Microsoft’s Homegrown AI Models, Trump’s AI Executive Order, OpenAI to Merge Codex & ChatGPT

0:13Stephanie Palazzolo:Welcome everyone to the information's TI TV. My name is Akash Pasricha. It is Wednesday, June 3rd. First up on the show today, a new wave of announcements out of Microsoft's Build Conference will break down all the new models, agents and quantum chips. and more with our Microsoft reporter who is at the event. We'll then get to Marvell's skyrocketing shares after Jensen Huang's glowing endorsement. And we'll also bring on our politics reporter to unpack the White House's AI executive order. We'll then unpack OpenAI's decision to combine codecs with ChatGPT and how customers are most likely to react to it.

0:53Stephanie Palazzolo:And finally, we will end the show with our meta reporter who broke the news that the company is limiting some of its employee tracking tools. It's going to be a fun show, so let's get right on into it. Microsoft unveiled a staggering number of new products at its Build Conference on Tuesday. There were models, chips, agents, and more. To unpack all of it, I want to bring on Aaron Holmes, our Microsoft reporter who is at the event. Aaron, welcome back to the show. It's great to have you here. Where do you want to start? I mean, what was the biggest release for us out of all of them? You know, I think the most significant thing that Microsoft released was these seven new models that were all, you know, trained from scratch by Microsoft.

1:35And specifically, they showed off a new coding model, which is going to go in their GitHub Copilot AI tool, as well as a reasoning model. And I think what was interesting about these is that Microsoft said that, you know, once you are able to sort of customize these models to your company's needs, they can actually be, you know, more useful in some cases than leading models from Anthropic or OpenAI at a fraction of the cost. Obviously, we need to wait and see whether customers agree with that evaluation. But if it's true, that's a pretty compelling pitch that Microsoft is making to developer customers.

2:11And then, you know, on top of that, there were a slew of other announcements ranging from new kind of open claw features within Windows and Office and new devices. And sort of across the board, you know, Microsoft making this push to essentially be a place where companies can use AI in the workplace. And according to Microsoft, in a way that is cheaper and also potentially more trustworthy than AI from Foundation Model Labs.

2:40Stephanie Palazzolo:So Microsoft is saying that these models will be cheaper than what's out there in some cases. With the agents that they unveiled, are those included in the bundles that they already have? Is it something they have to pay extra for? Yeah, so on agents, it's still a little bit unclear how they're going to be priced. But there were a couple of announcements. So one is Microsoft said that they're going to release this product called Scout, which is essentially OpenClaw, which is this open source agent tool with a sort of wrapper around it that makes it more trustworthy. And according to Microsoft, we'll prevent it from going rogue with your company data.

3:24And as I've reported, that's something they've been working on for several months. They promoted this corporate vice president, Omar Shaheen, to specifically build OpenClaw tools for Office. and this was kind of the first real look that they've given at what that will look like. And so, you know, the idea is that Scout can function like this autonomous open claw style agent that is just basically monitoring your Outlook inbox all the time or, you know, your team's messages and will suggest ways to reschedule meetings that conflict with other things or, you know, will send you reminders if you ask it to when you're about to go into a meeting with what you need to know.

4:04So it's also interesting, you know, whether companies have a strong need for always on agents like this, I think is a little bit untested. But Microsoft is definitely trying to capitalize on some of the developer hype that we've seen around agents like that.

4:22Stephanie Palazzolo:On the always on piece, did you get any sense from Microsoft or from the presentation how they might plan to control the costs that customers might incur with this? And I ask this because this is the risk with these types of agents is having a budget for token use and then basically running through all your tokens without even knowing it because an agent is on in the background. Did they address that at all in the presentation? You know, that didn't get too much airtime, although they did kind of point to these tools that they've already previewed called Agents 365, which is essentially software that's meant for companies, IT departments to keep track of how employees are using agents and also potentially put controls onto how those agents are being used.

5:15And, you know, as we wrote earlier this week, a lot of companies are starting to essentially throttle how employees can use AI agents to make sure they aren't chewing up too many tokens and running up the budget. And Microsoft didn't really get into this explicitly because I think they would love, you know, companies to spend heavily on it.

5:33Stephanie Palazzolo:I mean, what about the customers on the ground? I mean, what was the general reaction from the customers to all this? You know, I think developers who I spoke to were definitely curious to test out these new coding models and agent tools. I didn't hear too many specific use cases for the OpenClaw style agent, but I did hear, you know, a lot of folks immediately saying that, you know, these coding models, especially, and the reasoning models could be a big deal when used in their applications. and Microsoft is starting to make those available in a limited preview. So we'll have to see kind of how they play out directly in the enterprise.

6:10Stephanie Palazzolo:Okay, let's go to the handheld devices, the operating system for these devices. This was kind of interesting. It wasn't something that I think we were expecting, although I don't know, maybe you've heard this and maybe I just missed it. What was your reaction to that? Yeah, no, it was interesting. So Microsoft showed off these prototypes of potential reference devices that could use this new operating system that they've built, which is essentially meant to run on, you know, smart devices. So smart glasses, handheld badges, you know, little desktop touchscreen consoles. And the idea is Microsoft wants to basically sell this operating system to device manufacturers so that people can use those devices to talk to Copilot or to any other AI agent to remotely control their meetings or other pieces of software.

7:07And it's interesting because Microsoft actually showed a physical prototype of one of these devices. And at first, a lot of people thought that that meant Microsoft was going to start selling this device. It was a bad. But I've confirmed they don't have any plans to actually sell a device themselves. They're essentially hoping that OEMs, meaning manual manufacturers will step in and start making these devices and they can just sell the software layer

7:32Stephanie Palazzolo:that's like such a downer like so it's it's a bit of a tease right it's like you bring out this handheld device i mean you know they're in the business of xbox it's not like they don't know how to make devices i kind of wish that maybe that and and by the way then you know open ai would have really started panicking then because then you know they would have said are you serious you're doing those too, but okay, fine. I, you know, my understanding is that Microsoft is very aware that OpenAI is moving into this market as are lots of other AI companies. And I think that by essentially, you know, stepping into this race a little early and saying, we don't have devices, but we are building the software that would power such devices.

8:11They're starting to plant their flag in what they know is going to be a very crowded space in the coming years. Right.

8:18Stephanie Palazzolo:Last question for you, the desktop computers that were unveiled from Microsoft. It felt like this was an extension of the news we saw earlier this week. Microsoft and NVIDIA, they unveiled this chip together. And I mean, they described it as what, like a data center that is available on your desktop? Is that the idea? Yeah. So this is called the DevBox. It's part of Microsoft's Surface line, which is their in-house PC line, which up until now has mostly just been tablets and laptops. But this is a desktop box. It's roughly the same size as a Mac Mini and I think very much competing for the same types of customers.

9:00And I think the idea is this is going to be powered by the new NVIDIA processor. It can run pretty mid-size AI models locally on your own device. And the idea is that by spending to buy this device, we don't know how much it'll cost yet, but spending on this top-tier hardware, you can then save costs by running more AI locally instead of having to pay per token, which kind of factors into this broader pitch that Microsoft is making to developers, which is essentially trying to give them more flexibility to control the cost of AI rather than just relying on foundation models from OpenAI and Anthropik.

9:40Stephanie Palazzolo:Right. Well, Aaron, I want to thank you for coming on. That is Aaron Holmes, our Microsoft reporter at the Build Conference here on TI TV. Marvell shares closed yesterday, up more than 30%. They are continuing to ride up today after Jensen Huang dubbed it to be the next trillion dollar company. Meanwhile, there have been a flurry of other chip-related headlines over the past few days. Microsoft, at its Build Conference, also unveiled its latest quantum chip. We'll get to that in a minute. But Intel is also shipping a new inference-focused chip. And so to break down all of that news, I want to bring on John Vinn, semiconductor analyst at KeyBank Capital Markets.

10:17Stephanie Palazzolo:John, welcome back to the show. It's great to have you here. Thanks for having me. Okay, let's start with Marvell. So Jensen Huang has dubbed it the next trillion dollar company. Does John Vinn agree with that? I think when Jensen said that, I don't know if he really meant it literally. I think he meant it figuratively. We know he has a very high opinion of Marvell, right? They recently signed a strategic agreement in which NVIDIA invested$2 billion in Marvell, I think primarily to secure networking capacity over the next few years or so. So obviously, Marvell is very important and a strategic partner to NVIDIA.

11:03We think Marvell is a great company. We have an overrated rating on it. But in terms of the specific comment, I think it was more a figurative comment than a literal comment from Jen.

11:16Stephanie Palazzolo:So sounds like you don't think it can get to a trillion dollars. I think right now it would be kind of a stretch, right? It's, you know, market cap is about$250 billion right now. So for it to get to a trillion, it would be a fourfold increase. So that might be a little bit of a stretch. Now, remind us, Marvell, where does it fit in the chip ecosystem? It's squarely in the networking chips game, so competing with Broadcom, is that the idea? um you know it fits in a couple areas right um you know networking is really the area where it excels it acquired the optical interconnect business from infi it also acquired celestial ai which gives it access to next generation direct optical connect technology which is increasingly more and more important as you see networking proliferate through the data center And then on the other side, they also have a very competitive AI ASIC business where they compete with Broadcom on the custom ASIC side.

12:20And then they also obviously compete against some of the merchant suppliers, such as NVIDIA, on the merchant GPU side for AI and also AMD.

12:30Stephanie Palazzolo:So, I mean, where does Marvell have its own traction then? And where is it really fighting headwinds in this? I mean, look, I know any chips in this day and age is good, but I guess what I'm wondering is, what is the case to invest in Marvell over investing in a company like Broadcom? Yeah, I would say where it has traction today is AWS uses their AI ASIC services for Tranium. They also are going to be providing or building the next generation ASIC for Microsoft. You heard them announce and talk about the traction they're getting with MITO 200. Marvell is actually helping them build MITO 300, which is going to ramp next year.

13:20Stephanie Palazzolo:And which chip, Maya 200, 300, which chips for those of us who aren't as familiar with them? Yeah, Marvell's helping them with the next generation Maya 300. Maya 200 was discussed yesterday at the Build conference. Which is what? What type of chip is that? It's a custom AI ASIC. Got it. Got it. Okay. So, okay. Sounds like there's some traction there. Speaking of Microsoft, I do want to ask you, though. So, I mean, that was one announcement that they talked about yesterday. I want to ask you about the quantum chip that Microsoft announced yesterday, the second iteration of it. What stood out to you about the differences here with this quantum chip and how bullish are you on that?

14:05I think the quantum market in general is extremely early days. I think we're all trying to get our arms wrapped around, you know, what are going to be kind of the killer applications around it. Obviously, all the major compute providers and ecosystem partners are actively evaluating their quantum strategy. Certainly, Microsoft is obviously moving the ball forward there. But I think it's still pretty early days in terms of assessing what and where it fits going forward.

14:38Stephanie Palazzolo:So when you hear 2029 as their goal for a commercial quantum chip, do you think they'll get there by 2029? I think it's a possibility. I don't think it's out of the question. I think just the bigger question, as I mentioned, is just how widely adopted it's going to be in 2029. Right. And then I guess where my head is going is that IBM is also squarely in the quantum game. And I think they've set a similar timeline for 2029. They've invested something like$10 billion in this initiative. I guess I wonder how you think about the competition in quantum, if there is any right now for such an emerging field between IBM and Microsoft?

15:25Certainly, it's a very disruptive technology. I don't know if we have enough information and details to assess where they stand versus each other, but certainly it's something to keep an eye on. Right.

15:39Stephanie Palazzolo:Okay, let's move then to Intel. And I'm kind of chugging through all the chip headlines that we've seen over the last week or so because there have been a couple here. So Intel is launching this new inference-focused chip, Crescent Island. Can you explain to us what exactly is different about this chip and how it's planning to stand out? So Crescent Island is an AI GPU. um you know they have previously released gpu products before they haven't gotten very much traction with it so crescent island is is kind of that next generation chip i i think it was you know announced on the roadmap from intel a couple years ago um certainly you know my observations if you look at crescent island um in terms of intel strategy there's two things that kind of stand out to me.

16:31One is they're focused on the inferencing market. And the second thing is they've chosen to use LP, DDR, DRAM instead of HBM memory, right? And that makes sense to me, right? It's really hard to compete with NVIDIA on the training side. You know, NVIDIA just has a stranglehold there. And that's kind of where they're initially trying to target getting some traction in the market.

16:59Stephanie Palazzolo:And put that a little more simply for us. I mean, higher level, what is the... A lot of acronyms there. Just break down for us simply what the L, whatever you just said, Watson. Oh, sorry. Yeah. So most AI chips on the market use a type of memory called HVM, which stands for high bandwidth memory. It's very expensive. It's usually about three to four times more expensive than traditional commodity DRAM. The memory that Intel is using in Crescent Island is LPDDR, which stands for low power DRAM. It's a similar type of memory that is used in smartphones, such as your iPhone. So it is significantly less expensive.

17:57It consumes lower power. That's the main difference between the two types of memories that are being used here.

18:04Stephanie Palazzolo:And so why has Intel opted for that approach here? I think two things, right? I think, one, they're trying to focus on inference. So when you're focusing on inference, the need for HBM memory is lesser. And then cost-wise, right? It's obviously not going to be able to compete with NVIDIA on a performance basis. So by using low-power DRAM, it lowers the cost of the chips and makes it just more cost-competitive. And so in terms of who this is competing against, this is Intel's newest crack at an AI GPU. Where does this, I mean, do you think this is a viable challenger to NVIDIA? Not in the sense of taking significant share, but is it a product that you think people will take seriously and look to, you know, in this supply shortage of GPUs that we have?

19:10I'm somewhat skeptical of that. I think in this market, if you look at the major providers, on the merchant side, it's NVIDIA, then you have AMD. And then on the custom basic side, it's Broadcom and Marvell. I think the AI chip strategy for Intel is very much a work in progress. We very much like Intel as a stock, but for other reasons, such as Foundry and Advanced Packaging. But I think the AI strategy for them is a work in progress. I think their sweet spot right now in terms of the data center and AI in terms of where they have good traction is on the agentic AI workloads where there's more of a need for traditional server CPU architectures where they're kind of an entrenched incumbent there.

20:01Stephanie Palazzolo:Now, there was an interesting interview that the FT had with a top Intel executive talking about the playbook for this chip. And one of the things that they mentioned is that they are assessing whether or not they could sell this Crescent Island chip in China. And maybe that is an opportunity for them in an environment where NVIDIA and AMD have kind of been closed out of the China market. Does that feel like a viable future to you? Could you see that happening? does that give you any more confidence in intel here i think a lot of that depends on just the geopolitical relations between the u.s and china we did see recently that the u.s government had approved nvidia to sell its h200 gpus and did the china market right so the h200 gpus is a lower performance chip that's been cut down so that it can meet some of the requirements to sell into the China market.

21:08I would imagine that the Crescent Island GPU would be similar in performance to the H200. So assuming that geopolitical relations continue to normalize there, then yeah, I think that could be an opportunity for them. But I don't know if the China market over the longer term represents kind of a viable market for Intel or anybody else at this point.

21:33Stephanie Palazzolo:So last question for you, John, you know, I want to go back to what you were saying. You're slightly skeptical of this chip and how effective it'll be for Intel, certainly with respect to the AI GPU opportunity. What inevitably do you think happens here with this effort? I mean, you think they launched this chip, you know, does it kind of just putter along and then if it doesn't get traction, they discontinue it? Do you think this is ultimately going to be an area that Intel invests even more into over time? Play this out for us a year or two down the line. Yeah, I think it's part of the learning curve for them, right?

22:11I think they're going to take it to market. I think they're going to put it out there. I think they're going to try their best to get customers to adopt it. And then I think they're going to get some feedback on that. And then I think they're going to try to iterate. If you look at their strategy, Intel is approaching this AI market on a number of fronts. It's got its foundational core with server CPU, which as we talked about, they're entrenched there. It's an up-and-comer on AI ASICs and on merchant GPUs. They're also partnering with NVIDIA on the chiplet side. So they've announced a partnership agreement to integrate their x86 cores on a chiplet basis using NVLink to some of the NVIDIA GPUs.

23:01So I think if you're Intel, you're going to continue to integrate. I think you're going to try to move the ball forward and figure out what sticks. And then in the meantime, in the background, they're getting very good traction with advanced packaging, which is going to be a really core part of the AI chip ecosystem because increasingly we think more and more hyperscalers are going to use their advanced packaging, which is EMT, to integrate high bandwidth memory, which we talked about as an essential part of developing a custom AI ASIC chip for these hyperscalers.

23:39Stephanie Palazzolo:Great. John, I want to thank you for coming on. That is John Vane from eBank Capital Markets here on TITV. The White House has released an executive order on AI and cybersecurity guardrail. though it's not exactly the executive order many thought was going to be released last month. To break it all down, I want to bring on Leo Schwartz, our tech and politics reporter. Leo, welcome back to the show. It is great to have you here. What is in this executive order? Break it down for us. Yeah, it's good to be back. It feels like I was just here talking about the executive order. When it was delayed.

24:12We were so close that it didn't happen. It was delayed at the last minute. There was supposed to be this big, fancy signing ceremony at the White House with tech executives that didn't happen. And instead, two weeks later, we got basically the same order. Not much has changed. there's one small difference which is basically this controversial component of it where ai companies would have to or could share versions of their models with the government before releasing them to other partners and then the public it used to be up to 90 days before release which was untenable for a lot of companies now it's up to 30 days besides that the order is largely the same and makes clear to denunciate, including in an entire paragraph dedicated to this, that it is voluntary.

25:03It's not mandatory. It's not a licensing system. But at the same time, it does create for the first time this federal regulation that's compelling companies to be sharing their models with the government ahead of release.