HUMAIN is breaking ground while AI maps a route to Alpha Centauri. Pick your signal. If you're joining us mid-arc, here's the short version: Neocloud and specialist AI infrastructure providers are turning AI demand into contracted GPU and data-center buildouts. Before Oxagon, that included Together AI’s HUMAIN agreement: a 250 MW Saudi data-center path for 120,000 chips. The company projects more than $5 billion in gross annualized first-year revenue once that facility is operating. This is AI Daily Briefing. Today: how to tell an actual compute order book from a very expensive promise. Let’s start at Oxagon. We'll keep tracking this story — Neocloud spending and backlog boom. Follow the show so the next update finds you. From Sindhu V Kashyap at GEC Newswire:
Crews are on site at Oxagon, the industrial city on NEOM’s Red Sea coast, where HUMAIN and DataVolt have started building an AI data centre designed for the compute densities that advanced model training and inference now require. The two companies announced an expanded partnership at LEAP to jointly develop 100MW of the 360MW first phase, with that initial capacity expected to be ready for service in 2028.
Crews are on site, 360 megawatts are under construction, and service is targeted for 2028—that’s finally a project you can hold to a delivery test. The next number I want is contracted load for that first 100MW, because concrete and substations don’t generate inference revenue on their own. We’ve got another hard-power milestone in the HUMAIN buildout: crews at Oxagon are building the first 100 megawatts. Three straight days of Saudi compute announcements, and this one has actual construction, a 360MW opening phase, and a 1.5GW end state. DataVolt is handling power procurement, engineering, financing, construction, and operations—the unglamorous work that matters. A 2028 date can still slip—but at least HUMAIN and DataVolt are past the stage where the whole project fits on a LEAP keynote slide. The model press releases will keep arriving, sure. But those 1.5GW campuses mean years of grid work, chip orders, cooling, and a balance sheet waiting for customers to show up. MIT Technology Review writes:
A nonprofit organization called the Fermi Explorer Mission announced today that it intends to launch a spacecraft to our nearest star system by the end of 2029. It’s a hugely ambitious mission—if all goes well, the spacecraft could take up to 80,000 years to arrive at Alpha Centauri, which is 4.4 light-years away. And the spacecraft will follow a novel trajectory discovered by an AI system developed by Physical Superintelligence (PSI), an AI physics research lab.
A 2029 launch target and a $15 million mission budget: those are refreshingly concrete. The 80,000-year delivery window does make the usual roadmap-slippage conversation feel a little petty. PSI’s AI found the trajectory, and MIT Technology Review says the underlying paper hasn’t been peer reviewed. It’s a fascinating physics lead, but it still needs validation before anyone calls it an interstellar navigation breakthrough. And they’ve got $58 million from Breakthrough Energy, versus Breakthrough Starshot’s $100 million pledge in 2016—and, ten years later, no launch. “Actually launching” is a much better benchmark than a gorgeous trajectory plot. Put the technical work out there, let physicists attack it, then celebrate. A spacecraft carrying the Golden Record toward Alpha Centauri deserves more than an AI-generated headline and a press-release victory lap. Here's Hang Li at alphaXiv:
Across scene-level 3D object detection, object pose estimation, and scene reconstruction, Lucida improves mAP over Boxer by 69% on R2S-Scene, raises from 57.8% to 83.4% on CA-1M, and increases scene F-Score from 0.794 for SAM3D to 0.924.
Lucida has the right target: take a messy indoor video and turn it into editable objects a robot simulator can actually use. And I like that placement is closed-loop—GizmoAct keeps moving the object until it decides the alignment is right. The headline numbers are enormous: 69% better mAP than Boxer on R2S-Scene, and pose accuracy from 57.8% to 83.4% on CA-1M. Fine. What matters more is that they’re measuring the whole scene-reconstruction pipeline, not celebrating one pretty generated chair. I do want the ugly-capture test. Clutter, occlusion, bad lighting—the paper says that’s the problem it solves, and that’s exactly where multi-turn visual control tends to get weird. But getting from scene parsing through generation to physical placement is a much more useful robot stack than another chatbot that can describe a kitchen. Why are AI companies locking up so much data-center capacity now? And after all the bubble talk, how do we know these are real computing needs rather than a land rush for power? The immediate reason is the sheer scale of the computing buildout. Carnegie Endowment researchers Alasdair Phillips-Robins, Teddy Tawil, and Sam Winter-Levy estimate that America’s biggest technology companies will spend about $670 billion this year building compute clusters, while global companies and governments will put almost $1 trillion into data centers. And these commitments aren’t all vague reservations. Data Center Frontier’s David Chernicoff reports that Anthropic signed a 20-year lease with TeraWulf for about 401 megawatts of critical IT capacity in Kentucky, expected to generate $19 billion in contracted revenue, with initial capacity slated for late 2027. A long-duration contract with a named customer, a power figure, and an operating date is stronger evidence of demand than a generic announcement. But the distinction matters: Reuters reports that data centers have requested roughly as much electricity across the middle of the United States as every home in the country uses—and much of that demand may be illusory. Texas has frozen new grid connections for data centers while it investigates developers’ plans. Regulators are trying to separate concrete projects from speculative power requests. So a company can ask for power in several places without necessarily building all those facilities? That’s the risk behind what Reuters calls “ghost” demand: grid planners may be reacting to requests that never turn into operating data centers. For builders and investors, look for a named customer and contracted revenue, plus a specific megawatt figure and a credible delivery date—not just a large grid-capacity request. Watch Texas and other states’ connection reviews, because they could determine which announced projects actually get the power to move ahead. From Blankline:
We report a single, fully verified instance in which a frozen language model, coupled to an external memory that stores distilled records of its previous attempts, produced a configuration that extends beyond the published solutions and the search path represented by its own incumbent. Our analysis shows that the configuration is non-isomorphic to the published record and cannot be obtained through interpolation or local descent from the incumbent.
Blankline ran 92 attempts, got one verified win, and published the other 91. The archive matters as much as the final configuration. Yes. They used a frozen model with external memory and an exact verifier, and put the code and coordinates on GitHub—they call it an existence result. Labs somehow forget this level of humility when the demo video goes live. Their claim stays narrow: one case that cleared three tests—no retrieval match, no interpolation, no local descent from its own incumbent. Run the scripts, inspect the failures, then argue with the verifier. That’s a much healthier fight than watching an agent magically succeed on take one. And it sharpens the disclosure question from the Alpha Centauri item: a big claim earns trust when independent checking is boringly possible. Blankline handed people Node 18, three scripts, and the whole paper trail. If you want more on the intersection of AI and government, check out Anthropic Pentagon Watch—a daily briefing on Anthropic’s fight with the DoD over Claude, military AI use, autonomous weapons, and AI procurement blacklisting. Find it wherever you listen to podcasts.
On the calendar, Oxagon’s initial AI data-center capacity is expected to be ready for service in 2028, while the Fermi Explorer Mission says it intends to launch its Alpha Centauri-bound spacecraft by the end of 2029.
Links to every story are in the show notes if you want to go deeper. That’s AI Daily Briefing for today. This is a Lantern Podcast.