Anthropic just made compute somebody else’s credit problem. Who’s holding the bag? This is AI Daily Briefing. Today: bank-backed compute, a twenty-gigawatt promise, and one research result that actually shows the misses. Start in Norway. Anthropic committed $10 billion over six years to Volta Infra Holdings, and JPMorgan supplied $1.3 billion in letters of credit. This one's from Bytes Europe:
Anthropic, the maker of Claude, has signed a $10 billion, six-year compute procurement agreement with Volta Infra Holdings — a startup founded in January 2026 that has been in existence for roughly seven months — giving the frontier AI lab access to 121 IT megawatts of Nvidia Vera Rubin capacity at Bitdeer Technologies Group’s Tydal data center campus in Norway, according to Bloomberg.
Volta was founded in January. Seven months later, it now sits between Anthropic and a $10 billion, six-year GPU commitment. That’s a credit wrapper with a very expensive day job. And JPMorgan affiliates are putting up $1.3 billion in letters of credit for 121 IT megawatts at Bitdeer’s Tydal campus. Anthropic’s independence pitch keeps picking up outside balance sheets. Bitdeer still owns the site; Volta supplies the capital stack; Anthropic buys the compute. Fine—until workload forecasts miss. Then I want to know who eats the cost of 121 megawatts sitting under a called credit facility. Bitdeer’s shares jumped as much as 23% Tuesday, so investors got the structure right away. The frontier-lab buildout is becoming project finance with Nvidia racks attached. This one's from Nile1:
SpaceX is targeting up to 20 gigawatts of artificial intelligence computing capacity by the end of 2027, accelerating a massive infrastructure buildout backed by $18.4 billion in second-quarter capital spending. Chief Executive Officer Elon Musk disclosed the expansion targets during the company’s debut earnings call as a public company on Tuesday.
SpaceX has 1.4 gigawatts of nameplate compute today and says 20 by late 2027. That’s a fifteen-fold build in roughly eighteen months, and “tentative” carries a lot of weight there. The hard number is $18.4 billion in quarterly capex—nearly $16 billion for AI infrastructure. The 20-gigawatt figure came from Musk on an earnings call, with no filing-level build schedule or interconnection detail attached, so keep it in the target column. And SpaceX says it’ll get a “very significant percentage” of Nvidia’s GPU shipments next year. Fine—then show us how 20 gigawatts gets power, cooling, permits, and substations. GPUs are the easy noun in that sentence. SpaceX bought xAI in February; six months later it has $14.1 billion in contracted cloud agreements with Anthropic and Google among the customers. This is a rocket company becoming a hyperscaler at hyperscaler speed. This one's from Lumida News:
CoreWeave announced plans to build three data centers in Indonesia with a combined 360 megawatts of capacity — its first entry into the Asian market — targeting labs, startups, and enterprise customers across Southeast Asia; the facilities are expected to come online in 2028 and will be operated by a local team; CoreWeave did not provide a specific cost estimate but said it will spend “billions of dollars” on the project.
CoreWeave says three Indonesian sites, 360 megawatts, online in 2028—and “billions” in spending. I want to see the interconnection plan, because a 2028 target means nothing until those substations have a date and a signed path to power. And Indonesia is a meaningful choice: data-localization rules make local capacity more valuable than serving the region from Singapore or the U.S. CoreWeave is effectively sold out for 2026 across 49 sites, so demand is chasing physical infrastructure again. Their stock jumped 19.49% on the announcement, so investors heard 360 megawatts and filled in plenty of blanks themselves. I’d like to see who takes the construction and power risk—and what happens if Southeast Asia demand arrives later than the GPUs. Indonesia makes one thing clear: this buildout now reaches far beyond North America, Europe, and Korea. Grid queues are a global constraint. IT Brief Asia, with Sean Mitchell:
Sharon AI has signed a five-year AI cloud services agreement worth USD $373 million with a global artificial intelligence platform, increasing the share of its AI Factory capacity contracted to customers.
Sharon AI has now contracted 120 of its 132 megawatts in Australia. With just twelve megawatts left, this operator is nearly sold out. And the $373 million deal runs five years, with revenue starting in Q1 2027. Good—there's an actual customer commitment behind the build. But contracted capacity doesn’t tell us chip utilization; I still want to know how busy those B300s are at 3 a.m. The initial deployment is 2,048 Nvidia Blackwell Ultra B300s, while Sharon says it plans for 64,000 GPUs by mid-2027. Keep the scale straight: the first tranche is concrete; the larger fleet is still a build-out plan tied to demand. Australia has joined the same reservation rush we just saw in Indonesia. The customer is only identified as a global AI platform, but 120 megawatts under contract says somebody decided capacity was more valuable than waiting for a better price. arXiv, with Hao Shen:
We present MechGeo, a Mathlib native agentic framework that jointly addresses faithful autoformalization and certified proof construction for Euclidean geometry. In this framework, GeoFormalizer represents informal problems in GeoIR, deterministically translates them into Lean 4, and iteratively repairs candidate statements using structural diagnostics and semantic evaluation.
MechGeo gives us the denominator everybody likes to leave out in agent demos: 29 of 43 historical IMO geometry problems proved, and the other 14 got Lean-checked counterexamples. A system admitting where it failed is already more useful than most victory laps. And every proof is kernel-checked in Lean 4, including the repaired statements after expert correction. We just spent a segment on ten-billion-dollar compute commitments; this is what a technical result looks like when somebody brings the receipts. Twelve of 14 on Lean-IMO-Bench, two formally refuted—great. Now the hard part: can this method keep that pass rate when the problems stop being a curated set of 43 Olympiad geometries? This is exactly the kind of workload where more compute could eventually earn its keep. If you’re enjoying AI Daily Briefing, please leave us a review or subscribe wherever you’re listening. Reviews help other people find the show, and your support helps us keep bringing you the latest in AI.
We’ll be watching whether SpaceX surpasses 2 gigawatts of nameplate compute by the end of 2026. We’ll also track CoreWeave’s three Indonesian data centers as they head toward 2028, and Sharon AI’s cloud-services revenue as it begins in the first quarter of 2027.
Links to every story are in the show notes. Take a look at the ones that caught your attention. That’s AI Daily Briefing for today. This is a Lantern Podcast.