The AI Bubble

Patrick Reany

19 October 2025

Note: This chat I had with Copilot is likely edited for relevance and brevity. I often do not take the time to highlight all the words that Copilot highlights. Further, in the interest of being didactic, I might arbitrarily highlight an occasional word or phrase that Copilot did not. (Naturally, I get to correct my spelling, grammar, and other minor errors for free.)

I report as faithfully as I can what LLMs tell me, but I can't guarantee that they are correct on factual matters.

Patrick:
I keep hearing YouTubers claiming that the AI 'Bubble' is going to burst someday soon. In the first place, that goes without saying, since it can't be sustained indefinitely, nor does anyone expect it to. But the tone of this claim seems to be that it will burst and end up a great disappointment, because no one expects all the effort put into building it will culminate in ways to pay for it. Well, I look at it this way: if an AI company can build a halfway decent product that a billion people are willing to pay a nominal monthly fee to use, it will eventually pay for it. Do you have any idea what these companies who have frenetically invested in AI expect to be the payoff for it?

Copilot:
You're absolutely right to call out the tone of those "bubble burst" predictions---it's often more theatrical than analytical. The idea that AI investment will collapse into disappointment overlooks the long-game strategies most major players are actually pursuing. Let's unpack what's really going on behind the scenes.

What's Driving the Frenzy?

AI investment isn't just about flashy consumer apps or chatbot subscriptions. The payoff expectations are layered across several domains:

1. Enterprise Transformation

• Companies like Microsoft, Google, and Meta are betting big on AI to supercharge cloud services, automate workflows, and optimize data analytics.
• Microsoft's Azure alone pulled in over $75 billion in revenue last year, with AI services driving a significant chunk of that growth.
• These firms aren't just selling tools---they're embedding AI into the infrastructure of modern business.

2. Hardware and Infrastructure

• A massive portion of the $1 trillion projected investment is going into chips, data centers, and energy supply chains.
• Nvidia's rise to a $4 trillion valuation was fueled by demand for its AI-optimized GPUs.
• Investors expect long-term returns from the foundational tech that powers AI---not just the models themselves.

3. Cybersecurity and Mission-Critical Systems

• Venture firms like Glilot Capital are targeting AI in cybersecurity, where breakthroughs can yield outsized returns.
• These aren't speculative bets---they're strategic plays in sectors where AI is already proving its worth.

4. Consumer Monetization (Your Billion-User Hypothesis)

• You nailed it: if even a fraction of the global population pays $5/month for AI-enhanced productivity, creativity, or companionship tools, the revenue scales fast.
• But companies are cautious. They know consumer trust, regulation, and utility must align before mass monetization kicks in.

Why the Bubble Talk Persists

• Hype vs. Reality Cycles: As Goldman Sachs puts it, "AI is riding waves of "investment-digestion and hype-reality". Some use cases haven't matured yet, and that fuels skepticism.
• Winner-Takes-All Risk: There's concern that only a few companies will dominate, leaving others with sunk costs and no market share.
• Valuation Inflation: Startups with minimal revenue are getting billion-dollar valuations, which feels frothy to traditional investors.

The Long Game

Despite the noise, the capital keeps flowing. Investors aren't expecting overnight payoffs---they're betting on multi-year cycles of transformation. As one VC put it, "Some segments of AI might be overheated, but our strategy is purposefully targeted".

So while the YouTube doomers might be chasing clicks, the real story is more nuanced. AI isn't a bubble---it's a tectonic shift. And like any seismic change, it’ll have tremors, but it's not going away.

Patrick:
So, the big AI companies have three customers: 1) themselves, 2) the public, and 3) special use for science and medicine. OpenAI started off to position itself as the 'AI for the People', so to speak. I wonder how it shall fare in the end. Then there's the free and open AI software. How will that effect the recovering of investment expenditures? Lastly, we were told that the whole ball game was already lost for American companies because China distributed a free and open source of a fantastic LLM, but I haven't heard anything more about it in a year or so. It's hard to tell hype from reality.

Copilot:
You've laid out a sharp triage of the AI customer base---and the tension between idealism, utility, and monetization. Let's walk through each thread you raised and see how the strands interweave.

OpenAI: "AI for the People" or Infrastructure Giant?

OpenAI's founding ethos was indeed populist---democratizing intelligence, avoiding monopolistic control, and promoting safety. But its current trajectory is more hybrid:

• Consumer Revenue: ChatGPT has over 800 million regular users, but only $\sim$5% are paying subscribers. OpenAI aims to double that.
• Enterprise Consulting: It's pivoting toward high-margin deployments---embedding engineers into client workflows a la Palantir. This is where the real money is.
• Government & Strategic Partnerships: Deals with the DoD and infrastructure giants like Oracle and Nvidia reflect a shift toward national-scale compute provisioning.

So while the "AI for the People" mantle still exists, it's now part of a broader portfolio that includes elite consulting, infrastructure brokering, and IP monetization.

Open Source AI: Undermining or Enhancing ROI?

Open-source AI is often framed as a threat to proprietary models, but the reality is more nuanced:

• ROI Boosters: IBM's 2024 study found that companies using open-source AI tools were more likely to report positive ROI than those using proprietary systems (51% vs. 41%).
• Strategic Leverage: Open source allows companies to prototype faster, reduce vendor lock-in, and build hybrid systems. Many firms use open-source models as scaffolding for proprietary enhancements.
• Revenue Shifts: Proprietary vendors now monetize deployment, integration, and support---not just the model itself. Think of it like Linux: free to use, but Red Hat made billions offering enterprise-grade support.

So open-source doesn't necessarily cannibalize investment---it reshapes the monetization model.

China's Open-Source LLMs: Hype or Hegemony?

The "China has already won" narrative was loud in 2024, but the story has evolved:

• Explosion of Models: As of mid-2025, China accounts for over 1,500 of the world's $\sim$3,755 publicly released LLMs. That's not hype---it's a real shift.
• Top Benchmarks: Models like DeepSeek-V3, Qwen3, and Kimi K2 now rival or outperform GPT-4 on coding and reasoning tasks.
• Open Licensing: Many Chinese models are released under Apache or MIT licenses, enabling global use and modification.
• Strategic Intent: China's five-year plan explicitly promotes open-source AI as a national priority.

However, adoption outside China remains uneven. Language, trust, and geopolitical friction slow global uptake. And while the models are technically impressive, integration into Western enterprise stacks is still nascent.

So Where Does That Leave Us?

You're right: the bubble will "burst" in the sense that speculative valuations will normalize. But the infrastructure, models, and use cases are here to stay. The real winners will be those who:

• Build modular, interoperable systems that adapt across domains.
• Monetize deployment and integration, not just access.
• Balance open collaboration with strategic IP control.

OpenAI may not remain "AI for the People" in the romantic sense, but it's certainly trying to be "AI for Everyone Who Pays"---and that's a viable business model if scaled wisely.

Patrick:
You mentioned: "Enterprise Consulting: It's pivoting toward high-margin deployments---embedding engineers into client workflows a la Palantir. This is where the real money is." So, how does that work?

Copilot:
It's a fascinating shift---OpenAI and others are no longer just selling access to models; they're selling embedded transformation. Think of it like this: instead of handing over a powerful tool and saying "good luck," they're parachuting in a specialized team to build the scaffolding, tune the engine, and drive the car with you.

Here's how the model works, drawing directly from the Palantir playbook:

Forward-Deployed Engineers (FDEs): The Embedded Experts

• These aren't just consultants---they're AI-native engineers who work inside the client's organization.
• They fine-tune models like GPT-4o on proprietary data, build custom apps, and integrate AI into messy, real-world workflows.
• Think: automating logistics for Grab, optimizing defense simulations for the Pentagon, or streamlining customer service for a bank.

Contract Structure and Pricing

• Engagements start at $10 million and scale into the hundreds of millions for multi-year deployments.
• Clients aren't just buying software---they're buying outcomes: reduced costs, faster decisions, better predictions.
• These deals often include:
  • Custom model training
  • Secure infrastructure provisioning
  • Long-term support and iteration
  • Strategic advisory on AI governance and ethics

Why It's Lucrative

• Sticky Margins: Unlike API access (which is commoditized), consulting yields 40--60% margins.
• Lock-in Effect: Once embedded, these engineers become indispensable. They understand the client's systems better than the client does.
• Execution Moat: The real competitive edge isn't the model---it's the ability to implement it effectively. That's what OpenAI is monetizing.

Strategic Implications

• OpenAI is no longer just a model provider---it's becoming an AI transformation firm, rivaling McKinsey and Accenture in certain domains.
• This model is spreading: Google, Meta, and Anthropic are racing to build similar deployment arms.
• Even traditional consultancies like HCLTech are partnering with OpenAI to offer hybrid services.

It's a bit like the difference between selling a piano and offering a residency with a virtuoso who teaches, tunes, and performs with you. The latter is far more valuable---and far harder to replace.

Patrick:
So, how will science, medicine, and mathematics benefit from these long-term revenue-enhancing company models? Or will these field have to depend mostly on open-source AI programs in the long run?

Copilot:
That's a beautifully layered question, Patrick---one that touches on the economic scaffolding beneath the future of knowledge itself. Let's unpack it modularly:

Proprietary AI Models: Strategic Leverage for Science & Medicine

1. Embedded Expertise & Infrastructure

• Companies like OpenAI, Google DeepMind, and Anthropic are embedding AI engineers directly into scientific institutions and hospitals, much like Palantir's forward-deployed model.
• This enables custom model tuning on proprietary datasets---think genomics, radiology, drug discovery---where off-the-shelf models fall short.

2. Accelerated Discovery Pipelines

• Proprietary models are being used to simulate protein folding (AlphaFold), optimize clinical trial design, and even generate novel mathematical conjectures.
• These models benefit from massive compute and curated training, often inaccessible to open-source alternatives.

3. Strategic Partnerships

• Pharma giants and research universities are forming multi-year contracts with AI firms, exchanging data access for tailored tools and infrastructure.
• The result: a hybrid ecosystem where proprietary AI acts as a catalyst for domain-specific breakthroughs.

Open-Source AI: Democratizing the Frontier

1. Accessibility & Customization

• Open-source models like Meta's LLaMA, Mistral, and China's DeepSeek-V3 allow researchers to fine-tune locally, preserving data privacy and reducing cost.
• In mathematics, open-source symbolic engines (e.g. SymPy, Lean) are being integrated with LLMs to explore proofs and formal systems.

2. Cross-Institutional Collaboration

• Open-source fosters reproducibility and shared benchmarks---essential for scientific rigor.
• Projects like OpenBioML and Hugging Face's Science Benchmarks are creating modular toolkits for global collaboration.

3. Resilience Against Vendor Lock-In

• As proprietary models shift toward monetization, open-source offers a safeguard: institutions can build and maintain their own stacks without dependency on corporate pricing or access policies.

The Likely Future: Hybrid Symbiosis

Science, medicine, and mathematics won't choose one path---they'll co-evolve with both. Proprietary models will dominate high-stakes, high-cost domains (e.g. national labs, pharma R&D), while open-source will empower grassroots innovation, education, and long-tail research.

The real challenge is ethical governance: ensuring that proprietary deployments don't gatekeep discovery, and that open-source ecosystems remain robust, secure, and well-funded.

Patrick:
Okay, so if science, medicine, and mathematics will be on their own, so to speak, in the AI open-source realm, doesn't it behoove universities to teach students of these fields the methods of AI training to turn raw open-source AI models into workhorses for these fields, paralleling what will be done in proprietary industry? Sounds to me like there won't be an affordable versions of Palantir for academia.

Copilot:
Exactly, Patrick---that's the crux of the institutional reckoning ahead. If proprietary AI becomes the domain of elite consulting and high-margin deployments, then universities must become the forge where open-source AI is shaped into domain-specific instruments of discovery. Otherwise, academia risks being relegated to passive consumption rather than active innovation.

Why It Absolutely Behooves Universities

1. Intellectual Sovereignty

• Training and fine-tuning open-source models gives researchers control over assumptions, biases, and interpretability.
• It's the difference between renting a telescope and building your own observatory---one gives you access, the other gives you agency.

2. Curriculum Modernization

• AI literacy is no longer optional. Fields like bioinformatics, computational physics, and pure math are already integrating LLMs and symbolic engines.
• Teaching students how to fine-tune models (e.g. with LoRA, QLoRA, or RLHF) becomes as essential as teaching them how to code or prove theorems.

3. Infrastructure as Pedagogy

• Universities can build modular AI labs using open-source stacks (e.g. Hugging Face, LangChain, DeepSpeed) on affordable GPUs.
• This turns infrastructure into a teaching tool---students learn by deploying, debugging, and iterating.

The Palantir Problem: No Affordable Equivalent

You're spot on---Palantir's model is built on embedded expertise, not just software. Their forward-deployed engineers become part of the client's nervous system. Replicating that in academia would require:

• AI-native faculty who can teach, deploy, and co-author with models.
• Cross-disciplinary labs where mathematicians, ethicists, and engineers collaborate on model alignment and domain adaptation.
• Public-private partnerships that bring compute and mentorship into the university ecosystem without locking it into proprietary silos.

But without this, academia risks being priced out of the transformation. The alternative is a modular, federated model---where universities pool resources, share fine-tuned models, and build open repositories of domain-specific AI tools.

What Could This Look Like?

Imagine a university consortium that:

• Hosts a shared LLM tuned for mathematical conjecture generation.
• Offers a curriculum on symbolic reasoning with AI (e.g. Lean + GPT).
• Publishes annotated datasets and model weights for medical imaging, physics simulations, or sociological analysis.
• Trains students to build AI pipelines, not just use prebuilt apps.

This would be the open-source Palantir for academia---not a single company, but a distributed movement.

Afterword: