The AI bubble as an operational risk: what Nvidia, OpenAI and Oracle reveal about your AI subscription

Most people read the AI bubble debate as a stock market story: Nvidia, OpenAI, Oracle, credit ratings, all of it far away from the desk where actual work happens. In his latest video, Sasch the Tasch flips the perspective. For him the bubble is primarily an operational risk for anyone who has built a business on a single US subscription. We worked through the video, checked its core numbers against tagesschau, Der Spiegel and a market commentary, and sorted out what it means for anyone running agents in production. The short version: this is not about picking a crash date, it is about foundations. And foundations can be checked without reading a ticker.

The core of the video is the financial entanglement of the big AI players. According to the video, Nvidia invests in OpenAI, OpenAI needs datacenter capacity and buys it from Oracle, and Oracle in turn buys Nvidia chips. Inside this circle money moves from left to right and back at enormous scale, a volume of roughly 800 billion dollars. German reporting tells the same story: tagesschau describes how Nvidia pushes back against AI worries while stressing that Nvidia and OpenAI remain deeply intertwined, and Der Spiegel reports growing fear on the markets that the bubble bursts, precisely because companies like OpenAI and Nvidia are valued at dizzying levels.

The concrete numbers in the video sketch the shape of the risk: CoreWeave, a cloud provider built for exactly this infrastructure, was already under heavy debt pressure in 2024 and now carries 21.6 billion dollars in total debt, based on last year's figures. Oracle's credit rating has slipped toward BBB territory, close to junk, and its order backlog sits at 638 billion dollars. According to the video, more than half of that backlog depends on OpenAI. If that single customer disappeared, half of the backlog would disappear with it. This is the kind of dependency people usually associate with supply chains, not with software.

The second concentration sits inside the model companies themselves. According to the video, roughly 80 percent of revenue at OpenAI and Anthropic comes from one percent of customers, and at Anthropic, Cursor and GitHub together are said to account for about a quarter of total revenue. Losing two or three large customers can erase a massive revenue block, and the pressure then travels down to subscribers: higher prices, tighter limits, vanished features, discontinued models. That is why Sasch calls a single subscription not a foundation but a risk you either carry consciously or spread out.

The second part of the video deals with a trend Sasch considers already underway: open source and open weight models. His central indicator comes from the reports of OpenRouter, a service that provides API access to a large number of models. The share of US models in token consumption there has fallen from around 70 percent to 30 percent within roughly eight months. People keep using ChatGPT and Claude, but the heavy token throughput increasingly runs through Chinese open weight models. Sasch runs his own stack the same way: frontier models from the US for local work, Chinese open weight models for almost everything in the cloud.

The price gap makes the shift mechanical. In the OpenRouter comparison shown in the video, Claude Opus costs 25 dollars per million output tokens, while GLM charges 7.5 cents per million input tokens and 25 cents per million output tokens, roughly a hundredfold difference on the output side. Sasch compares new models every two weeks in the OpenRouter compare view along three axes: intelligence, coding and agentic capability. In his reading the quality gap has become small while the price gap remains enormous. He also names companies that have already moved: Airbnb reportedly works on a Qwen basis and cut its revision time from three hours to six seconds, and Coinbase is said to run on GLM and Kimi, a coding focused model that is clearly cheaper than Opus.

The trend has a dark side the video does not hide: a hacker group reportedly retrained the open Qwen model, stripped its safety guardrails and turned it into an efficient tool for writing exploits. Open weight therefore does not automatically mean safe, it is a decision per process, not a creed. On data protection Sasch stays level headed: caution is warranted with Chinese API providers, and he suspects similar practices on the US side. His answers are local AI on your own machine and European providers that host their models in Frankfurt, keeping data out of foreign jurisdictions.

The most practical part of the video is a simple framework. Sasch separates reactive agents from proactive agents. A reactive agent is a harness you sit in front of, issuing instructions, and he counts Claude Code, Codex, Cursor and Hermes in that group. Proactive agents are workflows that run on a cron job, a trigger or an API call, execute on their own and make their own decisions. His setup mirrors the split: reactive work runs on a US cloud subscription, proactive processes run on Ollama or OpenRouter with his own harness. The point is not the specific tools, it is the separation itself: two operating modes with different requirements get different model bases.

His second recommendation sounds conservative on purpose: stay with your harness. There are many harnesses by now, and constant switching costs energy because every switch rebuilds your setup. A harness you know inside out produces more output than a newer one you are still learning, even when the newcomer looks better on paper, because you know the pitfalls of your own system. The cloud harness should still be modular so that different models can be attached, and local operation is desirable but secondary. And one filter rule for model selection: if a model does not fit into your harness, it does not get evaluated at all.

At the end of the video there are four steps for anyone who wants to act now. First, list your own processes and calculate what happens if prices rise sharply: would the workflow still be worth it? Second, test those processes with different models, OpenRouter is a practical place to start, and check first whether your harness can even run the model. Third, use different models for different steps inside proactive agents. His example: a coding agent that updates a website all day runs on Kimi Code, while a research and SEO agent that needs context and intelligence runs on GLM. Fourth, keep a switching document that describes how a change between two models works, which he considers relatively easy to implement.

The golden nugget at the margin targets API heavy workflows: batching. Instead of sending many small requests, you bundle many requests into one, which is significantly cheaper. The catch: batching usually only works through the providers' original first layer APIs, not through wrappers like OpenRouter. Sasch reports cutting 50 percent of the costs of one of his tools this way, while stressing that this only affects a small share of users. For anyone operating their own AI powered tools it is still the fastest lever in the entire video.

The real message of the video is not that the bubble bursts on a specific date. It is concentration: a few large customers carry the revenue of the model providers, a few frontier models carry the workflows of many businesses, and a few datacenter operators carry the infrastructure. We know the same pattern from our own agent operation and solve it differently: instead of a subscription without limits there is a request budget with a weekly cap, instead of one model for everything there is an exchangeable local core, and instead of blind trust there is a measurement chain that shows what every run costs. Reducing dependency before it gets expensive, that is the whole playbook.

Fairness demands the counterpoint. A market commentary by DJE Kapital calculates that Nvidia trades at an expected price to earnings ratio of about 28 for 2026, clearly below the valuation level of a leading network equipment vendor. That does not invalidate the entanglement, it weighs it differently: not every high market cap is bubble air by default. Crash dates remain speculation, checking your own foundation remains a duty. The video sits exactly in that middle ground: no panic, four concrete steps you can start on a single afternoon.

Honesty first: this is a video breakdown, not a market audit. The critical numbers, the CoreWeave debt, the order backlog, the 800 billion dollar circle, the revenue concentration at OpenAI and Anthropic, the OpenRouter shares, the Airbnb and Coinbase examples and the 50 percent batching saving all come from the video and its own research, not from our own primary checks. tagesschau and Der Spiegel confirm the mood, the entanglement and the rising market anxiety, but not every individual figure. The Nvidia valuation multiple comes from a fund commentary, not from our own balance sheet reading. ETF rankings and credit ratings were not independently verified. What still holds is the direction: a single subscription is a fragile foundation for an AI business, and the levers against that fragility are concrete. That is what makes the video worth reading about.

Our own assessment from agent operations, as of 20.09.2026