Earth Sciences Update: The AI-Geology Mashup, Weird Funding, and More

AI Meets Geology: It's About Time

You've heard it a thousand times: AI is transforming everything. But in Earth sciences, it's not hype. This week, I saw stories that made me think, 'Okay, this is real.' AI is now helping scientists spot minerals, track deforestation, and even predict volcanic eruptions—stuff that was pure sci-fi a decade ago.

Take Nvidia's move. They quietly bought a chunk of SpaceX—about $21 billion worth. Why should a geologist care? Because SpaceX's satellites are the eyes in the sky for Earth monitoring. That investment isn't just about space tourism; it's about getting better data on what's happening to our planet. And that data is gold for anyone trying to model climate or find new deposits.

Token Loans: Banks Finally Get It

Here's a head-scratcher: the Bank of China just gave out 'Token Loans' to AI companies. Wait, tokens? Turns out, they're lending money based on how much computing power a company uses—not on property or machinery. They've already handed out 8 million yuan, with another 20 million on the way.

For Earth science startups, this is a lifeline. Most of them don't own fancy offices; they own servers and datasets. Traditional banks would laugh them out the door. But now, if you can show your models are being used, you can get funding. It's a shift from valuing what you have to valuing what you do. That's a big deal for a field that's data-heavy but asset-light.

Robots in the Field: Not Just for Car Factories

Unitree Robotics just went public in Shanghai, and people went nuts—the lottery rate was 0.018%, with 9.78 million investors fighting for shares. Why does a robot company matter for Earth sciences? Because these robots are the ones crawling into volcanoes and diving into deep-sea trenches to collect samples. They're doing the dirty work that humans shouldn't do.

One winner said they'd keep it quiet because they didn't want envy. Fine, but think about it: that robot you're deploying on a volcano's edge? It's connected to this boom. The hype is real, and it's funding the tech that makes fieldwork safer and data more precise.

Data Security: The Unseen Backbone

ByteDance, the TikTok parent, just set up a new AI data security department. Sounds boring, right? But for Earth sciences, it's crucial. Satellite images, climate records, geological surveys—these are sensitive. As we digitize everything and run AI on it, we need to keep it safe from bad actors. The new unit, led by Wang Yinglei, will consolidate data teams to provide cross-modal services. For researchers, that might mean better access to clean, secure datasets—if they can navigate the corporate labyrinth.

The 90-Hour Week: A Cautionary Tale

Let's talk about the grind. Insiders at OpenAI and Anthropic say 90-hour weeks are normal during 'sprints.' One ex-OpenAI employee told me he regularly hit 70 hours, and it was worse at launch. That intensity is bleeding into Earth science startups. I've seen climate modeling companies where the founders are running on fumes because investors want results yesterday.

A UC Berkeley study found that AI tools speed up tasks, but managers just pile on more work. So you're not going home earlier; you're doing more. If you're thinking about entering this field, know what you're getting into. It's exciting, but it's not a 9-to-5.

Anthropic's Big Numbers—and a Reality Check

Anthropic, the Claude people, are projecting $190-200 billion in revenue by 2028. That's a lot of zeros. And it's driving talk of a $2 trillion IPO. But here's the Earth science angle: Anthropic is all about AI safety, which matters when you're using AI to predict floods or assess climate risk. If the model goes wobbly, you get false alarms or missed warnings.

They just bought Decart for $6 billion and released a risk report that talks about 'hallucinations' and dirty data. For scientists, that's a reminder: AI is only as good as what you feed it. Garbage in, garbage out. So before you trust that climate projection, check the training data.

DeepMind's Pivot: Faster, Cheaper, but Not Always Better

Google DeepMind is shifting away from chasing the biggest models to focus on 'Flash' models—smaller, faster, cheaper. That could mean a third of the staff gets reassigned. For Earth sciences, this is a mixed bag. Flash models can run climate simulations quicker, making them accessible to more researchers. But they might not handle the high-resolution, complex projections we need for local predictions.

On a lighter note, they also released SL2T, a sign language-to-text model. It's now in Pixel 11 phones. It might seem unrelated, but it could make Earth science education more inclusive. That's a win, even if it's off to the side.

DeepSeek's Rollercoaster: Price Drops and Open Source

DeepSeek had a wild week. They put out a pro version of V4, then pulled it within 24 hours. Meanwhile, they're introducing peak/off-peak pricing for their API, with off-peak at half price. That's a boon for researchers running big simulations on a budget—you can schedule your jobs for off-hours and save big.

And they open-sourced their Harness framework under MIT. It's a plugin-everything approach that lets you customize AI agents for your specific task—like analyzing satellite imagery or predicting soil erosion. This is exactly what scientists need: tools they can adapt, not black boxes.

Zhipu's Growth Spurt

Zhipu AI, a Chinese startup, is on fire. Their ARR is expected to hit $2 billion by year-end, and they've added 2 million API users since July, hitting 7 million total. Their ZCode tool, similar to OpenAI's Codex, got 1 million users in a month. For Earth sciences, Zhipu's focus on domestic chips is interesting—they built a 1GW data center using only Chinese chips. That could reduce dependency on foreign tech and make advanced computing more accessible for environmental research in China.

Wrapping Up: The AI-Earth Sciences Tango

Look, Earth sciences and AI are now joined at the hip. This week showed that in financing, in tools, in the way we work. Whether it's banks valuing data over bricks, or open-source frameworks that let you tweak an AI agent for your specific problem, the field is moving fast.

For researchers and enthusiasts, the key is to stay flexible and keep learning. The tools are changing, the funding models are shifting, and the possibilities are expanding. It's an exciting time to be in this space—just be ready for the long hours and the constant evolution.