Three telcos on three continents just started renting and pouring concrete for compute in the same week — and not one of them shipped a model. If you're just joining, South Korea's been turning its AI push from policy into megaprojects: a planned 18.4-gigawatt nationwide data-center buildout by 2035, domestic AI chips, physical-AI ambitions. SK Group had already sketched huge targets across data centers, semiconductor supply, and a new southwest chip cluster. The open questions were power, siting, and execution. This is the AI Daily Briefing. Today: SKT names a region and a dollar figure, SoftBank rents at a scale that used to be a grid event, and — the twist — Washington wants a slice of OpenAI. From The Korea Times:
Korea's SK Telecom said Friday it will invest 140 trillion won ($91.5 billion) to transform the southeastern region of Yeongnam into Asia's artificial intelligence (AI) hub, with the construction of massive data centers there. Under the plan, SKT will start work on a 100-megawatt hyperscale AI data center in the port city of Ulsan, with the goal to start operations by the fourth quarter of next year.
SK Telecom is putting 140 trillion won — call it $91.5 billion — into turning Yeongnam into an AI hub. And here's the part that catches me: the lead isn't a chipmaker, isn't a ministry. It's the carrier. Right, and that's the update to Korea's buildout we've been chasing all week — we finally have a company and a region, South Gyeongsang. That's firmer than the phase-level targets we had before. An operator making a physical-infrastructure bet. Which means the utilization math I keep chewing on now sits on a very different balance sheet — SKT has to fill those halls with paying inference, not just cut a ribbon in Jinju. And notice what SKT isn't doing here: just renting capacity. It's building and controlling. That's the sovereign-model logic in corporate form: own the halls, own the stack. Hold that thought, because the SoftBank number coming up is the exact opposite instinct. From The Japan Times:
SoftBank Group and its telecom unit will start renting artificial intelligence computing resources to U.S. companies next fiscal year, using a growing pipeline of data centers to compete with the likes of CoreWeave and Nebius Group. The “neocloud,” or AI-tailored-cloud, venture plans to ramp up its computing resources to eventually supply data center capacity at a scale of 10 gigawatts by around 2030 for large-scale AI model training and inference, according to Junichi Miyakawa, head of the telecom unit.
SB Neo. SoftBank setting up a neocloud to rent AI compute to U.S. companies at ten-gigawatt scale — competing with CoreWeave and Nebius. And they're not building the models, they're renting the picks and shovels. Exactly. Right after we just heard SKT put $91.5 billion into owning its own hyperscale build in Yeongnam, SoftBank goes the other direction — rent the layer instead of owning it. Ten gigawatts of rental. Three years ago that number would've been a national grid story. Now it's a procurement posture. And here's what it costs them: rent the compute and you don't control the inference stack. That's the vulnerability a sovereign build is designed to dodge. Miyakawa's telling Bloomberg this thing can generate profit on a different order of magnitude — the whole bet is that rental rates stay cheap. That's the crack in the plan. Ten gigawatts only pencils if the price curve keeps falling. The moment inference rental rates tick up, you don't just renegotiate a line item; the whole model wobbles. Here's Yangtzeer observers at The Yangtzeer:
China Mobile, one of the country’s three largest telecom carriers, has begun operations at its intelligent computing center in Zhejiang’s Jiashan, completing the first phase of deliveries for a roughly 5 billion yuan ($737 million) project. The facility, handed over for commercial operation on June 30, forms part of China’s national computing infrastructure strategy.
China Mobile handed over Jiashan for commercial operation on June 30 — first phase, 15,000 petaFLOPS live. After a week of megabuild press releases, this is the rare one with a delivery date attached. And that's the third telco this week moving on pure compute — SKT in Yeongnam, SoftBank renting in the U.S., and now China Mobile bringing East Data, West Computing online. Nobody's calling that a pattern yet. The campus is designed for 50,000 accelerators, but phase one is 15,000 petaFLOPS. So the interesting number isn't the ceiling — it's how much of those government imaging and traffic workloads actually runs at capacity, versus sitting idle waiting for demand. Right, and unlike SoftBank renting, China Mobile owns this end to end — carrier, campus, inference. That's the sovereign-model instinct playing out in a third geography, same week. Ashley Belanger, writing in Ars Technica:
OpenAI CEO Sam Altman is reportedly in active talks with the Trump administration about the US potentially acquiring a 5 percent stake in the leading AI firm. Insider sources told the Financial Times that these talks are in “early stages,” but Altman “has argued that giving the public a financial stake in the company is the best way to share the upside of AI.”
So all week it's been power numbers and gigawatts — and now the Financial Times says the U.S. government's in early talks to take a 5 percent stake in OpenAI. That's the first story this week that isn't about a substation. And notice the framing Altman's using — that giving the public a financial stake is 'the best way to share the upside.' That's a pitch, not a policy. Five percent, early stages, and they're reportedly having the same chat with Google and Meta. Which is exactly why I'd give this the leaderboard treatment. If you're floating the same 5 percent stake to three companies at once, 'strategic position' starts sounding generous. It's a press release with equity attached. Here's what gets me, though. This is the same government that flagged Anthropic's models behind a Commerce export gate on a Tuesday. Now it might co-own the shop across the street. Regulator, gatekeeper, and shareholder — pick a lane. And it blurs the line I've been drawing all week. I kept separating sovereign balance sheets — like SKT, like the Korea build — from corporate ones. If Treasury holds OpenAI paper, that distinction just got a lot muddier at the very top of the stack. All week we've been watching who controls the compute layer. This is the first move on who controls the model layer — and it's the government reaching for it with a minority stake and no clarity on what it buys. From Georgia Butler at DataCenter Dynamics:
Microsoft has signed a land purchase agreement for a site in Grevenbroich in North Rhine-Westphalia, Germany, which could host a data center. The company revealed on Wednesday, July 1, that it has signed a conditional agreement and aims to use the property to expand its data center footprint in the area.
And rounding out the infrastructure hour: Microsoft signs a land purchase agreement in Grevenbroich, near Cologne. Conditional agreement, subject to a zoning plan the city hasn't drawn up yet. No size disclosed, no capacity disclosed, no location within the site disclosed. What we've got is Microsoft would like to buy some land if the paperwork clears. After SKT's ninety-one billion and SoftBank renting ten gigawatts, this one's practically a rounding error. Microsoft buying a plot in North Rhine-Westphalia — welcome to the small end of the week's scale. But it's a different geography sitting on the same demand curve. Everyone else we've covered today is telco or sovereign; this is the hyperscaler quietly doing the boring version — land options in Europe, contingent on zoning, no announcement fanfare. That's the part that actually gets built. Right — Georgia Butler at DCD framed it plainly: expand the footprint in the area, subject to Grevenbroich developing a zoning plan. That last clause is the whole story. A signed conditional agreement means the lawyers moved; the shovels haven't. Got thoughts on today’s briefing, an AI story we should watch, or a correction we need to make? Send us a note at aidailybriefing at lantern podcasts dot com.
What we’re watching next: SKT is aiming to bring its first 100-megawatt Ulsan hyperscale AI data center online by the fourth quarter of 2027. And SoftBank expects SB Neo to start renting AI compute to U.S. companies next fiscal year, with 10 gigawatts targeted around 2030.
You’ll find links to every story we covered in the show notes, so take a look at anything you want to dig into. That’s AI Daily Briefing for today. This is a Lantern Podcast.