AI Data Centers Bid Away Power from American Factories as Deleting Code Hits $10K Weekly

01The bottleneck in robotics moved from models to data, so labs started generating fake worlds to dodge the cost

General Intuition, a startup profiled by TechCrunch, is wagering that millions of hours of video game footage can serve as foundation training data for physical AI. The pitch is that robots can learn to move and react without the slow, expensive work of gathering real-world demonstrations. That premise rests on a problem three separate research groups are now attacking from different angles.

The constraint is physical. Training a robot policy means collecting trajectory data, and today that data comes from teleoperation: a human operator drives a real robot through a task, one demonstration at a time. The authors of RynnWorld-Teleop describe the cost directly. Every demonstration binds operator time to specific hardware and a specific workspace, which caps how much diverse data any lab can produce.

Their proposed fix removes the robot. RynnWorld-Teleop replaces the physical machine with a generative world model, so an operator's hand-pose stream drives a synthesized robot instead of a real one. The paper calls this digital teleoperation, a way to decouple data collection from hardware. The claim is that demonstrations no longer require a functioning robot in a room.

Evaluation carries the same tax. GigaWorld-1 opens by noting that language models get assessed on digital benchmarks, while robotic policies require slow, costly real-world rollouts limited by hardware and human supervision. Its authors position world models as surrogate evaluators that could score a policy without deploying it. They also concede the open question: what makes a simulated world reliable enough to trust as a stand-in for reality remains poorly understood.

RynnWorld-4D targets what those simulated worlds must capture to be useful. Recognizing what a scene looks like is not enough; a manipulation model must anticipate how a scene's 3D structure moves under interaction. The paper pairs RGB video with depth and optical flow, arguing that this combination grounds appearance in geometry and motion. The goal is a representation that predicts physical change, not just pixels.

None of this comes from the largest labs. A venture-backed startup and a set of academic and industry papers converged on the same wager: if real robot data stays scarce, the substitute is a model that manufactures the world instead.

02One company is raising $300M to generate code faster. Another charges $10,000 a week to delete it.

Lovable is reportedly in talks to raise about $300 million at a $13.2 billion valuation, double its previous mark, in a round Sifted reported would be led by Menlo Ventures. The pitch is speed: describe an app in plain language, and AI writes the code. Investors are pricing that speed as the whole product.

The bill shows up downstream. A firm called slopfix charges $10,000 a week to delete AI-generated code that runs but resists change. Its sales page names the failure directly: a "vibecoded" project grows past a certain size, the agent stops seeing the full codebase, and it duplicates logic instead of reusing it. Adding one feature takes days and breaks two other things.

The remedy reads like a demolition quote. Slopfix commits to a fixed reduction target, "100,000 lines down to 35,000, same functionality," then spends one week collapsing the redundancy. Fourteen date formatters become one. A hand-rolled framework becomes a library. The firm says it runs Claude Code too, "on a very short leash," and hands clients a CLAUDE.md, lint rules, and CI checks to slow the next batch of slop.

Sitting between the valuation and the cleanup invoice is a measurement problem. OpenAI published an analysis finding issues in SWE-Bench Pro, a benchmark widely used to rank AI coding models, and said the problems raise concerns about the reliability of its scores. The number buyers cite to justify AI-generated code is itself contested.

Slopfix analyzes a codebase for free before quoting and warranties its cuts for two weeks. That warranty covers the code it touched. It does not cover the next feature a client's own agent writes back into the same pile.

03The Power Meant for "Made in America" Is Getting Bid Away by AI Data Centers

The higher power bill arrived, and the reason sat a few miles down the road in a windowless hall of servers. Ars Technica reported that US manufacturers' electricity costs are climbing as AI data centers pull on the same grid, and the squeeze falls hardest on the Rust Belt, the industrial heartland Trump's manufacturing agenda was built to revive.

A factory and a data center draw from one regional grid. When demand outruns supply, the price of every kilowatt rises for both. Data center operators sign long contracts and can absorb the cost. A steel or auto-parts plant on thin margins feels it line by line, then passes what it can into the goods it was supposed to make cheaply at home.

According to the report, that competition threatens the "Made in America" plan directly. The policy promises factories, jobs, and domestic production, and it assumes affordable power. When the same electrons feeding a reshored plant get bid up by a server farm running AI models, the math that justified building at home begins to bend.

For a plant weighing whether to expand at home, power is not a footnote. It is one of the largest recurring costs of running a furnace or a stamping line. A rate increase does not close a factory overnight. It shifts the calculation on the next plant, the next shift, the next hire.

The demand driving those bills shows no sign of easing. NVIDIA introduced Vera, a CPU line it describes as built for the agentic AI era. The company says the processor sits on the critical path for reasoning, response time, and learning as AI systems call tools and run code. The framing is about compute scale, not power prices. It points one direction all the same: the industry is planning for more processors running, not fewer.

Every one of those processors runs on electricity, pulled from the same regional grids manufacturers depend on. The report frames the contest plainly. Cheap domestic power was the premise of reshoring, and AI data centers are now bidding against it.