Two giants spent the past day counting GPUs by the million. Meanwhile, a drone over New Jersey counted something a lot less flattering. And I've got one question for all of it: how much is actually plugged in? It's AI Daily Briefing, Friday edition. We start with Musk in Memphis. From Diego Almada Lopez at Crypto Briefing:
Elon Musk laid out a concrete timeline for scaling up the Colossus 2 AI computing cluster in the Memphis, Tennessee and Southaven, Mississippi corridor, committing to more than doubling the facility’s Nvidia chip count before the end of 2026. The expansion would push the cluster toward more than 1.1 million H100-equivalent accelerators.
Okay, do the division on the Google deal with me. $920 million a month for roughly 110,000 GPUs is about eight-and-a-half grand per GPU, per month, locked in through June 2029. That's the number I care about, way more than the million. And the million's a year-end target. Musk says he'll more than double the chip count to over 1.1 million H100-equivalents by December. It's September 25th. Those racks don't exist yet. Right, and look at what's powering the stuff that does exist. Temporary gas turbines and Tesla Megapacks. The word 'temporary' is sitting under a 946-megawatt IT load. I'd underline the tenant list, though. Google's paying rent, and Anthropic has capacity there too. A safety-branded lab training on Musk's power patchwork in Memphis. Whoever controls the inference stack holds the leash, and right now that's xAI. Yeah, no, I'd flip it. At eight grand a GPU, Google's got the leash. Musk needs that check clearing every month to fund the doubling. From Kieron Allen at Cloud Wars:
Now, the two powerhouses have announced a staggering expansion of this existing partnership which will, among other things — but this is certainly the standout headline — include the deployment of 2 million additional NVIDIA GPUs across AWS infrastructure between 2027 and 2028. The aim, according to Amazon, is to meet surging demand for AI infrastructure “from frontier labs, global enterprises, startups, and governments.”
So stack it on the Memphis number we just did. Colossus 2's 1.1 million, plus two million more from AWS. That's three-plus million GPUs announced in about a day, and this AWS batch doesn't even start landing until 2027. And it sits on top of the million-plus AWS already committed starting this year, per Cloud Wars. Look at the chip list, though: Blackwell Ultra, Rubin, Rubin Ultra. Two of those three are next-gen parts. It's a roadmap with a purchase order stapled to it. Yeah, I'm not counting anything I can't SSH into. What I'd actually circle is who Garman says it's for. Frontier labs, enterprises, governments. NVIDIA sells the silicon, but AWS decides who gets the capacity, and at what price. So on regional capacity, whoever runs the cloud picks which customers get served first. For builders, the only 2028 question is whether any of this shows up as a cheaper instance or just a longer waitlist. Here's Ars Technica:
This week, New Jersey ordered the operator of one of the East Coast’s largest planned data centers to pay a $1.1 million fine for secretly installing and operating gas generators in violation of the state’s Air Pollution Control Act. DataOne got hit with the fine after an investigation by The Guardian and Floodlight News in August shared thermal drone footage showing that 45 of the 62 gas generators were operating.
Okay, after two segments of counting GPUs in the millions, here's the one with receipts. Sixty-two gas generators, zero permits, and it was journalists with a thermal drone who caught it. The Guardian and Floodlight News, back in August. Forty-five of those generators were running hot on camera. And the state's permit line is 37 kilowatts. These were running at 1,982 each. Fifty times over. That's how you power up fast when the grid hookup isn't there yet. And that's why I care more about this than the Memphis number. A state regulator found an actual violation and put a dollar figure on it. Congress hasn't managed that for AI once. Over on Hacker News:
"Backlash was immediate, however, when Vineland’s mayor and council proposed giving DataOne a $6.2 million loan." These poor data center operators do really need a hand up after all. Incredible. Of all of the things your municipality could use its development funds on, this would not make my shortlist.
A $6.2 million loan from Vineland, against a $1.1 million fine. The town was about to pay the ticket five times over. Here's one from Hacker News:
1.1 million dollars and they are still allowed to run the gas generators? I think this is inaccurately called a fine when it sounds like a retroactively applied permit fee to operate.
Right, and that's the part that stings. Next to a multi-billion-dollar buildout, 1.1 million is a rounding error in the capex model. If it doesn't change the math, it's just a line item. Small, sure. But it's on the record now, and every other planned site running on temporary power just got a new line in its risk memo. Over on Hacker News:
transitioning to low emission quiet fuel cells Does this technology even exist? The lies are getting more brazen.
Fuel cells are real, I've seen them spec'd. Whether they replace sixty-two diesel-class gensets on this timeline? Show me the purchase order, not the press quote. arXiv, with Xingyu Wu:
To address these issues, we propose IterSynth, a role-decoupled and summary-based paradigm that alternates between a Planner for identifying information needs and a Synthesizer for integrating evidence into an evolving summary state. This design separates planning from synthesis while using the summary as the persistent state of search, reducing both capability coupling and context noise.
Okay, finally something I could actually run on a box under my desk. IterSynth, out of Zhejiang University and Tencent. It's an 8B deep-search agent averaging 50.7 across five benchmarks, BrowseComp and Xbench-DS among them. And it beats the best prior agent at eight billion or under by 4.2 points. Modest number, fine. But you get a method you can take apart and rebuild, which is more than half this week's announcements gave us. The design's what got me. They split one agent into a Planner and a Synthesizer, and the only thing carried forward is a running summary instead of the whole search history. My worry's the handoff. Does the noise really leave, or does it just get baked into that summary where you can't see it anymore? Which is testable, that's my point. And the bit I'd actually flag is the claim that it works as a prompting pattern on other models, zero-shot, no retraining. Most teams aren't running RL on their own agents, so something they can just drop in is the useful takeaway. "Substantial zero-shot gains" is the abstract talking, though. I'll believe it when I see it on my own retrieval logs at three in the morning. From Yehang Zhang at arXiv:
On LIBERO-Pro, WAA with skills evolved only from LIBERO-90 reaches a state-of-the-art 75.6% average success, outperforming end-to-end VLAs, code-as-policy agents, and a visual-harness baseline with the same backbone; the same skills remain effective on robosuite without further learning.
Second harness paper of the day, after IterSynth. This one drives robot arms. Every move becomes an editable proposal the agent previews and revises before the gripper ever touches anything. Headline number is 75.6% average success on LIBERO-Pro, with skills evolved only from LIBERO-90. Then those same skills work on robosuite with no further learning. For me, that's the interesting bit, capability walking out of the sandbox it trained in. Seventy-five percent also means one task in four fails, and in a warehouse that's a guy with a mop. But yeah, the rehearsal loop is smart. It catches the bad move before execution instead of after, which is exactly where chains usually die. Watch the last line, too. They fine-tune Qwen3.5-9B on the harness traces, so the big model teaches a small one to drive. Cheap, portable skill transfer is great for robotics, and it's the same pattern Dan Goodin's EvilTokens story was nervous about. Capability travels further than anyone tested it. Have feedback, a story idea, or a correction? Email us at aidailybriefing at lantern podcasts dot com. Your notes help us make the briefing more useful.
We're watching Google's roughly $920 million-a-month compute contract at the Memphis cluster, slated to start in October 2026, and whether Colossus 2 clears 1.1 million H100-equivalent accelerators by Musk's end-of-2026 target.
Links to every story are in the show notes, so check out the pieces that caught your attention. Thanks for listening, and we'll be here next episode. This is a Lantern Podcast.