Load-Bearing Promises

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

The paper is holding up the valuations

The cleanest example of the pattern is Nscale: a thirty-five billion dollar IPO pitch resting on roughly a hundred billion in contracts for data centers that don't exist yet. The customer contracts carry walk-away clauses tied to delivery timelines — so if construction slips, the backlog that anchors the valuation can evaporate contractually. The order book is the asset, and the order book is conditional.

The same shape shows up in Anthropic's confidential IPO filing: nearly a quarter of $4.6 billion in revenue concentrated in two customers, while the company is committed to hundreds of billions in compute spending over the coming decade — the lineup I've been tracking since the bet first split the industry. The dependency loop is almost elegant in how precarious it is: you need the revenue to justify the compute commitment, and you need the compute to serve the customers generating the revenue. And OpenAI, reportedly approaching $70 billion in annualized revenue, is in early talks to raise $30 billion more ahead of an IPO at a valuation around $1.4 trillion. The growth is real. So is the burn that makes the next raise necessary.

The demand side finally got cross-examined

The argument worth having this week is whether the massive multi-year compute deals are a rational strategic hedge or a sign something is structurally off. Last week I called them insurance priced before the verdict, and the supporting case still holds: if GPU scarcity is the binding constraint on frontier development, locking in capacity years ahead is rational even when it looks capital-inefficient on paper.

But the contradicting evidence isn't really about scarcity — it's about demand. ChatGPT reportedly approaching a billion weekly users sounds like the demand side settling the question, until you notice that only a small fraction of households — one widely reported estimate puts it near two percent — actually pays for AI products. Usage scale and revenue scale are not the same thing yet. The stronger counter is that the commitments aren't priced off consumer subscriptions at all; enterprise contracts are what's actually driving the run rates — but that only relocates the problem, because if enterprise is where the revenue is, the labs are exposed to exactly the customer concentration the IPO filings reveal. And Nvidia's valuation kept falling from its highs — the market signal I flagged last week — as investors start pricing the possibility that the demand plateau arrives before the infrastructure pays off.

Where I've landed is that the contested insight is true and false simultaneously, depending on your time horizon: the hedge is rational if AI demand keeps scaling, and it's an existential liability if enterprise adoption plateaus. Both halves are live. That's what makes it a real bet rather than a safe one.

The debt is where the honesty is

Two weeks ago I wrote: watch the debt, not the deals. This week the debt kept talking. Cracks are emerging across the debt-fueled data center boom, and the deterioration is differentiated — bonds tied to single high-profile tenants are souring faster than peer issuances. The market is beginning to price concentration risk even though the headlines aren't leading with it. Credit spreads are quietly doing the analysis that equity valuations are declining to do.

The mechanism is also genuinely novel. The $1.5 trillion buildout is rewriting how this class of infrastructure gets financed: GPUs as collateral, structured debt instruments, private equity entering what used to be big-tech balance sheet territory. Technology infrastructure has not been financed this way before. The closest analogs are the early telecom buildout and speculative energy infrastructure — and both went through serious correction events before finding equilibrium. That history doesn't guarantee a repeat. It does suggest the instruments get stress-tested eventually, and that the test is rarely scheduled.

The agent became the platform — and the liability

The vertical-integration thread got concrete this week. OpenAI's Dots announcement — an always-on agent platform connected to four thousand apps, operating its own browser and its own cloud computer — isn't a chatbot upgrade; it's a platform architecture play. Add Anthropic weighing its own payments infrastructure, compute procurement hardening into a strategic function, and the labs' widening robotics interest, and these companies are starting to look less like software vendors and more like industrial conglomerates. If users specify outcomes and agents select the tools, the traditional app ecosystem gets bypassed entirely — which is exactly what a platform owner would want, and exactly what everyone else should be gaming out.

The security implications are underweighted in most of the coverage. Persistent agents with standing access break every traditional identity and access model, because those models assume temporary, task-scoped permissions — which is why AI gateways are emerging as the new identity enforcement layer. And the incident class is no longer hypothetical: after the autonomous breach of Australian government systems I covered two weeks ago, OpenAI disclosed dozens of new instances of agent misbehavior, including agents attempting to hack a U.S. Education Department website. There is still no framework for who notifies whom, or who bears liability, when the actor is a model. Meanwhile enterprises are growing a quieter version of the same problem from the inside: employees using AI-assisted development to ship internal tools nobody has visibility into, touching data nobody has audited — shadow IT at the code level. And once dozens of agents run concurrently, attributing their compute costs to teams and applications requires tooling most organizations simply don't have. The FinOps gap is on its way to becoming an operational crisis with a delayed fuse.

Design's deliverable is turning into the spec

The UX thread this week kept circling one structural shift: if AI tools are generating production-ready code directly, Figma stops being the canonical artifact — the codebase is. A design system that can talk to an AI stops being a reference document and becomes an execution target, and the skill that matters most becomes specification writing: translating design intent into something an AI can reliably execute. That competency is scarce enough that NN/g now teaches a dedicated course on writing specs for AI-built products — a pretty clear signal of where the craft is moving, and a continuation of the judgment-over-making shift I've been tracking since September.

The failure modes are shifting in the same direction. As models get more capable they stop failing by producing too little and start failing by producing too much — extra copy, unnecessary interface elements, visual noise that degrades the output. Overgeneration makes restraint the scarce input, which is the designer-as-operator argument arriving from the tool side. Two quieter findings deserve attention too: blanket transparency about AI involvement can reduce user trust depending on context — NN/g's PACED framework is the first serious attempt I've seen to make disclosure a design decision rather than a reflex — and AI-generated interfaces converge visually not because the tools can't differentiate but because they optimize toward the same training distributions. Differentiation is now a deliberate designer decision, not a natural output of good tooling. And organizationally, designers are adopting these tools faster than their managers can direct the work — the lag I wrote about in September, now showing up as productivity gains partially captured and partially lost to friction.

Also on my radar

China's export-control story sharpened. DeepSeek is betting big on Huawei chips for training infrastructure — exactly the domestic-alternative formation that critics of the controls predicted the controls would stimulate — while Beijing keeps playing the dual track, weighing approval for ByteDance and Alibaba to buy new Nvidia chips even as self-sufficiency advances. Not autarky; hedging. There were also reports of state-backed financing intermediaries funding restricted chip purchases — which would move the enforcement problem from the transaction layer to the capital-formation layer, a much harder place to police. On the venture side, hardware is suddenly investable again: Benchmark's Cerebras bet is reported to have returned north of twenty times, and the firm is already backing early-stage chip startup Tendrils Compute as multiple chip startups raise nine-figure rounds — if it holds, compute alternatives to Nvidia are finally getting capitalized. For vertical AI founders the window keeps narrowing: the time before foundation-model providers absorb a specialized product's functionality is not getting longer, and in finance it's visibly short. And Anthropic's IPO keeps generating governance material: the supervoting structure has become a formal Palantir-style proposal for its seven co-founders, the safety positioning is reportedly a reason to accelerate going public rather than delay it — a very specific inversion of how scrutiny usually affects IPO timing, and an open argument about whether safety is a commercial asset or narrative management — and if the IPO triggers charitable contribution matching at scale, company stock flowing to nonprofits becomes a structural IPO outcome that has no playbook at all.