Microsoft · Nscale · AI infrastructure
The AI data center needs electricity. The grid has an opening sometime in the 2030s.
Microsoft has reportedly terminated its agreement for Nscale’s proposed data center in Loughton, Essex. Reuters, citing the Telegraph on October 18, also describes repeated delays and a reported wait until the 2030s for power. The reason for Microsoft’s withdrawal was not specified; the power problem should not be presented as its confirmed cause. Nscale told Reuters that its Microsoft relationship remains strong, that the grid issue does not affect its other contracts, and that it remains committed to Loughton while exploring alternative power supplies and UK sites.
The practical distinction is between acquiring computing hardware and supplying a dependable service. A processor cannot spend a grid delay reasoning more efficiently, because it first needs somewhere operational to run. Our broader point is that deployment joins together many timetables: the equipment, the building, the electrical connection and the customer expecting delivery. A failure in one can disrupt the whole arrangement without changing the technical abilities of the hardware. Software may arrive quickly; the physical systems supporting it have their own constraints.
The report concerns one agreement, rather than a demonstrated retreat from AI investment generally or proof that the entire project has been abandoned. It is still an instructive warning about treating intended capacity as delivered capacity. A contract describes what parties expect to provide; the operational result depends on completing the supporting work. Our question is how much of the industry's anticipated computing supply can arrive on its advertised schedule. One report cannot answer that for the whole industry. The computers may be ready to think. The power supply would appreciate a little more time.
Source: Reuters on the reported withdrawal, power delays and Nscale’s response ↗Autonomous agents · Web access · Digital authorization
The AI can browse the internet. The internet would prefer that it make an appointment.
The Verge’s October 17 report describes services restricting autonomous agents: Yelp requires participation in its paid data-licensing program, Delta blocks agent bookings, and JPMorgan Chase and Charles Schwab cite security concerns in blocking account access. The report also describes agents encountering ordinary anti-bot defenses and retailers cooperating through partnerships. Those are different explanations for a similar outcome: the assistant reaches a service and cannot complete its task. A commercial access restriction and a security control should not be treated as the same policy.
A customer may reasonably want an assistant to handle a tedious transaction. The service receiving that transaction also needs to know who is acting, what has been authorized and how mistakes can be corrected. Consider a hypothetical travel assistant allowed to compare flights: that permission would not automatically authorize buying a ticket, changing an existing reservation or accepting every condition encountered along the way. Our assessment is that delegated authority needs an explicit scope. A capable agent cannot settle the receiving service’s requirements simply by asserting that it represents a customer.
The disagreement reaches beyond whether an agent is clever enough to operate a browser. Controlled interfaces and commercial agreements may make some tasks more reliable, while limiting which services an assistant can use. Blanket blocking may frustrate legitimate users, yet removing defenses without a trustworthy alternative could create other problems. The engineering challenge is to distinguish permitted assistance from unwanted automation and make the result understandable to the person requesting help. The model can understand the assignment. The website still controls whether it will accept the visitor.
Source: The Verge on agent restrictions, partnerships and anti-bot friction ↗SpaceX · Starlink Mobile · Telecommunications
SpaceX bought the airwaves. Unfortunately, the ground is considerably more expensive.
Reuters’ October 16 analysis examines the distance between SpaceX’s spectrum deal and a national mobile service. A source put the Grain 800 MHz transaction near $8 billion; TD Cowen estimates that building a terrestrial network would cost more than $80 billion. SpaceX describes a hybrid satellite-and-ground architecture. The $80 billion figure is an analyst estimate, not an announced SpaceX budget or completed expenditure. Spectrum rights, satellite authorization and a functioning nationwide phone service are different milestones.
The analysis notes the difficulty of connecting small, low-power phone antennas to distant satellites, especially in cities. Ground infrastructure remains important for capacity and service quality. Our connection to the first story is straightforward: an extraordinary component does not eliminate the ordinary systems needed around it. A network must work where customers actually use their phones, rather than only where its most impressive technology has an advantage. The company that can launch a satellite still has to deliver an acceptable connection inside an unremarkable building.
Reuters also identifies customer service, handset subsidies and promotions as substantial requirements in a market already served by established carriers. Our view is that the meaningful result will be dependable service at a sustainable cost, rather than the size of an acquisition announcement. SpaceX’s aerospace experience makes the ambition worth following, but does not establish how quickly the broader operation can be assembled. This is a deployment and business challenge still to be demonstrated. Conquering space was one part. The next challenge is getting five bars in the grocery store.
Source: Reuters on the hybrid plan and estimated terrestrial buildout cost ↗