The latest edition of The Download lands at a moment when the AI industry is confronting one of its most consequential tensions: how to preserve the openness that has helped fuel rapid progress while limiting the risks that come with making powerful models widely available. Open models have become central to the frontier AI ecosystem because they lower barriers to experimentation, enable independent research, and allow companies and developers to build on top of shared infrastructure. But the same accessibility that makes them valuable also makes them harder to contain.
Anthropic's recent report, which argues that a powerful new model released by the Chinese company Z.ai could pose a major cyberattack risk, has sharpened that debate. The concern is not simply that advanced models can be used for harmful purposes in theory. Rather, the issue is that increasingly capable systems may lower the technical threshold for cyber abuse by helping users automate reconnaissance, generate malicious code, or scale phishing and intrusion workflows. That possibility has become a central policy and industry question as model performance improves faster than many safeguards.
Security Versus Openness
Open models are often praised for democratizing access to frontier AI. They can accelerate scientific discovery, support startups that cannot afford closed proprietary systems, and allow researchers to audit model behavior more directly. In a field where a handful of large companies dominate the most advanced closed systems, open releases are also seen as a counterweight to concentration of power. Yet the same qualities that make open models attractive can also make them difficult to govern once they are released into the wild.
The core challenge is that openness is not a binary virtue. A model can be open enough to support innovation while still being powerful enough to be repurposed for harmful use. Once weights, code, or deployment methods are broadly available, the ability to enforce usage restrictions diminishes sharply. That is why the current debate is less about whether open models should exist and more about what level of capability should trigger additional safeguards, staged release policies, or tighter evaluation before publication.
Anthropic's warning about Z.ai arrives amid a broader shift in the industry toward more serious model risk assessments. Frontier AI developers are increasingly treating cyber capability as a first-order safety issue, alongside biosecurity, misinformation, and autonomous agent behavior. The concern is especially acute because cyber tools can be deployed at scale, tested iteratively, and adapted quickly. A model that meaningfully improves offensive cyber productivity could have effects far beyond its original release context.
Geothermal's Quiet Promise
The same newsletter edition also points to geothermal power, a reminder that the AI boom is not only a software story but an infrastructure story. As data centers expand and model training grows more energy-intensive, the search for reliable, low-carbon power sources has become strategically important. Geothermal energy stands out because it can provide steady baseload electricity, unlike intermittent wind or solar, and can potentially support the round-the-clock demands of compute-heavy AI systems.
That matters because the economics of frontier AI increasingly depend on energy availability, grid stability, and long-term operating costs. If AI infrastructure is to scale sustainably, the sector will need more than chips and model improvements. It will need power sources that can support dense clusters of servers without forcing operators into a trade-off between growth and emissions. Geothermal is not a universal solution, but it is gaining attention precisely because it offers a rare combination of reliability and cleaner generation.
For technology companies, the convergence of AI and energy is becoming impossible to ignore. The same firms racing to build larger models are also looking for ways to secure their own power supply, reduce exposure to volatile electricity markets, and meet climate commitments. In that sense, geothermal is part of the broader industrial base of AI, just as much as semiconductors, cloud infrastructure, and networking equipment.
The Policy Pressure Builds
The emerging picture is one of dual pressure on the AI sector. On one side, there is the push to keep models open enough to sustain innovation and competition. On the other, there is mounting evidence that frontier capabilities can be misused in ways that are difficult to reverse once released. Policymakers, researchers, and companies are now being forced to ask whether existing evaluation frameworks are sufficient for models that may be useful to defenders and attackers alike.
That question is likely to shape the next phase of AI governance. If open models continue to improve, the industry may need clearer norms around pre-release testing, red-teaming, and disclosure of dangerous capabilities. At the same time, the energy demands of AI will keep pushing attention toward the physical systems that make the digital revolution possible. The result is a more mature conversation about frontier AI: one that is no longer focused only on model quality, but also on risk, resilience, and the infrastructure required to sustain the field's ambitions.