Google's Carbon Math Lies and the Canvas School Breach

The Carbon Books Were Cooked

Google's planning documents for two proposed UK AI datacentres understated carbon emissions by a factor of five. Not fifty percent. Five times. The kind of error that, in any other industry, would trigger a regulator with a clipboard and a grudge. Here it will trigger a press statement about methodology and a renewed commitment to net zero by some year that ends in zero.

The context matters. Datacentre buildouts are the physical substrate of the AI boom, and the boom's defenders have spent two years insisting that efficiency gains and clean power purchases will absorb the load. Then the actual numbers leak out of a planning application and the gap between the marketing deck and the spreadsheet turns out to be 500 percent. If the climate accounting is this loose at the application stage, assume the operational accounting is worse.

Meanwhile a research team is pitching heat-storing molecules as a path to decarbonising buildings. Lovely. We will need every molecule we can get, because the GPUs are not slowing down.

ByteDance Buys Compute, Everyone Else Buys Insurance

ByteDance bumped AI infrastructure spending to $29 billion, a 25 percent jump, mostly to absorb chip costs that keep climbing. This is the shape of the market now. The hyperscalers and the Chinese platform giants are committing capital at a scale that makes the rest of the industry into either suppliers, customers, or roadkill.

The ECB noticed. Central bankers issued a warning that AI poses risks to financial infrastructure resilience, lumped in with stablecoin concerns for good measure. When the ECB starts using the word resilience in the same paragraph as a technology, it means the supervisory teams have been quietly running scenarios and the scenarios are not pretty. Concentration risk in model providers, opaque automated decisioning, correlated failure modes across institutions running the same three foundation models. The warning is not alarmist. It is overdue.

Nine Thousand Schools, One Breach

A cyberattack on Canvas compromised personally identifying data from nearly nine thousand schools. Edtech has spent a decade vacuuming up student records under the banner of personalised learning, and the security posture has not kept pace with the data hoard. Children's names, addresses, and academic histories are now circulating somewhere they should not be, and the parents will find out via a letter written by a lawyer.

The Internet of Things continues to make this worse from below. Researchers disclosed a vulnerability in a connected robot lawn mower, which sounds like a punchline until you remember that every one of these devices is a foothold on someone's home network. Meta is simultaneously killing encrypted Instagram DMs, which removes one of the few defaults that protected ordinary users from ordinary surveillance. The attack surface expands, the defensive surface shrinks, and the AI layer on top promises to automate both sides of that exchange.

Sovereignty Theatre

Louis Mosley, Palantir's UK chief, has become the public face of British tech sovereignty anxiety, defending the company against a population that has correctly noticed it is American, deeply embedded in the NHS, and run by people with strong views about how societies should be organised. The defence is the standard one. We are partners, we are accountable, we follow the rules of whichever country we are in. The anxiety persists because the anxiety is well founded. Critical infrastructure dependencies on foreign software vendors are a strategic posture, not a procurement decision, and the British public has worked this out faster than the British government.

In unrelated empire news, the Daily Wire is laying off staff as Ben Shapiro's media operation contracts under the weight of internal ideological feuds and declining relevance. Mentioned here only because it is the rare story this week where the doom number is genuinely low and someone deserves to enjoy it.

The Through-Line

Four stories about disclosure failures, one about systemic financial risk, one about a mass data breach, one about the consolidation of AI capital into a handful of hands. The individual doom scores are moderate. The pattern is that the institutions meant to catch these problems are either issuing warnings after the fact or issuing planning applications with the wrong numbers in them. Routine motion, accumulating.