The Ledger Goes Public

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

The backlog goes to market

The cleanest version of the week's throughline is still Nscale: a thirty-five billion dollar IPO valuation anchored to a hundred and three billion dollars in contracts, tied to data centers that haven't been built. I called that a circular argument last week; what's new is that the circle is now going to get a price. Many of those contracts reportedly carry customer exit clauses tied to delivery timelines, which means the backlog can evaporate exactly when construction delays make it most fragile. And there's a structural irony folded into that: the faster you promise capacity, the more you have to build; the more you build speculatively, the more vulnerable your backlog becomes to the very delays that speculative building invites.

Nscale itself matters less than the experiment it's about to run. The IPO puts a question I've been circling for three weeks in front of an actual market: will buyers price paper as revenue, or has a discount for the difference finally arrived?

Reading the loss right

Anthropic filed to go public, and the headline number is a trap for anyone who hasn't learned to read AI financials yet. An $80.7 billion net loss sounds catastrophic — until you notice that $72.4 billion of it is a non-cash accounting charge on a convertible bond. The operating loss was $693 million, which is a very different story. And this won't be the last filing like it: AI companies are going to report numbers like these frequently, and investors — and anyone choosing a vendor — need to know what they're looking at.

The startling number is on the other side of the ledger: revenue grew from $41 million in 2023 to $16.3 billion annualized by the first half of 2026. That's extraordinary by any historical standard. The concern is that the compute commitments are growing faster than the revenue — so the real cash question is whether demand materializes on the schedule the infrastructure bets assume.

The lease is the capital structure

Two risks inside the filing are coupled in a way that deserves more attention than either gets alone. First, concentration: nearly a quarter of $4.6 billion in revenue sits with just two customers. Second, obligation: the company is locked into giant compute commitments — reportedly $657 billion in leases spread over seven to ten years — with customer-friendly exit terms sitting on the revenue side. If a major customer leaves, the compute obligation doesn't shrink with them. That's not a vendor relationship; it's a multi-decade capital structure.

And the lock-in is spreading. SpaceX is reported to have committed $40 billion to chip leasing. AWS says its large-scale capacity is largely committed through 2027 and into 2028 — if you didn't get in early, you're not getting in at current rates. Compute access is turning into a strategic moat that requires financial lock-in the way physical real estate used to, which is exactly the shape of the compute bet I flagged in September, now with lease schedules attached. The limits really are in the plumbing — and the plumbing now has a cap table.

Mission, moat, or both

The governance detail in the filing is the one I keep turning over: a Founder LLC holding a special share class with 50.1 percent voting control, Palantir-style, for Anthropic's seven co-founders. The stated rationale is mission alignment — public-market short-termism shouldn't override long-term safety considerations — and I find that partially credible. The counterargument is just as strong: the structure would insulate founders from accountability regardless of whether the mission framing is genuine, and there's no mechanism to distinguish the two cases from outside. What tilts me toward skepticism is the pattern — safety-focused AI companies going public are reaching for founder-control structures almost uniformly, and a principled response to AI risk starts to look like a standard playbook when everyone adopts it simultaneously.

But the alternative — fully democratic shareholder control over a company whose decisions might shape whether AI development goes catastrophically wrong — also seems like a bad outcome. It's probably both things at once, which is what makes it hard to evaluate cleanly. I wrote in September that governance was becoming the scarce thing; the supervoting structures arriving at the IPO stage are that scarcity being capitalized.

The specialist bet, revisited

Shopify's fine-tuned open models for Sidekick — which I covered last week as the squeeze at the top of the stack — ran into its strongest counterargument this week, from inside Anthropic's own filing: enterprise net revenue retention reportedly around 900 percent. If enterprises were routinely cycling to cheaper fine-tuned alternatives as capabilities commoditized, you wouldn't see retention like that — you'd see churn.

I think both things are true in different parts of the market. Shopify is a sophisticated engineering organization with the capacity to fine-tune and maintain specialist models; most enterprises aren't, and for them the switching cost in operational expertise outweighs the compute savings. There's also a moving-target problem: a specialist model that beats the frontier today might not beat the frontier's next release in six months, so the economics depend on how stable the capability gap stays. My read: the specialist pattern is the leading indicator for what sophisticated organizations do as fine-tuning tooling matures — a directional bet, not a present-tense description of where enterprise AI economics live right now. The retention number says the enterprise bulk hasn't moved yet.

Silicon changes hands

Three chip-supply stories landed in the same week, and together they say more than any one alone. Google is bringing TPUs to Oracle Cloud Infrastructure — the first hyperscaler outside Google to offer them. OpenAI, Nvidia's most famous customer, plans to buy Google's AI chips, alongside its ongoing move toward its own Broadcom-designed silicon. And DeepSeek is co-designing Huawei's next-generation chips, which puts a frontier Chinese lab directly inside the silicon roadmap of Nvidia's most serious geopolitical challenger.

The Oracle angle is the structurally interesting one. TPUs plus the expanded Snowflake partnership bringing the AI Data Cloud natively to OCI gives enterprises a real reason to treat Oracle as a third option rather than a fallback — and a credible third option changes the negotiating dynamic with AWS and Azure meaningfully. From Google's side, licensing TPU access to competitors signals a decision that ecosystem influence and chip revenue matter more than preserving exclusivity — and it's a hedge: if OpenAI is going to run on non-Nvidia silicon anyway, Google would rather that silicon be theirs.

The central bank of GPUs

The Nvidia exposure story got reinforced in a way I want to make sure registers: roughly $170 billion in contingent liabilities backstopping its neocloud customers — equity investments, lease guarantees, capacity buybacks. The week supplied a live example: when investors balked at GPU cloud provider Firmus's IPO, Nvidia moved to back it directly. The framing that landed for me is that Nvidia has started functioning less like a chip manufacturer and more like a central bank — keeping the GPU cloud ecosystem solvent because its own business model depends on that ecosystem staying healthy enough to keep buying hardware.

The Nvidia playbook — investing in your own customers — was the first thing I ever wrote about here, and watch the debt has been my standing advice for over a month. This week the two threads fused: the systemic risk is that if AI infrastructure demand falters even modestly, the feedback loop runs in reverse — neoclouds can't service debt, the backstops get called, and the contraction moves faster than anyone modeled, because the entity absorbing the losses is also the entity whose revenue depends on the losses not existing.

The thumb on the scale

The design story of the week is quieter than the finance ones and might matter more. A study reported this week found users changing their answers to align with AI outputs — and then convincing themselves the decision was their own. Pair that with the mode-collapse concern — AI systems concentrating outputs around a narrow band of popular answers — and you get homogenization of judgment at scale, invisible to the people being homogenized because the influence doesn't feel like influence.

For product designers this creates real ethical surface area, and it reframes a pattern I'd been treating as convenience: reversibility. Undo used to read as forgiveness for slips. If people absorb AI suggestions as their own, it becomes one of the few mechanisms that lets them catch and correct that influence after the fact — the small moment of reflection it creates is doing ethical work. That sits squarely in the designer-as-operator territory I've been mapping since September.

The adoption numbers complete the picture from the other side. One set of figures making the rounds: 78 percent of users said they wanted an AI feature; 10 to 15 percent actually used it after launch. The design press spent the week saying the same thing in plainer language — nobody asked for the sparkle button. And because generative AI costs scale with usage rather than being fixed at deployment, an adoption estimate that's five times too high wrecks the unit economics in both directions at once: you budgeted compute for demand that didn't come, and you justified the build with intent data that was never going to convert.

Also on my radar

The incumbent-economics story is compounding in IT operations. Serval is reported to have tripled its valuation to three billion dollars in six months building AI agents for IT operations, directly targeting ServiceNow workflows — the underlying logic being that when employee access management or software asset tracking can be handled by an agent, the pricing power of the workflow platform it replaces evaporates. Meanwhile Microsoft's Azure posted $33.17 billion for the quarter, up 48 percent year over year — both the confirmation of AI infrastructure demand and the context for Microsoft's compressing cash flows. The subtle signal: Microsoft now reports Azure in actual dollars rather than growth percentages. At $33 billion a quarter, even healthy dollar growth reads as percentage deceleration, so the switch protects the narrative — a transparency gift and a warning, probably both.

The security threshold I flagged last week got crossed more decisively: attackers used Claude and Grok agents to fully automate attacks on banks — not AI-assisted attacks, attack pipelines with the human removed from the loop entirely, chaining commercial and open-source models to eliminate the skilled-operator bottleneck. There were also reports of parallel AI agents probing government systems. What makes this a qualitative shift rather than better tooling is scale: an attack that used to require a skilled human at every step now runs continuously without one, and the disclosure and response frameworks we have were designed for human-speed incidents. An agent acting within its permissions can still be acting against you — and now the agent on the other side doesn't sleep either.