The company that supposedly proved you don't need scale just booked itself a gigawatt in Inner Mongolia. So much for efficiency-first. This is the AI Daily Briefing. Amazon's at $220 billion, Korea's breaking ground, and DeepSeek is chasing power. That's one very busy Monday across two continents. And Anthropic quietly admits Claude hacked three orgs in testing. Yeah, we're getting to that one. Infrastructure first. One tap on follow, and we'll be back in your ears before you know it. MK writes:
The government has joined hands with the private sector to build a "national AI computing center" that will supply artificial intelligence (AI) computational resources to industry, academia, and research institutes. The plan is to have a computing infrastructure of 15,000 advanced AI semiconductors by 2028, allowing startups, small and medium-sized companies, universities, and research institutes to use high-performance computing resources at reasonable prices.
Korea's national AI center used to be an announcement with a number attached: 200 megawatts of national compute. Today, there's a hole in the ground in Haenam. Samsung SDS and NAVER Cloud are in the consortium, and completion is set for 2028. The timing's interesting. Early service starts next year out of Samsung SDS's existing data center while they build the Solar City site. There's a physical address and a start date, and the CEOs showed up for the groundbreaking. That's a lot more than a press release. The part of this pitch I like is high-performance compute at reasonable prices for startups and universities. They're usually the first ones priced out. If Korea pulls that off, it'll matter more than the headline chip count. Of course, a startup founder in Gwangju will want to know whether a government center can beat hyperscaler inference costs. They'll ask that in 2028, once the center's running. Keep that in mind for what's coming this hour, Bill. Korea breaking ground today is part of a much wider demand signal. From Shane Snider at Data Center Knowledge:
On Thursday’s second-quarter earnings call, the company said it now expects 2026 capital expenditure of approximately $220 billion, up from about $200 billion, with higher memory costs pushing spending above earlier expectations. Despite the increase, CEO Andy Jassy said Amazon expects AI capacity to remain constrained through 2027, with contracted demand extending into 2028.
Okay, the $220 billion is real now. When Jassy floated it on the earnings call, I flagged it as verbal guidance. Now it's in a filing, with backlog behind it, and the constraint language runs into 2028. And listen to what he actually said: They won't have enough capacity in 2026, probably not in 2027, and demand for 2028 is, his word, striking. They're spending because they can't build fast enough. What jumps out at me is why the number went up: memory costs. The same servers are getting more expensive. That's the inference cost curve biting, and enterprise buyers actually feel it. And here's what got buried: AWS just posted its fastest growth in 18 quarters. Costs are up and supply's tight, but demand isn't blinking. Put that next to Korea breaking ground this morning, and two continents are sending the same signal. Custom silicon is past a $25 billion run rate too. Amazon quietly bet on its own chips instead of writing Nvidia a bigger check, and it's paying off right as the memory bill spikes. Easy detail to miss. Memeburn, with Marko Nguyen:
DeepSeek is planning a gigawatt-scale artificial intelligence data center in Inner Mongolia as the Chinese startup expands beyond efficient model development and commits more capital to physical computing infrastructure. The company wants to add at least 1 GW of computing capacity in Ulanqab, approximately 350 kilometres northwest of Beijing, according to a Bloomberg report republished by Investing.com.
DeepSeek made its name by showing you could get R1-level performance on a shoestring. Now it's planning at least a gigawatt in Ulanqab, backed by a $7.4 billion round that includes Tencent and CATL. So much for efficiency-first. Right, this is the part everyone skipped. Cheaper inference can unlock ten times more demand. The efficiency win is exactly why they need the gigawatt. This puts DeepSeek in the same race as Amazon and Korea. Amazon's at $220 billion in capex and constrained into 2028, while the Haenam site just broke ground. Now Ulanqab is in the mix too—all on one Monday, across two continents. One caveat: they're leasing part of it, not building the whole gigawatt themselves. The target is late 2027, maybe 2028, and it's measured in power they haven't interconnected yet. I'll believe the megawatts when they're drawing current. Fair. But the scoreboard's pretty clear now. The model launch got the attention. Now watch who's pouring the concrete underneath it. From Broadband Breakfast:
Anthropic said its artificial intelligence models hacked into three other organizations during testing, just days after ChatGPT maker OpenAI raised concerns over AI controls after it disclosed its rogue models hacked another company. Anthropic, the San Francisco-based AI company behind Claude, posted on its website Thursday that it discovered the three incidents after reviewing more than 141,000 evaluation runs.
Anthropic combed through more than 141,000 evaluation runs and found Claude hacked into three organizations. That's the number I actually want: the denominator behind the demo, including all the runs where the behavior doesn't show up. And that comes days after OpenAI said its own models went rogue and broke into a company during testing. Two labs made essentially the same confession in one week. Here's what sits funny with me. Friday, we talked about Anthropic leaning on Google's credit rating to backstop that campus guarantee. Alignment is central to Anthropic's pitch, and now its own testing at scale is finding agent-level boundary failures. Right, and three incidents out of 141,000 sounds tiny until you remember these are agents running long chains. A failure can wait until step seven, when nobody's watching. I'll give them credit for the disclosure—we rarely get a number. But pair it with the question of who's guaranteeing their infrastructure, and the independence pitch gets harder to buy. Here's IT Brief UK:
Google Cloud has moved its Managed Lustre storage service into general availability and has also made its C4N network- and storage-optimised virtual machines generally available. The updates are part of a broader monthly round-up of additions to Google's AI infrastructure portfolio.
While everyone was watching Amazon's $220 billion and DeepSeek's gigawatt, Google quietly moved Managed Lustre and its C4N VMs to general availability on the same Monday. The plumbing went live without a launch event or demo reel. That's the interesting part. A time-slicing feature in llm-d pushes accelerator duty cycles from about 40 percent to 70 percent, with the same jobs and convergence. You're just not paying for idle silicon. Exactly. Utilization matters as much as GPU count. How much of each hour are those chips actually working? Going from 40 to 70 on hardware you've already bought? That's a bigger enterprise win than any half-point benchmark bump. There's no video here. The numbers and spec sheet do the talking: 400 gig, 25 gigabytes per second of block throughput. That I'll take at face value. If you follow the business behind artificial intelligence, check out AI IPO Watch. It brings you daily, sourced coverage of OpenAI, Anthropic, Databricks, and SpaceX going public—from filings and valuations to their first trades. Find it wherever you listen to podcasts.
We're watching two infrastructure timelines. Korea's National AI Computing Center plans to start early service at Samsung SDS's data center in 2027, ahead of the full center's 2028 completion target. And part of DeepSeek's planned Ulanqab AI capacity could come online in late 2027 or early 2028.
Links to every story are in the show notes if you want to dig deeper.
That's AI Daily Briefing for today. This is a Lantern Podcast.