FSD data games, Hetzner price shock, and what Nvidia's filing tells us

00:00:04 Tesla sent self-published safety statistics to Swedish and Dutch regulators as part of its Full Self-Driving approval process. Traffic-safety researchers in both countries called the data misleading. The Dutch RDW approved FSD anyway, back in April. Here, the company building the system is also the one certifying it — self-published numbers for a safety assessment on an autonomous driving platform.

00:00:31 The gap between what traffic-safety researchers see when they look at those metrics and what regulators approve carries real weight. Someone's getting behind the wheel of a car that makes its own steering decisions, and the data justifying the approval came from the same company selling the product.

00:00:52 It's not unusual for autonomous vehicle developers to present proprietary data packages to regulators. What stands out is how fast researchers flagged issues — and how those flags didn't change the outcome in the Netherlands.

00:01:08 Over on HN this afternoon, a thread about Hetzner's latest dedicated server pricing changes caught my eye. They raised prices again — three to four times what they were just a few months ago. The AX102 went from €124 to €454. The AX162 with 256 gigabytes of RAM jumped from €244 to €844.

00:01:31 That follows about a thirty percent increase earlier this year. Commenters point to RAM and disk scarcity as the root cause. It connects to something bigger than any single provider's pricing page: the hardware supply chain for AI infrastructure is already constraining capacity downstream of the hyperscalers.

00:01:56 Nvidia reported nearly $194 billion in data center revenue last fiscal year, and people are writing checks for chips, complete systems, memory, and racks. That demand ripples through every layer — manufacturers allocate first to the biggest buyers, and secondary market pricing reflects whatever's left.

00:02:21 For developers running inference workloads on dedicated hardware, costs are climbing even as the capability curve gets steeper.

00:02:31 Wired reports today that the US government plans to let the Federal Data Center Enhancement Act expire in September with no replacement. That rule sets standards for data center use and operations across federal facilities. Put another way, regulatory oversight of federal data centers is disappearing this fall, with no announced plan to extend or replace the framework.

00:02:58 The timing is sharp. AI compute demand has reached levels where data centers are the bottleneck for everything from training runs to inference capacity. Power availability, cooling requirements, and security architecture all carried standards under this act. Those standards vanish just as commercial cloud and AI infrastructure demand accelerates.

00:03:23 What gets regulated matters. When oversight drops out, the gap between what federal operations follow and what commercial providers do widens. In a sector where supply chain constraints are tightening at every layer, that regulatory vacuum creates uncertainty for everyone downstream.

00:03:44 Earlier today, Nvidia filed a 424B5 prospectus supplement with the SEC. That document finalizes pricing details for registered securities offerings. Even after nearly $194 billion in data center revenue last fiscal year, Nvidia is raising capital through public markets.

00:04:03 That tells us something about the infrastructure buildout: even the company selling the chips needs outside financing to keep pace with demand. The hyperscalers are spending roughly seven hundred billion dollars this year on AI infrastructure. Some are raising debt.

00:04:22 Some are issuing stock. Power's tight, memory is expensive, and data centers take longer to build. Nate B Jones put out a video on this today that makes a useful distinction worth carrying forward. He distinguishes between two ideas: whether Nvidia's chip sales validate real demand and whether any part of the supply chain is overbuilt.

00:04:45 Here's what sticks: agent-driven inference workloads changed the cost math. An agent doesn't ask one question and stop. It loops, reads files, calls tools, writes code, checks results, fixes failures, searches again, runs again. That's not a conversation. It's a production job burning through tokens.

00:05:06 Once you see AI working that way, the infrastructure buildout starts to make more sense. Tokens are manufactured. Behind every agent tool call sits chips, memory, networking, power, cooling, construction, and ops. Jones frames it this way: you can have weak ROI in a corporate pilot and still face massive demand for coding agents, research automation, customer support systems, and enterprise tools that replace actual hours of work.

00:05:37 The infrastructure stays necessary even as some individual investments land at the wrong time. The bottleneck is timing. Cash flows need to arrive in the right place for the companies financing the buildout through the supply chain.