Chennai just went from an AI market to an AI-factory hub—and the bill is getting very, very real. Quick catch-up before we dig in: CoreWeave set the latest marker for the neocloud boom ahead of this edition, lifting 2026 capital spending to $35 billion to $39 billion after stronger second-quarter results and reporting $104.2 billion in backlog. The question is whether demand for rented AI capacity can keep funding ever-larger buildouts when power, GPUs, and customer commitments are the constraint. This is AI Daily Briefing. Today, we're looking at a giant Chennai cluster and Cerebras putting a huge backlog next to real megawatts. Then: a benchmark that asks whether vision models can keep their bearings past step one. From Supriya Roy at Times of India:
Bengaluru: Larsen & Toubro (L&T) has bagged a large artificial intelligence infrastructure order from US-based Together AI, marking its entry into the AI factory business.L&T did not disclose the value of the order. It classified it as a “mega” order, which for the company means a contract worth between Rs 10,000 crore and Rs 15,000 crore.
The neocloud-capex story now goes beyond CoreWeave: Together AI is tied to a mega L&T build in Chennai. By L&T’s own classification, the contract runs ₹10,000 to ₹15,000 crore for a single cluster of 10,000 Nvidia B300s. Okay, this is the physical proof I wanted. Together AI’s B300 bet now has an EPC contractor, a campus, and a 250-megawatt first phase—not just a cloud-spend number floating around on a slide deck. And L&T is explicitly entering the AI-factory business through this deal, with Vyoma.AI’s Chennai campus and LTN Compute executing it. This puts Chennai on the serious infrastructure map; it’s way beyond a boutique regional deployment. Ten thousand B300s against 250 megawatts is a useful pair of numbers. Together says it serves 400 trillion tokens a month. When this comes online, we’ll see whether that throughput target holds up under actual power, latency, and customer workloads. From Amit Chowdhry at Pulse2:
Cerebras Systems is rapidly expanding the physical infrastructure supporting its AI inference business, with more than 600 megawatts of data center capacity now live or under contract for delivery by the end of 2027 and manufacturing capacity expected to increase more than 10x during 2026.
Cerebras has 600 megawatts live or under contract through 2027, with $25.4 billion in remaining performance obligations behind it. That’s a real backlog-and-capacity disclosure. Now I want the contract vintages, because 2027 is a long time to promise per-token pricing. Put that next to the Chennai build: 250 megawatts for Together AI, 600 for Cerebras. At least we’re getting numbers to audit, not just glossy data-center renderings. The manufacturing side is where the hedge gets interesting. Cerebras says it’ll more than 10x capacity in 2026, with Flex, Sanmina, Rocket EMS, and TSMC wafer supply lined up—and its systems avoid HBM, CoWoS, and 3-nanometer constraints. If those claims hold, they’ve picked a much less crowded supply chain. They also raised $6.4 billion in the IPO and have $8.6 billion in cash and short-term investments, so this isn’t a capacity target stapled to a fundraising deck. It still comes down to delivery: turning that 600-megawatt contract pile into deployed inference before the economics shift underneath it. This one's from Microsoft Research Blog:
MindTopo is a new benchmark for testing topological reasoning in AI, evaluating whether multimodal models can understand concepts such as connectivity, enclosure, order, separation, and knots. The benchmark measures both reasoning and planning, testing not only whether models can recognize topological relationships in static images but also whether they can preserve and manipulate those relationships through a sequence of actions.
Okay, palate cleanser: Microsoft Research’s MindTopo asks whether a vision-language model can keep track of what’s connected, enclosed, or knotted after it starts changing the scene. That’s where agents break—by step seven, after the model looked brilliant in the screenshot. And Microsoft’s finding is refreshingly specific: current models recognize these relationships far better than they plan through them. Change a fence, wall, or rope, and the internal map starts slipping. Which is a pretty serious problem if the next stop is robotics. You don’t want a robot that can identify a knot with confidence, then propose pulling the rope through itself. A useful benchmark gives you a technical report and a concrete failure mode. Crucially, it tests whether a model can do more than see—it has to maintain a coherent model of the world over time. Got feedback, a story idea, or a correction? Email us at aidailybriefing at lantern podcasts dot com. Your notes help us make AI Daily Briefing sharper and more useful.
We’ll be watching Cerebras’ rollout of more than 600 megawatts of data-center capacity, which the company says is live or under contract for delivery by the end of 2027.
Links to every story are in the show notes if you want to dig deeper. That’s AI Daily Briefing for today. This is a Lantern Podcast.