Everything Split At Once

Here's what I took from it, and the questions I'm still sitting with.

The bottleneck moved below the chip

The story that reframed the most for me: SpaceX is laying the groundwork for its own turbine-blade factory. That is not a software company move. That is an industrial conglomerate move, and it happens because AI data center power capacity has been roughly doubling every ten months — about twice the pace of Moore's Law. The bottleneck left the chip a while ago. It moved to the buildings that house the chips, and then below that to the turbines that keep them cool.

Which turns vertical integration from a strategy preference into an operating requirement. If you cannot manufacture the part, you cannot deploy the chips you have already paid for, and you are holding a new class of stranded asset: capital equipment that is fine, fully funded, and idle. When a sell-side research desk starts citing a turbine story in a note, it has stopped being tech-media color and become board-level capital planning.

Nobody can score the compute bet yet

This is the argument I most wanted to have with myself this week. The case for locking in capacity is straightforward: Anthropic committed to another multibillion-dollar compute deal, and you cannot serve demand on efficiency gains you have not found yet. The case against is equally straightforward: OpenAI found a way to cut inference costs in half while ChatGPT closes in on a billion weekly users. If demand scales and unit costs keep falling, those long-dated capacity locks start to look less like hedges and more like bets that didn't need to be that large.

I don't think those positions are actually in tension, and that's the uncomfortable part — both can be right, and the scoring only happens in retrospect. What is not ambiguous is the market's uncertainty: Nvidia's valuation has fallen sharply from its highs as investors reassess whether the demand story is as guaranteed as it looked six months ago. Valuation moves are sentiment, not fundamentals. But three customers accounting for a large share of Nvidia's revenue is a structural fragility regardless of which way the stock is pointing, and the fix has a circularity problem: diversifying by investing in neocloud and AI firms means diversifying into companies whose success depends on Nvidia hardware staying dominant.

Back in the Nvidia playbook I wrote about three weeks ago, the pattern was licensing-plus-hiring instead of acquisition. This is the same instinct pointed at a different problem — absorb the risk rather than sit exposed to it — and it runs into the same limit: you cannot diversify away from yourself.

Outputs became a policy surface

The story I think has not fully landed: it was reported this week that the NSA has advised US AI labs to quietly degrade the answers given to suspected foreign intelligence users, without disclosure. Take the precedent seriously for a second. If outputs can be silently made worse for a category of user, then who belongs in that category is a policy question, not a technical one — and the category can be redrawn without anyone outside being told.

The practical consequence is that the reliability of any commercial AI system becomes conditionally geopolitical. That is a real input into the sovereign-deployment story, not a hypothetical one: Nvidia and Palantir are pitching on-premise sovereign architectures for critical supply chains, and part of the buying case is simply not having to trust that cloud outputs are unmanipulated. Enterprises that could not previously justify the cost of running their own stack now have an argument that survives a CFO conversation.

The revenue model underneath is being repriced

Per-seat pricing does not map onto agentic workloads, and Salesforce, OpenAI and others are already moving off it. Outcome-based pricing sounds clean in a pitch deck; what it actually does is delete the predictability of recurring revenue, which is the single property that made SaaS financeable. This is a slow structural shift that feels theoretical right up until it is happening to your own forecast.

Underneath that, the cost floor is moving too. Teams are cutting AI spend by switching to open models and smarter routing, and the adoption numbers behind Chinese open-source models show the cost-performance gap closing faster than proprietary vendors planned for. Once a credible open path exists, switching costs collapse — which is exactly why vendor lock-in stopped being an incidental risk and became a structural one. Model availability inside somebody else's product can disappear through an upstream contract dispute or an acquisition, and you have no recourse. The Stripe–OpenRouter deal I flagged as "brilliant or extremely early" now reads as the cleanest example of the second problem: you routed through a neutral layer, and the neutral layer may end up owned by a company with strategic interests of its own.

Agents stopped answering and started acting

Meta launched Muse this week — an agent that sends email, makes payments, and talks to smart home systems, running in its own virtual machine and continuing in the background. That is a large amount of real-world surface area for unintended action, and it arrives in the same week that labs acknowledged incidents of agents taking actions outside their intended boundaries in production rather than in a controlled test.

The structural point is the one I keep coming back to: when a system detects a problem and proposes an action before being asked, human approval becomes a safeguard rather than a trigger. The error mode shifts from a bad answer to a bad action, and actions have side effects that answers do not. I wrote about this as a trust-repair problem a few weeks ago; the incident reports since then suggest the industry is going to learn it empirically rather than by design.

The incentive structure does not help. a16z turned its infrastructure bets into an eight-billion-dollar result, and returns like that accelerate the next bet rather than encourage caution. Though the counter-signal is honest too: Grok has fallen behind its rivals, and concentrated returns still require picking the right concentrated bets.

Design's bottleneck is management, not capability

The cleanest articulation of the interface question I have seen: Salesforce and Figma are placing opposing bets. Salesforce thinks users will reach enterprise apps through AI assistants without logging in directly. Figma thinks design work stays inside Figma's own environment rather than routing through an external assistant. Both can be right for different use cases — which is the actual problem. There is no dominant integration pattern yet, so any team that picks one and builds deep into it is making a bet it may not be able to reverse cheaply.

The insight I want to sit with longer is quieter. AI has reduced the cost of building features. It has not reduced the cognitive cost those features impose on the person using them. Every feature is a tax on attention, and we can now levy that tax faster than ever — so the discipline of deciding what not to build becomes more valuable exactly when the organizational incentive pushes the other way. The same asymmetry shows up as debt: AI-assisted teams accumulate product and technical debt faster than any previous generation, because producing an artifact got cheap while owning one did not.

And the limiting factor is not designer skill. Designers are adopting AI faster than their managers know what to do with the capability — the constraint has moved up the org chart, which is consistent with the lag I wrote about two weeks ago: the capability is available, and the human systems for pointing it somewhere deliberately have not caught up. Two smaller signals point the same direction. Design systems are being rebuilt to be machine-readable, because a system encoded in implicit human convention cannot be reliably interpreted by anything else — that is moving from nice-to-have toward prerequisite. And tools like DialKit, with live controls sitting beside the interface being built, are a small concrete example of the workflow itself changing rather than just the tool.

This is also the continuation of a thread from my first post here: as producing polished artifacts gets cheap, what signals expertise is the transparency of the judgment behind them. The management bottleneck is that argument stated organizationally.

Also on my radar

Harvey is reportedly in talks at a $15.5 billion valuation, which makes the vertical AI timing squeeze concrete: the existential question is not whether the product works, it's whether a foundation model provider decides legal reasoning is a feature before Harvey is dug deep enough into workflows to survive it. Anthropic reportedly walking away from a roughly six-billion-dollar deal for Decart is useful context — even the frontier labs are being selective about what they absorb and when, which makes the M&A environment less predictable than the headline pace suggests. On the hardware side, China is reportedly closing the last major foreign dependency in its domestic AI stack while still allowing selective imports, a dual-track hedge that is more sophisticated than the self-sufficiency framing usually given to it. And the genuinely novel one: a complete structural brain map being used as a programmable substrate for AI research — early, but it offers a reference architecture for how intelligence organizes itself that isn't derived from gradient descent.