Apple-Broadcom AI Chip Deal, IBM’s Mainframe Push and China’s CUDA Challenge

Hey everyone. Happy Tuesday. Welcome to Techstrong Gang.

I got a confession. I thought my director, our engineers, designed the Techstrong Gang logo in blue and orange to celebrate my Knicks, but now it's been a couple weeks after the championship, and I think it's not gotten changed. So maybe it wasn't for the Knicks after all.

I don't know. But I still appreciate the blue and orange there. Welcome to a great gang here on this Tuesday.

Let me introduce you to our gang members, and we'll jump into things. He's up in Canada where they're celebrating their two weeks of summer, and not sure what he's wearing beyond that shirt, but we've speculated. Our own Chris Blask.

Joining us also is Alex Porter and our friend Sig Nag. Alex, Sig, welcome. And I guess up in New Hampshire, it maybe is not quite as hot, or it is, because you guys are all getting hot up there.

Kate Scarcella. Hi, Kate. And then the man in the blue light on the-- He was the man in the high mountain, but he now is the man in the blue light still in some- Nothing I seem to do about this matters ...

Adirondack getaway or something, it looks to me like, but he won't tell us. Mike Bizard. I'm in hiding.

I'm in the- You're in hiding. You're on the lam. I'm in the AI protection list program, and here I am.

Exactly. AI will find you. But gang, welcome.

We've got a great show to talk about. So Mike, I wanted to jump right into it and talk about, I don't know, this sounds like an oxymoron. Apple and servers?

Apple and AI servers? I know. I didn't know they had servers.

It struck me as odd too, but when you think about it, it makes sense. So apparently there's a story out talking about how Apple has extended its relationship with Broadcom through 2031 for silicon, some of which will be used to create what are being described as servers. And these servers will be used to talk to your mobile computing device, which will be running some local AI inference capability, but there'll be some need for caching.

Kate, I'm going to toss this to you to start with, but it kind of also feels like what's old is new again, in the sense that we are using distributed computing to deliver AI capabilities. But does this mean we're going to have servers everywhere, and will the Wintel community do the same thing? What's going to happen?

Wintel. So I think so, and as you say, the old is new. I think it's one of the reasons why, even though I know later in the block, we're going to be talking about IBM, but man, I can't step away from distributed regardless of how old and all the things that I have seen throughout my life.

So look, I love Apple, and that's not a surprise, as I've said many times on this show. So I love basics, I love precision instruments. I love the way that we're talking about having this special silicon.

So I know it goes a-- And I think it's important because the workload can dramatically improve what? Lower latency. And the whole entire lower cost, which is what we're talking about when we're talking about Edge AI.

So yes, so with servers being at the edge, I think we've seen it coming for a while, and we know that it needs to do more. So the silicon matters here, and we don't have the luxury anymore of having the data center being centralized, especially when we have all these wearables from phones. And think about even from a distributed point of view of having pacemakers.

All of us, we will be so connected and that we will become the edge. We will become the edge. Listen to this.

Yes. So Apple has always believed hardware and software should be designed together, and I still believe this. From a cybersecurity perspective, I find that custom silicon interesting because specialization can reduce unnecessary complexity.

So- Alex, I'm going to toss this to you. How far does this go? Do I eventually wind up with a very tasteful AI server in my living room next to the television and the stereo components, and that's how this is going to play out?

Absolutely. I fully believe that we will all have our own little AI micro boxes. I know that we're already actually doing that in our house for work and for play.

And where our company sits, actually, we have a very specific view on this because we're seeing, I consider it sort of a pendulum, right? We went hardcore cloud, and now we're going back to hardcore on-prem. And on-prem is going to change.

It's going to look a little different than just these monolithic server systems inside of data centers because we have a lot more flexibility with these little micro systems that we can install in our homes that are going to be more power efficient, to be honest, and they'll make those GPUs more accessible for all of our AI needs. Yeah. So I got a problem here.

When you tell me you extend a custom chip partnership, I take it that that means you already had a custom chip partnership, which I didn't know anything about. And I'm trying to think, what kind of chips were they making? Were they making memory chips?

Were they making the The A series or the M series chips that Apple designs in California, but I thought they were made in China, but not by Broadcom. But yet, reading the article, that Apple represents about 20%, give or take, of Broadcom's total revenue, which is huge. It makes them their biggest customer.

Apple actually announced a partnership with NVIDIA on June 9th. Yeah. No, that I knew.

That was around that. But no, this is an extension of an ongoing relationship. Yeah.

So what exactly has Broadcom been making? Is this some skunk works thing that Apple never talked about? Or what was Broadcom doing?

It's conceivable that Broadcom was giving them chips that Apple was using in its cloud service, because that's where there's networking and servers and all kinds of stuff that Apple builds out. It may not necessarily be in an Apple product per se, but they have a relationship probably for the services that Apple provides. They got to run on something.

Well, and the other thing, similar to Daniel Newman's quote in the article. This is an article by John, Mike, is that the story? Yeah, I think it's John, yeah.

John Schwartz. Daniel says this is really a lifeline for Broadcom, right? You got a 20% customer who represents 20% of your business extending five years, and I would think even growing that potential pot.

And Broadcom can use it, right? Broadcom's very overextended to OpenAI. Yeah.

I think it's important. I think the whole type of application systems, integrated technology is important, and I know that we're talking about Broadcom, and yes, this is really important for Broadcom, but I think, at the end of the day, it's a great partnership between the two because Apple hasn't been getting it right, right? We've all talked about how I think bringing on the new hardware CEO for Apple has been important, and I think that this is an important decision to move forward with Broadcom.

Yeah. But when it comes to chips, their M-series chips, prophetically, are more optimized for AI than just about any other endpoint- Yeah ... device chips out there.

They got chip design right, right? That's true. First the H series and now the M series.

But they haven't utilized it, right? Because if they had utilized it properly, when you're going through tunnels and things like this, I think what's going to happen with distributed is that you're going to be able to go into a connectionless type of environment and still everything work really well. Yeah, but I'll tell you something.

When OpenCL was all the rage, and it still is, whether it's NanoCL, MiniCL, that CL or this CL, if you're going to host your own CL, not in the cloud, the preferred footprint is an Apple Mini running M4, M5 chips. Yeah. They already had that.

To Alex's point about having an AI server at the house, those Mac Minis and Mac Studios were the choice. Now, is this going to be something that cannibalizes that? Something different?

I think it has to become more wearable. I mean, frankly. We are becoming the edge.

We are. With all the digital devices from our rings to watches to even glasses. Everything that we have is going to have to communicate as a data center.

We will become the data center. Chris, talk to me. Yeah.

I'm entirely with Kate and Alex on where this is going, right? Put this in context. You remember the '70s, my dad was working at IBM, and there's this whole thing that everyone is going to have a personal computer at home, right?

And that took a decade or two to settle in. Things move faster these days. Yeah.

And you look at the representative on the screen, you have a good intelligentsia, slice of the intelligentsia, of this technology space right now, and we have a hard time getting our heads around it because no, you're not going to have a big AI server next to your stereo in your house. But you're going to have more AI compute in your house total than you can possibly imagine, and it'll be distributed across, Kate, you just went there, with wearables and your phone and your clock. Now I know why the refrigerator had to be connected.

I had forgotten. Right. Your refrigerator, absolutely.

Well, the refrigerator is a perfect case, right? Because conceptually, back to the '70s, people saying smart refrigerator, oh, that'll help me with my shopping. Turns out it's more complicated.

But if you have a smart AI refrigerator that knows you backwards and is connected to an AI mesh in your house long before it gets to the cloud, internet, quad, even Apple world, that actually works. Better, faster, cheaper, save you money on food, and you'll eat better. Yeah.

I'm already envisioning having this conversation with my youngest, and I'm going to come home and I'm going to say, "Look at this awesome AI server we got," and he's going to turn his nose up and say, "I want my own. I don't want to share with you. " Well, I think you guys are over-indexing on the fact that Apple's only going after the consumer market.

They could be going after the enterprise market where you got- So that, to me, is almost inconceivable. Right? And I agree with you, Sidd.

I can't believe that Apple would- Aren't there Apple Minis within the enterprise? People don't use Apple computers within their enterprise? No, but they have said over and over for as long as- I can remember- Two jobs.

Back to the two jobs Well, the world is changing ... but that's not their market. And it's the funny thing, right?

" Oh, take the MacBook into the office with you. Right? " You've got your watch with you, your iPhone.

Right. " But for them to publicly say, "It's our market now," it would be huge. Pigs would be flying.

Yeah. " Yeah, exactly. Well, the other thing is the expansion beyond the Nvidia ecosystem, right?

Everyone's so tethered to the Nvidia GPU as a AI chip for training. Right. But now you've got options, right?

Now you've got options- Well, no, the fact is Broadcom's probably the biggest inference chip maker. There you go. That is their market.

Yeah, exactly. So if the future is going to be about smaller models running in enterprise environments rather than hyperscalers where training is being done, then you need that server at the branch office or the enterprise data center which can do inferencing on a smaller model. And you don't need a, I keep saying this, I sound like a broken record, you don't need a H100 at the cost of a Honda Civic to run that environment, right?

You need a smaller chip to do that inferencing at that edge, and I think that's what this market is about. Options. Mm-hmm.

I agree on that, and then the option side of things, the other piece I think is differentiation and who you're partnering with so that if something goes awry with sourcing, raw materials, and chip creation, you still have other options. And I think you can't put all of your eggs in one basket, and we've seen that over the last five years, many times with different tariffs, different issues that have created sourcing issues for just raw materials. And then ultimately a lot of the focus on the reshoring effort, right?

We know that has changed the entire chip market pretty massively. So I think that that is also a contributing factor for why the extension and why this is meaningful. Exactly.

Another story on CNBC just a couple of days ago, where Nvidia is pushing out their Kyber system- A year ... in 2028 because of supply chain issues, right? So yesterday they denied that whole thing, so then Nvidia said that that's not the case.

Well, they denied it here. Okay. There you go.

But anyway, I want to ask Alex this. So let's say I have this server and it's in my house or in my office, and then I want to go travel to Vegas. Am I going to network back to my server, or is there some way for my server to come with me and be cached in some server somewhere that will give me local latency access to my AI server?

And is the AI server in a box, or does it kind of travel with me? I will say that we have some infrastructure that is in Buffalo, New York, and we are based in Austin, and most of our work takes place on those servers in Buffalo. So, it's definitely possible.

How that makes it to a consumer market in a really attainable, simple way, there's some tooling missing in between, I think, to make it really accessible. I'm looking at you, Alan. I thought you said it was going to be wearable.

Yeah. If it's wearable, that's the other piece, is how can you network all of your devices together? And then the question is really about what do you want to allow into your network?

And so I think being able to go to Vegas and you have your server at home, what networks do you want to get onto, right? And what information is accessible or secure? And I think that's a piece that hasn't really been thought about necessarily in the long run on this.

Mike, do you take your Apple TV hockey puck when you travel? No. Well, I got my work- Sometimes I do, I got to be honest.

Yeah, sometimes I do, too. If I'm going to an Airbnb and I'm going to be a week, I take it. Yeah.

Yeah. Yeah. So ultimately what you're saying is that the big gold chain on my neck is going to be the server, and then all these other little things are connected to that on my fingers and everywhere else.

Is that how that works? So that's why you've been wearing the big gold chain. I thought it was just a style statement on your part.

That's right. You know the thing, you can take the boy out of the Bronx, but you can't take the Bronx out of the boy. I'm just saying, man, that it's going to be a fashion issue, I'm sure.

Hey, but you know what? We're over time for this one, guys. We got to jump to the next one here.

You know what? And I love this because in spite of everything and all the AI and wearables and all this stuff, we're still talking mainframes, Mike. What's this we got?

So today, IBM added a couple of additions to its mainframe portfolio, which is the z17 that they launched last fall, I think. And if you've been following IBM, and IBM quibbles about this, but basically the mainframes that matter are the odd number ones. So, it's 13, 15, 17 are the ones that they sell.

The other ones kind of fill in a gap. So every three years or so, there's a pop on mainframe sales. Now, what's happening here that's different, though, is IBM's making these things easier to drop into existing data centers, and there's multiple ways of doing that.

So Sid, my question to you is, are these just more boxes that the existing IBM mainframe customers are going to buy, or is IBM starting to expand the footprint of mainframe customers, potentially, because the cost of these things is dropping and they're easier to- ... kick into my existing data center environment? Am I looking at a renaissance of the mainframe here?

Absolutely. I think IBM's done a phenomenal job of repositioning the mainframe, which used to be a stodgy little box running somewhere in the back closet of a data center, and bringing it to the forefront and making it a true AI machine, right? They're not only going after the existing client base, but they're also going after net new clients with this strategy, where we know that AI is not going to be contained to one state or the other, like a hyperscaler data center with massive GPU clusters, but it's going to be, as Kate and Chris and Alex, all three of you said, it's going to be highly pervasive and distributed, right?

So, and I think IBM has done a great job in positioning the z17 and the new LinuxONE platforms, both from a perspective of doing a mainframe inference, but more so about the hybrid story, where these things are not going to be like the Apple story. They're not going to be standalone systems, but they will be hybrid, connected to cloud estates, connected to other edge environments. And they're even talking about revamping a new version of COBOL to run on these mainframe devices, right?

A system. So this is phenomenal. I think this is going to be the new world of the AI architecture, where IBM's going to have a big role to play.

I know. Kate, you used to work at IBM. Are you getting a little teary-eyed here?

Is this all coming back? Well, you know what's funny is I feel like I grew up in data centers with IBM, and I started doing break-fix with the IBM's mainframe Z series. So everybody every year predicts the death of the mainframe, and IBM always seems to resurrect itself.

So I agree with what Sid is saying. IBM always seems to be around, but we have to remember, IBM runs critical infrastructure, and it has forever, and so needing a new application, or COBOL, is not surprising considering how many of these applications run on it. And it's also becoming more affordable, right?

It was like in the millions, and now it's getting down to the hundreds of thousands for these things, for a true type of mainframe data center thing. So it's going to be interesting to see how things end up. I really, really, really, really, really believe in distributed.

So where does mainframe ... I think mainframes will go even closer to the edge, and that's what it's talking about in the article. It was talking about an on-prem.

Mm-hmm. More on-prem, and I think more on-prem. Who knows?

Maybe we'll all have a mainframe in our home. So Alex, let me throw this at you, because this is the next logical thing to what Kate was talking about. So IBM is also talking about building a hybrid platform, and they're using ARM chips that will eventually go into this, and they'll sit alongside the mainframe, and you can run all these workloads on the same machine is the thing.

But to Kate's point, is the mainframe just another node now in a distributed computing environment? It's just like one more thing. This idea that there's a mainframe and other, maybe it's just all one thing.

Absolutely. Hyper-networked, hybrid environments that utilize all available compute effectively. That's really sort of the next wave, right?

It's about the optimization and the effectiveness of the existing infrastructure, but it's also, this clearly is about modernization, right? Not just the mainframe, but also COBOL, right? And Kate is 100% correct, and this is very clear to me, right?

IBM owns enterprise, on-prem, secure markets, right? We're talking finance, we're talking defense, we're talking a ton of other things of that nature, and all of those things are ripe for an update. So, I'm seeing an absolute clear path to this next wave that actually can support AI more effectively, which I think has been obviously one of the big stopping points or barriers within the existing systems.

It just was never built for AI. So we really have to push updates on the hardware side, updates on the software side, and see what the next wave of modernization looks like to support it more effectively. And I know Chris, I'm like Chris is emphatically excited to add to that.

Oh, probably, right? And Mike, I have to note, this is one of these moments in time, the thin edge of the wedge is in. Mike is starting to use node and mesh taxonomy because yeah, a failure state of revolution is revolutionary changes.

Like everything's going to change. It's like, no, Unisys is still selling mainframes. Never stopped.

Univac, Burroughs is still there, right? Big, huge nodes are necessary, right? I got to spend a lot of the last decade in that Unisys space, and Citibank and the global economy relies on that stuff.

Does this get replaced by everyone running apps on our phones? No. I am on the record as not being a fan of the centralization powers in this.

We have Apple in the last block. We have IBM in this one. That's on the plate.

Maybe three to five companies control the entire information space going forward. I don't think so. But they don't go away.

We need mainframes. Someone has to have that big node. It just doesn't get to be the center of the universe.

Agreed. Yeah. Look, I think first of all, to Sid's point, and Alex, Chris, Kate, you've got to acknowledge the staying power of the Z systems, right?

Through it all, through the DevOps and agile movements, through cloud native. And cloud native, in many ways, was a diametrically opposed death star to mainframe, right? Let's run everything.

But you know what? Mainframes were running computers before Kubernetes was a twinkle in Google's eye, and they still run containers. And they were doing predictive analysis and pattern matching before Sam Altman was in diapers, and here they are still.

So you've got to give credit to just the whole continuing genius of the mainframe market, the mainframe powers that be. With that being said, it is increasing. So now it's got a place in the data center.

It doesn't need its own data center anymore, right? But it's increasingly being kind of pigeonholed into big banking, and part of it is price and all of that. Right.

You're not going to use it for the kind of things you use an AWS S3 bucket or something, you know what I mean, or what have you. But it's still nice to see it staying relevant, even if it's relevancy within its territory, if you will. Yeah.

Yeah. I think also the fact that IBM operates heavily in a regulated industry environment and a sovereign environment, I think Alex touched upon it, in industries such as banking, insurance, healthcare. This has a huge role to play because when you start talking about hardware enforced security, confidential computing, cyber resilience, the mainframes have massive, massive amounts of compute power to do those kind of tasks, right?

Those are not easy computing tasks. So that's why I think the mainframe has a huge role to play. Sorry, Katie, you were going to make a point.

Go ahead. No, the only thing I want us to be aware of, though, is that while, yes, what you're saying is true, what we need to be concerned about and understand is that as we have broken up the way we do business, think about Uber, think about Airbnb. I never would've thought that the New York City cabs could be broken up with Uber and Lyft and these other things, meaning that there has been disruptions in business.

And I think the disruptions in business will, like people with mainframes, we have to think about this, and we have to look at it because just because they run these critical infrastructure applications, the way applications are being run are different. And they're going to be more different as things become more distributed. So it's interesting, since we're quoting movies, "Jurassic Park," life finds a way, so do mainframes.

So I think that they'll always be sort of relevant, but I just think that we have change, a lot of change. Yeah. And yet- But the continuity- Go ahead.

The continuity, Alan, all three of you have touched on this. What I see in mainframes is Grace Hopper, right? When I worked at Unisys, I looked around, I couldn't quite verify it, but I guarantee there are lines of code that Grace Hopper wrote that are still working, still running.

And when we talk critical infrastructure, to me, as an OT geek- Right ... that means power, water, real critical infrastructure, and you see continuity over decades. In IT, continuity over decades is mainframes.

Why? Because that's the kind of thinking that's required for real critical infrastructure. True.

I think what's interesting is how far we haven't come, right? So if I go look up my bank balance on my iPhone or whatever it is, it's kicking off a mainframe transaction to this day. Yep.

So we have not come very far when you think about it. No. No, you're right.

But I do think that a lot of change is on the horizon. Because of the way AI code is even being written, there's a lot of change coming and a big disruption. Well, let me say something.

Go ahead. No, Syed. Go ahead.

No, no, Syed, you go. I was going to make one quick point. I don't know, this is my hypothesis.

I'd love to hear your guys' thoughts on this. The only challenge I think that one needs to think about is IBM mainframes still run what I call legacy applications. These are hardened applications.

Those are much more deterministic, right? Whereas in the world of agentic, the workloads are much more probabilistic. It's unclear to me whether the mainframes are architected to handle probabilistic environments like agentic.

That's the one piece I don't quite- So they're not- ... have done enough research on. Yeah.

So they're not there yet, but they have been playing with this AI chip that they embedded into the latest generation of the mainframes. And so I think that their plan is to offload more of that agentic workflow onto whatever the next generation of that AI chip. I think the first one was pretty much machine learning optimized, but they'll have one that's optimized for agentic, and watch this space is what I think.

I will. So let me make my point here. I live in South Florida.

A more plastic place, maybe Las Vegas is the only thing that comes to mind in terms of plasticity. And here in Florida, we got this thing, they tear everything down every 20 years and rebuild it just because they can, right? Instead of cleaning it, they just rebuild it.

It's almost ephemeral, right? I thought you were praying for hurricanes. I mean...

Well, no, I'm not praying for hurricanes. I never pray for hurricanes, but it's one of the things I dislike about Exactly. Mike, to your point about when you check your bank balance, you know why it goes to the mainframe to get that information?

Because it's still the cheapest, best, most stable, most secure way of doing it. Why fix what's not broke? Why tear down something that works great just for the sake of saying, "I want something new.

I want this shiny new technology, and it doesn't run on this thing that works great"? " Well, I think because of speed, might be one reason. So now mainframes are slow?

Is that it? Well, no, but if we are going to distribute it at the end of the day, and everything is distributed, it's not that mainframes are necessarily slow, it's- So we should go to distributed, and it doesn't work as well as what I already got, but we're going to distribute it, so we're going to- No ... move off.

Thank you, Alan. Go ahead, Chris. Yeah, let me try this tack.

I think in the end, the mainframes are going to take over the world. That doesn't mean the mainframes will be everywhere, but the kind of approach you have to have to build and run mainframes or a power grid. We have not applied that to the internet or IT at all.

It's this disposable, ephemeral- Yeah ... crap that we replace every couple of years. That's not mainframe thinking, that's not OT thinking, that's not infrastructure thinking.

And I think the kind of lessons that are required to build infrastructure work really, really well with AI at the small scale, at the huge distributed scale. So we've got a macro versus general relativity kind of argument going on here. Well, I think the argument is distributed- Quantum, I mean.

I'm sorry, yeah ... distributed computing doesn't have one form factor. That's the argument, right?

It can be everything. Yeah. If I want to run a COBOL application, rendering my bank information on a mainframe, and I want to run AI training on a hyperscaler server, I should be able to do both.

Doesn't matter, right? Right. And if you're starting out building something, and you're looking at an x86 server for, I don't know, say 10 grand, versus the starting point for an IBM mainframe, I think they quoted $167,000.

And this is the only list price on the entire mainframe portfolio, because everything else is pay for it as you negotiate for it. People are going to say, "That's a lot up front. " Everything else has to get paid on top of it.

Okay, let's run it on some Nvidia racks. Mm-hmm. That got pushed out a year, but now Nvidia get pushed out a year.

They ain't cheap either. But we're over time on this segment. We've got to jump.

This should be a good conversation, Mike. China shows its AI chops. First, I wasn't sure.

Isn't chops something they call when you do a signature in China? Maybe, I don't know. Like with a stone or something.

But anyway, I've got thoughts on this, but I'll let you kick it. But here in the US, chops refers to acumen. So- Mm-hmm ...

China has been making some significant advances, and now they're basically saying, "Hey, we're not sure we need you guys at all, because we're going to reverse engineer Cuda. We're going to use our own AI models. We're going to have caching capabilities and skillsets, and we even have our own AI coding tools now.

" And then it got weirder today because now the Chinese government, and these reports are out there saying that, "You know what? " Sounds familiar, doesn't it? And I think this is all kind of interesting.

" And if they happen to come from China, so be it. " So we'll see how that all turns out. But Alan, you have some thoughts you want to throw in here by all means.

Sure. So I think I mentioned I've been working on this book I'm writing, and there's a whole chapter dedicated to this. It's called "The Two Loops," the China strategy of two loops.

There's two parts to their strategy. And like much of what China does, it's playing a long game, right? Not for the short term, but it's not a dumb strategy.

The two loops are this. Number one, they're combating what was perceived as a gap, a skills gap, a performance gap between the leading US models, OpenAI, Anthropic, et cetera, Gemini, versus the rest of the world. And how China attacked it is using the open source, right?

They're saying, "We're the more open people. We're going to let you have access to this to do as you want. " Back to my July 4th stuff.

It's free as in freedom, not just free as in beer. Though the US model is expensive as in champagne And not very free when the government could kibosh it at any moment. So, that was the one loop, right?

And this is kind of ironic when you think about it. China's playing the freedom card to the rest of the world versus the US model. The other loop was, Trump and Jensen Huang made a big deal of saying, "We're going to let China use the H," was it the H200 or the H20 chips?

" Well, you know what China said? That's what China said. " You know how many of those chips they've imported into China?

Zero. Mm-hmm. They said, "That's okay.

Keep your chip. " And necessity's the mother of invention. And what we did by doing this was force China or accelerate China's capacity, and to go build their own models.

And they've got this one, I can't even pronounce it, but I actually did put it in my book manuscript. What's it called? Huangfu's- Yunfeng Zu ...

Yunfeng Zu. I got a tough time with this one. Yunfeng Zu.

I love it. Sid, you've had a little Mandarin training, huh? Just a little.

But, the point is, they have now, whether you want to say they started it with stolen IP or not, I don't buy into all of that. They steal our IP. That's how they got this.

I think give them their due, they're doing work. But they're building their own foundations from the chip up, to the inference, to everything else. And this was to be expected.

This is part of this whole thing. " Right? Mm-hmm.

Everybody wants in on this, right? The French, Mistral, took the open source route and saw their revenue increase 20X. This is real.

Now, the question is how do we respond? Do we need a response at the US level or the West level, right? Well, I think we do because I think this is too important a global market to cede or let it go.

But this is a classic kind of race, Cold War era kind of thing, where there's one model here, another model there. And now Mike, as you're saying, the Chinese are saying, "Well, maybe we don't want that open source. It's too good.

We've got to keep control over it as a response to the US," because the US is so clamping down on it that that gives China more breathing room to still be perceived as the more stable, lenient, better deal than dealing with the US. We've got to decide how we do this. Go ahead, Sid.

Yeah, I think you're absolutely right. There's this whole hardware conversation and the competition that's going to be ongoing between China and the US, and this protectionism, and some for good reason, some for not so good reasons. But I think there's another angle to this conversation, which is beyond the hardware, beyond the chip war, and beyond the model wars, like DeepSeek versus OpenAI and Claude and whatever, Anthropic, et cetera.

It's about software, right? This is also, I mean, platforms like that Yunfeng Zu are also designed to essentially reduce the dependency on things like CUDA, right? So there's a whole software battle that's brewing.

It's the whole stack. It's the stack. Stack, right.

Make no mistake. So that's one piece. And then I don't know if you read the whole story, there's also a framework that Alibaba has come up, what's called SkillWeaver.

Yes. Which is again, a software challenge, intelligence, orchestration platform. And then thirdly, there's something called, I was reading ZAI, Z Code, coding assistant for software developer and the whole entire DevOps space.

So I think this is not just a hardware fight. It's also reducing dependency on software that's part of the ecosystem that the US chip builders have developed, which is giving them a moat, and China wants to destroy that moat, right? So there's a multifaceted battle that's going on here.

" And the premise, Sid, is something that you kicked on here. Whenever something appears to be going to be indispensable, phone system, electrification, the internet itself, right? The whole way the internet works, TCP/IP, fiber, all that.

Governments can't let that go. No one's going to give someone, in today's world especially, a monopoly without regulating it. It's going to be regulated.

No one's going to get a monopoly, and no nation state today wants to be at the mercy of another nation state for something that is truly indispensable. And so, having a single source is kind of a cardinal sin in the indispensability trap. It's just, they can't let you have it.

And, that's what we're seeing play out here. China is going to give the world a choice A choice Is there some irony in this in that a technology that was funded by the US taxpayer and then created using data grabbed from all over the web, regardless of who created it, is now being moved over to China under a more open model? " It is what it is.

Maybe the president could call the AI foundation and see if he can get that changed. Well, this is a good follow-up- It did the US team a lot of good Yeah. It might be easier to rescind a red card.

Right. Exactly. I'm just saying that every time Trump interferes with some sort of major sporting event, it doesn't go well.

I wouldn't let his face look at those dice, as they say in the Bronx tale. But- Hard to progression dice ... let's go on.

Chris, you got something to say. Go ahead. Chris, yeah.

Yeah. Annalyn, you started with this one, right? And it's a bit of a bizarro world, right?

Because we have these two big superpowers, and unless you are one of them, you need to take them both with a grain of salt. But they're doing the opposite things. If we look at the B block, in the US right now, the mainframe lesson seems to be at the nation-state level that we need to, Washington, and specifically Pennsylvania Avenue, needs to own the mainframe that everything else is a slave to.

And the Chinese, of all nation-states, are doing the open model. We need to have an ecosystem around mainframes so the mainframes can exist. And I'm not a big fan of either, to centralization control things, but it is ironic that of those two superpowers, mine is arguing for what used to be the other team, and the other team seems to be arguing our argument.

Well, because- Okay, fair ... you're 100% right, but because this is the, and I hate to quote this word, but this is the yin and yang of it. " Sun Tzu.

Sun Tzu. It's a jujitsu move, right? To turn the opponent's energy around.

And so the open source to counter the US thing was, it was the thing to do, right? That you can't fault them for doing it. Can't we just open an AI institute in Switzerland and call it even?

Come on. Yeah, no, make it right here at the UN. In the UN.

I think one of the- I think the UN would go for it. Go ahead. I'm sorry, Kate.

Right. No, I know you're good. I think one of the things to look at, too, though, is the way that the markets work in the US versus the markets in China, right?

The government has a lot more control in China, so they're profiting a lot more off of any data that is acquired, any profits that are made, versus private companies that are being held and making returns. Well, it's not just profit. I'll tell you, Alex, in part of doing the research for the book, what they're training their AIs on is not, in spite of what we say here, stolen from Anthropic, they are actually using all of the data that all their factories are generating on production.

Everything that gets produced there, right, is kept in database and so forth. And the government has given the okay for the Chinese AI models to be trained on all of that industrial data. Absolutely.

Which 90% of which here in the US is probably behind a firewall. Absolutely. They're going to have much richer datasets because they have a much more- Yeah, and that's an issue.

Oh, that's a huge issue for the US- Yeah ... in general, whether that's on the sort of defense side or the private market side. That's a big challenge.

And the other thing is, right, what they seem to be doing right, which is exactly what you just mentioned, is that orchestration, translation, data provenance piece, they're just putting it all into one big bucket, essentially, because they've, from what we can tell, sort of removed the walls, and they're aggregating all that data into a more sort of centralized thing, which is, as Chris mentioned, not the best idea. However, it's going to be very powerful for them while we're over here working in silos and not sharing this data-rich information to protect profits. And ultimately, it is a really big issue, and that's the piece that I see that's really missing in the US markets right now.

Everyone's focused on the application layer, right? And they're not necessarily looking way down in the stack, right? You have to look at that orchestration, that translation, the data provenance.

Where is it coming from? Can we validate it? Is there a trail?

Is it auditable? How can we actually functionally do that? That's a huge area that we focus on at our company, and it's an area that I think the conversation is not turning to enough.

And this goes directly into what's happening with the SkillWeaver system that they have. " The right tool for the right job is something I say frequently. And I'm like, "Oh, cool.

" They're doing that at scale, which is interesting, but your data is completely subject to whatever they want to do with it. So there's no data sovereignty, there's no protection, there's no security, and unfortunately, I think that there will be people globally that don't think about that piece of it just to get the output and the function of what they want to use it for right now. Well, it's a bargain they make.

Sid, go ahead. I think this whole conversation about protectionism is a hopeless exercise. TCP/IP, you talked about TCP/IP, established the fact that trying to protect connections across nations, across borders, is a hopeless exercise.

It's useless. I agree. And I think AI frontier models are going to do that to data, right?

Trying to protect data across borders, across models, is a hopeless exercise. So, might as well democratize the whole thing and live with it. Sir, this is my last statement on this.

So I'm predicting that there will be two open source AI models in the US. One will be named after Madami, and the other one will be named after Rand Paul, and they'll both be the same, but they'll have different names, and everything will be fine. All right.

Wait. Let's not end on politics. But hey, what a great way to end it.

You know what? I have another ending. Are we going to build AI for the benefit of a few small men, billionaires?

Are we going to build it for the benefit of all mankind? And that's a good way to end this one because it's all about AI, it seems. But we're over time here.

We went way over. Alex, Sid, Chris, Kate, thank you. Mike, as always, thank you for leading the charge.

Thank you for watching. Of course, you can watch this on Techstrong TV, on Techstrong TV OTT app, on any screen you like, or our Techstrong TV YouTube channel. For now, though, this is Alan Shum.

We're out.