# AI Chip Race Heats Up From Microsoft to Musk

An episode of [Bloomberg Tech](https://invisiblenetwork.ai/podcast/1611767434), 12 Oct 2026, 44 min.

Audio: https://podtrac.com/pts/redirect.mp3/traffic.omny.fm/d/clips/e73c998e-6e60-432f-8610-ae210140c5b1/41764a4f-fc64-4e11-89ba-ae7c0030ab5e/8421ffa3-0093-47ea-a045-b4e1010d6a10/audio.mp3?utm_source=Podcast&in_playlist=9caafc41-289c-4115-995d-ae7c0030ab75

Bloomberg’s Ed Ludlow breaks down Microsoft's latest homegrown Maia chips, its latest effort to rely less on outside chips to power its AI ambitions. Plus, Elon Musk wants to put chip development on fast-forward, testing as many as ten new designs a year; and Apple is finally ready to unveil its new smart home strategy, seeking to plug a major hole in its ecosystem. See omnystudio.com/listener for privacy information.

## Transcript

[0:02] Speaker 1: Bloomberg Audio Studios.
[0:04] Speaker 2: Podcasts.
[0:05] Speaker 1: Radio.
[0:06] Speaker 2: News.
[0:13] Speaker 3: Bloomberg Tech is live from the heart of Silicon Valley
[0:17] Speaker 3: with Ed Ludlow in San Francisco.
[0:22] Speaker 1: This is Bloomberg Tech. Coming up, Microsoft wants to rely
[0:25] Speaker 1: less on outside chips to power its AI ambitions and
[0:28] Speaker 1: rolls out its latest homegrown Maya chips across data centers. Plus,
[0:32] Speaker 1: Microsoft isn't the only tech giant rethinking the chip playbook.
[0:35] Speaker 1: Elon Musk wants to put chip development on fast forward,
[0:38] Speaker 1: testing as many as 10 new designs a year. And
[0:42] Speaker 1: Apple is finally ready to unveil its new smart home strategy,
[0:45] Speaker 1: seeking to plug a major hole. in its ecosystem. Happy
[0:50] Speaker 1: Monday and welcome to the week. I'm going to start
[0:52] Speaker 1: in the chip space. There is a report out of
[0:54] Speaker 1: China about the domestic memory name CXMT coming with a
[0:58] Speaker 1: new generation of chip, more data. That's putting pressure on
[1:02] Speaker 1: memory stocks here in the United States. micron case in
[1:05] Speaker 1: point the decline in it trading at its lowest level
[1:07] Speaker 1: in three weeks is not much to write home about
[1:09] Speaker 1: but the philadelphia semiconductor index or socks is down or
[1:13] Speaker 1: on track to be down for its fourth straight session
[1:15] Speaker 1: longest run of declines going back to july there's been
[1:18] Speaker 1: some pressure there while tech overall uh has been near
[1:21] Speaker 1: record highs and we'll get to that in a little
[1:23] Speaker 1: bit our top story today microsoft is rolling out its
[1:26] Speaker 1: latest homegrown ai chip the maya 300 at gigawatt scale
[1:30] Speaker 1: starting next year Well, part of the push to cut costs,
[1:34] Speaker 1: but rely less on NVIDIA GPUs. Bloomberg's Matt Day has
[1:37] Speaker 1: the story. There's lots of readership and interest in this.
[1:40] Speaker 1: The basics of the Maya 300 generation, Matt, and their
[1:43] Speaker 1: plan to deploy it at scale.
[1:46] Speaker 4: So this generation, it's a system on a chip. It's
[1:49] Speaker 4: designed for AI tasks. It's designed for AI inference, though.
[1:52] Speaker 4: We're not talking about training new models. It's about running
[1:54] Speaker 4: existing ones efficiently. Microsoft thinks they can plug enough of
[1:58] Speaker 4: these into their data centers to make a meaningful difference
[2:00] Speaker 4: on their costs, on performance. And ultimately, they say these
[2:02] Speaker 4: things are going to be backing most of their AI
[2:05] Speaker 4: workloads internally.
[2:07] Speaker 1: My 300 promises 40% more tokens per megawatt than what
[2:11] Speaker 1: Microsoft said was leading GPU systems. But because it uses HBM4,
[2:15] Speaker 1: I'm pretty sure they're testing against NVIDIA's benchmarking against Vera
[2:20] Speaker 1: Rubin as the kind of comparator. What's the history and
[2:23] Speaker 1: backstory on Microsoft's chip program, right? Like all of the
[2:26] Speaker 1: hyperscales have custom silicon programs. The different players put different
[2:31] Speaker 1: emphasis on how much this is about diversifying compute away
[2:34] Speaker 1: from NVIDIA. Some of them put a focus more on
[2:37] Speaker 1: the economics or tokenomics of throughput.
[2:41] Speaker 4: And that's exactly where Microsoft has landed. I know they
[2:44] Speaker 4: got their start a little bit after their big hyperscale competitors.
[2:47] Speaker 4: And Google got a start on their TPUs years earlier. Amazon,
[2:50] Speaker 4: with its Onapurna acquisition and its Terranium and Inferentia chip lines,
[2:54] Speaker 4: got going. So Microsoft was a bit of a latecomer
[2:56] Speaker 4: to this. But I think what people were looking for
[2:58] Speaker 4: in this third generation of chip is whether it's at
[3:00] Speaker 4: the scale where they're going to start maturing the thing
[3:02] Speaker 4: and getting it out there in force, or if they're
[3:04] Speaker 4: going to go a different direction. And this signal multi-gigawatt deployment,
[3:08] Speaker 4: that's a ton. They probably only got about 12 gigawatts
[3:10] Speaker 4: in their data center today. That's a real indicator that
[3:13] Speaker 4: they want to continue pushing for it, just because they
[3:16] Speaker 4: think they can build a better mousetrap than NVIDIA when
[3:18] Speaker 4: it comes to efficiency.
[3:20] Speaker 1: It is interesting. You know, the stock's up more than
[3:22] Speaker 1: a percentage point, but again, near fresh record highs itself.
[3:26] Speaker 1: I think we need to remind the audience, Matt, that
[3:28] Speaker 1: the main point is that Microsoft is at hyperscale. It
[3:31] Speaker 1: is a huge buyer of accelerators or GPUs. It is
[3:36] Speaker 1: working on its own chips. Try and sort of explain
[3:38] Speaker 1: that scale and what their commitment is to data centers
[3:41] Speaker 1: for AI.
[3:43] Speaker 4: So they're spending hundreds of billions of dollars, like their
[3:45] Speaker 4: big peers, Amazon and Google and Meta, building data centers
[3:49] Speaker 4: and then most expensively fitting them out with NVIDIA rigs
[3:52] Speaker 4: to make the AI inside tick. So one of the
[3:54] Speaker 4: surprising things about Microsoft's announcement this morning is that they
[3:57] Speaker 4: are not going to rent the computing power of their
[3:59] Speaker 4: chips to third parties. They think there is enough internal
[4:02] Speaker 4: demand for things like Copilot, for things like GitHub, that
[4:05] Speaker 4: they can just use the chips to satisfy some of
[4:07] Speaker 4: that work and then save those NVIDIA precious chips for customers.
[4:11] Speaker 1: That's interesting. Meta has a very similar strategy with MTIA.
[4:14] Speaker 1: It's custom silicon. It's about internal workloads as well. Again,
[4:17] Speaker 1: Microsoft up a percentage point very near record highs. Thank
[4:21] Speaker 1: you very much to Bloomberg's Matt Day. Some fresh warnings
[4:24] Speaker 1: about the flood of money into AI. Speaking at a
[4:27] Speaker 1: DBS conference in Singapore. Broadcom CEO Hock Tan said some
[4:31] Speaker 1: investors have no clue what they're betting on. And he
[4:34] Speaker 1: compared the frenzy to gamblers at a casino. And there
[4:37] Speaker 1: are signs the market's getting a little bit more cautious.
[4:39] Speaker 1: DBS CEO Tan Su Shan says investors are now demanding
[4:43] Speaker 1: higher returns but taking a harder look at the risks
[4:46] Speaker 1: behind AI financing. Here's what she told Bloomberg.
[4:51] Speaker 5: And it's not a bad thing because now people are
[4:53] Speaker 5: asking for higher returns and people are also looking at
[4:56] Speaker 5: structures and they're making sure that whatever money is now
[5:00] Speaker 5: being hoovered into the whole AI CapEx ecosystem will lead
[5:05] Speaker 5: to some kind of long-term ROI that can then be
[5:09] Speaker 5: repaid back to debt or bond or equity holders.
[5:13] Speaker 1: Okay, so we're starting the week, the NASDAQ 100 lower,
[5:16] Speaker 1: still near record highs after a fourth straight weekly gain
[5:21] Speaker 1: last week. And heading into earnings season, there's this interesting
[5:24] Speaker 1: setup in tech. So despite this rally, the sector now
[5:28] Speaker 1: trades at the biggest discount rate. to its own historical
[5:31] Speaker 1: valuations of any group in the S & P 500.
[5:34] Speaker 1: That lines up with research from Fidelity's Denise Chisholm, who
[5:37] Speaker 1: since the summer really has been writing that tech's gains
[5:41] Speaker 1: have been driven by rising earnings expectations, not rising valuations.
[5:44] Speaker 1: And I'm really delighted to say Denise is back with
[5:46] Speaker 1: us here on Bloomberg Tech. That seems to be like
[5:49] Speaker 1: the math right now and the calculus is earnings, valuations.
[5:54] Speaker 1: And you've been writing a lot about that since July, really.
[5:57] Speaker 1: Where are we at?
[5:58] Speaker 6: Yeah, it's fairly rare to see a 30% multiple contraction
[6:02] Speaker 6: over any year. When you look from 1990, the technology
[6:06] Speaker 6: sector has only experienced that less than 5% of the time.
[6:10] Speaker 6: And the interesting part, you highlighted, you know, the stock
[6:13] Speaker 6: market is a discounting mechanism and technology stocks have a
[6:16] Speaker 6: very clear relationship when you look through the data. The
[6:19] Speaker 6: bigger the multiple contraction, and now they're only trading at
[6:21] Speaker 6: a 10% premium to the S & P 500, which
[6:24] Speaker 6: is in the bottom quartile of their historical range since
[6:26] Speaker 6: the 60s. When you see that, it actually skews your
[6:30] Speaker 6: risk reward, meaning the bigger the multiple correction, the more
[6:33] Speaker 6: likely the stocks are to outperform.
[6:36] Speaker 1: That is even true.
[6:38] Speaker 6: If earnings.
[6:39] Speaker 1: Decline, right?
[6:40] Speaker 6: So if you look at any time you've seen multiple
[6:42] Speaker 6: corrections like this, it's actually most of the time is
[6:45] Speaker 6: on strong earnings. So the biggest question I get is that, yes,
[6:49] Speaker 6: but they should be cheap if we are seeing peak earnings.
[6:52] Speaker 6: And there's some truth to that. So all peak earnings
[6:54] Speaker 6: usually have multiple contractions. The question is, is our multiple
[6:58] Speaker 6: contraction reflective of peak earnings? When you look back historically, 77%
[7:02] Speaker 6: of the time, earnings do continue to grow after the
[7:05] Speaker 6: multiple correction we've seen. And in that instance, you get
[7:08] Speaker 6: very strong odds of outperformance, but you even get strong
[7:11] Speaker 6: odds of outperformance in the 70 percentile range, even if
[7:15] Speaker 6: earnings decline, which is kind of a long-winded quantitative way
[7:18] Speaker 6: of saying, look, whatever you're afraid of in terms of CapEx, ROI,
[7:22] Speaker 6: is there too much spend? Is the declining free cash
[7:24] Speaker 6: flow going to be persistent? A lot of that might
[7:26] Speaker 6: be priced into the stocks.
[7:29] Speaker 1: The long-winded quantitative answer is why I love to have
[7:31] Speaker 1: you on the show, Denise. Don't worry about that. I
[7:34] Speaker 1: want to bring back the table that I showed at
[7:36] Speaker 1: the start of our conversation because it was fresh on
[7:38] Speaker 1: the Bloomberg this morning. So this is the Z-score, right?
[7:41] Speaker 1: The table calculates the Z-score, forward price-to-earnings ratios for the
[7:44] Speaker 1: main S & P 500 sectors. And then it uses
[7:47] Speaker 1: weekly data with a five-year look back. And the main
[7:49] Speaker 1: point is that tech has the largest D rating, right,
[7:53] Speaker 1: relative to its five-year norm. But that is ahead of earnings.
[7:57] Speaker 1: That's why we're talking. And there are strong fundamentals. So
[8:00] Speaker 1: everyone's expectation is you see outperformance in tech going forward.
[8:06] Speaker 1: Is that what you see?
[8:08] Speaker 6: Certainly the risk reward is skewed. Once you see that
[8:10] Speaker 6: multiple correction, I mean, that is usually a better signal
[8:14] Speaker 6: than some noise, right? So the market almost always gets
[8:18] Speaker 6: it right from that perspective in that earnings do continue
[8:21] Speaker 6: to grow. And that valuation support has more often than
[8:25] Speaker 6: not actually priced in a fair amount of bad news.
[8:28] Speaker 6: So it's kind of akin to saying that the technology
[8:31] Speaker 6: sector can still climb the wall of worry.
[8:33] Speaker 5: Right.
[8:34] Speaker 6: And I think that that's true for the overall U.S.
[8:36] Speaker 7: Stock market.
[8:37] Speaker 6: But I think technology looks more like leadership when you
[8:40] Speaker 6: look at the valuation data and the consistency of earnings
[8:43] Speaker 6: through time. Even when you look at the government data
[8:45] Speaker 6: in terms of CapEx spending and technology, and they do
[8:48] Speaker 6: break it out. We even have data centers going back
[8:50] Speaker 6: to 2020. When you look at it, yes, it's slowing
[8:53] Speaker 6: from a second derivative perspective, and that was even on
[8:55] Speaker 6: the June NIPA accounts or the GDP data, and it's
[8:58] Speaker 6: probably going to slow down further. And some equity investors
[9:00] Speaker 6: would say, oh, well, the second derivative slowdown is here.
[9:03] Speaker 6: That has not really been a problem for the stocks
[9:06] Speaker 6: historically because the vast majority of slowdowns in tech CapEx,
[9:10] Speaker 6: because it's been a secular trend, don't end in a contraction.
[9:14] Speaker 6: And in fact, the better the starting point, even in
[9:16] Speaker 6: a slowdown, the more likely technology stocks have been to outperform,
[9:20] Speaker 6: which kind of gets back to the by the time
[9:22] Speaker 6: you see the data, the stocks might have already discounted
[9:25] Speaker 6: it historically.
[9:27] Speaker 1: You are talking about and you've been writing about how
[9:30] Speaker 1: important tech spending has become to the U.S. economy. And
[9:33] Speaker 1: in that answer, you're kind of speaking to what happens
[9:35] Speaker 1: if that spending would slow down. What are some of
[9:38] Speaker 1: the headwinds in your models for the tech sector?
[9:42] Speaker 2: Right.
[9:42] Speaker 1: What would what is something that would would facilitate a
[9:45] Speaker 1: slowdown in their spending?
[9:47] Speaker 6: I mean, you're already seeing the slowdown, so I don't
[9:49] Speaker 6: even think we have to debate whether or not it's
[9:51] Speaker 6: going to slow. It's here in the accounts already. Again,
[9:54] Speaker 6: I think you see a secular trend from technology spending
[9:57] Speaker 6: or CapEx being 3% of GDP back in the 60s
[10:00] Speaker 6: to now in excess of 7% now. So people would
[10:03] Speaker 6: say that that's elevated and we are dependent on tech CapEx.
[10:07] Speaker 6: And the interesting part is, look, you can always argue
[10:09] Speaker 6: that it's going to be different this time, but we
[10:11] Speaker 6: have all this historical data. And you can see, one,
[10:13] Speaker 6: this trend is very secular, unlike residential investment, which kind
[10:17] Speaker 6: of got us into the crosshairs and the financial crisis.
[10:20] Speaker 6: But two is that slowdowns don't usually end in a contraction.
[10:23] Speaker 6: There's only been four contractions, the financial crisis, the dot-com bubble.
[10:27] Speaker 6: and then two recessions in the 70s, that's how strong
[10:30] Speaker 6: the secular trend is. Now, the interesting part is, let's
[10:33] Speaker 6: just assume that it slows from here, which, again, we're
[10:36] Speaker 6: already seeing that in the data. Then you can say,
[10:38] Speaker 6: all right, let's assess the dependency of the U.S. economy
[10:43] Speaker 6: on this secular trend. And what you find is that
[10:46] Speaker 6: when you see the slowdown, usually technology, usually CapEx and
[10:50] Speaker 6: GDP growth for the rest of the sectors actually accelerate.
[10:55] Speaker 6: Which almost tells you what I think we kind of
[10:57] Speaker 6: know as investors is that CapEx is more often than not,
[11:01] Speaker 6: which is not to say every time, a leading indicator
[11:04] Speaker 6: for the U.S.
[11:05] Speaker 1: Economy.
[11:05] Speaker 6: So as much as we are afraid of the fact
[11:09] Speaker 6: that the U.S. economy is dependent on technology CapEx, it
[11:12] Speaker 6: might actually be signaling future growth.
[11:17] Speaker 1: Denise Chisholm of Fidelity, really great to have you back
[11:20] Speaker 1: on the program. Thank you very much indeed. Now coming up,
[11:23] Speaker 1: NVIDIA and Cisco are teaming up on a new device
[11:26] Speaker 1: to address the AI infrastructure bottleneck. We're going to be
[11:28] Speaker 1: joined by Cisco's president and chief product officer, G.T. Patel.
[11:31] Speaker 1: That's next. This is Bloomberg Tech. Bloomberg Tech Okay, Cisco
[11:48] Speaker 1: and NVIDIA are partnering to introduce a new plugin to
[11:51] Speaker 1: help data centers move data faster. This is a new
[11:55] Speaker 1: optical transceiver. It's aimed at boosting network capacity as the
[11:59] Speaker 1: demand for AI infrastructure surges. Basically, in the world of AI,
[12:02] Speaker 1: you're in a situation, right, where you have thousands of GPUs.
[12:06] Speaker 1: In the future, hundreds of thousands of GPUs. And they
[12:10] Speaker 1: need to be able to move quickly, move and talk
[12:12] Speaker 1: to each other in real time. there is a brilliant
[12:15] Speaker 1: voice to discuss this technology and this partnership, and that
[12:18] Speaker 1: is Cisco's president and chief product officer, Jitu Patel. Jitu,
[12:22] Speaker 1: it's great to have you back on the show. I'm
[12:24] Speaker 1: trying to get into the sort of balance between the
[12:26] Speaker 1: layman's and the going into the detail of USFP, but
[12:31] Speaker 1: a new obstacle transceiver format, right? What's the problem you're
[12:34] Speaker 1: solving for here?
[12:36] Speaker 8: So the problem we're solving, so first, if we just
[12:38] Speaker 8: take a step back and say, what is happening right
[12:40] Speaker 8: now in AI? the amount of traffic that's going to
[12:47] Speaker 8: get generated with these agents, which is the second big
[12:50] Speaker 8: phase of AI that we are in now, especially with
[12:52] Speaker 8: these personal agents, is going to be growing exponentially. And
[12:56] Speaker 8: so what you're starting to see is you no longer
[12:59] Speaker 8: just have to make sure that you have the rack
[13:04] Speaker 8: itself scaling, but now what you're starting to see in
[13:06] Speaker 8: data centers is you're going to need to have hundreds
[13:09] Speaker 8: of thousands of GPUs over time within a data center
[13:12] Speaker 8: that need to act as one coherent cluster. And so
[13:16] Speaker 8: what you need to do is you need to make
[13:17] Speaker 8: sure that those transmissions can't just happen with a wire.
[13:22] Speaker 8: They need to happen with optical kind of transmitters, which
[13:25] Speaker 8: is basically light rays that go across, go over the fiber.
[13:31] Speaker 8: And so what we've done is we have partnered with
[13:34] Speaker 8: Nvidia to have the next generation pluggable optics It's a
[13:38] Speaker 8: standard called ultra-dense small form factor plug, which is USFP.
[13:44] Speaker 8: And what that allows you to do is think about
[13:46] Speaker 8: this as it's just going to have the exact same
[13:50] Speaker 8: highway where you can actually have many, many more cars
[13:53] Speaker 8: go through. And so it's going to have at least
[13:57] Speaker 8: 4X the bandwidth, and you're going to be 1.3X more
[13:59] Speaker 8: power efficient than what you used to have.
[14:01] Speaker 3: Yes.
[14:06] Speaker 8: So much, much more traffic patterns that can actually be
[14:08] Speaker 8: accommodated with this new collaboration that we're doing with NVIDIA.
[14:13] Speaker 1: You know, GC, there are so many people watching the
[14:15] Speaker 1: show that always want to be inside the data center,
[14:18] Speaker 1: and it's really hard to comprehend what's happening. The module
[14:22] Speaker 1: converts that electrical signal from the network switch into light.
[14:26] Speaker 1: That's the technology. Huge amounts of data traveling over fiber,
[14:31] Speaker 1: critically important. I'm trying to get to what's new in
[14:34] Speaker 1: the USFP. 400 gigabits per lane versus prior generation technologies
[14:39] Speaker 1: that weren't at that level. Just go with that a
[14:41] Speaker 1: little bit.
[14:43] Speaker 8: Yeah, so I think the other standards that are emerging
[14:46] Speaker 8: don't actually, what this allows you to do is, you know,
[14:51] Speaker 8: it's built and deployed and kind of serviced and managed
[14:55] Speaker 8: in the same way as pluggables are. And so you
[14:58] Speaker 8: don't have to have a full re-architecture of your data centers.
[15:02] Speaker 8: That's a huge advantage of this kind of standard and
[15:05] Speaker 8: this format. And you have up to, you know, kind
[15:07] Speaker 8: of 16 lanes or 32 lanes of 400 gigabits per
[15:12] Speaker 8: second transmission that's going through.
[15:14] Speaker 2: So it's just a lot more bandwidth. It's much faster.
[15:17] Speaker 8: And I think these switches, if you take a step back,
[15:20] Speaker 8: what you need to make sure that you have is
[15:22] Speaker 8: these switches within a data center need to transmit, you know,
[15:27] Speaker 8: traffic at very low latency so that your GPUs don't
[15:31] Speaker 8: remain idle. And that's what we're trying to do over
[15:33] Speaker 8: here is make sure that you actually have extremely fast
[15:36] Speaker 8: transmission at very, very low latency and very high power efficiency.
[15:41] Speaker 1: Why did Cisco need to work with NVIDIA on this,
[15:44] Speaker 1: NVIDIA specifically?
[15:46] Speaker 8: Look, I think this is the one thing that you
[15:49] Speaker 8: have to keep in mind is this is an ecosystem
[15:52] Speaker 8: play in general in this market. It's very hard for
[15:56] Speaker 8: a single company to just go out and do everything.
[15:58] Speaker 8: We need to make sure that we're working with partners.
[16:01] Speaker 8: And NVIDIA makes GPUs. They make … They make these
[16:05] Speaker 8: rack systems, and they really go out and equip the
[16:08] Speaker 8: data centers. And we make the networking that goes out
[16:11] Speaker 8: and connects these data centers. And we have these coherent
[16:15] Speaker 8: pluggable optics modules. And we wanted to make sure that
[16:18] Speaker 8: the standards that we're coming up with, we're actually collaborating
[16:20] Speaker 8: with others to do that. And so we're really looking
[16:24] Speaker 8: forward to collaborating with NVIDIA. The nice part about NVIDIA
[16:26] Speaker 8: is we have a great partnership with them. We've worked
[16:29] Speaker 8: with them for many, many years now. And we continue
[16:32] Speaker 8: to keep finding different avenues where we can keep pushing
[16:35] Speaker 8: the kind of challenging the limits of physics and keep
[16:38] Speaker 8: pushing this forward.
[16:40] Speaker 2: Right.
[16:40] Speaker 1: Your starting position was sort of that, you know, AI
[16:43] Speaker 1: is supply constrained, right? But supply relative to the demand,
[16:46] Speaker 1: I find where the demand from a tokens perspective is
[16:50] Speaker 1: coming from so interesting. Like agents are using 5X the
[16:53] Speaker 1: tokens that a human uses. power user is, right? How
[16:57] Speaker 1: is that sort of exacerbating the problem that you're trying
[17:00] Speaker 1: to solve for?
[17:02] Speaker 8: So it is a very interesting problem that we actually
[17:06] Speaker 8: face right now, which is as we move from this
[17:08] Speaker 8: era of chatbots to agents, these agents are, like you said,
[17:12] Speaker 8: far more consumptive on network bandwidth than even a human.
[17:15] Speaker 2: For conducting the same exact task. It takes about 500% more.
[17:20] Speaker 8: For an agent to conduct the same task with network
[17:23] Speaker 8: bandwidth compared to.
[17:23] Speaker 2: What a human would take to conduct that same task.
[17:26] Speaker 8: And so what you're starting to find is a tremendous
[17:29] Speaker 8: amount of need for more and more kind of technology
[17:34] Speaker 8: innovations that can push the envelope, specifically in data centers.
[17:38] Speaker 8: And so now what we are at the point of
[17:41] Speaker 8: is these– Today, sub 2% of human beings within the
[17:45] Speaker 8: world are using agents in any kind of power user capacity.
[17:48] Speaker 8: And my prediction is that within the next, I don't know,
[17:52] Speaker 8: 15 months or so, you'll probably have 30 to 40%
[17:56] Speaker 8: of humans use these agents in a power user capacity.
[17:59] Speaker 2: And so there's going to be an exponential rise.
[18:01] Speaker 8: And with 2% of the humans using these agents, we
[18:04] Speaker 8: are still completely saturated in supply. So you have to
[18:07] Speaker 8: make sure that you don't just build more data centers,
[18:10] Speaker 8: you make the existing data centers more effective.
[18:13] Speaker 1: So let me jump in on that, GT. The basics are,
[18:17] Speaker 1: you know, USFP is a super fast plug, connects all
[18:21] Speaker 1: the AI supercomputers in the data center. But to your point,
[18:24] Speaker 1: you also have to discuss the economics. So this technology
[18:27] Speaker 1: that you've worked on with NVIDIA, how does it change
[18:30] Speaker 1: the sort of speed to get something online, but also
[18:34] Speaker 1: the economics of running that data center?
[18:37] Speaker 8: The economics of running the data center, the total cost
[18:40] Speaker 8: of ownership actually keep going down as you start to
[18:44] Speaker 8: think about cost of tokens. So the cost of tokens
[18:47] Speaker 8: actually gets far more efficient than what it's.
[18:51] Speaker 7: Been.
[18:52] Speaker 8: And what you also see over here is the power
[18:54] Speaker 8: efficiency is 1.3x. So it's 1.3x more power efficient than
[18:59] Speaker 8: what you would have, which then would have a direct
[19:00] Speaker 8: correlation to what your cost of tokens would be within
[19:03] Speaker 8: this entire kind of equation.
[19:06] Speaker 2: So 4x more bandwidth density, 1.3x more power efficiency, 16
[19:09] Speaker 2: lanes of 400 gigabits.
[19:16] Speaker 8: Of scalable traffic, or you can go up to 32 lanes.
[19:20] Speaker 8: And so that's the, those are the stats. And the
[19:23] Speaker 8: key point of here is it's built, it's deployed and
[19:25] Speaker 8: serviced and managed the same way as today's pluggables. So
[19:29] Speaker 8: customers can get more out of the way that they've
[19:32] Speaker 8: already architected their data centers.
[19:34] Speaker 1: Cisco's president and chief product officer, Jitu Patel, back with
[19:37] Speaker 1: us on the show. Thank you so much. Coming up,
[19:39] Speaker 1: Elon Musk wants to dramatically speed up chip development and
[19:42] Speaker 1: thinks his companies can eventually beat the industry at its
[19:45] Speaker 1: own game. We've got that story next. This is Bloomberg Tech.
[19:59] Speaker 1: Elon Musk wants to put the chip industry on fast forward.
[20:02] Speaker 1: His goal, test as many as 10 new designs a year,
[20:05] Speaker 1: and eventually, he says, make chips better than anyone else.
[20:09] Speaker 1: Bloomberg's Kyle Porter has the story. Okay, so some more
[20:12] Speaker 1: kind of fighting talk from Musk on partly the TerraFab initiative,
[20:16] Speaker 1: but his chip-making ambitions. What's he saying essentially here, both
[20:21] Speaker 1: sort of his ambition and the reasons for it?
[20:26] Speaker 9: Well, he was certainly very verbose, the ex-taker on Saturday.
[20:30] Speaker 9: He outlined that he currently knows nothing about chips, but
[20:33] Speaker 9: he's going to learn very quickly. And he's convinced that
[20:37] Speaker 9: via Tesla and SpaceX skill customization, they can produce more
[20:41] Speaker 9: and better chips than anyone.
[20:42] Speaker 1: Else on the planet. He talks about something that's kind
[20:47] Speaker 1: of core to how Tesla used to run, SpaceX runs,
[20:49] Speaker 1: that he just wants teams to move quickly. So there's
[20:52] Speaker 1: this idea of the 10 new designs a year and
[20:54] Speaker 1: an iterative process. Just in normal people speak, what's he
[20:58] Speaker 1: talking about?
[21:00] Speaker 9: He's convinced that he can apply what he's done in
[21:02] Speaker 9: his previous companies to chip making. Now, they have brought
[21:05] Speaker 9: Antal on board for this, so it's not completely virgin
[21:09] Speaker 9: territory for the companies. But he made much of the
[21:13] Speaker 9: fact that when you walk around your competitors' chip factories,
[21:16] Speaker 9: You don't see custom equipment. He was seeing the same
[21:18] Speaker 9: equipment everywhere that he goes. And he's convinced that the
[21:21] Speaker 9: way to improve the manufacturing process is to have your
[21:25] Speaker 9: own kit so you can do it better, simpler, faster
[21:27] Speaker 9: than your competitors. And he's definitely putting his money where
[21:30] Speaker 9: his mouth is. The new Terraflap factory has an initial
[21:33] Speaker 9: investment of about $ 16 billion. And you'll get multitudes of
[21:37] Speaker 9: that if they eventually take over the, I think it's
[21:39] Speaker 9: 13,000 acres of square space they've got earmarked for this plant.
[21:43] Speaker 1: He did acknowledge that to take one design from design
[21:47] Speaker 1: to fabrication in two years is good going. He wants
[21:49] Speaker 1: to do 10. And he seems to think that his
[21:53] Speaker 1: teams are capable of that.
[21:55] Speaker 9: Yeah, he certainly does. And, you know, he has a
[21:57] Speaker 9: track record of having established and built businesses. But, you know,
[22:00] Speaker 9: anyone who's done anything with semiconductors will tell you that
[22:03] Speaker 9: it is extremely difficult to do even a percentage of
[22:08] Speaker 9: what he's proposing or doing now and doing it from
[22:10] Speaker 9: a standing start. So the jury is very much out.
[22:12] Speaker 9: And to Elon's credit, he said it may take us
[22:14] Speaker 9: years to do this.
[22:17] Speaker 1: Bloomberg's Kyle Porter on deck with the story. Thank you
[22:20] Speaker 1: very much indeed. Coming up in the show, hyperscalers show
[22:23] Speaker 1: no signs of slowing, especially on the AI build-out. But
[22:26] Speaker 1: the cost of building data centers is ballooning. We have
[22:30] Speaker 1: this kind of exclusive breakdown of the expenses. That's next.
[22:34] Speaker 1: This is Bloomberg Tech. Welcome back to Bloomberg Tech. Let's
[22:54] Speaker 1: get out to New York where Bloomberg's Yahira Anand is
[22:56] Speaker 1: starring by. Hi, Yahira.
[22:57] Speaker 2: Hi, Ed.
[22:58] Speaker 10: It's time now for Talking Tech. First up, Amazon could
[23:01] Speaker 10: be closing in on another big AI deal. The company
[23:04] Speaker 10: is said to be in talks to buy AI startup
[23:06] Speaker 10: Descartes for around $ 7 billion. That's according to the Wall
[23:10] Speaker 10: Street Journal. That comes after Anthropic walked away from a
[23:13] Speaker 10: potential $ 6 billion deal for Descartes last month. Plus, IonQ
[23:18] Speaker 10: is making another big bet on cybersecurity. The quantum computing
[23:22] Speaker 10: firm is buying Elicity for up to $ 1. 3 billion in stock,
[23:27] Speaker 10: expanding its security business as companies prepare for emerging AI
[23:32] Speaker 10: and quantum threats. The deal follows IonQ's $ 1. 8 billion acquisition
[23:37] Speaker 10: of chipmaker Skywater earlier this year. And DeepSeek is making
[23:42] Speaker 10: a bigger bet on robotics. A Chinese AI firm is
[23:45] Speaker 10: looking to lead a funding round for Winverse AI, a
[23:49] Speaker 10: months-old startup for building AI robots. The deal would value
[23:53] Speaker 10: the company at about $ 400 million. Winverse was founded by
[23:57] Speaker 10: Hu Han, a former top AI scientist over.
[24:00] Speaker 1: At Tencent, etc. Okay, thank you, Yohara. Let's take a
[24:04] Speaker 1: look at today's big number, $ 30 billion. That's the estimated
[24:08] Speaker 1: cost of building a gigawatt data center in Texas right now.
[24:12] Speaker 1: And as the costs of materials, power, and labor rise,
[24:16] Speaker 1: that number is expected to get even bigger. Bloomberg has
[24:18] Speaker 1: this exclusive breakdown of expenses. to construct an AI data
[24:22] Speaker 1: center in Abilene, Texas, based on data from Cushman and
[24:25] Speaker 1: Wakefield and estimates from Bloomberg Intelligence. Our managing editor for
[24:29] Speaker 1: data centers, Lynn Dunn, is here to break it down
[24:31] Speaker 1: with us. The biggest piece, and consistent with the theme
[24:34] Speaker 1: of today's show, is the chips. That's a big chunk
[24:36] Speaker 1: of it. But yeah, break down $ 30 billion for a
[24:40] Speaker 1: gigawatt data center. What's it made up of?
[24:43] Speaker 3: Yeah, that's exactly right, Ed, and thank you for having me.
[24:46] Speaker 3: The obvious big piece here are the servers, the chips,
[24:50] Speaker 3: the networking that goes into them. And not just those
[24:52] Speaker 3: fancy AI chips, the GPUs, the TPUs, but also the
[24:56] Speaker 3: memory chips, which have soared in cost because of the
[24:59] Speaker 3: data center boom. But this project was really enlightening because
[25:04] Speaker 3: it shed light. On the other 50% to 60% of
[25:08] Speaker 3: the data center costs, which the bulk of it is
[25:11] Speaker 3: the land, I mean power land in particular, the kind
[25:14] Speaker 3: that's connected to the power grids, has gotten uncontrollably expensive,
[25:19] Speaker 3: the design and engineering, the construction of the shell itself.
[25:23] Speaker 3: It's a very broad overview of exactly why we're seeing
[25:26] Speaker 3: these hyperscalers spend $ 850 billion in a single year.
[25:32] Speaker 1: This is a must-read, by the way. Go to Bloomberg.com
[25:34] Speaker 1: and check it out, how it's presented, broken down visually.
[25:38] Speaker 1: It really helps to understand what goes into making a
[25:41] Speaker 1: data center. The start of our conversation was that that
[25:44] Speaker 1: big number of $ 30 billion is going to go up.
[25:47] Speaker 1: Where are we seeing inflation for the business of data
[25:51] Speaker 1: center construction?
[25:52] Speaker 3: Where are we not seeing inflation, Ed? It's literally baked
[25:56] Speaker 3: into every single category of this breakdown. I just mentioned
[26:01] Speaker 3: that power land is getting super expensive. The designers, the engineers,
[26:06] Speaker 3: the workforce labor is getting very expensive. The construction, I'm
[26:10] Speaker 3: talking about the raw materials like copper and concrete and steel.
[26:14] Speaker 3: are all through the roof. The electricity, the power turbines,
[26:17] Speaker 3: as we cover in the story itself, the demand for
[26:21] Speaker 3: gas-fired power generators, the turbines themselves that make the electricity
[26:26] Speaker 3: that power the data centers on site is so high
[26:29] Speaker 3: that they're talking about a lead time of like two
[26:32] Speaker 3: and a half to, I want to say, 10 years,
[26:35] Speaker 3: you know, going out into like 2035 for delivery of
[26:38] Speaker 3: one of these things.
[26:40] Speaker 1: You opened the story with the point that just four
[26:42] Speaker 1: companies have expected spending of about $ 750 billion. And I
[26:46] Speaker 1: don't know why, but Ireland and Belgium's GDP gets a
[26:50] Speaker 1: bit of a shot at the second paragraph, which is good.
[26:52] Speaker 1: But the point is the scale is enormous. Give us
[26:56] Speaker 1: that bigger picture. What's happening across industry?
[26:59] Speaker 3: That's a compliment, Ed. I threw that in there on purpose.
[27:03] Speaker 3: So I wanted to give some size and scope because
[27:07] Speaker 3: this is a build-out that The world has honestly never
[27:10] Speaker 3: seen before. I think the highway system in the United
[27:14] Speaker 3: States is really the only one that people have thrown
[27:17] Speaker 3: out there as a comparable infrastructure project. The scale is
[27:22] Speaker 3: difficult to wrap your head around. And so I wanted
[27:25] Speaker 3: to give some figures. One of them was, you know,
[27:28] Speaker 3: this is more than the entire GDP of Ireland or Belgium.
[27:31] Speaker 3: The other is that more than half of what all
[27:34] Speaker 3: of the companies in the S & P 500 are.
[27:37] Speaker 3: spend in a single year. Another one was that you
[27:39] Speaker 3: could cut a $ 5, 500 check for every household in the
[27:44] Speaker 3: US with that kind of money.
[27:45] Speaker 1: Yes, it really is a must-read. Check it out on
[27:48] Speaker 1: the Bloomberg terminal, but also on.
[27:49] Speaker 7: Com.
[27:50] Speaker 1: It's a very visual breakdown of what $ 30 billion gets
[27:53] Speaker 1: you or is needed to build a gigawatt in Texas.
[27:57] Speaker 1: Bloomberg's Lindor, and thank you very much indeed. Some news
[28:00] Speaker 1: coming from over the weekend. Anthropic has released a report
[28:04] Speaker 1: on more rogue actions by various generations of its clawed
[28:07] Speaker 1: AI models. Recently, its AI submitted a fake tip to
[28:12] Speaker 1: Philadelphia police about a homicide. It also sent in 20
[28:16] Speaker 1: visa applications through a publicly available form on the State
[28:20] Speaker 1: Department's website. In its report, Anthropic said that the latest
[28:23] Speaker 1: cases had, quote, minimal real-world impact. But Anthropic has restricted
[28:27] Speaker 1: some types of internet access for its AI models during testing.
[28:31] Speaker 1: The Trump administration said it now requires AI companies to
[28:34] Speaker 1: notify affected parties and address security incidents involving their models. Okay,
[28:41] Speaker 1: coming up, we're going to have a conversation with Vista
[28:43] Speaker 1: Shares about what the market is getting right and what
[28:47] Speaker 1: the market is getting wrong when it comes to AI investment.
[28:50] Speaker 1: You heard some very interesting comments at the top of
[28:51] Speaker 1: the hour from Hock Tan about a lot of people
[28:55] Speaker 1: don't have a clue what they're investing in when it
[28:57] Speaker 1: comes to AI right now. That conversation next. This is
[28:59] Speaker 1: Bloomberg 10. Okay, with earnings season getting underway and the
[29:08] Speaker 1: market kind of reassessing the risks surrounding AI investments, questions
[29:12] Speaker 1: about the vulnerabilities of the AI boom grow louder. There
[29:14] Speaker 1: are more headlines. David Feverson-Harr, VistaShares Portfolio Manager and DVX
[29:19] Speaker 1: Ventures Principal brings this kind of unique perspective, combining a
[29:21] Speaker 1: venture background with public market investing through thematic ETFs. And David,
[29:26] Speaker 1: it's great to have you. Yeah, thanks for having me.
[29:28] Speaker 1: We kind of started the program with this reflection from
[29:31] Speaker 1: from the Broadcom CEO saying that actually investors don't have
[29:36] Speaker 1: a very good grasp of what it is they're investing
[29:39] Speaker 1: in in the field. of AI. I found that to
[29:41] Speaker 1: be a surprise.
[29:45] Speaker 11: Yeah, I think it kind of hits on three elements
[29:47] Speaker 11: that we kind of come back to quite frequently, which
[29:50] Speaker 11: is the first is there's a lot of noise in
[29:52] Speaker 11: the market and it's really separated. The whole AI trade
[29:56] Speaker 11: can be The demand side, the monetization piece, which is
[29:59] Speaker 11: really three layers. It's just, are the labs growing? Can
[30:02] Speaker 11: hyperscalers monetize? And can you actually prove the application layer
[30:07] Speaker 11: and the ROI within enterprises? All that kind of changes
[30:10] Speaker 11: depending on where you are in the layer as to
[30:12] Speaker 11: the mix of what you spend it on in the infrastructure.
[30:15] Speaker 11: But what's really interesting about his comments, he's pretty much saying,
[30:18] Speaker 11: layer one is not public. The whole market is kind
[30:21] Speaker 11: of operating on this opacity. No one can see what's
[30:23] Speaker 11: going on. We saw it last week. You kind of
[30:25] Speaker 11: covered it last week with the OpenAI accounting announcement and
[30:28] Speaker 11: their ARR numbers.
[30:29] Speaker 1: That's right. Just the recap on that is that there
[30:32] Speaker 1: was a discrepancy on where they were on a calendar
[30:35] Speaker 1: basis with ARR, but also whether they were reporting gross
[30:38] Speaker 1: or net revenues. You go back and watch that episode
[30:40] Speaker 1: from last week. You basically believe it's a good thing, therefore,
[30:44] Speaker 1: when OpenAI and Anthropic go public. It makes it more transparent.
[30:49] Speaker 11: For the whole ecosystem, you can decouple valuation and kind
[30:53] Speaker 11: of the private investor lens of when people had their
[30:56] Speaker 11: entry points. But from an ecosystem perspective, the ROI story,
[31:00] Speaker 11: as I kind of mentioned out, is the lab ARR numbers.
[31:03] Speaker 11: The second layer is the ROI from the hyperscaler perspective.
[31:05] Speaker 11: And the third being the monetization layer for apps. You
[31:09] Speaker 11: don't really have the first layer highly visible. So what
[31:14] Speaker 11: people end up doing is grasping at the next layer.
[31:15] Speaker 11: But the next layer is ROI hyperscaler story, which is
[31:18] Speaker 11: a multi-quarter story. People just started getting comfortable with it
[31:21] Speaker 11: last quarter. It's not just one quarter. The next quarter
[31:24] Speaker 11: is always going to be the most important quarter, or
[31:26] Speaker 11: this quarter is the most important quarter. It's always going
[31:28] Speaker 11: to be that way. And so you can actually show
[31:29] Speaker 11: consistent growth. As even your last guest alluded to, there's
[31:34] Speaker 11: just the scale of what is going into spend right
[31:37] Speaker 11: now is getting people a little bit nervous, and they
[31:39] Speaker 11: want to see ROI.
[31:40] Speaker 1: Right.
[31:41] Speaker 11: So if they can't get ROI from the first layer,
[31:43] Speaker 11: there's kind of this tension being placed on something that
[31:45] Speaker 11: is taking time to develop and should take time to develop,
[31:49] Speaker 11: but kind of giving people a little bit of time
[31:51] Speaker 11: to pause.
[31:53] Speaker 1: Jensen Huang, the NVIDIA CEO, talks about the five-layer cake
[31:57] Speaker 1: of AI, right? And you're talking about something similar, which
[32:00] Speaker 1: is for some people it is a choice. To their mind,
[32:03] Speaker 1: you're in one or another layer. It sounds like to you,
[32:08] Speaker 1: you feel comfortable being invested across the five layers. Am
[32:12] Speaker 1: I reading that right?
[32:13] Speaker 11: Yeah, I think the way that we kind of interpret it,
[32:16] Speaker 11: and it's a great follow-up again to your last segment,
[32:18] Speaker 11: is for the most part we follow where the capital
[32:21] Speaker 11: expenditure is going in the industry, where it is today
[32:23] Speaker 11: and where we think it's going to go.
[32:24] Speaker 1: In terms of where it's being deployed?
[32:26] Speaker 11: Where it's being deployed, exactly. So if we think that a.
[32:28] Speaker 1: Specific … The net recipients of capex.
[32:30] Speaker 2: Exactly.
[32:30] Speaker 11: We want to actually be investing in the net earners.
[32:33] Speaker 11: as part of what reinforces that flywheel of what I
[32:36] Speaker 11: kind of view as like demand and delivery, you want
[32:39] Speaker 11: to see that there's positive signals on monetization to get
[32:41] Speaker 11: really comfortable. Because as we've seen, the scale of what's
[32:44] Speaker 11: going into the earners of CapEx is just so large.
[32:49] Speaker 11: So people are starting to look now for the ROI
[32:51] Speaker 11: on the actual demand side of things. But for the
[32:53] Speaker 11: most part, all it changes really is when you move
[32:56] Speaker 11: from frontier lab ARR to hyperscaler ROI to app monetization,
[33:02] Speaker 11: it reinforces the delivery mechanisms, which is all the chips,
[33:06] Speaker 11: the networking, the cooling. Even the debate of open source
[33:09] Speaker 11: versus closed source is kind of a delivery mechanism, but
[33:13] Speaker 11: it just changes the composition of where you are and
[33:16] Speaker 11: what you need as part of the actual physical makeup.
[33:20] Speaker 11: It doesn't actually replace the need for the physical makeup.
[33:23] Speaker 11: It's really a reinforcing loop. So that's why there's constant
[33:27] Speaker 11: tension that you're seeing within the scale of the investment
[33:30] Speaker 11: dollars of what's going into the industry, and people looking
[33:33] Speaker 11: to make sure that there's an actual ROI to it.
[33:35] Speaker 1: I was going to ask you about that. You said dollars.
[33:38] Speaker 1: Everything that we talk about in this context on the
[33:40] Speaker 1: show is largely focused on what's happening here in the U.S.,
[33:43] Speaker 1: I'm not even talking about China, really. I'm talking about,
[33:46] Speaker 1: you know, we hardly ever talk about Europe, other jurisdictions
[33:49] Speaker 1: from an investment opportunity standpoint. Do you look outside of
[33:53] Speaker 1: the U.S.
[33:53] Speaker 11: Yeah, most of, close to a majority of our holdings
[33:57] Speaker 11: are international. And that was kind of part of our
[33:58] Speaker 11: thesis originally. When we initially launched the fund in December 2024,
[34:03] Speaker 11: it was really around this idea.
[34:04] Speaker 2: That a lot of U.S.
[34:05] Speaker 11: Investors couldn't actually get exposure to some of the really critical,
[34:08] Speaker 11: kind of like what we call the crown jewels of
[34:10] Speaker 11: the AI trade at the time. Those are the companies
[34:13] Speaker 11: in Taiwan, South Korea. And a lot of providing exposure
[34:17] Speaker 11: to that was, again, on a proportional basis of what
[34:19] Speaker 11: was going into a data center. We kind of viewed
[34:22] Speaker 11: that memory was kind of this very strong bottleneck. There
[34:25] Speaker 11: was no real price alleviation that was happening. And you can.
[34:28] Speaker 1: Still see it. A lot of the earnings. That was
[34:31] Speaker 1: back in 2024. That was back in 2024.
[34:32] Speaker 11: A lot of the improvement in the underlying fundamentals of
[34:34] Speaker 11: the company has actually not been on quantity. It's been
[34:36] Speaker 11: on price. So a lot of margin expansion. You can
[34:39] Speaker 11: even see with Micron's earnings.
[34:40] Speaker 1: Yes.
[34:41] Speaker 11: In exceptionally strong gross margins. So we've kind of been
[34:45] Speaker 11: tracking a lot of this global supply chain trade and
[34:48] Speaker 11: really following the dollars of where we think it's going
[34:50] Speaker 11: to be today, but not only say where we think
[34:51] Speaker 11: it's going to go in the future. And that's kind
[34:54] Speaker 11: of the critical piece of understanding the different layers of monetization.
[34:57] Speaker 1: David Febstenhoff from Vista Shares and DVX Ventures. Great to
[35:00] Speaker 1: have you with us.
[35:00] Speaker 2: Yeah, thanks for having me.
[35:02] Speaker 1: Chinese EV maker BYD is planning to go on a
[35:04] Speaker 1: hiring spree in Europe and increase its workforce there to
[35:07] Speaker 1: about 40,000 people by the end of the decade. This
[35:10] Speaker 1: would make BYD one of the biggest non-European auto employers
[35:13] Speaker 1: in the region. It comes days after China and the
[35:16] Speaker 1: EU agreed to curb the number of hybrid vehicles that
[35:19] Speaker 1: would be imported from the Asian country. BYD Executive Vice
[35:23] Speaker 1: President Stella Lee says the company is still assessing the
[35:26] Speaker 1: implications of the new EU restrictions. She spoke with Bloomberg's
[35:29] Speaker 1: Caroline Conan. at the Paris Motor Show.
[35:33] Speaker 12: We need to spend time to understand what does this mean.
[35:36] Speaker 12: So we don't have details yet. But I want to
[35:38] Speaker 12: assure BYD commitment in Europe is very deep. It's long term.
[35:43] Speaker 12: So we already built up our Hungary facility now. After
[35:47] Speaker 12: challenge ramp up by Q4, we're going to start mass production.
[35:52] Speaker 12: And now we continue to invest more in Europe. In
[35:55] Speaker 12: the past two years, we qualified more than 250 European
[36:00] Speaker 12: vendors to be BYD global vendors. And now we're building
[36:05] Speaker 12: our R & D in Europe. So step by step
[36:08] Speaker 12: BYD is going to invest more and not only for manufacturing,
[36:12] Speaker 12: bring the job, but also invest more for R &
[36:15] Speaker 12: D and really building the ecosystem in Europe.
[36:20] Speaker 13: BYD obviously has an increasing market share in Europe. The
[36:25] Speaker 13: latest numbers for Chinese cars, including BYD in August, were
[36:31] Speaker 13: about 12% of the sales of the new car registered
[36:35] Speaker 13: in Europe. Where do you see BYD going forward in
[36:38] Speaker 13: terms of market share in Europe?
[36:41] Speaker 12: I think BYD is in the front running and fast
[36:45] Speaker 12: growing here because now a lot of consumers, they very
[36:48] Speaker 12: much appreciate BYD is not just a car, it's a
[36:51] Speaker 12: technology company. Then they enjoy the car not only because
[36:55] Speaker 12: saving the money, especially have the Middle East conflict, the
[36:59] Speaker 12: oil price increasing. Now more people is really serious consider
[37:05] Speaker 12: BYD like an electric car. and then also stand by
[37:10] Speaker 12: here today and in the back on the stage we
[37:13] Speaker 12: introduce Atoll 8. Atoll 8 will be the first BYD
[37:18] Speaker 12: model in Europe to carry the flash charging technology.
[37:21] Speaker 13: But obviously your price competitiveness is one of your competitive advantages.
[37:27] Speaker 13: With these new restrictions, these new tariffs on EVs, these
[37:31] Speaker 13: new restrictions on hybrids, are.
[37:33] Speaker 1: Your prices going to go up? In Europe.
[37:37] Speaker 12: But you see what we're doing now, in past several years,
[37:39] Speaker 12: we're turning BYD international company to be more local company.
[37:45] Speaker 1: BYD Executive Vice President Stella Lee, along with our own
[37:48] Speaker 1: Caroline Connan from the Paris Motor Show. Coming up in
[37:51] Speaker 1: the show, Apple is set to release new smart home devices,
[37:54] Speaker 1: including a smart lock, doorbell, and thermostat tomorrow. More details
[37:58] Speaker 1: on these new gadgets when we return. This is Bloomberg Tech.
[38:08] Speaker 1: Let's get back to our top story. Microsoft's plan to
[38:10] Speaker 1: deploy its Maya 300 chips at multi-gigawatt scale, which Bloomberg
[38:14] Speaker 1: Intelligence says is an important step towards reducing its reliance
[38:18] Speaker 1: on NVIDIA chips. We expect adoption to begin with high-volume
[38:21] Speaker 1: first-party applications such as Copilot Teams, GitHub Copilot, where Microsoft
[38:25] Speaker 1: controls the model's software stack and workload utilization rights. Bloomberg
[38:30] Speaker 1: Intelligence Analyst Anna Agrana, you know, this is a diversity
[38:33] Speaker 1: of compute in the first instance, right? But you got
[38:36] Speaker 1: right to the point that it's for internal workloads. That
[38:39] Speaker 1: was a key piece in the reporting not to give
[38:41] Speaker 1: the capacity to third parties.
[38:43] Speaker 14: Yeah, I mean, when you think about it, Microsoft has
[38:46] Speaker 14: a lot of AI workloads coming in right now just
[38:49] Speaker 14: for their own business. And, you know, you've got to
[38:51] Speaker 14: figure out what to do over there because it costs
[38:53] Speaker 14: a lot of money to take care of those, you know, workloads.
[38:58] Speaker 14: And we all know Microsoft has been able to absorb
[39:01] Speaker 14: so much of AI work without sacrificing their margins. So
[39:06] Speaker 14: they have to do whatever they can to protect that,
[39:08] Speaker 14: you know, whether that is model swapping or whether that
[39:12] Speaker 14: is their own silicon technology.
[39:13] Speaker 1: So with Maya 300 eventually being a multi-quad scale, this
[39:16] Speaker 1: is inference, right? So you're focused very much on the
[39:20] Speaker 1: cost per token. Tell me about that. But you also
[39:23] Speaker 1: acknowledge on the training side, frontier model side, NVIDIA stays
[39:28] Speaker 1: as the main compute platform.
[39:30] Speaker 14: I think that's going to be the case for a while. Now,
[39:32] Speaker 14: the question is, that's really coming. The demand is coming
[39:35] Speaker 14: more from the frontier labs than Microsoft or any other
[39:38] Speaker 14: vendor trying to push it down the throat of the
[39:41] Speaker 14: frontier labs. Now, in the case of Amazon, they have
[39:43] Speaker 14: been able to diversify and convince customers Anthropic, for example,
[39:47] Speaker 14: do use some of their chips to do training work
[39:50] Speaker 14: as well. But frankly speaking, given where Microsoft is, even
[39:53] Speaker 14: if they can get a portion of their inference on
[39:55] Speaker 14: their own chips, it's going to save them a lot
[39:57] Speaker 14: of money, which I think is one of the most
[39:59] Speaker 14: important metrics every investor is looking for is which firms
[40:03] Speaker 14: a few years down the road will still be free
[40:06] Speaker 14: cash flow positive because where we are standing today, we
[40:09] Speaker 14: don't see an end to this AI capex spending, at
[40:12] Speaker 14: least from the hyperscale cloud providers.
[40:15] Speaker 1: Anurag Rana of Bloomberg Intelligence with the React. Thank you
[40:18] Speaker 1: very much. Apple has a few new products set to
[40:21] Speaker 1: release very soon. The company's launching a revamped MacBook Pro,
[40:25] Speaker 1: a foldable iPhone, and even a new line of smart
[40:28] Speaker 1: home devices, which will be unveiled tomorrow. Here to give
[40:31] Speaker 1: us the details is Bloomberg's Mark Gurman, who leads our
[40:34] Speaker 1: coverage of consumer technology and is the author of Power On.
[40:38] Speaker 1: And tomorrow is a moment that you've done a lot
[40:41] Speaker 1: of reporting on. We know a lot of what's to come,
[40:43] Speaker 1: but just remind us the expectation.
[40:46] Speaker 7: Yeah, it's a busy time for Apple. The iPhone Duo
[40:49] Speaker 7: goes on pre-order this Friday. It goes on sale the
[40:53] Speaker 7: following Friday. At the end of this month, they're going
[40:56] Speaker 7: to announce a touchscreen MacBook Pro, M6 versions of the
[41:00] Speaker 7: low-end MacBook Pro and.
[41:01] Speaker 1: The iMac, as well as a revamped iPad mini.
[41:03] Speaker 7: And tomorrow is their big welcome home event in New
[41:06] Speaker 7: York City. They sent invitations out to last week. I
[41:10] Speaker 7: reported in recent weeks that they're going to announce a
[41:13] Speaker 7: new smart home hub. They're going to announce a new
[41:16] Speaker 7: HomePod mini. They're going to announce an Apple TV update.
[41:19] Speaker 7: And these are all intertwined by the new Siri AI.
[41:22] Speaker 7: So this is quite a big moment for their smart
[41:24] Speaker 7: home strategy and really a reboot for the strategy. And
[41:27] Speaker 7: it's been a long time coming. And may I say finally.
[41:31] Speaker 1: I'm going to ask the audience, go and read Power On,
[41:33] Speaker 1: you know, for the details on the products and what
[41:35] Speaker 1: the kind of releases to come. What I wanted to
[41:37] Speaker 1: ask you about, Mark, is something that you reported a
[41:39] Speaker 1: couple of weeks ago on what John Ternus is doing different.
[41:43] Speaker 1: And one of those elements was like not having such
[41:45] Speaker 1: a fixed schedule on the events themselves and the release
[41:49] Speaker 1: of products being more fluid. How does tomorrow fit into
[41:53] Speaker 1: that change or not? It doesn't really.
[41:56] Speaker 7: I mean, this is more of a long-term thing. The
[41:58] Speaker 7: bigger thing with Ternus is trying to make Apple run leaner,
[42:01] Speaker 7: particularly the hardware engineering organization. There have been these small-scale
[42:04] Speaker 7: layoffs to remove engineering project managers and other people in
[42:08] Speaker 7: the organization, trying to reduce the number of people working
[42:11] Speaker 7: on these things, the number of people in the middle,
[42:13] Speaker 7: middle managers, so to speak, trying to make Apple run faster,
[42:17] Speaker 7: run leaner. So that's really the approach he's taking.
[42:21] Speaker 1: Very quick, the product that you, Mark Gurman, are most
[42:24] Speaker 1: eager to see?
[42:26] Speaker 7: Probably the iPhone Duo and how it's doing and how
[42:28] Speaker 7: many people are going to buy it and what the
[42:30] Speaker 7: reception is going to be. That's the one to keep
[42:32] Speaker 7: an eye on this fall.
[42:34] Speaker 1: Thanks, Mark Gurman. Thank you very much. And don't forget.
[42:37] Speaker 1: So there's the power on newsletter, but there's now also
[42:39] Speaker 1: the podcast, which you can tune in for on Thursdays,
[42:42] Speaker 1: wherever you get your podcast power on with Mark Gurman.
[42:46] Speaker 1: I heart Spotify, YouTube everywhere. There are podcasts real quick
[42:51] Speaker 1: before you go. A rough opening weekend for Aaron Sorkin's
[42:54] Speaker 1: return to the Facebook story. The Social Reckoning brought in
[42:57] Speaker 1: just $ 5. 5 million at the U.S. box office, finishing fifth
[43:01] Speaker 1: over the weekend. The film stars Jeremy Strong as Mark
[43:04] Speaker 1: Zuckerberg and is a companion to Sorkin's 2010 hit, The
[43:09] Speaker 1: Social Network. But yeah, $ 5. 5 million U.S. dollars at the U.S.
[43:13] Speaker 1: box office. I haven't seen it yet. You know, I
[43:16] Speaker 1: will do, but I don't know. If you are surprised
[43:20] Speaker 1: by that, find me on the socials. Let's talk about it.
[43:23] Speaker 1: Interesting start to the week. There's a lot to recap
[43:26] Speaker 1: from the show. Big emphasis on what's happening in the
[43:29] Speaker 1: world of semiconductors still. Some underperformance in chip stocks in
[43:32] Speaker 1: the market in the here and now. Recap the show
[43:35] Speaker 1: and the podcast. You know exactly where to find it.
[43:38] Speaker 1: On the Bloomberg terminal and other Bloomberg platforms as well
[43:40] Speaker 1: as online. on Apple, Spotify, and iHeart. From the city
[43:44] Speaker 1: of San Francisco, thank you so much for being with
[43:46] Speaker 1: us to start your week in the world of technology.
[43:48] Speaker 1: This is Bloomberg Tech.

Shows are listed from Apple Podcasts' charts and directory, and Podcast Index where it is used. Every show and episode belongs to its publisher and plays from the publisher's own feed; this site hosts none of them.