AI: Bubble or Breakthrough? What Enterprise Leaders Should Really Be Watching
By Nick Dimitrov
The Stakes Are Massive
Is the AI hype a trillion‑dollar house of cards, or the next internet gold rush? This question is not just academic for CEOs in betting, retail, fintech, crypto or travel – it’s a tactical dilemma that will shape product roadmaps, budget allocations and competitive advantage for years to come. On one side, contrarian investors such as Michael Burry have started flagging a bearish stance on high‑profile AI stocks, warning that soaring valuations may be disconnected from sustainable cash flows. On the other, technologists like Andrew Ng and Clem Delangue (Hugging Face) argue that the AI application layer is still under‑invested and that the real upside lies in specialised, domain‑specific models rather than the headline‑grabbing large language models (LLMs). The truth for enterprise marketers is not a binary “bubble vs. boom” but a mis‑allocation of capital that creates a ripe opportunity for teams that can build agentic AI workflows that directly move the needle on ROI.
What the Financiers Are Saying
Financial executives have been unusually vocal about the risk of an AI‑driven froth. Goldman Sachs CEO David Solomon warned that “a lot of capital that was deployed doesn’t deliver returns”. This line appears in the Yale Insights commentary on the AI hype and is echoed around the net. Jeff Bezos called the environment an “industrial bubble,” and Sam Altman of OpenAI cautioned that “people will over‑invest and lose money.”
Data backs their concern. JPMorgan’s Michael Cembalest notes that the Magnificent Seven – APL (Apple), MSFT(Microsoft), AMZN (Amazon), GOOGL or GOOG (Alphabet), META(Meta Platforms), NVDA (Nvidia), and TSLA (Tesla) have generated 75 % of S&P 500 returns and 90 % of AI‑related capital spending since ChatGPT’s launch in November 2022. Meanwhile, RBC’s Kelly Bogdanova points out a widening gap between the tech sector’s market‑cap and net‑income, a classic bubble symptom that began to converge in 2023‑24.
Michael Burry has taken a contrarian approach, publicly stating that he is short‑selling AI‑heavy equities because the “cap‑ex‑to‑revenue mismatch” looks unsustainable. The specific contract sizes are not disclosed, but the signal is clear: a leading value investor sees a material risk in the current AI valuation model.
Why the Tech Leaders Remain Optimistic
The same data that scares financiers also fuels a very different narrative among AI pioneers. Andrew Ng (DeepLearning.AI) repeatedly emphasizes that the application layer is starved of investment. A recent MIT report (2025) found that 95 % of 52 surveyed firms recorded zero ROI on generative‑AI pilots, despite $30‑$40 B of spend. Ng argues the real value will come from agentic workflows—software that combines LLM output with task‑specific logic, data integration and business rules.
Clem Delangue, co‑founder and CEO of Hugging Face, clarified that the current frenzy is really an “LLM bubble.” In a TechCrunch interview (Nov 18th, 2025) he explained that while LLMs dominate headlines, smaller, specialised models for finance, chemistry, video or audio will dominate many enterprise use cases. He cites the example of a banking‑chatbot that doesn’t need a general‑purpose LLM capable of philosophising about the meaning of life, but a compact, fast, on‑prem model that can be audited and run at low cost.
How are AI Models Being Commoditised
Google’s Gemini 3 Pro and Nano Banana Pro (released in November 2025) illustrate that frontier models are rapidly becoming commoditised services—they are free to try through their new IDE Antigravity or https://gemini.google.com, while being priced on demand per‑token through their APIs. The ability to be swapped with a single configuration change makes it trivial to be integrated in a massive line of specialised agentic flows, like digital assets generation, copy generation, competitions analysis, etc. For enterprises, this signals that model selection is becoming a plug‑and‑play decision, while the real engineering work lies in data governance, orchestration and ROI‑focused solutions.
The Gold Mine: Where Capital Meets Real‑World Value
The financial data tells us that most AI‑related spend is funnelled into infrastructure and model training – the “shovel” side of the classic gold‑rush analogy. JPMorgan estimates that 75 % of AI‑related market‑cap growth comes from the Magnificent Seven, yet only a fraction of that spend translates into incremental earnings.
For marketers, betting on the application layer is where the upside lives. Three tiers make up the AI stack:
- Infrastructure – GPUs, data‑centers, networking (dominated by Nvidia, AMD, cloud hyperscalers).
- Training – large‑scale model development (expensive, high‑risk, diminishing marginal returns).
- Applications – agentic workflows, domain‑specific models, low‑code AI orchestration (high‑impact, high-return, lower‑cost).
Because most capital is still pouring into tier 1 and 2, the application tier is comparatively starved – a classic “mis‑allocation” that creates a competitive moat for organisations that can rapidly prototype, test and ship AI‑enhanced features.
A Use Case Scenario
A retailer could build an edge‑AI recommendation engine using a tiny, fine‑tuned transformer hosted on their own Kubernetes cluster. Provided a 24-7 request saturation, the solution could cost 80 % less than a comparable cloud LLM API with a fine-tuned commercial model. The project requires simple change in the model provider configuration to use the local model configured with a set of tools to pull in product‑catalog data via MCP for example. This mirrors the “low‑code AI agents” that Andrew Ng champions and demonstrates the ROI upside when you focus on the application layer rather than the hardware.
Call to Action for Enterprise Leaders
The bubble is real at the top, but the floor is fertile for those who build the freight trains that run on the existing infrastructure tracks. Start by inventorying your AI spend: how much is on infrastructure vs. application? Pilot a low‑code orchestration platform that can swap models via a single config change (e.g., OpenRouter, LangChain or n8n which is amazing for prototyping). Layer in governance – automated PII redaction, audit logs and role‑based access – before you route data to multiple providers. Finally, measure impact on a core KPI (conversion, click‑through, fraud‑loss reduction, time saving) rather than on token count and you are off to a great start to integrate agentic AI into your daily work processes.
Nick Dimitrov, Chromeye CEO
Talk to us about the most promising AI‑application you’re experimenting with, and join the discussion for more insights that turn hype into tangible ROI.
Sources: Yale Insights (Oct 2025), Goldman Sachs, JPMorgan Cembalest, MIT State‑of‑AI‑in‑Business 2025, TechCrunch (Delangue, Nov 2025), Fortune (Solomon quote), CNBC, PitchBook, UBS AI‑spending forecast.