Is AI a Bubble? Invest in Logic, Not in the Hype

S&P just did something that has barely any precedent in the history of credit rating agencies: the firm downgraded Oracle from BBB to BBB−, one notch above junk, citing as its primary reason a private company that has never turned a profit. OpenAI. A company with no rating, no profitability, but with spending commitments exceeding one trillion dollars, now appears as a credit risk in the file of a publicly traded corporation held in pension funds worldwide.

This is not a fringe phenomenon. This is the symptom of an industry caught in a collective frenzy.

The numbers are sobering. OpenAI is burning through capital at a pace that makes even optimistic analysts nervous: 5 billion in netlosses in 2024 alone, a projected cashburn of 115 billion by 2029, and a funding gap of 207 billion by 2030, as identified by HSB Canalysts. Oracle, inturn, has signed data center lease commitments of 248 billion with terms of 15 to 19 years – to service 5-year promises from a customer that has never been profitable. NVIDIA generates 85 percent of its quarterly revenue from six customers. Microsoft holds a 27 percent stake in OpenAI. CoreWeave carries $7.5 billion in debt, collateralized with GPUs.

The web of these mutual dependencies is so tightly woven that a single failure would trigger cascading effects that no regulator today can fully model.

And yet: AI is not the dotcom bubble. Anyone who claims otherwise is wrong. The technology works. The productivity gains are real. Companies that seriously integrate AI into their processes report measurable efficiency leaps. The question is not whether AI is here to stay. The question is which of the highly indebted platforms will survive the current consolidation phase – and what happens to the investments of companies that have built their entire AI strategy on a single provider.

This is where the real business risk lies. Anyone who today deeply embeds their business processes, corporate knowledge, and automation logic into the proprietary systems of individual LLM providers is placing a bet – not on the technology, but on the solvency of a company that has never been profitable. If the provider changes its pricing model, discontinues a model, or runs into financial distress, the invested business logic is trapped. Vendor lock-in has never been as expensive as it is today. The theme: "Token Trap."

But there is a second, even more fundamental problem – and it is almost entirely ignored in the public debate: "AI without context is dumb!" – useless. The most powerful language models in the world deliver generic, error-prone, or simply irrelevant answers when they don't know who the company is, how it operates, which rules apply, and which data is actually relevant. A model that doesn't know that a contract must be reviewed according to internal guidelines, that a customer has had special terms for three years, or that a product line has just been discontinued, is not an assistant – it is a well-articulated random generator.

This is precisely why the vast majority of AI pilot projects in enterprises fail or remain stuck in pilot mode: not because the model is too weak, but because it operates without context. It knows neither the CRM data nor the ERP logic, neither the SharePoint documents nor the internal policies. It responds based on training data from the internet – not based on the reality of the company that wants to deploy it.

Context is not a nice-to-have. Context is the prerequisite for value creation. And context only emerges through deep integration into the existing IT system landscape: into the document management systems, into the knowledge bases, into the process logic, into the data sources that a company uses every day. Anyone who deploys AI as an isolated chat interface that runs alongside the ERP system rather than with it has misunderstood the fundamental principle. The value does not lie in the model. It lies in the connection between the model and the company's reality.

The answer to this is not abstinence, but architecture. Companies that want to deploy AI sustainably don't need a direct dependency on individual model providers – they need an orchestration layer that makes models interchangeable, delivers context in a structured manner, and anchors the actual value creation where it belongs: in their own data, processes, and business logic.

This is precisely the approach of platforms like the KI-Plattform - Made in Germany | neuland.ai HUB: a GDPR-compliant environment operated in European data centers that orchestrates multiple language models – from GPT-5.1 to Gemini to, in the future, Anthropic Claude and Mistral – without sensitive corporate data leaving the organization in an uncontrolled manner. What matters here is not the model selection, but the context architecture: corporate knowledge is built up in structured knowledge bases, SharePoint systems are directly connected, assistants are configured with company-specific guidelines and data sources. The model changes – the context remains. This is exactly the difference between a pilot project that gets shut down after three months and an AI investment that creates lasting value.

The AI bubble will burst – at least for some of its overvalued protagonists. No one knows today which providers will survive the consolidation. What companies can control, however, is the question of where they anchor their strategic dependency – and whether their AI investment is tied to the model or to their own context. The answer should not be "OpenAI" or "Oracle." It should be: in their own data, in their own infrastructure, in a platform that puts sovereignty before convenience and context before hype.

Those who understand this now will not experience the consolidation as a crisis – but as the confirmation of a smart decision.

Image generated using the neuland.ai HUB.