The Ground Beneath AI Infrastructure

00:00:04 Saturday morning brought a story that shifts the baseline — and it came from somewhere unexpected. Elon Musk personally ordered a throttle on Starshield in Ukraine. The Wall Street Journal reports this was unilateral, not an announced act of US government policy.

00:00:22 The effect was immediate: it tilted the war against Ukraine in Russia's favor. Steve Blank has been in London for the last month and he says the conversation in European policy circles shifted overnight. The doomsayer scenario — a US company operating critical military infrastructure able to cut off an American partner at will — went from paranoid fantasy to real budget planning.

00:00:49 The structural feature of how AI and cloud infrastructure work is what most people overlook: the kill-switch problem. Here's what Blank heard from European decision-makers, in spirit if not every word: What capabilities would Europe lose?" To a lot of European policymakers, it is no longer just plausible — it's becoming the expected risk.

00:01:26 The lesson they're drawing is that no nation can continue to outsource its security by depending solely on US data systems or intelligence — and the timing just got more urgent. What makes this concrete for us as builders is the second part of Blank's argument: AI actually collapses the cost of creating sovereign software.

00:01:49 The old way to build that stack — Office, Salesforce, Palantir, the works — was a multi-year grind. But today, AI puts it within reach. The cost is collapsing just as geopolitical risk is climbing. That's an unusual alignment — usually feasibility and urgency move in opposite directions.

00:02:09 So a US founder selling into European defense now needs to answer 'who holds the switch?' before the first meeting. And European founders just got sovereignty as a sellable feature. I'm wondering whether this is a one-off blowback from Musk's personal decisions, or what it reveals about how centralized infrastructure behaves under geopolitical stress.

00:02:34 There's also the smaller story of personnel churn. Two key members of Anthropic's Claude Code team — Thariq and Anthony Morris — both left recently. Thariq posted a departure note that got over 12,000 likes; the full text didn't load, but the gist is clear: he loved working at Anthropic and believes deeply in the mission.

00:02:57 That kind of departure usually means someone found something else. The Claude Code team's momentum makes those departures worth watching — not because two people leaving a company is newsworthy on its own, but because it intersects with this broader sovereignty question.

00:03:17 If US AI companies lose key engineering talent while European alternatives are being built, the competitive dynamics shift in ways that aren't captured by benchmarks.

00:03:29 If Musk's Starshield decision reveals the fragility of cloud infrastructure as a geopolitical concept, there's another story from Finland that makes it physical. Google cleared 816 acres of woodland in Northern Finland to make room for two data centers. They did it without completing the mandatory Environmental Impact Assessment that Finnish law requires before large-scale operations begin.

00:03:57 The result: an estimated 250,000 trees removed, leaving behind a lifeless mud wasteland with tire tracks and muddy puddles. The NY Post reports this from The Guardian's investigation. An environmental department head at the Finnish Licensing and Supervision Agency said the preparatory measures — tree removal, topsoil stripping, road construction — had already been completed before the EIA was done.

00:04:25 They gave Google's subsidiary a deadline of October 23rd to suspend all operations. The environmental damage is concrete and visible. Streams near the site turned brown from disturbed soil. Wolves that used the forest as habitat were forced into the nearby village of Otanmäki, about fifty miles from the clearing, looking for food.

00:04:49 A resident named Anni Ranta told The Guardian: 'We had wolves coming into the village because all the wild animals moved out of the forest.' She also noted that hard rock soil amplifies construction noise — she compared it to a drum when standing near the cleared area.

00:05:08 The financial scale puts it in perspective: this is part of Google's $15 billion investment in Finnish AI infrastructure, announced back in September. That's a lot of capital for a country with five and a half million people. Google's Finnish top chief, Antti Järvinen, claimed the company believed it had legislative approval to commence the tree cutting.

00:05:33 The conservation board chair called that 'pure arrogance.' She said: 'They bought the vast land area and they just started to cut it, just like that.' Google said it would protect water sources before planting back trees around 321 of the acres — but that's a partial restoration plan for land where ecological disruption has already happened.

00:06:02 It lands in a wider context. First, this isn't an isolated incident in Europe. New York Governor Kathy Hochul signed an executive order in July banning all large-scale AI data centers from the state — the first such ban in the US. Second, there's a pattern developing where scale is outpacing governance.

00:06:23 Google cleared trees before completing the required legal process. That suggests either deliberate circumvention or aggressive interpretation of what 'approval' means when you're under pressure to build at speed. What stands out isn't whether Google will face consequences — Finnish regulators have already suspended two sites.

00:06:47 It's how many other projects are operating on the same timeline gap between ambition and regulation. Data center industry has been growing fast, and environmental review processes move at different speeds. When you're building infrastructure at AI-scale velocity, there's pressure to start before the paperwork catches up — or to retroactively complete what you assumed was approved.

00:07:14 The next story lives at the intersection of academic publishing and AI hallucination. It starts with a paper that cites references that don't exist. A peer-reviewed paper in New Review of Academic Librarianship was published last month. It's titled 'Z-Library and Full-Text Access: Democratizing Knowledge or Undermining Copyright Ethics?' The two librarians who wrote it were examining Z-Library — one of the largest pirate ebook libraries on the Internet.

00:07:47 The paper itself doesn't pick a side. But TorrentFreak's Ernesto Van der Sar noticed something odd when he requested a copy from the lead author: at least ten of the 44 references in the paper don't exist. The errors look like this: The reference links only to the homepage.

00:08:24 The paper says the DOJ arrested individuals in Argentina in 2022 and gives a title for that press release. The title doesn't exist — the actual DOJ press release uses different language. A 2019 Harvard Journal of Law & Technology article by IP professor Daniel Gervais is cited, but the issue's table of contents shows no such article.

00:08:48 A 2019 College & Research Libraries article attributed to Carol Tenopir doesn't appear in the cited issue either. She did co-author a similar piece, but it appeared in a different journal under a different title. Missing references stretch from articles by Björn Brembs to pieces by authors just named Smith or Piper, plus a Kaspersky report on shadow libraries and works by Okike and Aronson.

00:09:16 The paper's acknowledgements may explain this: 'We confirm that Consensus AI was used in sifting through large volume of literature and extract relevant information.' They also added that ChatGPT was used to improve the language. Van der Sar ran the abstract and conclusion through Pangram, an AI detection tool.

00:09:38 Both sections were flagged as 100% likely AI-generated. The irony here is layered. Rightsholders are suing AI companies for using shadow libraries like Z-Library to train models. And this paper about a shadow library uses AI tools — producing hallucinated citations — to write about those same sites.

00:10:00 It also notes something real: the paper claims that Z-Library hurts publisher revenue, which funds peer review and editing. But it doesn't mention that peer reviewers generally aren't paid for their work, an issue central to a recent lawsuit against Taylor & Francis, which published this paper.

00:10:21 What matters going forward isn't whether the paper gets retracted — Taylor & Francis has referred it to ethics teams. It's what happens when AI tools enter the research pipeline and produce plausible but inaccurate references. This is a verification problem that becomes harder as AI becomes more useful for writing.

00:10:44 If your literature review tool generates citations that look real but don't exist, you need external methods to check them — and those methods are getting crowded with their own fictions. I'm not making an argument about AI-assisted research being bad. The paper's use of Consensus AI for literature screening is a legitimate workflow that many researchers would recognize.

00:11:11 The problem isn't the tool; it's that the hallucinated references are plausible enough to slip through initial review. What I am wondering is how academic publishers actually verify citations when both the claims and the tools used to verify them can be AI-generated.