Why I'm Skeptical of "Democratic" AI

The Open Source Dream

In 2024, the AI community rallied around a simple idea: if foundation models are going to reshape society, they should be open. Transparent weights, reproducible training runs, and community governance. The argument was compelling: monopolistic control over intelligence itself is too dangerous to leave to a handful of Silicon Valley labs.

Two years later, the dream is fracturing. Not because the vision was wrong—but because the physics of AI make true democratization nearly impossible.

The Compute Barrier

Training a frontier model in 2026 costs north of $500 million. That's not a typo. Between GPU clusters (H100s, now B200s), power draw measured in megawatts, and the engineering talent to prevent training collapse at scale, only a tiny group of actors can afford the entry ticket.

Meta released Llama 4 as "open weights." Mistral followed with their latest models. Both are incredible contributions. But neither organization is a scrappy underdog—they're billion-dollar entities with access to compute infrastructure that 99.9% of researchers will never touch.

This isn't criticism; it's physics. The hardware requirements for state-of-the-art AI create a natural oligopoly. Democratizing access to models is possible. Democratizing the ability to train them is not.

Data: The New Oil (And It's Already Claimed)

Compute is half the equation. The other half is data—and here, the concentration is even starker.

Google has 25 years of search queries, email, and YouTube transcripts. Meta has social graph data from 3 billion users. OpenAI, through its partnership with Microsoft, has access to enterprise datasets most startups will never see. The best models aren't just trained on scraped Common Crawl—they're fine-tuned on proprietary corpora that represent the actual distribution of human knowledge work.

"Open data" sounds great until you realize most valuable data is private by necessity: medical records, financial transactions, internal company docs. The organizations that control this data aren't going to hand it over, and synthetic data—while improving—still lags behind the real thing for specialized domains.

This creates a paradox: the most "democratic" path would be to centralize training in a few well-resourced labs and then distribute the resulting models broadly. But that's not democracy—that's charitable oligarchy.

The Alignment Tax on Openness

Even if we solve compute and data, there's a third problem: safety and misuse.

Open-weight models can be fine-tuned for harm. Jailbreaks, weaponization, misinformation at scale—these aren't hypotheticals. They're happening now. The more capable the model, the higher the risk surface.

Governments are already reacting. The EU AI Act imposes liability on model deployers. The US is debating compute thresholds for reporting requirements. China has outright restrictions on public model releases. The regulatory environment is moving toward gatekeeping, not openness.

This puts open-source advocates in a bind: do you release powerful models knowing they'll be misused? Or do you self-censor, undermining the entire premise of democratization?

Meta and Mistral are betting that model capability will plateau before catastrophic misuse becomes widespread. I'm not convinced. And neither are the regulators.

What Actually Gets Democratized

So if training frontier models stays centralized, where does that leave everyone else?

The actionable frontier isn't in training—it's in deployment, specialization, and orchestration.

• Fine-tuning: You don't need to train GPT-5 from scratch. You need to fine-tune a 13B model on your vertical (legal, medical, logistics) and beat the generalists at your specific task.
• Compound AI systems: The future isn't one mega-model. It's 20 small models orchestrated with retrieval, deterministic logic, and human-in-the-loop verification. This is where startups can compete.
• Inference optimization: Running models efficiently at scale is still an open problem. Quantization, distillation, edge deployment—these are areas where innovation doesn't require a $500M compute budget.
• Data moats: If you can't compete on model scale, compete on proprietary data. The best cybersecurity AI isn't trained on Common Crawl—it's trained on 10 years of attack traffic from Link11's scrubbing nodes. That's a dataset no one else has.

Access to *inference* can be democratic, even if training is not. The real question is whether that's enough—or whether it just creates a permanent dependency on a handful of model landlords.

The Realpolitik of AI Power

Here's the uncomfortable truth: concentration of AI capability isn't a bug. It's an emergent property of the technology itself.

Compute is expensive and centralizing. Data has network effects. Talent pools in a few hubs. Regulatory pressure favors large, compliant actors over scrappy challengers. Every force in the system pushes toward oligopoly.

We can still build on top of this concentrated base. APIs are cheap, models are improving, and fine-tuning is accessible. But let's not pretend we're in a world of "democratic AI." We're in a world of benevolent landlords—and the terms of the lease can change at any time.

Finding Leverage in a Centralized World

If you can't change the structure, you adapt to it. Here's how:

• Build vertical moats: Compete on domain expertise, not model scale. A 7B model trained on your proprietary data beats GPT-5 in your niche.
• Own the orchestration layer: The winners won't be the ones with the best single model. They'll be the ones who know how to chain 10 models, 3 databases, and 2 APIs into a reliable system.
• Hedge your dependencies: Don't build your entire company on one provider. Multi-model strategies (OpenAI + Anthropic + local inference) reduce your exposure to pricing changes and service outages.
• Contribute to the edges: If you can't train frontier models, contribute to tooling, evaluation, interpretability, and deployment infrastructure. The ecosystem matters, even if you're not training the core.

The Long Game

I'm not pessimistic about AI—I'm realistic about power.

The era of "anyone with a GPU and a dream" was always temporary. The next phase is professionalization: well-capitalized labs, regulated deployment, and a growing gap between the model-builders and the model-users.

That doesn't mean there's no opportunity. It just means the opportunity isn't in trying to out-scale Google. It's in finding the leverage points where small teams can still build defensible, valuable businesses.

Democratic AI is a beautiful idea. But in practice, AI will be as democratic as the internet itself: open standards, but concentrated power. The question isn't whether we can change that structure—it's whether we can thrive within it.

And I think we can. But only if we stop pretending the playing field is level.

---