SpaceX put an $18.4 billion AI bet on the board—and Wall Street’s already asking who pays for the encore. Quick catch-up: SpaceX has been building an AI compute business alongside its space operations. It reported 1.4 gigawatts of nameplate compute at the end of the second quarter and is targeting more than 2 gigawatts by year-end. Longer term, it’s aiming for up to 20 gigawatts of AI power and cooling capacity by late 2027, built on Nvidia Vera Rubin systems. This is AI Daily Briefing. SpaceX says the payoff comes fast. Investors just put a price on that claim. Let’s start there. This story isn't over: SpaceX AI compute buildout. Follow us wherever you're listening, and the next chapter comes to you. From Ari Levy, Lora Kolodny at CNBC:
After SpaceX spent way more on its AI buildout than analysts expected, executives tried to convince investors on Tuesday that it's all worthwhile, claiming that the company is making its money back within a year. The message didn't resonate, as SpaceX shares sank following the company's first earnings report since its IPO in June.
SpaceX’s first post-IPO earnings report put an $18.4 billion quarterly capex bill in front of skeptical investors. That’s more than double quarterly sales, with over 80% aimed at AI. Wall Street got the one-year-payback pitch, then marked the stock down 7.5% after hours. That’s a live stress test: real money voting against the slide deck. CFO Bret Johnsen pointed to $6.7 billion in newly contracted cloud-services revenue, ramping in October. Fine—show me that ramp against an $18.4 billion quarter, because those are very different-sized numbers. Capex rose more than sixfold year over year. Glad they beat revenue estimates, but “efficient” has to survive 3 a.m. utilization, power bills, and customers actually consuming the compute—not just signing contracts. Alex K.T. Martin, writing in The Japan Times:
The country’s biggest data center operator, which is owned by government-backed NTT, intends to add roughly 750 megawatts of capacity over the next seven years to meet soaring demand from companies seeking to catch up in artificial intelligence, people familiar with the matter said.
NTT Data is eyeing a ¥1.5 trillion plan to reach one gigawatt in Japan by 2033—roughly 750 megawatts added over seven years. Seven years. The build schedule is setting the pace now, not the model-release calendar. That ¥2 billion to ¥2.5 billion per megawatt estimate doesn’t include the Nvidia accelerators. So the $9.6 billion headline covers building and power before you buy the expensive part that actually runs the workload. We just heard Wall Street push back on SpaceX’s one-year AI-capex payback. NTT’s plan at least puts the physical timeline out in the open: that capacity doesn’t arrive until 2033. A gigawatt is serious. But every enterprise buyer in Japan hoping to catch up on AI should notice the date attached to it: 2033. Hardware roadmaps move a little faster than construction permits. Dan Robinson, writing in The Register:
Amazon is now upping its forecast for how much capex it will spend this year on expanding its infrastructure, including that needed for those AI workloads. "Earlier this year, we said we plan to invest approximately $200 billion in cash capex in 2026, the majority of which to support AI and AWS," Jassy told analysts on a conference call about its financials.
Amazon just raised 2026 cash capex from $200 billion to $220 billion, partly because memory costs rose. And Jassy says even that won’t cover demand through 2027. At some point, “we can’t build fast enough” is a very expensive forecast to get wrong. The Register puts the aggregate at roughly $595 billion across Amazon, Google, and Microsoft. Grid connections and memory supply don’t bend to a model-launch calendar. Neither do construction crews. AWS pulled in $42.2 billion this quarter, up 36.7 percent, so it’s not spending on vapor. But fast-growing revenue and a quick payback on $220 billion are two different spreadsheets. SpaceX just got that lesson priced into its stock. Exactly. A $169 billion annualized AWS run rate is enormous—Fortune 500 scale on its own—but it doesn’t make the denominator go away. Wall Street is finally asking whether all this physical buildout has a revenue ramp behind it. Converge Digest, with Jim Carroll:
CoreWeave announced plans to expand its AI cloud platform into Indonesia, marking the company’s first data center presence in the Asia-Pacific region. The expansion will add three facilities with a combined 360 megawatts of contracted IT power, with the sites expected to come online in 2028.
CoreWeave has 360 megawatts of contracted IT power across three Indonesian sites, but the machines don’t arrive until 2028. Investors are pricing demand years before any hardware serves a single latency-sensitive workload. And CoreWeave will own and operate the compute environment itself. For Southeast Asian customers with data-locality rules, control over the inference stack may matter more than whose logo is on the accelerator. Wall Street just flinched at SpaceX’s one-year payoff pitch. CoreWeave has until 2028—but 360 megawatts is still an expensive bet on local demand showing up on schedule. NTT Data has seven years to pursue a gigawatt in Japan; CoreWeave’s first Asia-Pacific footprint comes online in 2028. Model cycles run in months. Grid capacity, permits, and local operating teams are setting the calendar. DataMagz writes:
Samsung SDS has launched South Korea’s first NPU-as-a-Service (NPUaaS) platform, providing enterprises with on-demand access to FuriosaAI’s RNGD artificial intelligence inference accelerators through the Samsung Cloud Platform. The service enables organisations to run AI inference workloads without deploying dedicated hardware.
Samsung SDS is putting FuriosaAI’s RNGD chips inside Samsung Cloud, with one-, two-, four-, and eight-card instances. That matters: a Korean enterprise can tune its deployment and run inference without handing the whole operating layer to an American hyperscaler. Furiosa has set a baseline: 512 FP8 teraflops at 180 watts per accelerator. Eight cards fit in a roughly 3-kilowatt server. Now Samsung SDS has to prove the boring but decisive part—throughput per dollar on an actual customer workload. After all the hundreds of billions in capex we just ran through, this is more useful than another glossy facility rendering. NPUs run finished models, and Samsung SDS is selling access to that inference layer as a product. Right. “Agentic workloads” sounds lovely until the eighth step fails and somebody gets paged. But if RNGD delivers comparable latency for less than the Nvidia cloud bill, Korean teams will switch fast. Enterprise procurement is never sentimental. Want to go deeper on the infrastructure behind AI? Try The Data Center Daily—a daily briefing on AI compute, hyperscaler capex, the power grid, chip supply, and energy markets reshaped by intelligence at scale. Find it wherever you get podcasts.
One thing to watch: SpaceX says its newly contracted $6.7 billion in cloud-services revenue begins ramping in October.
Links to every story are in the show notes. Check out the ones that caught your attention. That’s AI Daily Briefing for today. This is a Lantern Podcast.