Arch-rivals Samsung and SK hynix are teaming up with OpenAI to build local Stargate datacenters in South Korea as part of the chatbot maker’s global infrastructure push.
Stargate is a $500 billion OpenAI datacenter infrastructure project to build 20 AI-focused datacenters for OpenAI’s ChatGPT. The intent is to have them operating by 2029. OpenAI is partnering with SoftBank and Oracle in the project.
Reuters reported earlier in October that OpenAI was working with Samsung and SK hynix on building two datacenters in South Korea, with an initial capacity of 20 megawatts, “after OpenAI CEO Sam Altman met South Korean President Lee Jae Myung and the chairmen of Samsung Electronics and SK hynix at the presidential office in central Seoul.” There will be an SK hynix/OpenAI datacenter at Jeollanam-do and a Samsung/OpenAI one at Pohang, according to Korea’s Chosun Daily.
We understand the two Korean memory makers may help fund part of these projects.
Samsung and SK hynix will scale up memory chip production to a target of 900,000 DRAM wafer starts per month. OpenAI does not specify whether this will be standard DRAM or specialized HBM, the type of stacked memory that provides more bandwidth and capacity for closely connected GPUs than that provided by socketed DRAM to x86 CPUs.
The KED Global media outlet said it was for HBM. We understand that the Stargate orders could include server DRAM, graphics DRAM, and even SSDs as well.
KED Global says 900,000 wafers/month is double the current global production capacity. We understand that SK hynix currently operates at 160,000 DRAM and HBM wafer starts per month.
That means both Samsung and SK hynix will have to build new HBM fabs. A knock-on effect could be that they build less DRAM, which is not as profitable as HBM.
Last year, we reported that SK hynix was increasing its memory manufacturing capacity, and would invest about ₩9.4 trillion ($6.8 billion) in building an HBM fabrication plant at the Yongin Semiconductor Cluster in Gyeonggi Province, Korea. The fab construction would start in March 2025 and finish May 2027. The intention is then to add three more plants one after the other to the cluster. Another new SK hynix plant, its Cheongju M15X facility, should be complete by the end of this year.
The Stargate deals mean more than ₩100 trillion ($72 billion) of incremental demand for Samsung and SK hynix if the OpenAI order rate is consistent at that level. In effect, OpenAI is securing its HBM and other memory component supply chains for the two South Korean datacenters planned with them.
Altman said: “Korea has all the ingredients to be a global leader in AI – incredible tech talent, world-class infrastructure, strong government support, and a thriving AI ecosystem. We’re excited to work with Samsung Electronics, SK hynix, and the Ministry of Science and ICT through our global Stargate initiative to support Korea’s AI ambitions.”
Samsung and SK hynix will also look to deploy ChatGPT Enterprise and API capabilities into their operations to improve workflows and support innovation.
OpenAI also signed a series of agreements to explore developing AI datacenters in Korea. These include a Memorandum of Understanding with the Korean Ministry of Science and ICT (MSIT) to evaluate building AI datacenters outside the Seoul Metropolitan Area, supporting regional economic growth and job creation.
There is a separate partnership with SK Telecom to explore building another AI datacenter in Korea, and an agreement with Samsung C&T, Samsung Heavy Industries, and Samsung SDS to assess opportunities for additional datacenter capacity in Korea.
Sustained chatbot demand
Will there be the sustained AI chatbot demand that could sustain such vastly expensive datacenter building and memory fabrication projects? Altman spoke to analyst Ben Thompson in a Stratechery interview after “OpenAI launched a number of new initiatives, including Apps in ChatGPT.”
OpenAI is “introducing a new generation of apps you can chat with, right inside ChatGPT. Developers can start building them today with the new Apps SDK, available in preview.”
“Apps will be available to all logged-in ChatGPT users outside of the European Economic Area, Switzerland, and the United Kingdom on Free, Go, Plus and Pro plans. Our pilot partners–Booking.com, Canva, Coursera, Figma, Expedia, Spotify and Zillow are also available today in markets where their services are offered starting in English. More pilot partners will launch later this year and we expect to bring apps to EU users soon.”
We might imagine that enterprise apps could become available inside ChatGPT as well.
Altman told Thompson: “Most people will want to have one AI service, and that needs to be useful to them across their whole life. And so you’ll use ChatGPT, but you’ll want it to be integrated with other services and so you need to have other apps inside of ChatGPT.”
ChatGPT will act as the front end, with other apps integrated but still visible within it. Altman said: “We’ve been thinking about apps and ChatGPT and how we’re going to unify the API and the consumer business for a long time.”
He remarked: “My favorite analogy for AI, my favorite historical analogy, is the transistor… I think it will just kind of seep everywhere into every consumer product and every enterprise product too.”
Where the transistor underpinned the development of hardware underlying IT, as a base component, ChatGPT will overlay everything as a top layer, not a bottom layer, in this view.
If this works, consumer use of ChatGPT would be enormous, and keep these Stargate datacenters busy.
Thompson asked about Altman’s enormous deals with Nvidia and AMD and TSMC, and queried: “Who do you expect to pay for it? Is this a matter of what these deals are about, you guaranteeing you’ll buy the output of it and you need these companies to invest?”
To which Altman responded: “Yes, I expect OpenAI revenue to pay for it.”
We would imagine this applies to the Samsung, SK hynix, and other deals in South Korea as well.
Thompson: “Do you see yourself almost as a financial guarantor who’s helping them secure better interest rates?”
Altman: “Yes, but we will help with financing. We are working on plans to be able to help with the financing these companies need at this kind of scale ahead of revenue.”
Thompson also asked specifically about the Samsung and SK hynix deals: “One of the surprising deals, just to get in the weeds a little bit, was with Samsung and SK hynix. Obviously memory is a massive constraint as far as building out these chips in the future, is this sort of tied into the AMD deal?”
Altman: “Give us a few months and it’ll all make sense and we’ll be able to talk about the whole – we are not as crazy as it seems. There is a plan.”
Bootnote
Samsung has more good news. Wedbush analyst Matt Bryson told subscribers that Jensen Huang has personally informed Samsung that its HBM3e has passed Nvidia’s quality qualification tests and that Nvidia intends to place orders. Bryson says the two companies “are now reportedly fine-tuning details such as volume, pricing, and delivery schedules.”
