저자: Kasey/Nick·ChipStockInvestor · 외국인
요약 보기
브로드컴과 OpenAI의 커스텀 AI 칩(ASIC/XPU) 협력 논의가 보도되며 주가가 약 3% 상승. 발표자들은 이를 새 소식이 아닌 진행 중인 흐름으로 봄. 브로드컴은 구글(2015~), 메타 등과 커스텀 실리콘을 협업하며 AI 데이터센터 네트워킹 역량을 보유. 이에 따라 포트폴리오에서 마벨 비중을 줄이고 브로드컴 비중을 늘렸다고 언급. 엔비디아와 브로드컴은 서로 다른 방식으로 AI 인프라 수요를 취함. 근거는 주가 3% 상승, 구글과의 장기 협업, 3대 고객 구조.
다듬은 글
The big news yesterday was that Broadcom is in talks with OpenAI and that shot the stock up about 3 percent yesterday. So let's break this down. What exactly is going on? This episode is sponsored by our friends at public. com. Are you dissatisfied with your savings account? Check out public. com. Public.
com has just launched its new high yield cash account, offering an industry leading 5. 1 percent APY. There are no fees, no subscriptions, no minimums or maximums, just 5. 1 interest on your cash. You can transfer or withdraw cash as often as you like, and you get to Up to 5 million in FDIC insurance, grow your cash and an industry leading a 5.
1 percent APY with a high yield cash account at public go to public. com to learn more. Yeah, and it looks like Broadcom holding on to that 3 percent gain here on Friday morning. It's been a hot stock. It and NVIDIA have been, the poster children, of semiconductor AI movement, especially since the start of 2023.
Of course, NVIDIA is up an insane amount, but hey, not bad Broadcom, 100%, compound annual growth rate is, pretty good, I'd say. OpenAI and Sam Altman have certainly had their share of controversy this year. We've had multiple reports of him looking for funding to build chip fabs and AI chip factories.
You can see that one on the left there, five to seven trillion dollars, which is just absurd amounts of money. Where does OpenAI come into this story along with Broadcom? Yeah, first of all, I'd like to just emphasize Kasey, what you said. 5 to 7 trillion to build a network of chip fabs, 5 to 7 trillion, yeah.
Emphasis on absurd. We covered this earlier this year. There is a battle heating up for custom designed semiconductors. NVIDIA has been getting more and more into this game, reportedly, and, of course they are, because Broadcom is actually one of the leaders in custom chip design, especially in this new era of accelerated compute infrastructure, so it only makes sense that OpenAI would also want to hold talks with Broadcom.
No doubt they're also probably having talks with NVIDIA also, probably with Marvell Technology group. This is not about reducing reliance on NVIDIA as much as it is a way for them to plan for their infrastructure needs in the future and maybe save some money along the way as they do it. Now, Kasey, let's maybe, talk a little bit about what.
These custom chips are, or ASICs, application specific integrated circuits, versus just regular old semiconductors. Merchant silicon is quote unquote, off the shelf silicon semiconductors. ASICs are custom designed semiconductors for specific workloads. NVIDIA has made GPUs a very common mainstream name in semiconductor land.
GPUs fall somewhere in the middle between these two types of semiconductors, merchant and ASIC. GPU's are accelerated computing for specific applications and for some general purpose work. It makes sense that both NVIDIA and Broadcom would be delving further into the application specific semiconductor design department.
They both have very deep portfolios of IP and patents. They're both fabless chip designers, NVIDIA's totally fabless, broadcom's mostly fabless. And then they're increasingly building more than just integrated circuits. They're increasingly also delving into the full blown computing device segment. NVIDIA in particular, building, these large platforms, the latest being Blackwell that will start shipping to customers later this year.
And then along with that, they're packaging up enterprise software for their customers that actually use that hardware. It just makes perfect sense that, as these customers, need to purchase more and more of these very expensive, accelerated compute systems, they're utilizing Increasingly specific types of software running very specific applications in data centers, they would circle back to the same companies, NVIDIA, Broadcom, and maybe to a lesser extent, Marvell Technology and say, Hey, can we work with you to develop a specific chip that runs a specific workload that we need?
And of course, open AI backed by Microsoft, one of the world's largest data center operators with Microsoft Azure is going to have those conversations with these companies. And the media keeps latching onto this, because, NVIDIA is selling these expensive accelerated compute systems. But the real thing going on here with Broadcom they're helping these companies design these large data centers and figuring out how to network together currently tens of thousands of GPUs.
And at some point in the near future, perhaps 1 million or more accelerated compute devices together. Broadcom is a leader in custom silicon design, work. You can see in this chart that they have provided. They have 3 top customers that they have been working with. 2 of them since 2020. They've recently added a 3rd.
Two of these customers we know. They've been working with Alphabet since 2015, and they also added Meta in recent years, and that third customer we're not sure of, but it could be Microsoft slash Open A.I. What could a relationship with OpenAI actually look like? This is what Broadcom has been designing for example, for Google, the Tensor Processing Unit, or TPU, for its data centers.
It's an accelerated processing unit. That's their moniker, they've given these things. The XPU versus the GPU, which is a bit more general purpose. And they've been doing this as a way to embed the actual accelerated compute processors more closely with the network itself, which is all the components that hook all of the processors together in these data centers to process these very, very massive workloads.
You may recognize these slides from Broadcom. We pointed them out when they had an AI conference earlier this year, but Nick, you can explain to us why OpenAI would be interested in this type of processor. It seems, Sam Altman is pretty vehement in the belief that the world is not building enough accelerated computing or AI infrastructure.
That's probably a very self serving viewpoint for his company, OpenAI and for the chat GPT and other AI products. To his point, this is what Broadcom's infrastructure has looked like, for example, at Google. So they have these existing data centers, and tapping Broadcom to design things like computing offload.
So you have these data requests coming into the data center via the Internet, and most of those workloads can be handled by the existing data center infrastructure itself. And maybe some of those requests coming in are some sort of AI compute workload. So they split that request off to a different branch of the data center.
And that's where the XPUs and Broadcoms networking chips handle that AI work. And then the old compute workload continues on to the existing infrastructure. And so it seems very obvious that, OpenAI and Sam Altman have this vision of tapping into this future need for lots of more data centers that are purpose built just for data.
AI training and just for AI inference, which is when the AI has already been trained and it gets put into use and customers start actually using it. So it seems like that's probably the conversation they're having with Robcom to see if they can get some help building these AI specific data centers. We can come to the conclusion then that this isn't really news per se.
This has been an ongoing discussion with these companies like Broadcom and Marvell With, our portfolio specifically, we've shifted away from our Marvell Technology stock and put more of that position into Broadcom. Does that change still make sense with this news item? Yeah, you can catch that video right here.
Link to that. As well as the one you referenced from earlier this year, it was actually in March, it was the same week NVIDIA had its GTC, Broadcom had its own little AI event. I think ultimately, this is being pitched as a new thing, a new development. Really, this has been going on for a while now, folks.
Broadcom, not as well known a name as NVIDIA. Very much at the absolute forefront, cutting edge of developing these accelerated compute systems for especially large customers, but also providing plenty of merchant silicon off the shelf stuff, maybe appeals to smaller customers that can't shell out the big research and development budget that a Microsoft and an Open AI can to make a custom chip.
That merchant silicon very prevalent at Broadcom as more AI creeps into existing software that's out there already. We'll start seeing lots of new AI powered features showing up in your software products in the next couple of years, Broadcom is behind the scenes working on that. That I guess answers the question was it a good idea to divest ourselves of Marvell and put more money towards Broadcom.
Broadcom Broadcom is a leader in this custom silicon part of the market and obviously big companies are still continuing to partner with them. Another newsworthy item that we'll cover next week is what's happening with CrowdStrike and Microsoft. Make sure you check out the links below, we've got Semiconductor Insider for 5 a month that gets you access to our stream of consciousness on our Discord channel. We'll see you very soon on Chip Stock Investor.
Comments 0