Wall Street has started selling the AI buildout. Our stack was sized to be boring about it

We run a sovereign AI platform the unfashionable way. Five model families we train and serve ourselves — language, speech, voice, vision, retrieval — sized for production workloads rather than leaderboard trophies; one OpenAI-compatible gateway, Qevron, that fronts those models plus 43+ external providers with metering and cost analytics on every token; and deployment wherever your risk committee sleeps best — on-prem, isolated, or in the cloud. No moonshot datacenter, no gigawatt press release, and for three years running the same quiet answer when asked why: because the bill for the other approach would eventually come due, and we did not want it arriving addressed to our customers. This month it started arriving. On 6 November, ING put a dozen data-center loans — half a billion dollars’ worth — up for auction, Vantage and CyrusOne debt among them, plus a $25 million slice backing a New Mexico site of the Stargate venture itself; Oracle fell 7% and CoreWeave 10% the same day. The question of late 2026 is no longer how big your cluster is. It is whose balance sheet your AI is sitting on.

What actually went on sale

The mechanics matter more than the headline. The ING portfolio, with bids closing 19 November, is not a fire sale by a tourist — it is a core infrastructure lender choosing to shrink exposure to the asset class it helped create. Nor is it alone: JPMorgan, Morgan Stanley, SMBC and MUFG have all been quietly shopping pieces of the AI debt they underwrote, with some reportedly willing to sell at a discount to non-bank buyers. Underneath sits the chain every credit analyst now draws on a whiteboard: the collateral is ultimately compute, and nobody agrees whether a GPU is a three-year asset or an eight-year one; depreciation uncertainty flows into tenant economics, tenant stress flows into the special-purpose vehicles that financed the buildings, and the equity cushion between those links is thinner than the keynote slides implied. And the honest caveat belongs in the same paragraph: nothing in the auction has priced yet — no discount is public, and an offer to sell is not proof of distress. What is new, and what moved Oracle and CoreWeave on the day, is visibility: the lenders closest to the money want less of it, and they no longer mind being seen wanting less of it.

Renting intelligence means renting a balance sheet

Here is the part that lands on ordinary AI customers, far from Wall Street. When your product runs on rented frontier compute, you do not just inherit a model — you inherit the financing behind it. Capacity that was priced to win market share gets repriced when the debt behind it does; rate limits appear where capex plans stumble; models are deprecated on schedules set by someone else’s unit economics; and in the tail scenario, your critical dependency is a counterparty workout. None of this requires a crash to hurt — it only requires your vendor’s CFO to have a worse quarter than your roadmap assumed. We made this argument about lock-in long before the debt made it fashionable: a single-vendor AI dependency is a business risk wearing an API key. The buildout’s financing just added a line to that risk register — one your procurement team never signed.

Where we stand: a stack sized to be boring

Our answer has three layers, and none of them is a prediction. The models are ours and sized for production: Verinova for language, Blab for speech synthesis, Solab for transcription, Spook for vision, VeriRAG for retrieval — tuned for the work our products and customers actually run, on infrastructure whose costs we control and understand. The gateway is the hedge: Qevron fronts our families and 43+ external providers behind one OpenAI-compatible endpoint, with routing, caching and per-token cost attribution — so when a provider repricing, rationing or wobble arrives, the exit is a config change, not a rewrite, and you ride the industry’s price war instead of funding one side of it. And the deployment is your call: on-prem or isolated, the hardware is on your books, depreciating on your schedule, serving your workloads — a capex decision you own instead of a financing structure you cannot see. Sovereignty was never only about jurisdiction. It is also about whose debts your roadmap services.

• Five in-house model families sized for production workloads — not leaderboard trophies
• One OpenAI-compatible gateway over our models plus 43+ providers — the exit is a config change
• Per-token metering and cost attribution — unit economics you can see, not infer
• On-prem and isolated deployment — hardware on your books, depreciating on your schedule
• No moonshot exposure: a platform whose capex story is deliberately boring

What we will not claim

Three honest limits. First, we are not calling a bubble. An auction is not a crash, no price has printed, and the hyperscale bet may yet pay out exactly as its backers project — our claim is about asymmetry of exposure, not prophecy, and anyone selling you certainty about 2027 is selling. Second, small is not free. Sizing models for production means we do not chase frontier records, and some tasks genuinely want a frontier model — which is precisely why the gateway exists: route those calls out, metered and priced, and keep the rest on economics you own. That is the design, stated plainly, not a limitation discovered later. Third, on-prem moves risk; it does not delete it. Hardware on your books is your depreciation, your utilisation problem, your ops burden — the same honest trade we have described for every self-hosted product we ship. The difference is that it is your trade, made at your size, visible on your own ledger.

When the cluster is a trophy, the debt is the leash. Own what you must, meter what you rent — and keep the exit a config change.

For teams in Türkiye this week was less a surprise than a confirmation: the BDDK’s localisation rules have long required banks to keep primary systems and original data in-country, which made “where does it run, on whose terms” a regulatory question here before it was a fashionable one — and the buildout’s financing now adds the creditor’s version of the same question. Our platform was built so the answer stays yours: our models and gateway on your premises if you want them, 43+ providers behind one metered endpoint when you want reach, and a cost ledger that makes the unit economics visible before someone else’s balance sheet makes them visceral — from Arpanet Bilişim A.Ş., engineered for the KVKK from the first line. The AI buildout’s bill is starting to circulate. Make sure your name is not on someone else’s copy.