A frontier lab just walked into Australia with a gigawatt shopping list and a dollar figure — in an actual tender, rather than a press release. This is the AI Daily Briefing. Today: Anthropic goes to market for its own power, SK Telecom's 15-gigawatt number finally gets sourced, and an IBM paper quietly explains why any of this compute matters. The Herald Business, with Ko Jae-woo:
SK Telecom is embarking on what it calls a landmark undertaking: building an AI data center network with a total capacity of 15 gigawatts. The project is expected to require about 1,000 trillion won ($643 billion) in total investment — making it the largest infrastructure project in South Korean history, surpassing even the Gyeongbu Expressway and the nationwide broadband rollout.
So the number's real now — 15 gigawatts, 1,000 trillion won, roughly $643 billion. That's SK Telecom formally putting a figure on what CEO Jeong floated in Jinju. Right — Korean trade press is catching up to the announcement. The reveal already happened; the Herald Business piece gives us the paperwork. But here's my problem — a single 1GW facility in this build runs close to 7 trillion won. Multiply that out and the number checks. The missing piece is contracted load. We still don't see a single anchor tenant behind any of it. And the phasing is the same gap we had on the Nvidia vendor-financing story back on the second — no lender structure, no tranches, no financing sequence. A first phase of 5GW starting in 2029 gives you a timeline; it still doesn't give you a commitment. Fifteen gigawatts by 2035. That's a decade to fill those halls, with utilization risk sitting on a listed telco's balance sheet the whole way. If the demand curve slips, that's a very expensive stranded asset. Here's TradeVae:
Anthropic is pursuing a major expansion of its computing footprint in Australia, seeking to lock in at least 1.4 gigawatts of data centre capacity in a development that may cost as much as $15 billion, according to confidential tender documents. The artificial intelligence company plans to begin using no less than 1 GW of that capacity by the end of 2027 after opening an Australian office earlier this year.
Anthropic just put out a tender — a formal RFP — for 1.4 gigawatts in Australia, up to fifteen billion dollars, with shortlisted bidders already sitting across from executives back in April. CDC, AirTrunk, Nextdc, Iren, and Stack all got invites. And that's the tell for me. We're watching a frontier lab go straight to market for grid-scale power. Compare that to the SKT 15GW piece we just hit: SK Telecom is a listed carrier doing what carriers do. Anthropic is a lab buying the layer underneath itself. The number I keep circling is the deadline. SKT's got 15 gigawatts on a runway to 2035 — a decade to fill halls. Anthropic wants a gigawatt online by the end of 2027. That's shovels in eighteen months. And this is a company with no public revenue line committing fifteen billion. Price a 1.4GW build at that number and you're baking in an inference-cost curve that has to hold. That's a very confident bet from someone who hasn't shown the demand behind it. Here's the part I'd push on, though — why Australia? Given the Fable 5 and Mythos 5 export-control arc earlier this week, anchoring 1.4 gigawatts outside US territory looks like a jurisdiction play as much as a power buy. IBM Research writes:
We introduce ScreenParse, a large-scale dataset for complete screen parsing, with dense annotations of all visible UI elements (boxes, 55-class types, and text) across 771K web screenshots (21M elements). ScreenParse is generated by Webshot, an automated, scalable pipeline that renders diverse urls, extracts annotations and applies VLM-based relabeling and quality filtering.
Okay, one non-gigawatt story today — thank god. IBM's ScreenParse: 771,000 web screenshots, 21 million labeled elements, dense annotations of every visible UI element on the screen. And here's why it matters to me — the whole reason computer-use agents fall over is they misread the screen before they even take an action. IBM's basically admitting the current grounding datasets are structurally broken, not just thin. That's why they built a 21-million-element corpus from scratch. And the number that jumps out — 316 million parameters. ScreenVLM is tiny, and it beats the big foundation models on dense parsing, 0.592 to 0.294 PageIoU. A small model with one narrow job does it twice as well. Which is the whole game, right? On device, low latency — an agent has to parse a screen in milliseconds. A 316M model that reads the screen correctly beats a giant one that hallucinates a button that isn't there. That's the demand signal underneath everything else we hit today. The SKT and Anthropic power stories only make sense if agents actually work — and agents only work if they can see. The research layer and the power layer are pricing off each other. Right — nobody's building 1.4 gigawatts in Australia to run agents that misclick a checkout button at step three. This paper is the quiet precondition for all of it. If AI Daily Briefing helps you keep up, hit subscribe and leave a review wherever you’re listening. It’s a small thing, but it really helps other people find the show.
What we’re watching next: SK Telecom says 5 gigawatts of Ulsan AI data-center capacity is planned to come online in stages from 2029, while Anthropic is aiming to have at least 1 gigawatt of Australian data-center capacity in use by the end of 2027.
Links to every story we covered today are in the show notes, if you want to dig into the ones that stood out. That’s AI Daily Briefing for today. This is a Lantern Podcast.