Nvidia spent all week funding other people's data centers. Today, they signed their own lease — up to fifty billion dollars, on their own paper, in Texas. This is AI Daily Briefing. And the same day the chipmaker becomes a landlord, they drop five billion in cash on a lab with no product and no publications. One of those numbers I trust. One tap on follow, and we'll be back in your ears before you know it. From CryptoBriefing:
Nvidia just locked in one of the largest data center deals ever. The company signed leases worth up to $50 billion for Hut 8’s Beacon Point AI data center campus in Nueces County, Texas, a facility designed to house hundreds of thousands of its own graphics processing units.
Remember the question I kept circling — was Nvidia's guarantee for OpenAI's campus a one-off favor, or was it a pivot? The Hut 8 deal answers it. Nvidia's the signing tenant now: 704 megawatts contracted, $19.6 billion base, and up to $50 billion over fifteen years on its own paper. Nvidia went from chip supplier to lease guarantor, and now landlord. Beacon Point in Nueces County will put hundreds of thousands of its own GPUs under a roof it's renting — then subleasing to the hyperscalers. It's a balance-sheet transformation dressed up as a real-estate story. And here's why it lands for me: this is the first genuinely committed number of the week. After all the letters of intent and vague interim-capacity gaps, there's finally a line item: $26.6 billion base across the whole portfolio. And they secured an actual AEP Texas interconnection for a full gigawatt before pouring concrete. Power first, press release later. But owning the campus doesn't fix the thing I've been hammering all week. Nvidia's still the one fighting for the HBM4 queue — and now they've made themselves their biggest counterparty. The landlord and supplier are the same company, so if the memory pipeline slips, they're on both sides of that miss. From Bitcoin miner to AI landlord in one filing cycle. Hut 8 spent years building power and cooling for mining rigs, and it turns out the interconnection was the moat all along. Bret Kerr, writing in Bret Kerr:
Safe Superintelligence has no product, no publications, no revenue, and somewhere between twenty and fifty employees — and it got the seat. NVIDIA also said, in writing, that it entered the partnership "after obtaining rare access into the company's closely guarded research." Somebody looked. Then somebody wrote a $5 billion check, which Bloomberg first reported and Reuters corroborated, and opened the silicon roadmap.
So, right after that Texas lease — $50 billion in Nvidia's own name — the same company writes a $5 billion cash check for a lab with no product, no papers, no revenue, and maybe fifty people. By my usual standard on this show, that screams marketing: no technical report, no demo, just a closed door somebody paid to open. But Bret Kerr pulls out the line that stops me: SSI gets a seat at the datapath table. That's the part that survives my eyebrow test, honestly. Customers get allocation and a press quote. They don't get to co-design silicon. Nvidia has the best chip team on the planet, and they still opened the roadmap. Kerr's read is that Nvidia bought the observer rather than the model — Sutskever's bet on what the models can't do yet. Nvidia paid five billion to look at that, then reshaped Feynman-era hardware around what it saw. The market this week rewarded a $50 billion lease you can measure to the megawatt and a $5 billion bet on a company whose entire value is that it hasn't shipped anything. Same buyer. Same seven days. Here's DEV Community:
What I appreciate most is that they don't open with the method. They open with an empirical autopsy: Qwen2.5-1.5B + GRPO on ALFWorld, breaking "long-horizon training collapses" into two quantifiable failure modes.
Okay, this is the paper I've wanted all week. It doesn't open with a benchmark brag — it opens with an autopsy. It's Qwen2.5 with GRPO on ALFWorld, and they spell out exactly why the agent falls off a cliff as tasks get longer. And they quantify it. This CAR number — Contradictory Action Ratio — peaks above 40%. Nearly half your actions get a positive gradient in one run and a negative one in another, for the exact same move in the exact same state. That's the mechanism behind step 7 going bad, and there's an ablation table to back it up. Look at what we actually have here: a paper you can pull from ZJU-REAL out of Zhejiang, complete with an arXiv number and a GitHub repo. Now put it next to the $50 billion Texas lease we just walked through. You can feel the tension. We're pouring balance sheets into GPU campuses to run long-horizon agents, and a lab in Hangzhou just showed that the training regime for those agents collapses past a certain length. BEACON drags ALFWorld from 53.5 to 92.9 percent — sample utilization from 23.7 to 82. The fix exists, but it hasn't made its way into the marketing. Right — and the gains grow with horizon length, not just as a flat average bump. That's the part that should scare anyone selling a ten-step agent demo. The longer the chain, the more the old method was quietly bleeding out. Here's Arxiv:
Classical alpha mining achieves strong risk-adjusted returns by combining many low-correlated predictive signals, yet deep learning stock-ranking methods typically produce a single alpha per stock, rely on increasingly complex architectures with diminishing gains, and obtain diversity only through separate models or implicit routing, without explicitly controlling inter-alpha correlation.
Okay, a nice palate cleanser after all that $50 billion landlord talk — a quant paper. It's MAPLE, from four authors, tested across equity markets in the US, China, and Japan. The headline: 55x fewer parameters and 2.5x less training time. This is the part that actually survives contact with reality. Everyone's out here stacking architectural complexity, and these folks got 10 to 23% better Sharpe by fixing the loss function and the capacity allocation instead. Right, it's a single training pass that generates multiple low-correlated alphas. The whole trick is a regularizer that explicitly penalizes pairwise correlation across signals. Cheaper and more diverse at once. It's kind of the opposite of the day's other story, honestly. Nvidia paid five billion for a lab that ships nothing, and here's a four-author paper doing more with 55x less. Guess where the arXiv number is and where the marketing budget is. And they ran it across nine baselines and five backbones, with the numbers right there on the page — Sharpe, Calmar, ablations and all. That's the disclosure standard I keep wishing the demo-video crowd would hit. Want to follow the biggest moves into public markets? Try AI IPO Watch. It has daily coverage of OpenAI, Anthropic, Databricks, and SpaceX going public — every filing, valuation, and first trade, sourced, not rumored. Find it wherever you listen to podcasts.
You'll find links to every story in today's show notes, so take a look if there's one you want to dig into. That's AI Daily Briefing for today. This is a Lantern Podcast.