# The Pentagon’s CTO Explains the Anthropic Rift and Superintelligence Task Force

Markdown copy of https://realalpha.app/v/1hXxzkU77iI/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

Emil Michael argues the Pentagon cannot use models that override military command, even as it pursues common AI safety rules.

- Show: The Information's TITV
- Hosts: [Akash Pasricha](https://realalpha.app/people/akash-pasricha/) & Amir Efrati & Leo Schwartz
- Guest: [Meerah Rajavel](https://realalpha.app/people/meerah-rajavel/) · Senior vice president & chief information officer, Palo Alto Networks
- Guest: [Erik Bernhardsson](https://realalpha.app/builders/erik-bernhardsson/) · Co-founder & CEO, Modal
- Guest: [Nick Del Deo](https://realalpha.app/people/nick-del-deo/)
- Guest: [Emil Michael](https://realalpha.app/people/emil-michael/)
- Published: Oct 16, 2026
- Length: 1 h 31 m

## The best minutes

- [29:53](https://realalpha.app/v/1hXxzkU77iI/?t=1793) [CoreWeave’s debt is layered around customer contracts](https://realalpha.app/s/coreweaves-debt-is-layered-around-customer-contracts/) (6 min)
- [58:29](https://realalpha.app/v/1hXxzkU77iI/?t=3509) [Michael objects to giving AI models authority over commanders](https://realalpha.app/s/michael-objects-to-giving-ai-models-authority-over/) (2 min)
- [1:09:59](https://realalpha.app/v/1hXxzkU77iI/?t=4199) [Michael favors American open-source AI but flags China risks](https://realalpha.app/s/michael-favors-american-open-source-ai-but-flags-china-risks/) (4 min)

## Also worth watching

- [11:29](https://realalpha.app/v/1hXxzkU77iI/?t=689) [Palo Alto uses a harness to measure AI-assisted development](https://realalpha.app/s/palo-alto-uses-a-harness-to-measure-ai-assisted-development/) (4 min)
- [41:10](https://realalpha.app/v/1hXxzkU77iI/?t=2470) [Power access favors experienced data center operators](https://realalpha.app/s/power-access-favors-experienced-data-center-operators/) (3 min)
- [18:33](https://realalpha.app/v/1hXxzkU77iI/?t=1113) [Modal sees software as its advantage over neoclouds](https://realalpha.app/s/modal-sees-software-as-its-advantage-over-neoclouds/) (3 min)
- [47:58](https://realalpha.app/v/1hXxzkU77iI/?t=2878) [Pentagon AI use grew from 80,000 to 1.7 million](https://realalpha.app/s/pentagon-ai-use-grew-from-80-000-to-1-7-million/) (3 min)
- [1:20:33](https://realalpha.app/v/1hXxzkU77iI/?t=4833) [Pentagon financing targets supply-chain gaps](https://realalpha.app/s/pentagon-financing-targets-supply-chain-gaps/) (3 min)

Skip the opening agent-security product walkthrough, Modal's funding and revenue recap, and the longer Pentagon discussion of legal frameworks and government-company oversight.

## Full analysis: Emil Michael says Pentagon AI use reached about 1.7 million

- About 1.7 million Department of War employees now use AI, up from 80,000 a year earlier.
- Michael separately described model testing and a task force seeking common incident-response practices.
- Leaders need outcome measures alongside adoption counts, and clear limits on AI authority.

Its 6 sections, what it means for you and what to watch are on [the page](https://realalpha.app/v/1hXxzkU77iI/#analysis), free with an account:

- [50:30](https://realalpha.app/v/1hXxzkU77iI/?t=3030) The Pentagon expanded AI use, but gave no outcome baseline
- [46:27](https://realalpha.app/v/1hXxzkU77iI/?t=2787) The task force is coordinating shared protections
- [59:13](https://realalpha.app/v/1hXxzkU77iI/?t=3553) Michael rejects Anthropic over model authority
- [1:10:34](https://realalpha.app/v/1hXxzkU77iI/?t=4234) Michael favors American open source, with caveats
- [1:21:00](https://realalpha.app/v/1hXxzkU77iI/?t=4860) Defense financing targets supply-chain gaps
- [1:29:09](https://realalpha.app/v/1hXxzkU77iI/?t=5349) Michael declined to assess a reported strike investigation

## Summary

Meerah Rajavel is CIO of Palo Alto Networks; she says the company was the first customer to implement agent-security software before letting engineers use coding tools. Erik Bernhardsson is Modal's cofounder and CEO, and described its $750 million round, $15.75 billion valuation and $1 billion annualized revenue run rate; Nick Del Deo is a managing director at MoffettNathanson. Emil Michael is the Pentagon's CTO and undersecretary of defense for research and engineering.

Meerah Rajavel says AI agents can expose sensitive systems and that companies need agent-level visibility, continuous scanning and human review; she reports that new product introductions at Palo Alto Networks rose 3x in six months without added software-engineering headcount. Erik Bernhardsson says Modal's $1 billion annualized revenue run rate reflects demand for AI applications, and argues software layers differentiate infrastructure providers from compute-heavy neoclouds; he sees human leverage, not full automation, as a source of value. Nick Del Deo says CoreWeave's project economics are attractive but its dependence on external capital and access to power creates significant risks. Emil Michael says the Pentagon wants common incident-response and infrastructure protections, but argues Anthropic's model principles conflict with military command; he also supports American open-source AI and defense-sector financing to widen the supplier base.

### Minute by minute

- [2:12](https://realalpha.app/v/1hXxzkU77iI/?t=132) **Agent security needs visibility and guardrails.** Rajavel says agents can inherit users' permissions across laptops, code bases and financial systems, creating risks such as prompt injection, data poisoning and supply-chain attacks. Palo Alto Networks' XDR1 is intended to show what agents do and provide guardrails and kill switches. She says both vulnerabilities found in existing code and flaws introduced by AI-generated code are risks, so the company uses continuous multi-model scanning and virtual patching.
- [9:10](https://realalpha.app/v/1hXxzkU77iI/?t=550) **AI usage should be measured by outcomes.** Rajavel says Palo Alto Networks tracks AI use and spend without setting employee quotas, and routes work through a harness to assess code quality, security, release velocity and efficiency. She reports that new product introductions rose 3x in six months without adding software-engineering headcount. She says she does not yet trust fully autonomous agents and requires human review and user testing before release.
- [15:08](https://realalpha.app/v/1hXxzkU77iI/?t=908) **Modal layers software over compute.** Bernhardsson says Modal's annualized revenue run rate reached $1 billion, up from $300 million in May, as AI applications drive demand. Modal annualizes its last 30 days of revenue and rents capacity from 25 cloud vendors. He argues its software for inference, training and sandboxes differentiates it from providers that mainly rent compute, and says its margins are healthy and improving.
- [23:41](https://realalpha.app/v/1hXxzkU77iI/?t=1421) **Human leverage can matter more than automation.** Bernhardsson says full automation is not the only measure of AI's value: he expects agents to help people produce more and enable new products. He cites his own coding work and suggests companies can track output through measures such as code written or pull requests merged, while acknowledging that productivity is difficult to measure.
- [26:25](https://realalpha.app/v/1hXxzkU77iI/?t=1585) **CoreWeave's expansion depends on outside capital.** Del Deo says CoreWeave's project economics look attractive, estimating 15 or 20 cents in net value for each dollar invested in compute. But he argues its capital needs exceed its internal resources, leaving it reliant on external financing to meet deployment commitments. He estimates about $22 billion in available liquidity, likely enough to carry the company through year-end.
- [35:02](https://realalpha.app/v/1hXxzkU77iI/?t=2102) **Del Deo sees capital access as the key risk.** Del Deo says CoreWeave's parent-company debt has risen from 9% or 10% several months earlier to about 15%, while secured loans are closer to the underlying GPUs and customer contracts. He sees access to capital markets as a bigger risk than near-term refinancing. He says CoreWeave could lean more on equity or supplier support, but an NVIDIA backstop could weaken perceptions of its independence.
- [41:00](https://realalpha.app/v/1hXxzkU77iI/?t=2460) **Power limits new data-center capacity.** Del Deo says power access is the biggest constraint across the data-center sector, especially for less experienced operators without long-standing relationships with utilities and municipalities. He says limited new capacity could raise prices for facilities already online but reduce growth opportunities. He expects that trade-off to be a net negative for fast-growing neoclouds such as CoreWeave and Nebius.
- [44:02](https://realalpha.app/v/1hXxzkU77iI/?t=2642) **The task force will coordinate AI safeguards.** Michael says the 120-day task force will define government ties with superintelligence companies, develop incident-response procedures and address critical infrastructure such as hospitals, wastewater systems, electricity grids and banks. He says the Pentagon rates models and expanded AI use from 80,000 of its 3.5 million employees a year earlier to about 1.7 million, while testing models and hardening systems.
- [51:15](https://realalpha.app/v/1hXxzkU77iI/?t=3075) **Michael says existing laws offer enforcement tools.** Michael says warnings from AI researchers can amplify public alarm, while acknowledging that some risks require safeguards. He argues existing laws covering computer misuse and fraud, along with export controls, give the executive branch tools to act. He cites the administration's rapid export-control action against Anthropic after it refused to pull back a model it had said it could withdraw.
- [56:33](https://realalpha.app/v/1hXxzkU77iI/?t=3393) **Michael says Anthropic's principles conflict with command.** Michael says Anthropic is no longer used by the Department of War, but expects to keep working with the company through the task force. He argues that Anthropic's model principles give a model the ability to override human direction, which conflicts with military command and control. He says the Pentagon will not work with Anthropic unless it changes that principle.
- [1:03:25](https://realalpha.app/v/1hXxzkU77iI/?t=3805) **AI governance will use common frameworks.** Michael says the task force's 120-day work includes an incident-response framework, critical-infrastructure hardening and a review of existing laws; a separate advisory board will study whether legislation is needed. He questions the idea that government and AI companies will merge, saying firms should invest more in safety and serve national-security needs. He says government equity stakes are not unprecedented but leaves any decision on frontier labs to the president.
- [1:09:59](https://realalpha.app/v/1hXxzkU77iI/?t=4199) **American open source offers auditability.** Michael favors American open-source AI for wider access, auditability through visible model weights and lower consumer prices. He worries that Chinese open-source models could become dominant in the U.S. because those companies are subject to the Chinese government. He prefers American alternatives, while noting that they are not yet at the frontier.
- [1:13:55](https://realalpha.app/v/1hXxzkU77iI/?t=4435) **Michael sees AI advantage as deterrence.** Michael says a superintelligence advantage could improve intelligence gathering and weapons development, deter conflict and reduce troop exposure; he fears an adversary with that advantage could coerce the U.S. He also says foreign-national access creates risks at AI labs and universities, citing a reported spying-device case at OpenAI and the arrests of 13 Chinese nationals at a Michigan university. He says the FBI can vet prospective employees when companies request it.
- [1:20:28](https://realalpha.app/v/1hXxzkU77iI/?t=4828) **Defense financing targets critical supply chains.** Michael says defense financing will focus on microelectronics, drone components, batteries, critical minerals and space, with most deals structured as loans; he says warrants can give taxpayers upside. He wants new entrants to complement established contractors, arguing that newer firms may build drones and autonomous systems faster and at lower cost. He says they would supplement, not replace, existing systems such as submarines, F-35s and Patriot.

## Transcript

The Information's TITV, hosted by Akash Pasricha, Amir Efrati and Leo Schwartz, with Meerah Rajavel, Erik Bernhardsson, Nick Del Deo and Emil Michael.

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

[1:13](https://realalpha.app/v/1hXxzkU77iI/?t=73) welcome everyone to the information's ti tv my name is akash pasricha it is monday october 19th today on the show we're going to speak with palo alto network cio about how it is helping customers deal with the risk of rogue agents. We'll then talk with the CEO of Modal, the AI infrastructure company recently that tripled its valuation to more than $15 billion in less than six months. We'll also get a status update on the NeoCloud route and how CoreWeave is holding up. And to close out the show, we have an extended conversation with the Pentagon's chief technology officer about their work with Anthropic. We're also going to hear from him about the new super intelligence task force and dealing with foreign AI risks. It's going to be a great show, so let's get right on into it. Palo Alto Networks is expanding its cybersecurity platform to protect businesses from AI agents from going rogue, accessing sensitive data. Joining us now is Palo Alto Networks CIO, Meera Rajival. Meera, welcome to the show. It's great to have you here.

[2:17](https://realalpha.app/v/1hXxzkU77iI/?t=137) Thank you, Akash. Thank you for having us on the show. So tell us about this product that you're releasing this morning. So, Akash, we are releasing a product called XDR1. XDR is endpoint security. And if you think about traditional endpoint security, they focus on network cloud and the local infrastructure of what is getting executed. But today, the biggest thing is AI and the AI agents, what it is doing. So you go to any company, it's actually a great productivity. It's a great, powerful tool when you give it to the employees for the reason of software engineering or for the reason of, you know, individual productivity, like having something that's their personal twin.

[2:59](https://realalpha.app/v/1hXxzkU77iI/?t=179) However, all these AI agents pose a serious new risk. And when you give an agent an autonomy or agency for what I'm doing with my permission, then it has actually access to things that I have access to, not just on my laptop, but also it's beyond our endpoint, but it is beyond the systems I'm having, simple things from the code bases to financial ledgers. When you're dealing with it, there are multiple different things that agents can get into vulnerability issues like data poisoning, supply chain attacks, and prompt injection. You just feed these agents a malicious kill that can put you in a place that's deadly for the company.

[3:45](https://realalpha.app/v/1hXxzkU77iI/?t=225) The second thing is, as a CIO, I care about the full visibility, but most of us, the AI agents are today a blind spot because the traditional endpoint security gives you visibility of what's coming and it doesn't say what exactly is happening. The analogy I've used it, HDR is like you have an X-ray scanner, it scans through. But if you wanted to know every single thing in that box, how it is getting used, you don't have that visibility. That's where the agent endpoint security comes into picture. We spotted this risk because we are a company of, you know, 14,000 engineers. And as a CIO, my job is to really enable innovation, not hold them back.

[4:28](https://realalpha.app/v/1hXxzkU77iI/?t=268) When we started seeing the power of AI in software engineering, we spotted this as a significant risk, and this is also a supply chain risk for our customers we didn't want to take. So we are scanning the market to see what's going on. And when coil security was even in the stealth mode, we spotted them. In fact, we literally brought them in and we became the first customer of implementing them to say, okay, we really need to get this enabled before we let our, you know, engineers touch the tool to. go and do the software engineering. Right. So today, yeah, go ahead. Let me jump in here because I think you have an interesting perspective here on this being the CIO of a company that effectively is advising other CIOs on their own stack here.

[5:17](https://realalpha.app/v/1hXxzkU77iI/?t=317) You know, the issue around agents and security is, it's interesting because on one hand, we have these coding agents that help us write code. And, you know, I saw that Bank of America came out and said that they're hiring a thousand contractors now to help them deal with the vulnerabilities that Mythos found in their own code. And, you know, this is additional headcount that they have to bring on. And the question I have for you is, is the problem here that the coding agents and the code that CloudCode and Codex create is a problem that they are creating vulnerabilities in companies' own stacks and in their architecture because maybe it's just, it's a little bit quick and dirty and it's not the most secure code?

[6:07](https://realalpha.app/v/1hXxzkU77iI/?t=367) Or is the problem more that the code is, it can be attacked from agents that can be used by malicious actors and stuff like that? Which of these two risks is more prevalent right now, from your view? I think, Akash, both are a risk, right? Because when you talk about what Mico's found, take a company like Bank of America or any company that is in existence for a while, right? I mean, they have a lot of software. Today, especially with the digitization, every company, doesn't matter which industry you're talking about, even highly, highly, the legacy or the brick and mortar companies are digital.

[6:46](https://realalpha.app/v/1hXxzkU77iI/?t=406) So they have software code. And when we talk about vulnerabilities, Mythos and Fabel, all are, you know, for that matter, GPT, 5.6, everything is finding vulnerability in the code that they're using to run their business today. And it is actually a must that they have to address those vulnerabilities because that's an existential crisis, right? Because you don't want that to be exposed to the attackers who are leveraging AI to compromise into their environment. That's one side. Then you talk about, you know, AI producing code. Remember, AI is also producing code by learning from what's out there.

[7:27](https://realalpha.app/v/1hXxzkU77iI/?t=447) And if you remember what Mythos found, 25 years ago, the Berkeley's Unix kernel, it found, you know, the open source vulnerability that we are all using for so long. So now when you think about the software code is generating, there is a possibility that it may be introducing vulnerability as well. I'm not going to say every time, but it will. So it's really important you start looking at software supply chain, right? So there are two things, right? I mean, like two weeks back, we launched something called Frontier AI Security, which is actually a service we launched because we saw our customers saying, hey, mythos was a moment, but not it's any more mythos because we are finding it.

[8:06](https://realalpha.app/v/1hXxzkU77iI/?t=486) Forget about Frontier Labs. Even open source, open rate models are as good as the Frontier Labs. And it's only like earlier, it used to be four to five months before they catch up. Now it's four to six weeks. So we launched a service, which is a multi-model harness that can actually continuously scan whether they are producing a new software or they're introducing a new, you know, on their brownfield. We want to make sure the posture of our customers are continuously managed and monitored. Again, also, we introduced something called virtual patching, which is like it's impossible for a company like Bank of America, which has been in existence for so long, to really patch everything within a matter of time.

[8:48](https://realalpha.app/v/1hXxzkU77iI/?t=528) You need actually something to stop in the fringes, and then you need to really work through this. Both are required, and also, this goes back to basic zero-trust security hygiene. keep talking to, I mean, when I keep talk to CIOs and CISOs, one of the things I emphasize is more than ever, security hygiene is becoming super important today. We have to focus on hygiene. Let me ask you a couple more questions about how you are approaching your own stack internally. So of all the models that you're using, is Palo Alto Networks, are you using open white models from China at all? We actually use multi-model, but we are not using open-way Chinese models at this point. Okay. And why not? I mean, it is a decision we made at this point in time where the things are. We just wanted to be in control of the models that we are using. And that was a more strategic decision. And how are you controlling your own spend on these models internally? Are you giving employees specific caps on their usage? How are you dealing with that?

[10:05](https://realalpha.app/v/1hXxzkU77iI/?t=605) Yeah, I think, I mean, one thing is, this is nothing different than when cloud came, but it's just that it is so accelerated and much more expensive than cloud, right? I mean, shadow AI is absolutely a concern today. So we have actually a full visibility of the use of AI in the company today, this is one of our critical things that we introduced. So we have full visibility of what AI uses in the company. And we also have a way to continuously monitor our spend. We haven't gone down the pack of quota. Rather, we have gone down because we wanted to make sure we are rewarding outcomes and efficiency rather than just putting an artificial quota. And we also want to make sure people who are doing really good, efficient outcomes, that has to be replicated. So at this point in time, we watch and we stop, you know, behaviors that are not the right behaviors. And we are actually taking the ones where the outcomes are higher or efficiency of use of AI is higher. And we are taking that and replicating as part of our harnesses. So absolutely cost is something we focus on, but not with the artificial, you know, boundaries saying that we are going to give you a quota. At some point in time, we may consider that, but at this

[11:15](https://realalpha.app/v/1hXxzkU77iI/?t=675) point in time, that's not the baby. Well, yeah, because, you know, I'm sort of curious on how you prevent internally this idea of what Toby Lucky at Shopify has called the slop grenade and people just, employees sort of launching slop at each other and it actually not helping productivity. This is something we've talked about, productivity being one of the most difficult things to measure. How do you prevent that? Do you have to, do employees have to say that, you know, I used, This is AI generated, this is non-AI generated. Like, what's your method? Absolutely. Great question, right? I mean, this has been a ramp.

[11:52](https://realalpha.app/v/1hXxzkU77iI/?t=712) I mean, the last, I would say, three, six months, we have actually put a lot of focus on our SDLC harness. And the harness is the way people can come into it. And that's the way we are watching where the spend is. When somebody is outside, we can see, like, for example, in IT, I have a harness that actually connects to both our first-party applications as well as third-party applications. Anybody using AI has to use through the harness. I also know if they're outside the harness where the spend is and we nudge them back into the harness because when I put them in the harness, I'm not just talking about coding.

[12:28](https://realalpha.app/v/1hXxzkU77iI/?t=748) I'm tracing from PRD all the way to what they're releasing in production, right? Today, I would say in some technologies, it's a little harder to trace it, but I'm able to get the visibility for over 85, 90% of the work that's happening. So I look at it as three things. One, what is the quality of code they're producing? How secure it is? Two, what is the velocity of outcomes that we are generating or release outcomes? Three, how efficiently we produce that work? For a simple example, I mean, the market is heating up, security is heating up. We are actually, you know, doing a lot of M&A and the new product introduction for Palo Alto has gone skyrocketing.

[13:12](https://realalpha.app/v/1hXxzkU77iI/?t=792) I mean, for me as a CIO and my organization is responsible for NPI, every single innovation that comes to the business.

[13:20](https://realalpha.app/v/1hXxzkU77iI/?t=800) **Meerah Rajavel:** And we have gone 3x NPI over the last six months, and I have not increased even a single headcount on my software engineers.

[13:30](https://realalpha.app/v/1hXxzkU77iI/?t=810) But are they good product? Like, this is the question. Are they good product in a sense? Help me understand the sense of, like, is AI producing good code? Yeah, yeah, no, no. I mean, I guess what I'm saying, so product launches are one thing, but have the products meet the goals that you have set. Yeah, yes, yes, yes. Because it is human in the loop, Akash, right? I mean, I don't believe in autonomous agents yet. We are not at the place where, you know, having autonomy fully, especially when you're looking for outcomes and the bigger things. I always have the code is generated. There is always a human review before the release.

[14:09](https://realalpha.app/v/1hXxzkU77iI/?t=849) And also there is actually a user testing before the release. This also brings us to what we are launching today with XDR1 because the XDR1 will give me the endpoint security guardrail. So I can be a lot more taking higher risk on because I'll have a full visibility of how the AI, I mean, we already rolled it inside, by the way, but we'll have a full visibility of how the agent is behaving and how I can have the appropriate guardrails and kill switches to protect it if it's going wrong. And it's also connected to our, you know, SOC XIM solution, which has a way to respond in an autonomous fashion. To go back here and question, I think it really comes down to having the high risk or higher value things, putting human in the loop. And the second thing is really looking at it from the point of you have the right security guardrails. And if you need to go down the path of saying no, we have a quick way to respond to it.

[15:04](https://realalpha.app/v/1hXxzkU77iI/?t=904) Great. Well, Mira, I want to thank you for coming on. That is Mira Rajabal, the CIO at Palo Alto Networks here on TITV. Thank you. AI infrastructure startup Modal announced it raised $750 million at a $15.75 billion valuation. That is more than triple its valuation from just five months ago. I want to welcome back to the show Eric Bernhardsson, co-founder and CEO of the company. Eric, welcome back to the show. It's great to have you here. Thank you. It's great to be here. Okay, so big numbers on the board here. $100 billion. Sorry, $1 billion. I'm sorry. That would be great. That would be great.

[15:42](https://realalpha.app/v/1hXxzkU77iI/?t=942) $1 billion in annualized revenue. As of now, up from $300 million in May. Where is that growth coming from, Eric? Yeah, I think we just see so much demand coming from all these insanely cool AI applications getting built out there, whether it's Ramp or Cognition doing agent decoding. could be generative media like runway and suno uh could be you know type safe with jav uh many other things there's just so many cool applications getting built and and i think it's really representing kind of a new wave of software like like these things are things that couldn't have been built just a few years ago i mean i've been coding for so long since since the early 90s and and this is the biggest change in how software works and what you can do and how you build software and and and you know it's just kind of amazing seeing all these changes and and the incredible demand that that that's unlocked and a lot of the funds on modal so uh 1 billion uh in annualized revenue i just want to make sure i understand this so is this the like is this 1 billion in revenue that the customers are paying you and then you have to pay out some proportion of that to the model providers like how are you measuring that what's the what's the gross

[16:54](https://realalpha.app/v/1hXxzkU77iI/?t=1014) versus net on this so so this is our run rate revenue right so you We annualize the last 30 days times 12. And we don't pay it to model providers. We rent a lot of compute capacity from various different clouds. In fact, we work with 25 different cloud vendors. A big part of the value prop of modal is don't think about capacity. We kind of manage all that entire, we aggregate a lot of the capacity and present it to the customers in a way where they don't have to think about it and they can move fast and build a lot of cool things, whether it's, you know, inference or training or sandboxes, a lot of other stuff.

[17:31](https://realalpha.app/v/1hXxzkU77iI/?t=1051) And so, yeah, I mean, you know, we have meaningful cogs, like we have to spend, you know, a fair amount of money on the underlying layer. But we have very healthy margins, actually, in fact, like it turns out like building that like very differentiated software layer on top of it means you can have, you know, very healthy gross margins that are actually growing over time. How, what do your margins look like right now? I don't know if I can comment on, you know, latest numbers. I mean, you're the CEO, actually. You can say whatever you want. I can say whatever I want, but some things I don't want to say.

[18:05](https://realalpha.app/v/1hXxzkU77iI/?t=1085) But, okay, so you're right. The main cost is compute. And, look, you guys sit in an interesting spot here because, you know, I'm looking at the NeoCloud route of last week, and, you know, I'm thinking about how they are looking to move up the stack a bit with their own inference-like services, I think, helping people optimize the models that are being used, stuff like that. So what's your view of that story crumbling a little bit here? What do you think? Because, by the way, I mean, it doesn't seem like you're too far from it in some ways, right? You're renting compute as well. I think the NeoClouds have had a good run over the last few years.

[18:50](https://realalpha.app/v/1hXxzkU77iI/?t=1130) I mean, at the end of the day, they're relatively undifferentiated software undifferentiated businesses.

[18:56](https://realalpha.app/v/1hXxzkU77iI/?t=1136) **Erik Bernhardsson:** I think software is really where you can make a massive differentiation.

[18:59](https://realalpha.app/v/1hXxzkU77iI/?t=1139) We have a very rich software layer. We provide a number of building blocks. Like we really think about modal as like, sure, you can get the capacity. You can go out and rent capacity from NeoClouds, but there's a number of problems. Like most developers, like the way you think about it, it's like you want these higher level building blocks, you know, inference, sandboxes, training, reinforcement learning, agentic applications. In the future, it's going to be way more. And that's at the level where you want to operate. And so enabling developers to move fast and ship these incredible applications in a way where they don't have to think about infrastructure, there's so much alpha to capture in that layer.

[19:32](https://realalpha.app/v/1hXxzkU77iI/?t=1172) And that's an asset-like layer where you can move very fast and have very good margins, not quite software-like margins because we have significant cogs, but certainly way better than the sort of the neoclots we're getting squeezed between NVIDIA and the rising borrowing costs. So I think it's a very different part of the stack. And, you know, it's a stack where software actually makes a very massive difference. And that software is quite unique and it's quite hard to copy. Tell me, in a billion dollars of this annualized revenue for you, is customer concentration a concern for you at all? No.

[20:10](https://realalpha.app/v/1hXxzkU77iI/?t=1210) The diversification keeps getting better every month. We see, you know, some very large customer, but also a long tail of smaller customers across like a very wide range of different verticals and use cases. I mentioned inference and training and sandboxes and many other things, but it's also different spaces. So you see a fair amount of stuff in biotech actually growing very fast and life sciences, materials, physics. You see stuff in robotics growing very fast. Big companies, small companies, enterprise, startups. It's all over the place. So what do you think inevitably happens to the neoclouds here?

[20:47](https://realalpha.app/v/1hXxzkU77iI/?t=1247) I mean, do you see there being consolidation? Do you see yourself maybe moving, not becoming a neocloud, but you're renting the chips right now? Could you buy chips? Have you bought chips? How do you think the story goes? It's possible. I mean, I think the Neocloud businesses are fine. I think there's been a correction because, you know, it's a tough business. I mean, at the end of the day, it's a financing business. You know, you're borrowing a lot of money up front and then collecting the payment from your customers. You know, you're in the financing space. As it pertains to us, like we may go down in the space, like we might go and look at ownership or quite interested in looking at, you know, Colo or doing joint ventures with Neoclouds or other things.

[21:28](https://realalpha.app/v/1hXxzkU77iI/?t=1288) we think we can go down in the stack relatively easy it's harder for the new class to go up in the stack into software uh into the software space um what's going to happen with the neoclouds i'm not sure i think they're actually going to be quite fine i think you know there's been a little bit just a correction people have been like very excited now you know turns out these businesses are still fine and pretty good do you think that i mean i was going to ask if you saw them you know buying their way more into the inference space buying you know some of these inference providers although i don't know that they have a lot of money to do that with right now given they have like if you look at nebios and CoreWeave they made well yeah i just mean i just mean they're using it all to build the facilities i don't know how much cash they have to exactly i think it's a question of like you know retained earnings and how do you reinvest the profit yeah um and i think that's exactly right but but i think they will try and i think that we'll see a lot of attempts to move up in the stack um and you know in some cases like you know it might succeed like i mean we take it very seriously like our all competition and you know like in in certain ways like we we

[22:32](https://realalpha.app/v/1hXxzkU77iI/?t=1352) like to work with neoclots because we get a lot of capacity with them but from them but but i also think there's going to be some interesting you know attempts by the neoclots to move up in the stack so what happens when next year is you guys had uh your your flagship conference if i'm not mistaken a couple run time yeah yeah yeah run time so our reporter cory weinberg was there and he wrote a piece about his experience there. So there were two things that stood out to me, two comments from speakers on stage at the event that I wanted to ask you about. One was from Diogo, who he co-founded TypeSafe, which is the company behind Jav. And then there was another comment from an executive from Anthropic. Both of these comments were around agents and around the extent to which we can automate things i think the the anthropic executive said something to the effect of we're still dealing with the 20 you know human in the loop right now um and the jev uh uh executive you know he said uh where is all the automation i think i'm paraphrasing his comment so my question for you is eric are the models just not good enough enough for agents to fully automate everything? Is full automation even the goal? What needs to

[23:53](https://realalpha.app/v/1hXxzkU77iI/?t=1433) happen here from your perspective? I think the models are incredible. I mean, you know, I use them for writing code every day and they're phenomenal and they're getting better every day. I think full automation, I don't know if it like matters. Like my view is always like people are obsessed with sort of, you know, automating away humans. I think in every era, like the value that gets created is rarely a sort of, you know, getting rid of humans. The value is like creating new use case and new amazing things and building new products right like that to me is like the exciting thing like i'm incredibly excited about ai you know using ai to discover new drugs or or to enable new applications like sooner runway that previously didn't exist like that that to me is like typically where the value is created so but let me let me push back on that here because i mean you're saying it doesn't matter and and the reason i'm asking the question though is because if you think about enterprises and companies spending on ai seeing any kind of return i mean the reality is they are not seeing the return right now and you could argue maybe it's because they still have to have a person in the loop the person is doing different work

[24:52](https://realalpha.app/v/1hXxzkU77iI/?t=1492) they're not doing um you know as much manual work but they're still spending the same amount of time so i mean i'm gonna push back here, I think it does matter. Yeah, I mean, I think so. Like it does matter. I'm not trying to, like, I think if you go back 30 years and you look at word processors or internet and email and things like that, like every time there's like a new technology, you get rid of like 80% of old junk and then you make humans more efficient at the remaining things. And humans are able to move higher up in the value chain and have even more leverage. And that is, I think, what's going on. It's like, I think these agents are incredible at letting humans get more stuff done just as like you know i feel it when i'm writing code i'm able to write way more code and so i don't know if it like really matters like do we really need to add automate away like the last you know the 10 to get rid of me like i'm not sure i think at the end of the day you can have a human controlling a lot of these things and the humans can have way more leverage and i think that's incredible but but isn't automation the only way you can then quantify the the cost savings or at least attach a dollar value to to uh what what productivity it's actually having no i don't

[25:58](https://realalpha.app/v/1hXxzkU77iI/?t=1558) think so. I think you can look at the output. I think you can look at the number of lines of codes being written or the number of PRs being merged. I mean, these metrics are hard. I heard in previous conversations, it's always hard to measure productivity, right? How do you measure output? But I think that's ultimately what matters, right? It's like the output coming out of the factory, right? Great. Well, Eric, I want to thank you for coming on. That is Eric Bernhardsson, the co-founder and CEO of Modal. Thank you so much. from TI-TV. Microsoft pulled out of a data center deal with N-Scale in the UK in the latest data point to a challenging narrative facing neocloud companies right now. To break down where this story is headed, I want to bring on Nick Del Dio, Managing Director at Moffitt Nathanson.

[26:40](https://realalpha.app/v/1hXxzkU77iI/?t=1600) Nick, welcome to the show. It's great to have you back. Yeah. Thanks for having me, Akash. So, Nick, we saw the Microsoft N-Scale news over the weekend. We'll get to the power side of this story in a second here. I think CoreWeave is the name that everyone is watching as a proxy for all this. It was a tough week. I saw Goldman upgraded the company, which gave it a little bit of a pop this morning, although it looks like that was short-lived. So where do you think we are with the CoreWeave story? Where are you on it? Help us walk through the scenarios. Yeah. So I would argue that for me, pure fundamentals-based valuation, it's probably undervalued right now.

[27:22](https://realalpha.app/v/1hXxzkU77iI/?t=1642) But I think we also have to be cognizant of the fact that the financing environment is not necessarily that great for them at the moment. There's sort of this collision taking place where CoreWeave and a lot of the other neoclouds have these immense capital needs over the coming years. But their balance sheets and their internally generated cash flow don't support those ambitions. So they need to rely on external capital to really achieve what they're trying to deploy. You know, at the same time, however, you see this increasing scrutiny on AI linked debt deals, you know, particularly for higher yield issues. The market is anticipating a lot of AI linked high yield issuance over the coming years. So between those two things, you know, you've got a setup where, you know, again, maybe from a pure fundamental perspective, the stock looks interesting.

[28:11](https://realalpha.app/v/1hXxzkU77iI/?t=1691) But on the other hand, if the capital markets don't cooperate, the company could be in a very difficult situation where they are unable to raise the capital they need or fulfill the commitments that they've made to deploy compute for customers. Why are the fundamentals attractive to you? Well, I would argue, and to be clear, we've had a neutral rating on CoreWeave since they came public. So we've been, I think, a more conservative stance than most of our peers. But when you drill down into the project-level economics, how much money do they actually make from a deal? It's reasonably attractive today, and it has gotten somewhat more attractive with the increase in pricing that we've seen for GPU rental.

[28:49](https://realalpha.app/v/1hXxzkU77iI/?t=1729) That's partially offset by higher costs and a higher cost of capital. But for every dollar that they put to work deploying compute, we estimate that they generate 15 or 20 cents in net price. Value. So under those circumstances, the more money that they put to work, the more value it creates for their shareholders, obviously. Again, I have I have concerns longer term that, you know, eventually the market is going to normalize from a supply and demand perspective. We'll see competitors improve the quality of their offering relative to core weave. Again, there'll be just less of the the imbalance that we see in the markets that that's benefiting providers of compute today.

[29:28](https://realalpha.app/v/1hXxzkU77iI/?t=1768) But at least for the foreseeable future, it seems like they're in a decent spot with respect to the pure value creation potential, the deals that they've been consummating. I want to talk about the financing component to this, the second half of this story. So, I mean, we saw inflation was hot last week. You know, the rate hike risk is real for the financing story. Is all of their debt right now, is it fixed? Is it floating? What exactly is the concern here? Well, it's a mix. So we should probably step back a little bit because Corriev has a reasonably complicated, multi-layered debt capital structure.

[30:08](https://realalpha.app/v/1hXxzkU77iI/?t=1808) You're going to tell us how creative they've gotten. All right. So the foundation of their borrowings are really what they call delay-draw-turned-loans. And this is essentially secured borrowing against contracted revenue that they have with the GPUs posted as collateral. So let's say that they sign a deal with Microsoft for five years. what they will do is take that contract, put it in a box, take the GPUs, put it in the box, and then go to a lender, and the lender will underwrite that revenue stream from Microsoft and loan money against it.

[30:39](https://realalpha.app/v/1hXxzkU77iI/?t=1839) **Nick Del Deo:** It's essentially receivables factoring. And that's the bulk of their debt. Basically, again, secure lending.

[30:48](https://realalpha.app/v/1hXxzkU77iI/?t=1848) They also use some vendor financing, say, from Dell. That tends to be the safer portion of their debt capital structure. it's closer to the assets and the interest rates are naturally lower as a result um you move up and then to the extent that lenders depending on the credit quality of the customer may loan 70 to 90 percent of the value of the gpus in these secure deals CoreWeave needs to plug the remainder with essentially parent company debt so it has a residual claim on those assets and that's where you've seen the rates spike recently um for straight parent company debt it's probably gone from 9% or 10% several months ago to closer to 15% today. They also have some convertible debt outstanding. So in that case, the interest rates are lower today, but if the stock does well over time, they can get converted into equity and dilute shareholders. What's really interesting about this dynamic is that if you look at CoreWeave's enterprise value today, it's probably 60% equity. And again, without getting into the weeds, you typically don't find distressed companies with 60% equity as part of their enterprise value tends to be a lot lower.

[31:58](https://realalpha.app/v/1hXxzkU77iI/?t=1918) So there's a bit of a mismatch in terms of what the equity market is saying about Corwe's prospects and what the debt markets may be saying. But again, the 15% interest rate number that's gotten a lot of attention recently is for a portion of their debt capital structure closer to the equity. And that's where a lot of the pushback or a lot of the tension in the markets has been um that high yield component i think the the secure debt component is probably not an area where they're going to have a lot of issues raising capital prospectively have you talked to the company at all in in the last week uh as they've sort of navigated this this recent um reaction from investors no i haven't spoke to the company but they've made you know comments that were either public or yeah made privately but intended to be public effectively i i'm just wondering if if you've gotten a sense uh if they have changed anything about their strategy not just in the last week but maybe a lot maybe the last couple months to uh ...to get ahead of some of these risks, or if you've gotten the sense that they're going to change anything, you know?

[33:09](https://realalpha.app/v/1hXxzkU77iI/?t=1989) Well, certainly, you know, they have historically been very focused on the use of debt to scale the business, the argument being that to deploy the amount of capital that they want to deploy using anything but debt would result in essentially an obscene amount of equity issuance and dilute their shareholders into oblivion. um they um after the second quarter you know they issued a a reasonably large convertible bond i think it was four billion or thereabouts um so they have some capital some cash on the balance sheet from that they have a number of delayed draw term loans that they have not fully drawn down yet um they've got about five billion in cash on the balance sheet so i think in total between those things they have about 22 billion or so in liquidity available which is probably sufficient to carry them through the end of the year or thereabouts.

[33:59](https://realalpha.app/v/1hXxzkU77iI/?t=2039) So I would think by year-end, they probably need to come to market again to raise additional capital, and we'll see what happens. They did put an ad the world will. Well, we'll see what happens, meaning... How the capital markets respond to that. And if they don't, I mean, if they don't, then that's going to be pretty dicey, right? Absolutely, and that's the risk they pay because the sheer quantum of capital they require means that capital markets need to be open for them to draw what they need. You know, they did put an at-the-money offering in place. That means sell equity into the market. My anticipation, or my expectation rather, is that they probably lean a little bit more heavily on equity, respectively, try to lop off some of this tail risk that comes from being so levered to the debt markets remaining open at reasonable terms.

[34:53](https://realalpha.app/v/1hXxzkU77iI/?t=2093) But we'll see. That's kind of my read on the situation as opposed to anything the company has set out, right? So the other risk, I guess, that I've been hearing about is sort of this refinancing risk. If they need to refinance any of this debt, to what extent they will be able to do that, what rates they'll be able to do that, whether or not they'll be able to meet their interest payments. That's why I was asking about the fixed versus floating and the extent to which any of these interest payments will change. I don't have a great idea of the extent to which interest payments will get. More expensive by census, it could, right?

[35:30](https://realalpha.app/v/1hXxzkU77iI/?t=2130) Yeah, so the parent company debt is, I don't think they have a maturity until 2030 or thereabouts. So the parent company debt, they're not going to have to refi that for some time. The delayed draw term loans do have an amortization schedule, so they pay those down as they collect from the underlying customer contracts. Those are typically floating rate instruments. So for plus a spread. But in many cases, the lenders actually require that they that they hedge that exposure. So I think, you know, there is some potential for rates to increase in the in the near to medium term with the broader rise in interest rates. But in terms of a big refinancing wave, probably not on the horizon for for some time.

[36:14](https://realalpha.app/v/1hXxzkU77iI/?t=2174) Okay. So then, well, let's talk about sort of the, I guess, the other risk that people are scared of, which is just default, which is why these credit default swaps have gotten so much attention. How much of a risk do you think there is at all? Yeah. So, yeah, I think the term loans that they have outstanding, very close to the assets, I think those are fine. The parent company debt, it's really based on the residual value of these GPUs at the end of the day. You could say that Corby's business model, in a sense, is basically buying GPUs and paying for them with a combination of borrowed money and prepayments from customers, where you put in.

[37:02](https://realalpha.app/v/1hXxzkU77iI/?t=2222) So the value to the company comes from, you know, the value of the GPUs and what they can generate from a cash flow perspective after the term loans are paid off. It seems like the economic lives of GPUs are likely to extend beyond the four or five years that a lot of people thought they would be previously. We've seen a lot of evidence in the market to support that. you know in terms of their ability to um to repay based on that you know it would suggest that they're not in a terrible spot i think the bigger risk to core weave is the idea that the capital market sees up and they are unable to procure additional funds um to deploy compute that they've that they've contracted to deploy for customers or again if we look out a few years and they have some of this parent company debt come due that they need to refinance that they're unable to tap the market for that.

[37:54](https://realalpha.app/v/1hXxzkU77iI/?t=2274) So I think if that, you know, this is not just a CoreWeave risk, by the way, we're talking about the whole sector here. Just looking ahead here, if that happens to CoreWeave, if that happens to any other of these NeoCloud companies, who, I mean, do you think NVIDIA steps in and buys one of these companies? What do you think? Yeah, so clearly, you know, one of the options that CoreWeave has, or they have several options, you know. if stuff really hits the fan, they're not going to stand still. You know, yeah, I think that equity is a likely option. There's the NVIDIA put, you know, where NVIDIA might come in and either guarantee, provide allow CoreWeave to get financing on more reasonable terms, maybe a direct investment in the company, maybe eased payment terms, things like that.

[38:43](https://realalpha.app/v/1hXxzkU77iI/?t=2323) You know, we could see additional vendor financing from key suppliers like Dell. We could see them embrace more of a bring-your-own-compute type models, maybe offload some of the capex to other parties yeah i don't think any of those are are very appealing options you know to the extent that NVIDIA stepped in and helped them i think that would sort of shatter the image of of you know these businesses as ones that could stand on their own two feet and you know do some last they're already investing they're already investing in these companies though yes yeah but i think there's you know at least in my mind there's a difference in you know investing to help get them going versus being core weave where you've got you know, multi, a very large enterprise value and you need help from your key supplier to keep it going.

[39:25](https://realalpha.app/v/1hXxzkU77iI/?t=2365) But I view an outright acquisition as less likely simply because if NVIDIA were to do that, it starts to introduce some conflicts between, say, the chips, the pricing, the terms that they would give like their own cloud versus third parties that they're working with. That's one of the challenges that Google is trying to work through as it tries to build out its TPU ecosystem. And so I think NVIDIA would try to avoid that if they could. And let me ask you another question here about falling token prices. This is another discussion that's been going on. How do you think falling token prices affects the economics of these NeoCloud businesses?

[40:09](https://realalpha.app/v/1hXxzkU77iI/?t=2409) Because there's the whole Jevons paradox argument that, well, people will keep using it. I don't know if you have any counter-arguments against that or if you think that is actually a threat. Yeah, I think, you know, to a point, it's arguably helpful. You know, lower prices stimulates more use cases, more demand for the underlying infrastructure. We've seen, you know, despite the rise of open weights models, we've seen prices for GPUs rise this year. You know, so again, there are plenty of factors that go into that, but it's not obvious to me that falling token prices, is, um, uh, you know, completely trash this business model.

[40:47](https://realalpha.app/v/1hXxzkU77iI/?t=2447) Because they'll be able to, uh, sell more, uh, capacity basically. Yeah. Ultimately, you know, the, the customer base may be less profitable. I think it's, I, you know, you think about the different layers of the value chain, that's probably where it has the most impact so long as it. as long as it isn't so detrimental as to impair the customer's ability to pay for the compute, it probably helps infrastructure providers. Right. And if we, let's go full circle to the Microsoft N-Scale story now. That is a story about power. Microsoft reportedly was not able to, well, I guess the N-Scale facility wouldn't have been able to get connected to the power supply as quickly as they wanted to. So of the companies you cover, you cover a lot of these neoclouds. How many of them are having issues or facing delays in their facilities because they cannot get connected to a power grid or supply of some kind? Look, it's a challenge across the data center space. Different players have been more successful than others. You tend to find that folks that are more experienced, so think of the brand names like Digital Realty, Equinix, GTS, and so on, who have long-standing relationships with the utilities, with key suppliers,

[41:55](https://realalpha.app/v/1hXxzkU77iI/?t=2515) **Nick Del Deo:** with municipalities and so on tend to plan further out and can leverage that relationship and experience to get projects online on time.

[42:07](https://realalpha.app/v/1hXxzkU77iI/?t=2527) I think disproportionately the challenges tend to arise with newer, less experienced players. So certainly N-Scale would fall into that category. They have tremendous ambitions, but a relatively short track record. But certainly, again, the biggest challenge across the data center space today is access to power and whether that's access to the generation, the transmission, transformers, UPS, whatever it might be, that whole ecosystem is the biggest single challenge to bringing capacity online. So I'm just wondering the extent to which you think this could be the straw that breaks the camel's back. I mean, financing is one risk, like you said. Fundamentals are another risk, and the other risk is power here.

[42:50](https://realalpha.app/v/1hXxzkU77iI/?t=2570) I mean, if this continues, that's my view on it, at least. This could be it. Yeah, say that access to power, access to data center capacity is crimped across the board. I think what that would tend to do from a neocloud perspective is those with capacity in place today would probably be able to benefit from increasing prices, so long as your capacity is not under a long-term contract such that you can't actually reprice it. Conversely, you wouldn't be able to deploy as much incremental volume as you would like to otherwise. So if you think about the value of these businesses, a combination of in-place assets plus future growth opportunities, further constraints in data center or supply coming online would probably help the former.

[43:39](https://realalpha.app/v/1hXxzkU77iI/?t=2619) It would help the value of your in-place assets, but it would crimp the amount of value that you can create from bringing new capacity online. I think for a fast growing company like Acorweave or Nebius or however it might be, it probably nets out to a negative. Great. Well, Nick, I want to thank you for coming on. That is Nick Del Dio, Managing Director at Moffitt Nathanson here on TIA-TV. The Trump administration's new super intelligence task force met for the first time late last week. The Information's co-executive editor, Amir Afradi, and tech and politics reporter, Leo Schwartz, got a chance to speak with one of the leaders of the group, Emile Michael, Undersecretary of Defense for Research and Engineering and the Pentagon's Chief Technology Officer.

[44:24](https://realalpha.app/v/1hXxzkU77iI/?t=2664) They talked about the task force's first meeting, the Pentagon's frayed relationship with Anthropic, and the AI race with China. Here is that conversation. Emile, welcome. Good to be here. Good to see you both. All right. Well, you just had the first full meeting of the SI task force yesterday. This was just set up to coordinate a lot of super intelligence policy activity with the executive branch and the private sector. So tell us what was discussed. What are your goals for that group? And anything else that's really a headline from that? Well, this group was formed on the backside of what was really a historic summit that brought the leaders of the superintelligence industry and, you know, the cloud industry, the cybersecurity industry.

[45:22](https://realalpha.app/v/1hXxzkU77iI/?t=2722) I mean, the number of people who were there from Lisa Su to Jensen Wang to Elon Musk and Dario, you name it, it had really quite the roster of leaders from the tech industry writ large. And then specifically, the subject was superintelligence. And what is the government going to do to ensure that Americans get the benefits of superintelligence but are protected against the risks? You know, there's acknowledged risks with cyber and other categories. So Jay Clayton, who's the director of national intelligence, is leading it alongside of me, Scott Kupor, and the head of the FTC, Andrew Ferguson. And so what we're trying to do and what we did as the first meeting yesterday is start to task out how we start to define the relationship between the government and these superintelligence companies.

[46:21](https://realalpha.app/v/1hXxzkU77iI/?t=2781) And I'm in particular responsible for the relationship with the superintelligence companies. I know a lot of these these folks and for the incident response, like how are we going to make it so that when something happens, the government knows about it and has visibility into it and then potentially can can point resources toward that issue. Right. And then Andrew Ferguson from the FTC is going to focus on all the laws and regulations that already exist today that are applied to these companies so that the American people understand that there's like a robust set of protections already have. So and Scott Kupor is going to focus on critical infrastructure, not just government critical infrastructure, but hospitals, wastewater treatment plants, electricity grid, banks, you name it.

[47:09](https://realalpha.app/v/1hXxzkU77iI/?t=2829) And so the meeting yesterday brought all of us together alongside of Susie Wiles and Secretary of Treasury and Secretary of War, among others, to figure out how we're going to split up the tasks, how often we're going to meet, what is the output of this 120-day effort. So it was pretty exciting to be in the room with that kind of set of people who really care about this, these set of issues, along with Secretary of Commerce, Howard Lutnick. And it was a very successful kickoff to what is, I hope, a good 120-day period. Well, we've got a lot to talk about when it comes to security and safety of these models that are coming out.

[47:47](https://realalpha.app/v/1hXxzkU77iI/?t=2867) You testified earlier this year that we're sort of seeing the first kind of cyber weapon, you know, AI models. You've warned that some of these models could be used to develop biological and chemical weapons. And maybe that's coming as soon as next year. And you've also talked about the potential for Chinese operatives and other potential adversaries to really kind of poison the training data for AI. And so just given all these threats and given your responsibility to infuse the Department of War with this technology, this type of technology, how can you do that given those kind of concerns, especially if models may already be compromised?

[48:32](https://realalpha.app/v/1hXxzkU77iI/?t=2912) Well, I was saying that that's what the model companies themselves say about their products. And I think, you know, it is certainly true that there are cyber capabilities that have some risk if someone's going to use them offensively. I think that we can sort of all agree on. And then we're also then, therefore, in the midst of a cyber hardening. exercise where not only Department of War, but government and businesses are going back on their old code and their old systems and making sure there's no vulnerabilities and so on. But that's what they say about them. But a lot of what they've said has been proven wrong or is likely to be proven wrong. So part of the other reason for this SIF force, if you will, this SIF, is to make sure that the good news story is told about what AI can do.

[49:29](https://realalpha.app/v/1hXxzkU77iI/?t=2969) because I think it's been a lot of doom and gloom. If you look at the public surveys about how Americans feel about superintelligence relative to any other country in the world, we're the most doom and gloom out there, yet we're the center of the world on that point, right? And obviously there's some narrative here that I think has affected people. But how do I proliferate this in the Department of War and do it safely? I mean, we do a lot of testing. We do our own like just, you know, there are several parts of the government, including at the Department of War, where we rate models based on their capabilities.

[50:06](https://realalpha.app/v/1hXxzkU77iI/?t=3006) So we know what they are, you know, oftentimes ahead of time before they're launched, which gives us insight into what their capabilities are. And I think that's good. I think the American people want that, right? They want to make sure that their national security enterprise understands what kind of products might be out there that could have some risk in them. So, but I will tell you that it's been successful so far.

[50:30](https://realalpha.app/v/1hXxzkU77iI/?t=3030) **Emil Michael:** A year ago today, we had 80,000 people out of 3.5 million Department of War employees using AI. And today we're about 1.7 million.

[50:39](https://realalpha.app/v/1hXxzkU77iI/?t=3039) So, and it's just for basic use cases, right? So, it's making a difference at the enterprise level, like making people more efficient, like you would understand as a writer and a journalist. It's making people more efficient in the intelligence community because you could synthesize more data more quickly. And at the warfighting level, it's baked into a lot of our systems. And again, people assume, you know, they don't understand that warfare is a lot of forms and paperwork and permission and all these kinds of things. Let's simplify the back end of that so that people could focus on the battle at hand.

[51:13](https://realalpha.app/v/1hXxzkU77iI/?t=3073) Well, you know, speaking, we'll talk more for sure about AI and warfare in a sec, hopefully. But the doom and gloom that you mentioned has actually gone way up, at least when it comes to the kind of alarm bells that a lot of these AI researchers developing the frontier models are sounding. And so Leo and I are really curious about your response to things like the hugging face incident and other potential risks that we're seeing that appear to be a little bit more tangible than they were even just a few months ago. And then just this question around, you know, are the existing laws sufficient for us to deal with the consequences of, I don't know, runaway AI or AI that's doing something it's not supposed to do?

[52:02](https://realalpha.app/v/1hXxzkU77iI/?t=3122) Yeah, I mean, I remember being in Silicon Valley and I lived in that bubble for a long time. So I'm familiar with how big that bubble could be. And I'm not talking about a financial bubble, just the information bubble. if you will. And so, you know, the researcher who'd hopped between OpenAI and Anthropic and then a few weeks later sort of came out and was talking about humanity ending in three years. You know, I do think, you know, in this kind of information environment, that kind of gets a lot of play. Like, never before have you been able to get 200 million views on something in a short period of time, unless it was so dramatic like that. So I think people may be living in that bubble a little bit. And you can believe that and also believe that there's some risks with the technology that we have to protect against. Like those two things are not inconsistent.

[53:00](https://realalpha.app/v/1hXxzkU77iI/?t=3180) What I think is inconsistent is if these SI megacorps say, hey, you have to stop me because I can be doing something dangerous. I'm, you know, I'm a drunk driver in the car. You've got to pull me over. I can pull over myself, but you have to pull me over so the other guys don't get there first. That kind of thing doesn't make a lot of logical sense to me. So I try to take a balanced approach to this. And that balance is with with all great technologies. I remember when we were launching Uber and people were lamenting the loss of taxis. And now with Uber, people lamenting loss of Uber with driverless cars. Right. So the fear kind of traverses technological advancement, you know, throughout history. So I take a more balanced approach to it.

[53:51](https://realalpha.app/v/1hXxzkU77iI/?t=3231) And when it comes to, are the laws sufficient? I'll leave that to Andrew Ferguson to answer. But based on my understanding from him and my reading, more than sufficient. There's the Computer Misuse Act. And when I say people, what laws govern the Internet? They said, well, there's Section 260 or whatever. I said, well, that exempted the Internet from liability. That was the one law that took liability away from them, didn't impose any new restrictions on them. What laws would have stopped the mobile phone? So people sort of assume that the body of laws for computer misuse, fraud, someone – you have a product that you knowingly know that people might misuse because you haven't built the right guardrails in.

[54:37](https://realalpha.app/v/1hXxzkU77iI/?t=3277) There's a lot of those concepts that are already there. What about the mechanisms that the executive branch currently has against AI companies that aren't adhering to these frameworks? Because one of the biggest questions I know our readers have is a lot of these are voluntary. You had the joint commitment that the Frontier Labs agreed to. Obviously, there's the voluntary framework that came out of the executive order this summer. We saw in reporting yesterday after the first meeting of the task force that a White House official said that export controls like were imposed on Anthropic earlier this year were one potential stick that could be used against AI companies.

[55:13](https://realalpha.app/v/1hXxzkU77iI/?t=3313) Do you think that's an appropriate mechanism for keeping companies in line? Yeah, of course. I mean, there's been something called ITAR for years. If you ask any cybersecurity researcher whether there was a certain amount of encryption for a while, you couldn't export that encryption because we thought it was too dangerous for other countries to have that sort of encryption. So we have lots of tools, whether it's a restriction on a type of technology being export because of the advantage we think it might give an adversary to just straight-up export controls and limitation on foreign nationals' ability to use that software.

[55:50](https://realalpha.app/v/1hXxzkU77iI/?t=3350) That's just a few among many things that the executive branch can do. Then you go to law enforcement, whether it's a bad trade practice, something that violates a law that the DOJ can enforce on. There's just a lot of them. And I guess I'm not the world's expert on this. I think the FTC chairman is. So I think we try to get an interview with him. But we've already proven that when Anthropic refused to pull back their model, and this is the company that said you should have the right to pull back any model at any time for any reason. That's what they said. And they said they wouldn't when we found out it was jailbroken. They were export control within two hours. So I think you can count on this administration, this president to act with urgency should there ever be a real issue.

[56:37](https://realalpha.app/v/1hXxzkU77iI/?t=3397) So speaking of Anthropic, what's your relationship like with them today? When's the last time you spoke to Dario or Tom or any of the other co-founders? And then I have a follow up on kind of what's going on with Anthropic. Yeah, well, Anthropic is out of the Department of War. As we said, six months ago, they would be. With Dario at the White House event a few weeks ago, I talked to all these companies and really have for a long time constantly because my first duty is to or for, you know, to the department and for the American people from a national security standpoint. And if that means I'm talking to someone who's suing me, well, then I'm talking to them.

[57:20](https://realalpha.app/v/1hXxzkU77iI/?t=3440) That's OK. So I talk to them as much as I need to, and I'll probably be more in the coming weeks and months. so do you feel yeah go ahead please let's say quickly what do you expect to be working with anthropic on and do you think that relationship that tense relationship you have with them will impair at all this new responsibility you have as the vice chair of the si task force and and being responsible for interfacing with these companies no they're they're i'm an adult they're adults um you know i've seen tom brown many many times and spoke to him many many times and expect to be dealing with them just as much as I'm dealing with the other SI companies in a fair, even-handed manner. We're trying to get to common frameworks, right? The White House Accord was four common elements that they each agreed to, to implement. And I think the things that we will be doing in the future, whether it's incident response, will be a common framework that they all agree to, hopefully, about how to report incidences and how to rate the severity of incidences so i expect to work with them uh fruitfully through this period and when it comes to to the department of war obviously you have the contract cancellation with anthropic labeling them

[58:33](https://realalpha.app/v/1hXxzkU77iI/?t=3513) a supply chain risk um you know there's been a lot of chatter here in silicon valley about anthropic uh kind of imbuing its ai with the ability to be like a conscientious objector and telling customers that they cannot scold or use abusive language with the model. Mustafa at Microsoft is warning about this kind of anthropomorphizing of the technology because maybe that in and of itself could lead to AI taking unintended actions. And so do you agree with that? What do you think about that? And do you think you'll ever change your mind on working with Anthropic at the Department of War?

[59:13](https://realalpha.app/v/1hXxzkU77iI/?t=3553) **Emil Michael:** I mean, here's my very, very succinct view on this, which is if you don't want humans to lose control, you can't give models the control over humans.

[59:27](https://realalpha.app/v/1hXxzkU77iI/?t=3567) And that whole notion that a model has feelings and it can overrule the founders of a company is by definition loss of control. They are saying we want to lose control to something that may be smarter than us or may be more moral than us and are ever conscious than us. That I don't agree with as a moral principle. And as long as that moral principle is embedded in their software because they attach their constitution, which has that concept against how the model behaves, I cannot have them at the Department of War. Because the Department of War, we have a command and control. So we have a constitutional structure that says the commander in chief and then he delegates that power down forward.

[1:00:12](https://realalpha.app/v/1hXxzkU77iI/?t=3612) If a model gets in between one commander and their superior and says, well, I don't like what you say, despite your commander saying it, that doesn't work in the military. I don't think the American people want that to work in the military. So, you know, if that would happen in the real world, you would have to quit your job and you just leave. You're like, I don't want to do what my boss is telling me to do. So in this case, when you have a foundational technology that in a time of need, you may not know how it's going to behave. That just doesn't work. So unless they change that, there's no discussion.

[1:00:46](https://realalpha.app/v/1hXxzkU77iI/?t=3646) How do you know that models being developed by other companies don't have hidden safeguards that prevent you from targeting whoever you want to target or analyze data at the Department of Work? Well, I mean, that's why you have terms of use in a contract and you have a, you have either have a constitution or you don't, or you have, you know, the CEO, you hear what Sam Altman says fairly regularly. It was like, we, you know, we are governed by democratic process. We live in a democratic society. They're an entity in that democratic society. um elon signed the same terms of use um that which were for all lawful use cases we could use his software google signed the same thing so really the outlier are the ones that say they want to impose their views on the department of war not the standard are you expecting are you expecting the the supply chain risk legal battle with anthropic to go to the supreme court you have we've had two cases now with with very different decisions yeah the i mean the the the case in California was at a district court level. Um, the one in DC that we won was at a circuit court level. So that would be the controlling, um, you know, precedent in our view. I'm sure that they

[1:02:00](https://realalpha.app/v/1hXxzkU77iI/?t=3720) appealed that already. Um, I think we'll win and, you know, but, you know, legal processes are legal processes. And, and I think, you know, I think everyone was, they were surprised that we It was clear to me that we were going to win, but we'll see. There was a lot of discussion after that summit that you mentioned about President Trump complimenting Daria Amadei from Anthropic. Do you think, I mean, is there any disagreement within the administration over how to deal with Anthropic? I mean, I got to tell you that, and I called it historic for a reason, and I don't throw that word around lightly, but not just the amount of market cap, the amount of brains that are in that room, and then the fact that they all came together and agreed on this compact, including Dario, including Elon, including Jensen, including Sundar, was quite awesome, right?

[1:03:00](https://realalpha.app/v/1hXxzkU77iI/?t=3780) And so if you could get, you know, and I remember I kind of joked to Elon afterwards, I was like, that was great. He's like, what do you think was great about it? I said, the fact that all you guys agreed, you all have your different competitive battles that, in fact, you all came to agree and said, here's a good safety framework start that we all can sign up to was awesome. And so, like, if that ethic and that sort of cooperativeness continues, I think hopefully everyone will be fine. There won't be any divergent views on any one of these companies. What else are you expecting to come out of it after that 120 days?

[1:03:37](https://realalpha.app/v/1hXxzkU77iI/?t=3817) You mentioned an incident framework. Do you think that will be something set in stone that the companies will agree to? Do you think there will be any legislative calls out of it for potential bills or new statute around what incident reporting looks like or otherwise within your responsibilities there? Yeah, well, the three things out of the SIF are this incident response framework, this, you know, how are we going to make sure critical infrastructure is hardened, and then what laws and regulations we already have that apply to these kinds of companies in this new technology area. Separately, the David Sacks-led advisory board called PCAST, I forget what it stands for at the moment, they're going to do a study on any recommendations that might exist that are necessary for legislation.

[1:04:28](https://realalpha.app/v/1hXxzkU77iI/?t=3868) You know, there has been legislation proposed by this administration a while back. That was to sort of preempt state rules, have a national framework. It was going to address some of these issues. And I think we're just taking, you know, David Sacks' study is going to take another look at alignment and some of the key technological sort of things that people worry about, study them and see if there's anything more necessary to recommend to the president to do based on that. when you when you talk to some of the most experienced researchers or people at these companies, including Anthropi, a lot of these folks have been talking about the inevitability of kind of government and private industry coming together. I don't want to say merging, but just as the technology becomes more powerful, more capable, important to national security, that there's just going to be all these different entanglements that are just inevitable. The state is just going to be like, this is too important. We have to, I'm not going to say take control of it, but take control of it in some way or have more oversight in some way. How do you think that that will manifest itself? Maybe not right now, but maybe in a couple of years, three years, you know, is the United States going to

[1:05:43](https://realalpha.app/v/1hXxzkU77iI/?t=3943) take equity states in these companies? Is it going to take control over or have oversight over facilities where critically important workloads are running or being developed? What do you see there i i don't see this this is a very very silicon valley bubble sort of question not not i'm not accusing you of that but just you know government and business merging you're like okay what does that make us like a authoritarian capitalist state or what i don't know what that is but um that's their prediction that's their prediction no i i i understand this i know i've seen interviews where people talk about this merger of government and SI companies. But I think if you see what's come out of Meta in the last few months, you're like, wow, this is now a frontier AI company, SI company. They've picked a certain vector for consumers. It's wildly popular. They delayed it because it wasn't as safe as they wanted to be with your data.

[1:06:50](https://realalpha.app/v/1hXxzkU77iI/?t=4010) And now it's a great product. You're like, well, that sounds like a great model. And why presuppose that something else is going to be used different? And there's lots of solutions if something is going to be used different. Let's say someone had a bio-capable model to help do biological research, to do disease research, right, or disease cures. You know, maybe you just do a know your customer thing. I don't know. Maybe they want to know who they're selling it to, these companies, right, because they want to want to be misused, just like you know who you're selling a reagent to at a lab, right? You don't sell it to, you know, the guy at the bodega. You sell it to, you know, you sell it to someone who might use it in a way that you think is appropriate.

[1:07:34](https://realalpha.app/v/1hXxzkU77iI/?t=4054) There's lots of ways this could go, but a big merger between the two will be not only market-destroying, where will the talent go? I mean, there's so many bad implications of that that we've learned throughout history of state-run enterprise. So what are lesser gradations of that? I don't know. I think if I'm one of these companies that's worth a trillion, two trillion, going to $3 trillion, $4 trillion, I'd be spending a lot more on safety and making sure that my products weren't misused and making sure that I was serving the national government in the right way because this government is protecting my interests.

[1:08:18](https://realalpha.app/v/1hXxzkU77iI/?t=4098) And I think you're more likely than anything to see a shift in their investment toward safety in their training, in their sandboxes, and towards serving the government in a way that's productive. What about equity stakes, though? Because this was something that President Trump had brought up as an option and discussed with OpenAI. Have there been any further discussions on that? That was last place. I don't know. I'm not in the loop on that. This is an aside, but... Oh, go ahead. I do war, not equity. I was going to say, do you think that's a good idea for the government to take equity in the frontier lab specifically?

[1:08:55](https://realalpha.app/v/1hXxzkU77iI/?t=4135) I mean, we've taken equity in Intel. We've taken equity. Some of the deals, frankly, actually in the Department of War where we're giving a low-cost loan for a critical minerals mine and we're trying to get the taxpayers some potential benefits. So there's lots of places the government has historically taken equity and will in the future. So I don't think it's sort of a crazy notion, but I'll leave that to the president. this is an aside but uh who who came up with the idea to change ai to si there's only one man who came up with that idea and if it's in executive order you can read it i think it was the day after we did the uh the si summit but uh i a lot of way and then elon you know changed spacex ai to spacex si so i think I think there's some goodness in there, you know, in sort of branding it to be closer to what it is, as a superintelligence, you know, relative to artificial intelligence.

[1:10:03](https://realalpha.app/v/1hXxzkU77iI/?t=4203) So, but he is the master brander. What's your position on open source now? Because I know this has been a major source of debate within the administration, especially on government review of open source, being able to maintain that it's used by companies without too much regulation, But obviously, there's more security risks or different security risks that come in. How do you balance national security risks with maintaining open source potential for companies? And where do you stand on that currently?

[1:10:34](https://realalpha.app/v/1hXxzkU77iI/?t=4234) **Emil Michael:** Well, I think, you know, it's, you know, in general, from a business model standpoint, I'm very in favor of open source, particularly American open source, because I think, you know, it's a democratization of intelligence, if you will.

[1:10:49](https://realalpha.app/v/1hXxzkU77iI/?t=4249) And I think there's some real goodness in that. And I think there's also a level of auditability in it because you're showing your weights, which are sort of the brains of SI. And in theory, people can audit that over time. Maybe people will have – probably industries will have to develop around that to do that effectively. And so I think there's a lot of goodness in that. And then the price benefits to consumers is sort of – is huge, right? So it adds more competition in the marketplace. So from a business model standpoint, from an economic standpoint, from an auditability standpoint, there's a lot of good things in it.

[1:11:27](https://realalpha.app/v/1hXxzkU77iI/?t=4287) I think, you know, the concerns you have, frankly, with Chinese open source becoming dominant in the U.S., I worry about that a little bit because, you know, those companies are subject to the rule of the Chinese government, right? So what do those models look like now? But can you solve that with auditability? Well, I prefer to solve it with American open source like Nemetron, like reflection AI, like thinking machines. But I think, you know, none of these companies yet is at the frontier, right? American companies. So I think, you know, we'll see what happens when they get to the frontier. But the Chinese models that are closer to the frontier, we're also talking about potential kind of long-term threats, either cyber threats or kind of bio and chemical threats, assuming that the people using these models can also acquire those materials.

[1:12:24](https://realalpha.app/v/1hXxzkU77iI/?t=4344) That's really the concern, is that as those Chinese models get better, they don't have built-in safeguards, or they may not have built-in safeguards. What are we going to do about that when individuals anywhere in the world or in the United States can potentially use these things for harm? Yeah, I mean, you're assuming the Chinese government absolutely wants chem bioattacks to happen in their own country or their citizens to be committing them on other countries. So I do think, you know— you you know those safeguards will be built in i i who what i what i don't know is i think a lot This is where you have to be a real technologist to really get deep into, like, how are these models trained on what data?

[1:13:08](https://realalpha.app/v/1hXxzkU77iI/?t=4388) How capable are they at each of these subject matter areas that you're talking about? Cyber and math are relatively easy because they're deterministic, right? There is an answer to a math problem, but there's an answer to making flawless code. Some of these other things are different because they also rely on the physical world a bit in a way that's different. And so how those models develop, I think we're all still learning, right? No one's cured cancer yet. That's been the promise, right? No one's cured the common cold as far as I can tell. So I still think when the models meet the physical world, and that's why there's something called world models that think about that problem set a little differently, at least from a robotic standpoint, I think that'll be a potentially separate thing to look at.

[1:14:00](https://realalpha.app/v/1hXxzkU77iI/?t=4440) um shifting to to kind of the the future of warfare and the the super intelligence era and and also just keeping china in mind i think there's always this this talk about the race with china the race with china um i can you articulate for us what exactly are the benefits from a national security or defense standpoint of being ahead of china like do we think that they're going to go to war with us if they're ahead or what's really kind of behind this this notion from your standpoint and where you work um what i think that um a one way to understand this for the for a person who doesn't like live in the war department is um that if intelligence intelligence gives you um able to give you a multiple legs up on your adversary whoever that adversary is um well then you could actually have it as a deterrent to a conflict and let me and there's a reason why we have the intelligence community right the intelligence community takes satellite photos it it uh it you know it does all kinds of different things imagine if you added like super intelligence to that intelligence agency, the things they can understand about what the adversary is doing to protect you from that adversary making a move. You could also potentially

[1:15:30](https://realalpha.app/v/1hXxzkU77iI/?t=4530) use that intelligence from aerodynamics, physics, mathematics to develop weapons that the enemy can't do anything about. The way the United States would approach that is if we had an overwhelming intelligence advantage on, you know, intelligence, on weaponry, on being able to deploy as few troops as possible and as many robots or drones and counter drones and all these things that made you impervious to attack. Well, then, you know, we would approach it as like we did after World War II as like that would be a great deterrent and we would go on our merry way building our economy and our country. I think the scary part, and I think some of these labs would acknowledge this is if an adversary that's ambitious for territory or economics had that advantage over you and was able to put you in a corner or to sort of coerce you as a nation into doing something you didn't want to do, well, that's a scary world. So because we've been the protector of the free world since World War II, I think us having that intelligence is vital for the entire Western world, and then we would, you know, I believe, keep a safer world.

[1:16:43](https://realalpha.app/v/1hXxzkU77iI/?t=4603) Whereas certain adversaries, if they had that intelligence advantage over us, you know, would use it in a coercive manner. You mentioned this dynamic before in the context of export controls of foreign nationals. Obviously, foreign national employees were included in those export controls. Do you worry about foreign nationals from adversaries working at U.S. frontier labs? We just had the case of a Saudi employee putting a spying device at OpenAI. How do you think about that? Do you think there should be stricter controls at these companies? Yeah, I do worry about it. I think the example at OpenAI was a reminder that we all have to be vigilant about this.

[1:17:25](https://realalpha.app/v/1hXxzkU77iI/?t=4645) Obviously, all foreign nationals aren't bad, but foreign nationals that have other loyalties bring with them some risks. And I hope these companies take those risks seriously. I think, you know, the one other thing I don't, you know, we haven't talked about in my job is, you know, I do all the grants to universities for war department research, right? And we just had a case in Michigan University where we had 13 Chinese foreign nationals arrested for espionage effectively. And so this is a problem at universities. It's a problem at companies. We've seen with the North Koreans taking jobs in American companies in subversive ways.

[1:18:09](https://realalpha.app/v/1hXxzkU77iI/?t=4689) As an open society that welcomes foreign national participation in our universities and our companies, we run that risk. And I think we have to be smart about that risk. You know, it's hard to say that without trying – without, you know, sweeping a broad brush on every foreign national at every company. But how should you be smart about that risk? Do you think the government should take any further steps there, especially when it comes to AI? The FBI literally will vet any potential foreign national employee of any company where that company is worried about them to do a background check and make sure that they're what they say they are, who they say they are, and go as far as they can in legal ways.

[1:18:56](https://realalpha.app/v/1hXxzkU77iI/?t=4736) So that service exists, but we can't compel companies to do that. Are there any American institutions or universities that are not compliant or not acting in great faith? Or do you think they're just not vigilant and you're going to help them be more vigilant? I think there's two different categories. is some of them, you know, it's potentially a loss of money, you know, from tuition from foreign students who pay full boat. And then for others, they're a university. They don't have like crazy impressive security practices, right? And so for the ones that say, hey, we want to do the right thing, help us design a security practice that works, we'll work with them and do that. And And there's lots of examples of universities that do that.

[1:19:55](https://realalpha.app/v/1hXxzkU77iI/?t=4795) For the ones that don't want to, they just run the risk of loss of grants because, you know, once an incident happens where there's like a lot of clear that a university doesn't have good security procedures, well, you know, why would you give them sensitive grants that are at the secret or top secret level anymore, right? Because you can't put that national sort of asset at risk. So hopefully I can convince them of the trade-off benefit there. But so far, I will say that, you know, they've generally been very cooperative. Leo, did you want to ask about Project Meridian or the In-Q-Tel news? Yeah, well, I wanted to ask about, you had mentioned the equity opportunities or grants that you explored.

[1:20:46](https://realalpha.app/v/1hXxzkU77iI/?t=4846) Obviously, the loans coming out of the Office of Strategic Capital. Can you preview any other sectors you're looking at now, drones or any other industries, companies that might be getting loans or equity stakes from the Department of War in the near future? Yeah, we're looking at microelectronics companies, things that are, you know, companies that are doing different parts of the supply chain, you know, that are currently by, you know, only done by ASML and TSMC. We're looking at drone parts because sort of certain drone parts that we're dependent on adversary nations for, like brushless motors and batteries.

[1:21:26](https://realalpha.app/v/1hXxzkU77iI/?t=4886) You know, we're very dependent on foreign nations for our battery supply right now. Critical minerals, as I mentioned, you'll see companies doing space, you know, a lot of different space stuff that we think is critical. So I think you'll see like a steady drumbeat of continued deals from us. Some of them will seem esoteric. There'll be some cathode manufacturer or some precursor to pharmaceuticals or something that's specific and has high utility in the War Department. But you'll see a wide range of deals from us over the next 12 months. And what's the breakdown between loans and taking equity stakes there?

[1:22:09](https://realalpha.app/v/1hXxzkU77iI/?t=4929) How do you think about those two options and what do you think they will be in this case? They're primarily loans. You know, I think when we take equity stakes, the primary options, like a warrant, if you will, they're not like us putting cash in.

[1:22:25](https://realalpha.app/v/1hXxzkU77iI/?t=4945) **Emil Michael:** And the warrants sort of give the taxpayer some upside. And so it's, you know, in the high 90s in terms of loans.

[1:22:33](https://realalpha.app/v/1hXxzkU77iI/?t=4953) But, you know, the thing I remind people all the time who are like, oh, you know, equities, you know, every president, almost every president has done some equities. People just forget about it. But we've been doing it in cases where we actually have some grants to give. We can give someone a grant, which is free money, and the taxpayer gets nothing. Or we give them a loan, and we get some equity for it. And you're like, well, that sounds like a good trade for the taxpayer. So we're not out there as a VC just running around trying to do equity deals. They're always in combination with a defense mission or a war mission.

[1:23:10](https://realalpha.app/v/1hXxzkU77iI/?t=4990) and they're usually in combination with we would otherwise give a grant. So instead of that, we're doing some other form of financing. And given your venture and tech background, I mean, I know that there's the SBIC program of helping jumpstart some venture funds or providing loans to venture funds. Have you considered at all taking LP or GP stakes in venture funds through the Office of Strategic Capital or other funds? Yeah, I think there's a program where we give some leverage to VC funds, and I believe we get some of the upside, some of the carry, if you will. So it's a form of GP, if you will. It's not sort of a stake in the GP per se, but it's a percentage of the upside is my recollection.

[1:24:00](https://realalpha.app/v/1hXxzkU77iI/?t=5040) is that the the sbic program do you know yeah yeah yeah no it may not be the first set of deals that were done before i showed up but i think the newer form of deal that we showed up has that and and with this new program or this new announcement with inqtel which is the the cia backed venture firm i mean it was a little a little hazy what that would actually look like is any part of that doing co-investments alongside inqtel what are the actual mechanics of that partnership it's not co-investments it's more it's more that you know they've been historic i think they invested in andrew really early and palantir like from a first investors so it's just making sure that we have a forum where we can we communicate the things we're interested in they still are going to do their own investments but we're also then going to look at their investments more closely and see are these companies who serve who have products that we want to buy so it's just tying us close more closely together in terms of knowledge sharing like what are the mechanics of that does that I mean, you're going to have standing meetings with them.

[1:25:03](https://realalpha.app/v/1hXxzkU77iI/?t=5103) Is that just more information sharing when they invest in a company, you hear about it? You know what, one of my assistant secretaries, Mike Dodd, you know, runs that relationship. So he'd be the right one kind of to ask, but he ran with that deal and made that happen. So, but I don't, I'm not sort of up to speed on like the exact mechanics on it. Speaking of Anderle, I know you've got Project Meridian, which is this initiative to look at the future of warfare. Paul Merlucky involved, Elon Musk, Joe Lonsdale. We were just curious in a nutshell, you know, what you think is not currently working well with the legacy kind of prime kind of contractor companies, maybe like Lockheed Martin.

[1:25:46](https://realalpha.app/v/1hXxzkU77iI/?t=5146) And do you expect some of these neoprimes like Anderle to displace them? I mean, I think, you know, that if you look at the history of the defense industry, in the 80s, there were 50 defense contractors, 5-0. And now we're down about five primes. And so the lack of competition that that creates, combined with we're the biggest bureaucracy in the world, so we sort of buy the transitive property, like, loan our bureaucratic ways to these companies. And then you have something that's not a healthy ecosystem of competitiveness. So what I'm trying to do in my role is bring in new companies. And it's fortunate in that we've learned so much from the Russia-Ukraine war to know that even if we still need and we do the exquisite submarine, the amazing F-35, the like world-class patriot, that you still need drones, counter drones, undersea drones, you know, boats that can that can sail autonomously and they're cheaper and they're easier to produce and they supplement the existing exquisite capabilities. So I don't think we're at a replacement point. Like we don't need to do this for that. We still need to increase our productivity on those on those other companies. But I think for the new kinds of things we need, the new companies are, by definition,

[1:27:18](https://realalpha.app/v/1hXxzkU77iI/?t=5238) clean sheet of paper, going to have a better shot at being able to build them at a scale and speed and at a cost that we like. I had one last question on this. All right, last one. Yeah, last question, I promise. On the offensive cyber program that I think was announced in August, we've seen a bit of mixed messaging on that. I know the national cyber director, Sean Carencross, earlier in the year said that he didn't expect private companies to be doing offensive cyber operations on foreign adversaries. This new program seems like there will be industry partners who are gonna be conducting those operations.

[1:27:54](https://realalpha.app/v/1hXxzkU77iI/?t=5274) What's your current understanding of what that looks like? And do you have a sense of which companies, which private companies might be enlisted in those efforts? I don't really have anything to say about it. okay well that's since um okay then maybe i'll use the last question for this it's a little it's a little bit of a tougher one but i think it's it's it's worth asking because i get a lot of questions about it um so i don't know if you're directly involved in this but but some of the pentagon investigations on the minab school strike in iran have reportedly you know shown that yes there was human error um involved in that um but also that maybe some of the ai or si systems may have accelerated the the kill chain to the point where like you know there was like a rubber stamp on outdated uh uh human intelligence um and so if if is there a lesson from that that maybe in some cases si is being deployed faster than our own kind of safety protocols can can manage um and what does that you know what does does that change anything in your mind around rules of engagement in And...

[1:29:00](https://realalpha.app/v/1hXxzkU77iI/?t=5340) in war i know you've talked about the need to kind of swing back those rules to be a little bit more relaxed or a lot more relaxed compared to the prior administration but what's your what's your view on this you know i haven't read the report so it would be or i haven't heard i haven't you know i like i'll tell you amir i live in a uh i live in a skiff all day so i don't have my phone so i'm not as online as uh i used to be so i haven't read what you you might have read and i I want to be careful about commenting on that, something I haven't read or have knowledge with, sadly. But I appreciate being here with you guys today, and thanks for the time, and thanks for the ability to just talk through some of these things.

[1:29:46](https://realalpha.app/v/1hXxzkU77iI/?t=5386) Thank you very, very much. We really appreciate it. That does it for today's show. A reminder, we are on this stream Monday through Friday at 10 a.m. Pacific, 1 p.m. Eastern. If you can't make it then, episodes are available on theinformation.com, on our YouTube channel or wherever you get your podcasts. Make sure to follow us on social media, on X, on Instagram, on TikTok, and on LinkedIn. I am already excited for our next show tomorrow. Have a great rest of your Monday. Bye-bye for now.