The Download: open model trade-offs and geothermal power’s promise

The conversation around open-source AI is hitting a wall. For years, we have lived in this idealistic bubble where 'open' equals 'good.' As a builder, I have always leaned that way. You want to see the weights, you want to run it on your own hardware, and you definitely do not want to be beholden to a single API provider who can rug-pull your access at any moment. But the latest friction between Western labs like Anthropic and rising Chinese powerhouses like Z.ai is forcing a reality check.

The Security Paradox

Anthropic recently dropped a report that feels less like a technical whitepaper and more like a geopolitical warning shot. They are looking at the new model from Z.ai and essentially flagging it as a potential weapon for cyberwarfare. Their argument is simple: when you release powerful weights into the wild, you lose the ability to put guardrails on how that intelligence is applied. If a model is capable of identifying zero-day vulnerabilities or drafting sophisticated phishing campaigns, making it open-source is equivalent to handing out digital skeleton keys.

For those of us building in the trenches, this creates a massive conflict of interest. We need open models to keep the industry competitive and to avoid a monopoly by the big three cloud providers. However, if open models become the primary vehicle for state-sponsored cyberattacks, the regulatory hammer is going to fall hard on everyone—not just the bad actors.

The Founder's Dilemma

From a founder’s perspective, the 'safety' argument often feels like a moat-building exercise. When OpenAI or Anthropic talk about the dangers of open source, it is easy to assume they are just trying to protect their market share. But we have to look at the nuance here. There is a difference between a model that can write a decent marketing email and one that can autonomously navigate a complex network to inject malware.

The Z.ai situation is unique because it bridges the gap between high-level capability and total lack of oversight. Because the model originates in China, Western regulators have zero leverage over how it was trained or what safety fine-tuning was applied. Once those weights are on Hugging Face, the genie is out of the bottle. For builders, this means we are entering an era where the tools we use might come with ethical baggage we never asked for.

Why We Still Need Open Weights

Despite the warnings from Anthropic, we cannot abandon the open-source movement. If we move to a world where all high-level intelligence is behind a closed API, innovation will stall. We have seen what happens when platforms become too powerful; they start censoring output, increasing prices, and killing off third-party apps that get too successful. We need open models to act as a check on that power.

• Ownership: Running your own models means you own your stack.
• Privacy: You aren't feeding your proprietary data back into a competitor's training loop.
• Customization: Fine-tuning for specific industry use cases is significantly easier when you have full access to the architecture.

The Geothermal Connection

While the AI world is fighting over weights and safety, there is a physical reality we often ignore: power. All this talk of 'open vs. closed' doesn't matter if we run out of electricity to train these monsters. This is why the renewed focus on geothermal power is actually a crypto and AI story. We are seeing a shift toward energy sources that are consistent, unlike wind and solar which fluctuate.

For those building high-compute infrastructure, geothermal represents the 'base load' we have been looking for. It is the physical version of open source—reliable, decentralized in its potential, and harder for a single entity to monopolize once the infrastructure is in place. If we can solve the energy problem, the pressure to restrict AI development might ease, as the cost of 'safe' compute drops.

What This Means for You

If you are building a startup right now, you need to be platform-agnostic. Relying solely on a closed model puts you at risk of regulatory shifts. If a government decides that 'Model X' is too dangerous, and your entire business is built on that API, you are finished. At the same time, if you lean too heavily into unverified open models from adversarial regions, you might find yourself in a compliance nightmare when trying to sell to enterprise clients.

The goal is not to choose a side in the open vs. closed debate, but to build a system that can survive regardless of who wins.

A Skeptical Takeaway

Don't buy into the hype that open source is a death sentence for humanity, but don't ignore the risks either. Anthropic's report is a reminder that we are playing with tools that have real-world consequences. The 'move fast and break things' era is evolving into the 'move fast and secure things' era. The winners in the next three years won't just be the ones with the smartest agents; they will be the ones who can prove their agents aren't being used to dismantle the digital infrastructure we all rely on.

Keep your eyes on the energy sector and your models diverse. The friction between global powers and AI labs is only going to get louder. As founders, our job is to navigate the noise and keep shipping products that actually solve problems, rather than just adding to the chaos.

Read the original at MIT Technology Review →