to be spreading, and I’m here to try to stop it. I would say the source of the meme is this viral thread about Nvidia.
I’ll try to break the Nvidia is a fraud reasoning into a few points.
CoGs didn’t rise with revenue
CoreWeave is a large related party and is a channel stuff akin to Valeant
I’m not spending much time on this because it’s stupid, but they wrote down inventory during the previous year during the crypto demand bust. In their 2023 fiscal Q3, they wrote down a meaningful amount of inventory related to China.
So as we highlighted in our prepared remarks, we booked an entry of $702 million for inventory reserves within the quarter. Most of that, primarily, all of it is related to our data center business, just due to the change in expected demand looking forward for China. So when we look at the data center products, a good portion of this was also the A100, which we wrote down.
Now, they are selling the inventory they wrote to zero, hence zero inventory cost! So this explains the “zero increase in CoGs”; it’s explained by this pesky thing called accrual-based accounting. Just because you don’t understand something doesn’t make it nefarious.
CoreWeave is the weirder one. Nvidia did invest in CoreWeave, but it’s also invested in every single AI company. Part of the reason is that the invested companies try to convince them to use GPUs. Another reason is that CoreWeave is frankly taking the blood off the hands of Nvidia. Nvidia is at the most profitable and strategic intersection of the AI value chain while being a service provider like CoreWeave is not. Why would Nvidia do all the low-value margin services like customer support and keeping a data center up?
So they partnered with CoreWeave, which was previously a GPU cloud for crypto mining. CoreWeave is simply a new computer OEM with a different wrapper. In the current paradigm of data centers being the new computing unit, the 1000s of GPUs strung together are the product being sold, not individual cards. CoreWeave is the new SuperMicro for data center scale deployments of GPUs, the next strategic computing unit.
There’s also a lot of condemnation about the financing of the GPUs backed by the GPUs as collateral. The rationale is pretty simple, in my opinion, because the payback period of the H100 is under a year at current CoreWeave pricing, and they have availability. Pretty much the rationale of the underwriters is that if CoreWeave somehow defaults, they can seize the H100s and sell them. If they don’t, they can probably make the entirety of the investment back in the year, so it’s a win-win either way. The bet is that pricing is tight for half a year, and I think that’s a super-rational bet. Heads the underwriters win, tails the underwriters also win. It’s not as nefarious as it sounds.
CoreWeave is also a strategic snub to Amazon. I think this is amazing and clever strategic thinking, and to conflate a plan to neuter a cloud with “fraud” is a pretty astounding lack of understanding. The people who seem to be eating up this “fraud” narrative are the dumbest people I’ve seen. I think that Jensen will explode the channel, but it won’t be CoreWeave specifically that does it.
Moreover, the actual cost of egress for data to train models is pretty cheap. The entire model size is very easy to transfer from cloud to cloud, so it does make sense to move your model to a specialized cloud, train the hell out of it at the lowest cost of computing, and then bring back your final work to your preferred cloud provider.
Maybe, just maybe, it’s a lot simpler than this. What I think is happening is this amazing meme by @eliant_capital on Twitter. The thread's author was originally short, but he cannot change his mind, so he has to find a new way to justify it. Think of it as a doomsday cult whose doomsday never came; it’s time to pick another date.
Before I go, I want to say that Nvidia’s recent results are amazing. But so is everything that happens at the Internet scale. This is a perfect example of the sheer size and complexity of the internet being too large for humans to appreciate. So, if something is too big for someone to understand, rationalizing it and creating patterns that fit your viewpoint of the world is how your brain protects you. This is what’s happening with this odd conspiracy theory. On to the rest of the update, aka the valuable stuff, not the conspiracy theories.