No Theater Is Sealed Anymore
Taiwan: Compression Over Continuation
The Wall Street Journal reported China's navy tightening its grip around Taiwan this week, with destroyer deployments that regional defense attaches describe as operational positioning rather than exercise. Taiwan's defense ministry logged a concurrent surge in PLA (China's People's Liberation Army) activity across the strait. Shipping companies routing freight through those waters are watching the same data as the defense attaches. Beijing's official channels offered no explanation.
The case for calling this compression rather than continuation rests on density. Destroyer deployments close to Taiwan's maritime approaches, at frequencies and in formations that allied intelligence officials have described in background briefings as rehearsal-level coordination, suggest a party that has moved its internal clock forward. A naval logistics officer at a Southeast Asian port watching those same shipping lanes, which carry the LNG and container freight her country's economy runs on, now faces a practical question about alternative routing that did not feel urgent six months ago.
What has not changed is the deterrence architecture. U.S. carrier presence in the Western Pacific, Taiwan's own defense posture, and the commercial interdependence that makes any kinetic move catastrophically expensive for Beijing all remain in place. Senior regional analysts whose published work I have followed for two decades estimate the convergence window created by Beijing's internal political calendar, the PLA modernization cycle, and Taiwan's executive transition closes somewhere around late 2027. This column does not predict war. A regional bank with Taiwan-related credit exposure and a shipping insurer pricing the Taiwan Strait corridor are reading this week's deployments for different reasons, and neither has the information it needs to price the risk accurately.
The Bubble and the Hub Contradiction
Leaked financial documents reported by Ars Technica showed OpenAI losing several billion dollars annually against revenue that, while growing, has not yet approached break-even. This position differs from Anthropic's. Anthropic has reached gross-margin profitability, meaning it recovers its direct production costs on each dollar of API revenue even while spending heavily on research and infrastructure. OpenAI is spending to acquire market position and retain talent. Anthropic is spending to fund research while its core business turns contribution-positive. The distinction matters to any enterprise CTO deciding which AI provider to build on for the next three years.
SpaceX acquired Windsurf, an AI coding assistant, for approximately sixty billion dollars this week. Baseten, a company that runs the infrastructure enterprises use to deploy and serve AI models at scale, announced a $1.5 billion raise. Barret Zoph, OpenAI's VP of Research, departed after five months. Meta's AI workers, per Wired's reporting, are in open revolt over compensation and strategic direction. A CTO deciding this month whether to commit her engineering team to OpenAI's API or diversify across providers has more evidence for diversification this week than she had in May.
Hong Kong sits in this with a specific tension. RTHK covered the government's AI-enhanced typhoon prediction system and a university town proposal as indicators of AI development commitment. Both are substantive investments. The government has also moved to eliminate voter demographic data, the kind of granular, verified population-level dataset that allows AI systems to be trained and benchmarked against local conditions. A data scientist at a Hong Kong fintech building regulatory-compliant credit or risk models on local population data now faces a shrinking supply of exactly the data her models require, at the moment her government is promoting the city as a regional AI development center.
When the Channel Is the Weapon
North Korean state actors, operating under the cluster security researchers call Lazarus Group, this week poisoned 140 packages in npm (the package registry that JavaScript developers use to install pre-written software components, the shared infrastructure underlying most modern web and AI applications) in an attack targeting the Mastra AI framework, a toolkit for building AI-powered business applications. The packages appeared legitimate. Developers who ran a standard dependency install pulled the malware onto their machines without any visible indicator. Any development team that used the Mastra AI toolkit during that window needs to audit every system where the install ran.
FortiBleed, the vulnerability now confirmed across 86,644 FortiGate devices (FortiGate is Fortinet's enterprise network firewall, used by governments and financial institutions to inspect and control traffic between networks), reached devices that had already been patched. Operation Endgame, a coordinated law enforcement action across multiple jurisdictions, this week removed SocGholish malware infrastructure from 15,000 WordPress sites. SocGholish had already used those sites to reach downstream victims before the cleanup ran. A network administrator who applied FortiGate patches on schedule this spring may be running compromised firmware without knowing it.
The pattern is the same across all three. The trusted channel delivered the compromise. A security operations center lead at a financial institution whose developers use npm packages or whose network runs FortiGate hardware cannot treat these as separate incidents. The threat now arrives routinely through the channels her team manages and approves. A compliance officer reviewing her institution's vendor software inventory needs to treat that inventory as a threat surface, not a cleared list.
I counted back through this quarter's morning synthesis notes. Of the twelve weeks since late March, nine had Taiwan and AI in the first three items simultaneously. Seven included a cyber development in the same group. Four added a Hong Kong governance item in the top five. At nine weeks out of twelve, the editorial desk is behind if it still treats these as separate assignments. The correspondent covering the argument that runs through all four theaters is not yet assigned.