Liu Gong thinks the week’s main theme was compute expansion running into the hard edges of the physical world. In Mississippi, xAI moved to buy out homeowners in exchange for dropping noise litigation; in Texas, voters organized against data centers; and nationwide, opposition has turned bipartisan ahead of the 2026 midterms — three versions of one mechanism, in which expansion costs get externalized onto host communities. On the technology side, Liquid AI’s open LFM3 recipe compresses vertical multimodal delivery from three months to three weeks, Genie 4 pushes world models into real-time editing, and Anthropic’s 300-page S-1 shows what a foundation-model balance sheet actually looks like.
Signal 1: xAI Subsidiary Moves to Buy Mississippi Homes in Exchange for Dropping Noise Lawsuits (Oct 28). Liu Gong reads this as a case study in externalized compliance costs rather than a neighborhood spat. He expects the first public sale records in Southaven within six months, and at least one state legislature or county board to enter buy-the-house-waive-the-claim into hearing records; once that precedent lands, residents in other compute-heavy counties will use it as a negotiating template. The boundaries are narrower than they look: several states protect the right to sue by statute, class membership is hard to sign away in perpetuity with a purchase contract, and a homeowner who waives today can still be sued by a neighbor for nuisance tomorrow.
Signal 2: Liquid AI Releases LFM3 and Open-Sources Its Recipe for Specialized Multimodal Models (Oct 27). Liu Gong expects 1B-class bases like LFM3 to produce shippable vertical multimodal products first in industrial inspection, equipment patrols and report structuring within six months, not leaderboard wins. The reason is the recipe rather than raw capability: publishing it compresses delivery from three months to three weeks. Adapters carrying just 1.5%–3% of parameters over a frozen base map neatly onto plant floors with 5MP cameras, fewer than 200 defect samples per class, and a strict rule that images never leave the building.
Signal 3: Texas Voters Organize Against Data Centers as Local Approval Resistance Escalates (Oct 28). Liu Gong predicts at least one more Texas project that already cleared zoning will be pulled back into council review within six months, and that by 2027 community consent will displace power prices as North America’s first gate on siting. The math is concrete: a GB200 NVL72 rack draws roughly 120 kW, 12 to 24 times a traditional cloud rack and the equivalent of 80 to 100 US households; a 100 MW campus equals about 80,000 homes, and its substations and diesel gensets have to occupy real acreage.
Signal 4: DeepMind Ships Genie 4, Pushing Real-Time World Models Closer to Photorealism (Oct 27). Over the next 12 months, Liu Gong does not see real-time world models replacing Unreal or Isaac Sim for production-grade simulation; they will land first where errors are cheap, namely robot policy pretraining and game or film prototyping. Watch a six-month window: if DeepMind opens external access, the entry point will most likely sit under Gemini API or Vertex AI rather than a standalone product. The hard part was never image quality but hitting frame latency, long-horizon consistency and action controllability at once; at 720p/24fps the budget is about 41.7 ms per frame, leaving perhaps 20 to 30 ms for the forward pass after codecs and network round-trips.
Signal 5: Anthropic Files a 300-Plus-Page S-1, Disclosing IPO Details (Oct 26). Liu Gong expects an IPO to close within 12 months, but the first reported quarter’s gross margin to come in below what public markets expect of an AI software company, because inference costs still eat into subscription revenue. The 300-plus pages run 50%–100% longer than the 150 to 200 typical for tech issuers, and the extra length most likely sits in risk factors, compute purchase commitments and related-party disclosures rather than product narrative. A 200K-token context window is 25 times early GPT-4’s 8K, and long context burns more memory and bandwidth per call, so margin pressure is structural.
Signal 6: Microsoft Offers Buyouts to as Many as 60,000 Employees (Oct 26). Liu Gong expects Microsoft to funnel the freed-up payroll into AI infrastructure and Copilot commercialization over the next six to nine months, with hiring freezes on non-core functions extended. The 60,000 figure is a ceiling rather than a headcount target; voluntary buyouts trade one-time charges for organizational flexibility and spare the employer brand that forced layoffs would bruise. That the offer covers Europe and India teams as well suggests the cost work is being run along global function lines, not carved up by geography.
Signal 7: Google’s Genie 4 Adds Real-Time Editing and Audio to Its World Model (Oct 27). The release marks the shift from generating a world you can watch to generating one you can edit live. Liu Gong expects Google to put real-time language editing in front of a limited set of developers within six months rather than shipping it to consumers, since live editing stresses consistency and latency far beyond one-shot generation and real workflows are the cheapest way to pressure-test cost and stability. It also fits Google’s two-year pattern of trickling Gemini capabilities down through APIs and tools. Live natural-language editing is not an input box bolted on; it ties world state, instruction parsing and the render loop together.
Signal 8: Anthropic Pushes Ahead on IPO With a Structure to Shield It From Wall Street’s Short-Term Grip (Oct 27). Over the next six to 12 months, Liu Gong argues that what matters in a filing is not revenue growth but whether safety evaluations, release gates and board veto rights are written as clauses shareholders cannot easily rewrite; if they stay as mission language, expect the first governance friction within 12 months of listing. Constitutional AI baked into training and alignment is a technical asset, but public markets will ask whether those rules slow releases or raise costs, and they will want the safety budget and evaluation throughput disclosed, not just model capability.
Signal 9: Nvidia’s RTX Spark Laptops Go on Sale With Six OEMs at Launch (Oct 28). The point is not benchmarks; it is that VRAM quietly disappears from the PC spec sheet. Liu Gong expects 64GB of unified memory to become the entry line for premium AI laptops within 12 months, leaving 32GB-and-below machines with little pricing power among developers and creators. Unified memory solves data movement rather than compute: PCIe 5.0 x16 at roughly 128 GB/s bidirectional is not slow, but with tens of gigabytes of weights and high-resolution image batches, that copy step is where latency and power go, and moving it on-package drops the cost to on-die bus levels.
Signal 10: US Data Center Backlash Turns Bipartisan and Becomes a 2026 Midterm Issue (Oct 27). The binding constraint is no longer transformers and gas turbines; it is the county hearing. Liu Gong predicts at least three states will pass legislation within 12 months requiring large compute loads to bring their own generation or prepay grid upgrades, and that the first data center project above $1 billion will die to a local denial within six months. The root cause is a mismatch between load curves and grid build cycles: 300 MW to 1 GW campuses fill in 24 to 36 months, while 345 kV and 500 kV lines take seven to 10 years and large transformers now quote beyond 120 weeks.