Amazon hit the power wall. Now it's about what it’s willing to build—and burn through—to get around it. This is AI Daily Briefing. Amazon’s spending surge has a specific bottleneck, while a Texas project raises a much bigger question: who gets to approve the fix? If today's show was useful, follow us wherever you're listening — the next one will be waiting. South China Morning Post writes:
Amazon has confirmed it is financing a massive, private gas power plant in Texas that could become the single largest source of greenhouse gas emissions in the United States. It is the latest example of tech giants going off the grid to get their AI operations online faster. The news was first reported by market intelligence company Cleanview, which reviewed satellite imagery to connect three data centre construction permits filed this week by Amazon to the gas plant, which is being developed by Pacifico Energy and is known as GW Ranch.
Seven-point-six-five gigawatts at one Texas site. Amazon is financing generation on the scale of a utility because apparently waiting for somebody else's power plan wasn't an option. And Cleanview pieced it together through satellite imagery and permit records—not an Amazon announcement. It linked Amazon's three data-center permits to Pacifico Energy's GW Ranch. For a project this large, that’s a pretty bleak transparency standard. Thirty-five turbines, permitted for more than 30 million metric tonnes of emissions a year. Put the permit approval next to Amazon's sustainability accounting, because this is way past a cute cloud-computing footnote. Exactly. Even as a private project, GW Ranch’s emissions ceiling would top every other source in the country. Amazon can call it data-center infrastructure; regulators should give it the scrutiny a power project that size demands. Here's Merin Rebecca Thomas at IBTimes:
Amazon expects to spend about $220 billion in capital expenditures in 2026, up from its previous estimate of around $200 billion, as rising costs for artificial intelligence memory chips drive higher infrastructure spending. Speaking during the company's second-quarter earnings call, CEO Andy Jassy said the rising cost of memory has become a major factor in Amazon's infrastructure spending as the company races to expand capacity for artificial intelligence workloads, Yahoo Finance reported.
We just covered Texas. Now Andy Jassy says the extra $20 billion in 2026 capex is coming from memory costs. Amazon has two ceilings at once: getting electrons, and getting enough high-bandwidth memory to turn them into useful compute. Memory just moved the biggest infrastructure budget on Earth by ten percent. People obsess over GPU allocations, but a rack full of accelerators without the memory configuration your workload needs is a very expensive space heater. Jassy says $220 billion still won’t cover 2026 demand. He says 2027 will be tight, and 2028 demand is already “striking.” That’s a serious demand signal, but Amazon also needs to separate what customers are truly consuming from what it’s reserving for its own AI operation. The cloud bill is about to include a bottleneck tax for everything at once: power, chips, memory, networking. A model can get marginally better and still lose the deployment if its memory footprint blows up latency and cost. Kim Ji-young, writing in Seoul Economic Daily:
VESSL AI, a company specializing in artificial intelligence (AI) infrastructure, said on the 7th that it has secured more than 1,000 Nvidia Blackwell (B200) graphics processing units (GPUs). The company plans to supply the high-performance GPU infrastructure, obtained through cooperation with SK Telecom, to customers via its flagship service, "VESSL Cloud."
A thousand B200s through SK Telecom is a meaningful new supply route. Telcos are becoming GPU brokers, giving a company like VESSL a way into Blackwell without waiting for a hyperscaler to leave crumbs. And VESSL already has Upstage and Motif Technologies on its customer list, alongside U.S. startups. The fleet is going into VESSL Cloud as on-demand, spot, and reserved capacity. “AI factory” is still one of those phrases that makes my eye twitch. But if VESSL can keep B200 capacity available at a price founders can stomach, the branding won’t matter much. The useful signal is Blackwell supply reaching Korean infrastructure players in four-digit batches. That sits awkwardly next to Andy Jassy saying memory costs alone added $20 billion to Amazon’s capex. GPUs are getting easier to route; the rest of the stack keeps finding fresh ways to get expensive. Why are AI companies suddenly so concerned with gas plants, grid connections, and data-center power? Aren’t faster chips and better models still the main bottlenecks? They’re still bottlenecks, but none of it matters if the physical infrastructure underneath can’t get electricity when it needs it. Reuters’ Ron Bousso reports that Microsoft, Amazon, Alphabet, and Meta have announced plans to spend more than $600 billion on AI in 2026, while U.S. grids are struggling to keep pace with hyperscaler demand. So the spending now reaches beyond servers. Companies need sites and grid connections, and sometimes dependable generation close by. RMI says data centers and other large new users can struggle to secure enough power quickly, delaying facilities or pushing them toward expensive, polluting backup systems. TechCrunch’s Tim De Chant also reports a scramble for natural-gas supplies and equipment as companies race to lock down data-center power. The clearest warning sign is in PJM: Wharton’s Santiago Gallino writes that the grid operator fell nearly 6,600 megawatts short of its reserve target for summer 2027, with data centers accounting for 94% of projected load growth. So is power actually becoming the limiting factor for AI, or are companies just buying every scarce input because they’re afraid of being left behind? Both are happening. Models, chips, land, and construction capacity still matter, but a completed data center can sit idle if power isn’t available on schedule. Watch whether grid interconnection delays, reserve shortfalls, and new power deals keep determining where AI capacity gets built. If they do, electricity is a practical near-term constraint on AI expansion. Want to go deeper on the infrastructure behind AI? Check out The Data Center Daily, a daily briefing on AI compute, hyperscaler capex, the power grid, semiconductor supply, and energy markets reshaped by intelligence at scale. Find it wherever you listen to podcasts.
Every story is linked in the show notes if you want to dig deeper. That’s AI Daily Briefing for today. This is a Lantern Podcast.