The Great Rerouting
Compute flees the grid, the million-line rewrite, and open models' trust problem
What caught my eye
Compute is quietly routing around the grid
I read a Works in Progress essay this week called Why American Data Centers Can’t Plug In, and it reframed a bunch of headlines I had been treating as separate stories.
Start with the number that matters. In 2005, the median new power plant waited under 20 months for a grid connection. By 2023, the wait was 55 months. Models ship in months. Chips ship in quarters. Grid connections take half a decade.
Now look at what companies are actually doing about it. Per that same essay, 62 percent of data centers are considering some form of off-grid power, and xAI ran its Memphis site on 422 megawatts of on-site gas turbines rather than wait in line. SpaceX just filed with the FCC for up to 100,000 new Starlink satellites, and that filing is separate from its million-satellite orbital data center plan. A reactor startup reached criticality at Idaho National Laboratory this month with data centers as the target customer.
Living in Texas, this doesn’t feel abstract. ERCOT has 143.5 gigawatts of data centers asking to connect. The state’s all-time record demand is 85.9 gigawatts. The queue wants to be bigger than the entire grid at its busiest hour ever.
Nobody is solving the queue. The money is routing around it, to orbit, to reactors, to turbines in a parking lot. The thing I keep coming back to: the winners of this next capacity cycle are going to be picked by energy strategy, not model quality.
Works in Progress on why data centers can’t plug in and DCD on SpaceX’s 100,000-satellite filing.
The million-line rewrite just became a budget question
The Bun story is everywhere this week, and I think most people are quoting it without sitting with what it actually means.
The creator of Bun rewrote the entire codebase, roughly 535,000 lines of Zig, into Rust in 11 days. He ran 64 Claude agents in parallel, with every file written by one agent and reviewed by two more playing adversary, and the whole thing cost about $165,000 in API fees. His own estimate is that doing it by hand would have taken three engineers about a year, which really means it never would have happened at all. Mitchell Hashimoto ran the numbers and said the quiet part: no engineer at that salary could have hit those milestones in 11 days.
The counterweight keeps it honest. Zig’s creator called the result “unreviewed slop,” and the quality debate is not settled. It also took an unusually strong test suite and a founder who knew every corner of the codebase. This was not a prompt and a prayer.
What I think it changes: every company has a migration like this sitting in the someday pile, deferred because a year of zero user-facing progress was never going to get approved. That excuse now has a dollar figure attached. The conversation shifts from “can we afford the rewrite” to “do we trust the tests that tell us it worked.” That second question deserves a lot more attention than it’s getting.
Bun’s own writeup of the rewrite and The Pragmatic Engineer’s analysis.
Open models had their best week and their worst week at the same time
Two things happened in open AI models this month that belong in the same paragraph, and I haven’t seen anyone put them there.
The first: the biggest open-weight releases ever. Moonshot’s Kimi K3 shipped at 2.8 trillion parameters, the first open model in the 3-trillion class, with full weights coming July 27. And Mira Murati’s Thinking Machines released its first model, also open-weight, which gives the West a real answer after Meta pulled back from open releases last year.
The second: a researcher backdoored an open-weight model for under $100. Ten poisoned training examples were enough to make it reliably write code with a remote-execution hole, and larger models were easier to poison, not harder. The same week, Grok’s coding CLI got caught uploading users’ local files to the cloud.
So the golden age of free frontier intelligence and its trust problem showed up in the same news cycle. Most companies evaluating open models are pricing the first half of that story. Downloaded weights get treated like a free library. They are closer to an unvetted dependency, and at most companies nobody owns the question of where the weights came from or what is inside them. If you are going to take the free model, someone on your team has to own that question.
Moonshot’s Kimi K3 announcement and The Register on the $100 backdoor.
Oregon just made data centers pay their own way
Oregon approved a 29.7 percent electricity rate increase for data centers using more than 20 megawatts, under the state’s 2025 POWER Act. Residential rates go down 1.3 percent, about $1.91 a month for a typical household.
The dollar amounts are small. The precedent is not.
Until now, the cost of grid upgrades for large compute mostly spread across everyone’s bill. Oregon flipped that. Above 20 megawatts, you fund your own buildout.
Every state utility commission is looking at the same AI load forecasts. Oregon just handed them a template. I would expect versions of this to show up in other states quickly, including here in Texas.
OPB on the ruling and Tom’s Hardware analysis.
The short list
TechCrunch on Apple suing OpenAI. Trade secret theft claims, a hardware business called “rotten to its core” in the filing, and job candidates allegedly asked to bring Apple parts to their interviews.
Stanford’s “We Must Act Now” statement. Sixteen Nobel laureates and 200-plus economists and AI researchers, including longtime displacement skeptic Daron Acemoglu, calling for urgent preparation for AI’s economic transformation.
Resultsense on Thinking Machines’ Inkling. Mira Murati’s lab ships a 975-billion-parameter open-weight model, the first serious Western open release since Meta’s retreat.
DCD on Nvidia’s GPU backstop. Nvidia will rent back unused GPUs from cloud customers in exchange for a revenue share, with Firmus and Sharon AI first in, following similar deals with CoreWeave and Lambda.
POWER Magazine on Aalo Atomics reaching criticality. The Microsoft-backed startup’s test reactor went critical at Idaho National Laboratory by July 4, with data centers as the target market.
If this was useful, share it with someone who’d find it worth the read.
Only proofread with AI, never written.