AI models that predict evolution. AI’s big debt load. Moving backward on defense tech. All that and more in this week’s Prototype. To get it in your inbox, sign up here .

AI is already being used to forecast the weather, stock prices and protein structures. Now Ben Lamm, the co-founder of de-extinction company Colossal , wants to go bigger: building AI models that predict evolutionary changes in whole biological systems.

That’s the goal of Colossal spinoff Astromech, co-founded by Lamm and geneticist George Church, which announced yesterday it had raised a $20 million venture round led by Arch Ventures co-founder Bob Nelsen. The new investment brings the startup’s total funding to $60 million, and values the firm at $3.8 billion.

At the heart of Astromech’s models are the biological datasets that Colossal has collected during its research into how to bring species back from extinction. That provides useful insights into how species have changed over time. For example, where genes are similar between modern elephants and woolly mammoths–and where they aren’t. By working backwards, their models can predict how those genes will change in the future.

That’s not just an arcane question of interest to paleontologists. Understanding how and why genes change as they evolve can help farmers breed better crops, enable pharma companies to find better drugs and assist public health officials in identifying disease threats before they emerge, Lamm told me.

Initially, Astromech is focused on understanding longevity, using its model to map 46 genes associated with longer lifespans, cancer resistance and cell maintenance. The new capital, Lamm says, will be primarily focused on hiring research teams. Because its models are highly targeted for these problems, compute is not as much of an expense as it is for other AI companies, he added.

While Astromech certainly isn’t the only company building datasets to model biological changes, Lamm told me that the data Colossal has collected gives it an edge. “I don’t think anyone has the geographic distance and the time distance that we have,” he said.

The booming buildout of infrastructure for AI needs so much more money–around $2 trillion–that the bond market can’t cover it, Torsten Slok, the chief economist at Apollo, wrote last week .

And that’s what tech companies are expected to still need. According to a recent Wall Street Journal analysis, there’s already some $3 trillion in debt commitments related to AI that isn’t on the books of top tech companies. That’s on top of the around $600 billion in capital expenditures they did report.

The bet here is that the investment will pay off when all this hardware is used to meet future AI demand, but as I’ve discussed before , there are big questions about whether that will happen. It does look like AI revenue is rising , but not nearly fast enough yet to recoup expenditures.

Compounding the problem is that while AI tools can definitely be useful, companies are still struggling to turn that usefulness into durable productivity gains or profits–even when AI adoption boosts revenue . If that continues to be the case, just making useful tools isn’t enough. AI firms need to find ways to turn their tools into a real competitive advantage.

The Hot Take: We Should Be Talking About CPUs

Each week, I ask investors for their take on tech trends within their industries. Today the answers come from Nan Li , founder of Dimension Capital , which invests in frontier science and healthcare startups. I profiled the firm a few years ago, and last month it announced it has raised $800 million for its third fund.

What tech is being overhyped right now?

Consumer Robotics . Foundation models for robotics have started to show very exciting generalization and scaling laws, so it’s easy to extrapolate this progress straight to the scenario of a robot in every home. However consumer environments are the ultimate gauntlet for robotics with regards to unbounded task complexity, near-infinite environmental diversity, and a very high safety threshold. The rollout of robotics will likely go through industrial environments for the foreseeable future. Deploying robots even further across the industrial world represents a massive opportunity that can drive early commercial scale for the model developers and offers a compelling means to subsidize real-world data capture for further training runs.

What should more people be talking about today?

CPUs . For years the AI economy has scaled on the back of GPUs. No matter how the competition played out between the frontier labs, open models, application-layer wrappers, and post-trained private models, all roads still led back to matrix multiplication workflows run on specialized silicon. In the aftermath of this dominant run by Nvidia and many custom accelerator companies, CPUs are re-entering the zeitgeist. As data centers grow more complex and workloads shift from monolithic training runs toward long-horizon agentic inference, CPUs are increasingly called up to handle orchestration, tool calls, branching logic, and I/O. The rise in intelligence continues to scale through compute, but CPUs represent a growing part of the equation for the hyperscalers and neoclouds.

What are we all going to be talking about in five years?

Agentic Science . The overlapping waves of technological development in agentic reasoning, foundation models in biology/chemistry, lab automation, and experimental data generation will massively accelerate the pace of core scientific research across medicine, materials science, and fundamental physics. There are incredibly exciting developments across a variety of scientific capabilities and we are underestimating the compounding effects of these technologies to drive research abundance in the near future.

mRNA for cancer: On Wednesday, Moderna shared clinical trial results of a cancer vaccine it is developing with Merck for melanoma, which showed the medicine extended patient lifespans and prevented the skin cancer from both recurring and spreading. The stock has more than doubled since it announced the news.

Moving military tech backwards: The Trump administration is turning back the clock on some advanced defense technology . Last week, it ordered the Navy to abandon electromagnetic catapults on aircraft carriers for launching warships, a move that will cost billions of dollars and require hundreds more sailors to crew those vessels. And yesterday, the Army announced it’s shutting down a unit dedicated to learning drone warfare in Europe, even as the conflict with Iran–which has a large drone fleet–continues on.

Orbital data center race: Florida-based Starcloud, which is developing orbital data centers, raised $250 million this week , doubling the company’s valuation to $2.3 billion.

Big tech socialism: Libertarian think-tank Cato Institute notes that the federal government has now taken stakes in 30 companies related to emerging technologies, such as rare earth minerals, rockets, nuclear power and quantum computing. While the government has been no stranger to bailouts or subsidies of big companies, actually claiming partial ownership has been rare up until now, and raises questions about whether there can be a level playing field if the government suddenly has interest in protecting its own investments.

What’s Entertaining Me This Week

So, despite Sydney’s recommendation , I’ve yet to sit down and watch Masters of the Universe . It’s on my radar, though! However, I have been listening a lot to the score of the movie , by Daniel Pemberton (who also did the score for this year’s Project Hail Mary ). Pemberton composed it in collaboration with Queen guitarist Brian May, who plays on several of the tracks. It’s an amazing piece of music, reminding me of the great 1980s rock-infused scores like Flash Gordon (which Queen composed). I’ve been putting it on in the background a lot as I binge my way through the Dungeon Crawler Carl books (I’m on book seven, This Inevitable Ruin , now) and it makes a perfect pairing.