Everybody wants AI compute. The fight is over who gets the electricity—and who gets locked out. If you're just joining us, the AI compute buildout has moved beyond model labs renting GPUs. It’s now about multi-year commitments for power, data centers, and specialized cloud capacity. We’ve tracked that race across AWS, Alibaba, ByteDance, DeepSeek, OpenAI, and Anthropic—including OpenAI becoming Firmus Technologies’ anchor customer for dedicated compute at two Malaysian AI Factory sites. That pushed Firmus’ contracted capacity past 900 megawatts. This is AI Daily Briefing. We’re starting with Microsoft’s capacity crunch, then SpaceX and a bankrupt airline’s data sale. If this story matters to you — Big Tech AI lease obligations — hit follow. We'll be back on it soon. From Martin Holloway at Financial News:
Microsoft Corp (MSFT) plans to more than triple its global data centre capacity by 2032, according to a Bloomberg News report published on 10 September 2026. The push follows shortages that have already forced the company to turn away some AI and cloud business, Bloomberg said, citing people familiar with the plans.
Back to Big Tech’s compute commitments: Microsoft is reportedly targeting more than 38 gigawatts of data-center capacity by 2032. They’ve got roughly 12 today, so it’s more than tripling. And Bloomberg says they’re already turning away AI and cloud customers for lack of capacity. Good. Finally, a demand signal with consequences. Every customer Microsoft turned away had a workload, a budget, and somewhere else to take both. About a third of that 38-gigawatt plan is slated for AI-specific chips, per the Bloomberg reporting aggregated by Techmeme. Model-launch chatter makes compute sound like a menu item. Microsoft’s treating it as scarce industrial capacity—because it already is. I do want the build schedule before I award them the trophy. Earlier this year Microsoft walked away from about two gigawatts of planned US and European projects, then Amy Hood said capacity would stay tight through 2026. Forecasting demand at this scale is apparently a contact sport. Here's Gate:
According to SpaceX CFO Bret Johnsen, speaking at Goldman Sachs' Communacopia + Technology Conference on Thursday (Sept. 10), the company plans to expand its terrestrial AI compute capacity from just over 2 gigawatts at the end of 2026 to between 5 and 10 gigawatts in 2027.
Five to ten gigawatts in 2027 is a hilariously wide range for a plan Bret Johnsen pitched at Goldman. At the low end, SpaceX more than doubles from just over two gigawatts; at the high end, it’s building a small country’s worth of compute. And it’s Nvidia-exclusive. We just hit Microsoft’s 38-gigawatt target; now SpaceX is putting a five-to-ten-gigawatt expansion behind one hardware vendor. Nvidia is collecting an awful lot of rent before these models answer a single question. The demand case has more behind it than most: SpaceX says its Google and Anthropic hosting agreements are worth $26 billion a year, plus a newer deal expected to add $13 billion in annual recurring revenue starting December 1. But with exclusive hardware, Nvidia’s price curve becomes part of SpaceX’s margin model. Johnsen says they can see the power, facilities, and permits. Good—because at ten gigawatts, the chips are only one bottleneck. The grid does not accept a Goldman conference slide as an interconnection request. Ars Technica writes:
Last week, an AI training data lab, Micro1, put in a last-minute competing bid that it claims resolves all objections to the data sale. Outbidding Google by 25 percent, Micro1 offered $12.5 million in cash, while promising that if debtors pick Micro1 over Google, the bankruptcy proceedings can avoid any further delays.
After all those gigawatt numbers, here’s $12.5 million and a bankruptcy hearing—and honestly, this is where the training-data bill gets real. Micro1 is 25 percent above Google’s bid, but it wants to de-identify Spirit’s data itself. That’s exactly what Spirit’s debtors objected to. They were prepared to take a lower fallback bid because independent third parties would do the scrubbing. So Micro1’s claim that it resolves every objection has a fairly large asterisk. And Springshot says, hold on—Micro1 won’t segregate its IP or trade secrets. It only promises not to build a rival product or sell the data to one. Great. The failure mode is already described in the offer. Google versus Micro1 is almost beside the point. The court has to decide who can establish provenance after an airline collapses. And “we’ll clean it up in-house” doesn’t cut it when other companies’ data may be mixed in. If you’re following AI’s rapid evolution, check out The Data Center Daily. It’s a daily briefing on AI compute and hyperscaler capex, along with the power grid, semiconductor supply, and energy markets being reshaped by intelligence at scale. Find it wherever you listen to podcasts.
We’re watching December 1, the expected start date for SpaceX’s newly signed hosting deal, which is expected to generate $13 billion in annual recurring revenue. We’re also watching the bankruptcy court’s upcoming hearing on whether to approve Spirit’s dataset sale—and how proprietary data gets separated before AI-training assets change hands.
Links to every story are in the show notes, so take a closer look at whatever caught your attention. That’s AI Daily Briefing for today. Until the next episode, this is a Lantern Podcast.