How Cheap AI Could Derail OpenAI And Anthropic's IPOs

0:00 Are OpenAI and Anthropic actually worth $1 trillion each? Well, soon Wall Street is going to find out. Both companies are courting investors for IPOs at valuations north of $800 billion. People really believed in this AI revolution, and they wanted to put their money to work behind it. The pitch, they're the next Microsoft, the next Google tech giants with pricing power that will last for decades. In the first quarter of this year, we saw if you were to annualize it 80x growth per year.

0:33 But here's the problem that pricing power. It's already cracking Chinese open source models. They're eating the low end, American competitors are coming for the high end and the moat that these companies are selling to public investors. Well, that's shrinking in real time. The gap between the two countries, it's closing rapidly. I'm Deirdre Bosa. The biggest IPOs of the year. They're priced for a market that is already splitting underneath them.

1:13 So as enterprises go from a few experimental AI projects to now rolling out across entire workforces, they're beginning to ask, is it worth the cost? Where their spend is going? What's the nature of the task? And if other models may be more performant, and how much could be saved by shifting volume based on the right level of tasks? I think this is going to be a mega theme for the next year. Well, let's do the math.

1:37 Say your company has a $10 million AI budget. Run it on Claude opus. That's Anthropic's top model. You could burn through it in weeks. Run the same budget on Deepseek. That's China's open source model that may stretch across most of a year. The benchmarking firm Artificial Analysis crunched the numbers. Cloud costs nine times more than the cheapest Chinese alternative to do the same work. Even models that are out of date six months old, are perfectly performant for most of the nature of tasks being done.

2:08 And not just cheaper. They're competitive. Chinese labs, Moonshot, Xiaomi, Deepseek, ZHIPU they have shipped open source models in the last four months that match or nearly match American frontier models on the benchmarks that matter. Our models are better overall. OpenAI's is better, Anthropic's is better, Gemini is better. However, their open source models are well ahead of us. Adoption is following the price. On Open Router, the largest AI traffic aggregator, three of the top five models this month, hey're Chinese. And they went from 1% of usage in 2024 to more than 40% this year.

2:44 The reality is that so far, the ones dominating open source AI are Chinese company. For the first time in 2025, the volume of download of open models from Chinese providers was higher than from American providers, right? Even in the US. For Chinese labs, constraint became the strategy cut off from Nvidia's best chips amid export restrictions. Chinese labs. They had no choice but to get creative. So smaller models, cheaper training, more efficient inference.

3:17 The best AI researchers in the world because they are limited in compute, they also come up with extremely smart algorithms. If most of the advances came from algorithms and computer science and programing tell me that their army of AI researchers is not their fundamental advantage, and we see it deep seek is not inconsequential advance. Meanwhile, American Frontier Labs, they are spending hundreds of billions of dollars in AI infrastructure, training ever larger models on the most expensive chips that Nvidia sells on a power grid that cannot keep up, that gets passed to the consumer. So the edge that justified premium pricing that's eroding Chinese open source is closing the gap on capability while American competitors, they're coming for the sensitive high trust workloads where customers will pay up. American Frontier Labs, they do have one stronghold left trust banks, grid operators, defense, healthcare regulated industries that won't touch Chinese models no matter how cheap or good they get.

4:24 For governments, for regulated industries, critical industry. They're not going to be using and leveraging that technology. Chinese models just simply aren't an option. And so organizations that need to trust the models, the decisions that they're making, the code that they're writing are going to be willing to pay a premium to access Western democratically aligned technology. That is where premium pricing holds, but it's also where OpenAI and anthropic are about to get squeezed by American competitors, building exactly for this gap. Take O'hear, founded by Aidan Gomes, one of the authors of the paper that kicked off the modern AI era.

5:04 Cohere builds for that niche smaller, more efficient models specifically for regulated industries. I see things moving heavily in Coheres direction. We've seen that over the past year with our revenue 6x last year and continuing to grow very rapidly this year. And then there's Nvidia, a company that U.S. enterprise already trusts. It's now shipping its own open source models called Nemo Tron, positioning them as the alternative to both Chinese options and the closed Frontier Labs.

5:34 I think Nvidia recognizes that they need everyone to be able to build AI, not just a few players. And so that's why they're supporting open source AI so much. That's why they became, in my opinion, the American King. Palantir. Salesforce. Servicenow, CrowdStrike. They're all already adopting Nvidia open source. And then there's reflection AI, a startup that just raised at a multibillion dollar valuation, building open source frontier models as an American alternative to deep seek. All three are going after the same gap.

6:07 Capable models at a fraction of frontier prices on infrastructure U.S. enterprises already trust. If we can see more and more American startups contributing to open source, we can definitely catch up to Chinese open source. OpenAI and Anthropic. They don't just have a China problem. They have an America problem, too. Even Elon Musk isn't betting on a standalone AI lab anymore. In February, he merged his AI company xAI into SpaceX.

6:39 Elon Musk's rocket company SpaceX, acquiring Musk's artificial intelligence company xAI in what would be the largest M&A deal in history and ahead of a possible blockbuster IPO for SpaceX. The stated reason was more capital for xAI. Before the deal, xAI was reportedly burning roughly $1 billion a month. But after the deal, that burn would be backstopped by SpaceX's billions in annual revenue. So the man who built the largest AI supercomputer in history, who has bet more of his own money on AI than maybe anyone else, decided his AI lab needed to diversify other businesses to support it. Openai and anthropic. They do not have that option, so Wall Street will have to judge them on AI economics alone.

7:24 So here's where we end up. OpenAI and Anthropic. They're being priced as if they own enterprise AI, but the evidence says they own one slice of it. And even Elon Musk hedged his bet. The pitch was pricing power for decades, but the evidence is stacking up against it. Public investors, they're about to decide who's right. You heard from Aidan Gomez in the piece? He coauthored the paper that started the modern AI era, and he now runs Cohere Building for Regulated Industries. We think the full interview is worth your time. He gets into why he thinks the market is moving coherent direction, how he sees that China threat and where premium pricing actually holds is premium AI, frontier AI.

8:11 Is it still worth it? I think you can see in the demand that people are willing to pay, and you can see in the level of CapEx spend on the reduction in free cash flow, that companies can see the future demand ramping. So, yes, people will continue to spend to get access to extremely high quality models in terms of the enterprise and how their adoption of AI is being shaped to the core bottlenecks to that are cost and security.

8:43 And so trust is really deciding who they're going to choose. And to your point about Deep SEQ V4, I think that's an example where for governments, for regulated industries, critical industry, they're not going to be using and leveraging that technology. Chinese models just simply aren't an option. And so organizations that need to trust the models, the decisions that they're making, the code that they're writing are going to be willing to pay a premium to access Western democratically aligned technology.

9:16 Talk a little bit about coheres business model then, because my understanding you guys serve the enterprise and you're a model builder, how are you thinking about it and the cost payoff and benefits? Yeah, absolutely. So we, we build our models from scratch. Our business is exclusively on the enterprise side. So we focus in particular on the high security settings. So think grid operators financial services telco, uh, government.

9:47 Um, and in those settings, the data and the systems that these models are accessing are really national security concerns. And so there's just no way that there's going to be a reliance on a non democratically aligned technology stack. And what cohere does uniquely well among all the the labs is very secure deployments. So whether it's on prem or completely airgapped, we can deploy in a way that there is no cyber risk. We can deploy inside a customer's data center.

10:18 We can deploy into a submarine a kilometer under the surface of the ocean with no ability to talk to the internet. And so that level of security just gives us a completely unique value proposition. It also presents constraints. We have to deploy on extremely limited compute footprints. And so we talk about how these large models are driving massive build out and spend for the hyperscalers. But for these high security settings, they can't get their hands on enough chips either. And so we need to be able to deploy on 2 to 4 GPUs.

10:53 And so that presents a completely new technological constraint that basically rules out massive, massive models. And so we focus on something that is right sized for the market that we we serve. Right. You know, two of the biggest frontier labs, OpenAI and Anthropic, certainly turning their attention to security with mythos. And, you know, Sam Altman, OpenAI were out with, you know, big plan last night. How do you compete with those ones? Um, and those bigger models, you know, mythos, the whole narrative surrounding it is it's so powerful. They can't release it when they do.

11:29 How does cohere fit into that picture? Yeah. So the massive models, you know, uh, rumor is that it's something like 10 trillion parameters in scale that does not fit on 2 to 4 GPUs. And so it's not going to be served inside of these extremely compute constrained environments. So we don't really run into those models when we're serving the market. What I will say is that these massive models that are capable of very sophisticated cyber offense and defense use cases are a very interesting new capability that exists in the world. We can now, at scale find exploits inside very sensitive software. And so it reinforces the significance of private deployments, ensuring that our grid operators aren't putting the code that operates, the power that flows into all of our homes, all of our our businesses in a place that it could be accessed by a third party where that third party could use AI models to find exploits that they could use to shut off the grid, or the same thing in financial services or in healthcare, etc.

12:44 if we expose the infrastructure, the software that powers our economy, whether it's in finance, whether it's in healthcare, whether it's in energy, um, the risk now that someone will be able to exploit vulnerabilities and use them against us is higher than it's ever been. So I think the notion of private deployment is essential. Uh, we've known that these cyber capabilities were going to emerge in models for a while. And as we started to see software increasingly be written by these large language models instead of by humans, you're going to care a lot about which models you're using, whether they're who they're coming from, essentially.

13:25 So if you're going to be replacing your entire software engineering team with models that are writing your code, you probably don't want that coming from China because it might be subtly introducing vulnerabilities that you're not going to catch. And so the trust barrier that I was describing, the trust barrier to enterprise adoption continuously gets higher and higher. Right? And it sounds like, you know, coheres models have a very specific use case when you can have a model on prem and you need the utmost security.

13:55 Um, when you say though that, um, sort of the, the whole. So I think also what you're talking about is sort of the back door, right? That's been a concern with these Chinese models. You can self-host them, but you don't know, um, if there's a back door that can get into your data, why then does AWS, the hyperscalers, Michael Dell, why are they hosting and promoting these open source Chinese models? Well, for certain applications, I think they're great, right? Like in low sensitivity settings, it's pretty reasonable to, to use a Chinese model, I think for startups that are looking for the lowest cost option, uh, and aren't operating at scale yet.

14:36 It seems like a reasonable and very helpful tool. Of course, I think everyone would prefer to be using a democratically supported and aligned piece of technology, so I hope there is. And cohere is contributing to a much more democratic open source. We want to continue to put that out there. But, you know, there is a market for these and they are cheap, as you, as you say. And for the less secure settings, maybe there's a place for them.

15:06 Right. And, you know, you see sort of usage go up and here all the time about how we're compute constrained. So is that calculation for enterprises changing, especially if you're not in an industry that requires so much security. Is that calculation changing especially because some of these open source models are getting very close to the frontier. And that sort of lag is closing. Yeah cost control. And then the other thing is like the compute bottleneck, right? So there's simply not enough compute to support using these massive models.

15:42 So we need efficient small models that are good enough for the use cases that people are pursuing. Otherwise we're just not going to have sufficient compute to meet demand. Um so there is a shift in the market. I think everyone over the past year has been racing to, you know, at all costs, adopt AI. There's going to be another phase that we enter into where CFOs are looking at the expenses being spent on some of these models, and they're going to try to optimize. They certainly aren't going to pull back, and it's going to continue to grow. But we're going to need to find ways to use smaller models, more efficient models.

16:22 Um, I hope that doesn't mean shifting over to a Chinese tech stack to satisfy that. I think there's companies like cohere and others who are building, uh, you know, more aligned technology that satisfies that efficiency need. Right. But it does sort of beg the question, are the American bottom? Are the American models better than Chinese ones? Like what is the gap is deep seek V4 a real threat, especially when we see sort of it matching or surpassing some of the frontier ones on the benchmarks.

16:55 I think it's it's very close the gap between the two countries. Uh, in addition to Canada, in addition to France, Germany, um, it's closing rapidly. Right. And you're saying that there's options. I mean, cohere is working on it. Nvidia is working on it. The startup reflection is working on open source. Is there enough companies working on it in America? And what happens to OpenAI and Anthropic if they aren't working on this sort of open source model race?

17:30 I think there's a lot of folks contributing to open source, um, cohere. We've been investing in it for years now. I'm not so concerned about the number of players. I think there will continue to be good open source options coming from democratic nations like the US, like Canada, like Germany. But I do think that, um, the thing that China is doing that's setting it ahead is effectively distilling the large models, which is against the terms of service of a lot of these large models. Um, and so that gives them a shortcut.

18:05 They can very quickly catch up just from distillation of someone else's work. Whereas for the rest of us that are trying to build a completely independent stack from scratch, it requires more effort and more work to build competitive, great models. Um, but I do see things converging. I do see capabilities, capabilities converging, uh, between all the major providers. Mhm. So what does that tell us about sort of the huge infrastructure spend that we see from the hyperscalers.

18:35 And meta just talked about this coming off a massive earnings day. They're all planning to spend more than $700 billion in infrastructure this year. From where you sit, especially as cohere is building these very specialized models on just a few GPUs, does that money come back, or do you think that the trend is going towards these smaller, more specific models, the kind that cohere is doing? I definitely think it's pointed towards more efficient models. There's going to be a big wave in the market seeking to reduce costs and make things more efficient. Um, it's easy to do POCs on some of these large models, but then as soon as you push that into production, you're dealing with production level scale. Suddenly the price tag just changes the buying equation and cost becomes a huge constraint.

19:22 Uh, so I see things moving heavily in Cahiers direction. We've seen that over the past year with our revenue six x ING last year and continuing to grow very rapidly this year. So I do think there's going to need to be a lot more infrastructure because the demand is virtually insatiable. Um, but as that infrastructure comes online, we're also going to make a big push towards efficiency. So there will be much more AI in the world.

19:54 But those eyes will be using less GPUs, uh, to do the same work as they were doing six months ago. Are the hyperscalers plus OpenAI and Anthropic, are they building for that kind of world? Uh, the hyperscalers are definitely building the infrastructure necessary to serve, um, a significant amount of this demand, although it's important to say that beyond cloud on prem deployments, private deployments, they're growing massively as well. And this is becoming an increasingly well understood security threat that moving too much to the cloud is a mistake and exposes you to vulnerabilities, especially in this new frontier of potential cyber attacks.

20:39 Thanks to these models and so on prem deployments, folks building their own data centers, neo clouds, you're going to see a whole new array of compute providers to satisfy all this demand. So certainly the hyperscalers are building out the infrastructure that is needed, but they won't capture the entirety of the market. They'll probably capture half of it and the rest will be these on prem secure deployments. So where does that leave an OpenAI or Anthropic. That's not you know, the cloud business is not the main business. It's frontier models.

21:12 I think the demand for frontier models, I think the demand for all models is going to be extraordinary. I think there's going to be. So it's basically a rising tide lifts all boats. Everything is going to benefit from this. I'm not too worried about the demand for their products or our own. There is a huge market out there. The applications for AI in the enterprise side, it feels like we're just scratching the surface. We're still doing very simple things. We know the models are capable. We know the models from 18 months ago are capable of way more than we're doing actually out in the industry. So even if you don't update the models for a year and a half, the industry is still catching up. So right, the demand piece, I feel extremely confident.

21:57 There's a lot to go do in the global economy. Good question from Carson Allen, who's a regular viewer of our live streams asking, um, are is your pricing increasing as you see these compute constraints? No. So we don't actually charge our customers for the compute we deploy inside of their private deployment. So it's their own compute, uh, what we see with our customers purchasing their own compute. Yes, the the GPUs are becoming more expensive. There is a scarcity. Um, you know, it's, it's something that we try to help our customers with, but our model of sales is not to provide the compute with the model.

22:38 Instead we just provide the software. So are you in the server building business as well? Is that sort of what this looks like in the future? Yeah. No. Uh, no infrastructure from us. We just focus purely on the models, the platform that the models power, so that integrate with all the different tools and data that enterprises use to automate things, right? To automate workflows, uh, for big banks, you know, inside capital markets divisions, wealth management, investment banks, um, we automate work and augment employees inside those organizations and we don't build the actual data centers themselves.

23:17 Or help source it. They do that and you come in and implement the tools, AI tools. Got it. Um, no. Aiden. Um, part of the reason I want to talk to you so much as well is, of course, you coauthored that famous paper that started essentially the modern AI age. This is a really general question, watching where it's gone and, you know, a small number of years, you've got infighting at the top. Musk versus Altman's a big focus in San Francisco this week.

23:43 Um, you see growing backlash in the public. What do you think? Did you expect to be here? Do you think that the industry has mishandled the public image side of AI? I certainly, uh, I mean, it's a completely general technology. It's like computing, right? Like it can be deployed into any sector. It can be deployed against any use case. It's the most general piece of technology humanity has created. And so it's going to have very sweeping impacts and, and touch everyone's lives both professionally and as consumers. So of course, it was very important that that technology be developed in a way and communicated to the public in a way that was accessible, that they could try for themselves, that they could test and find the faults and give feedback in terms of how the sentiment has evolved throughout that process over the past three years.

24:44 I, I am concerned, I do think it's, um, you know, there are certain folks who feel skeptical of the technology, whether it's AI slop. And in the creative industry, there's a lot of resistance to this technology. I think we need to better compensate artists and ensure that they can contribute to the development of the next generation of models, um, or whether it's the energy crisis, right? Like to power all of this adoption. We need a ton of energy. And if that is making things more expensive for people. I think that's a hugely regrettable outcome.

25:21 So investing in energy infrastructure, ensuring that prices don't go up for consumers as a result is a is a massive priority. Um, I think people are rightly skeptical. I think criticism is a good thing. Being aware of the weaknesses, being aware of the shortcomings about how the technology is being rolled out and the potential consequences is net positive in the long run. It's a it's a tricky thing that both sides are figuring out both the public and the companies building the technology.

25:52 We both want it to go well. That's a really thoughtful answer. Um, do you think we have the right spokespeople thinking of Dario and Anthropic and Sam Altman at OpenAI, Elon Musk. Um, what's what's needed here? It is kind of like a narrative, an image problem where you did mention really real things, but also, you know, there is does feel like this narrative problem at the top. Yeah. You know, uh, since you're Canadian, I can say it. Uh, I try to bring a Canadian touch to to things.

26:23 I think we need to be empathetic, kind, uh, you know, thoughtful in this and not just, uh, steamroll people. Um, that's that's fair. I think it's. Yeah, it's really great that the conversation is happening. I think it's essential. Um, there should be criticism. Uh, we should respond to that criticism with action to try to mitigate the potential downsides for people. Well, we hope to hear a lot more from you, especially that sort of nuanced, thoughtful answer acknowledging both sides of this and the real criticisms. Uh, Aidan, thank you so much for taking the time. We really appreciate it. And from fellow Canadian to another, thanks again.

27:03 Thank you for having me.