# The foundation for the agentic AI era | Arm CEO keynote at COMPUTEX 2026

https://www.youtube.com/watch?v=aEvEoDJoJQY

[00:00] Heat.
[00:00] Heat.
[00:00] Heat.
[08:37] Wow.
[09:02] Come on.
[09:12] Wow.
[10:18] Heat.
[10:18] Heat.
[15:51] Hallelujah.
[16:38] Heat.
[16:38] Heat.
[18:54] N.
[19:31] The keynote are about to start.
[19:34] Please take your seats.
[19:50] Heat.
[19:50] Heat.
[20:13] Hey,
[20:20] love me.
[20:25] Hey, hey, hey.
[21:10] Hey,
[21:14] Hey, hey.
[21:19] Heat.
[21:46] Heat.
[21:46] Heat.
[27:33] Heat up here.
[27:59] Heat.
[28:21] Heat up.
[28:42] Here.
[28:47] Heat.
[28:47] Heat.
[28:47] Heat.
[29:21] There's a specific kind of silence,
[29:27] Right before the world changes,
[29:30] it's that split second of realization.
[29:34] The moment when you know that life as
[29:37] you've lived it is about to become
[29:39] something
[29:41] entirely new.
[29:51] For more than 35 years, we've learned to recognize that moment.
[29:57] From Cambridge to the US to Taiwan to the world, we partnered with our ecosystem, solving problems together.
[30:12] We've been here from the start of AI, preparing the world for what's next.
[30:20] powering intelligence everywhere in how we live, work, play, and move together.
[30:35] This is our moment.
[30:37] Because when you know you have the power to change everything, you step forward.
[31:06] Renee H. Please welcome Arms Chief Executive Officer Renee Hus.
[31:21] Nihow, welcome.
[31:25] I apologize for uh the delay, but we will uh get moving as quickly as possible.
[31:29] We have a love a lot of things to share with you uh this afternoon.
[31:33] June means Computex and June means a muggy evening and afternoon in Taipei.
[31:40] But it is wonderful to be back here.
[31:43] I think my first competex was 2004 2005ish.
[31:49] So it's 20 plus years plus or minus when when COVID hit.
[31:52] ARM started in 1990 and it was not long after 1990 that Taiwan and ARM started a relationship.
[32:01] Taiwan has built ARM.
[32:05] We are nowhere without the ecosystem and partners that
[32:08] exist inside Taiwan.
[32:12] Now going back in time, we think probably around 1993ish, a few years after we started, the first ARM chip was designed here.
[32:20] Those were early days.
[32:23] SOC's were a kind of a foreign thing.
[32:26] Design tools, physical design, EDA that could support SOC really didn't exist.
[32:32] But we were working with a with ETR back in the day, uh, who did some initial work with us to test out our IP, our methodologies.
[32:41] Not long after that, the first ARM chip manufactured in Taiwan.
[32:48] We believe it was TSMC.
[32:50] We're looking back.
[32:50] It may have been UMC.
[32:52] It was some early test chips.
[32:53] We didn't get into production really until later in the decade.
[32:57] But the first ARM chip was packaged and tested here.
[32:58] So really not long after ARM started, we were we were linked to Taiwan.
[33:03] Some of the significant volumes though that really embody what ARM is all about started in
[33:11] The 2000s.
[33:11] And this is before the iPod.
[33:14] If folks remember these little tiny MP3 players that had maybe 256 songs uh fit in your pocket from Creative, Diamond, Rio, companies like that.
[33:24] Those were all ARM based.
[33:27] And of course then the iPod which in many ways was uh the catapult for the ARM technology being uh everywhere was here and that was in a chip designed by portal player that went into the the very first MP3 player that took volume was the iPod.
[33:46] But it was really in 2008 when we had the revolution that was grown relative to mobile.
[33:55] Here we go.
[33:55] The mobile revolution was really launched in Taiwan.
[33:57] Now we were involved uh obviously with the early GSM phones as folks know from from Nokia and LG etc.
[34:05] But it was really the launch of the iPhone and then the Android phones.
[34:11] that soon followed and that revolution really really launched the growth of ARM into uh a set of volumes we've not seen before.
[34:19] So it was really that period that was the most significant for us.
[34:23] And today, and I'll talk more about ARM server CPUs, 100% of those CPUs are built here.
[34:29] And when we look in aggregate across everything that we've done in our history with all our partners, about 250 billion chips have been built in Taiwan, more than any other region in the planet.
[34:52] I I cannot tell you the gratitude we have as a company for the ecosystem, the people, the talent, the partners here.
[34:59] ARM is nowhere without the partners of Taiwan.
[35:04] Now, some very cool products have come out of the Taiwan ecosystem.
[35:06] When we look at the edge, products such as the
[35:11] Amazon Edge, Oppo Vivo phones, Apple MacBooks, a product I use constantly.
[35:18] I don't mean this as a promo, but these Meta Ray-B band glasses, they are amazing.
[35:21] I use them for phone calls, videos, messages, all here in Taiwan.
[35:28] physical AI, the humanoids, the most advanced in the world, whether it's Tesla, Figure, Techmen, all the chips here built in the Taiwan ecosystem.
[35:42] And then of course cloud AI, whether it's the TPU racks, the racks by Nvidia, Graviton, everything here, as I mentioned, 100% of our ecosystem is built in Taiwan.
[36:00] So without Taiwan, there really is no ARM.
[36:04] Thank you again.
[36:11] Now, what seems like a long time ago and in the world that we're living in with AI, uh we're living in lightspeed.
[36:19] We did an event uh called ARM everywhere back on March 24th.
[36:24] And at that time we were looking at what was going on relative to the growth of agents and agentic AI.
[36:33] And at that time and this is March 24th, not so long ago, showed a slide about the growth of open claw uh relative to Linux and Kubernetes.
[36:42] GitHub stars on the left are exactly what you think they are.
[36:47] They are stars that rate the popularity or the stickiness of a of a certain application.
[36:55] Open Claw reached levels almost beyond parabolic in terms of the takeoff.
[37:00] And this is back in March 24th.
[37:03] And what that told us was that the growth of these agentic platforms were driving demand for CPUs in a way we had not seen.
[37:13] before.
[37:16] And the logic behind that is quite simple.
[37:20] GPUs, XPUs are amazing at generating tokens.
[37:23] That is their purpose.
[37:25] Whether it's training to generate the learning or inference to deliver the tokens, the token machine, the token factory is the accelerator.
[37:32] But agents, unlike humans, don't sleep.
[37:36] And agents beget agents that beget agents.
[37:38] And all of those tokens that need to be distributed, managed, orchestrated, delivered to the destination, that's only a workload that CPUs can do.
[37:48] CPUs of course in conjunction with a with a full system design.
[37:53] So we made a comment back on March 24th and I think we were probably one of the very first to do this that said we believe going forward that four times the number of CPU cores needed in the same power envelope going forward.
[38:10] Now that multiplier I end up getting so many questions relative to show me the math
[38:15] and how do you figure that out and not long after that you started hearing numbers of 4x 8x 10x it's a hard number to predict just based upon the growth rate of these agents but what we do know is as follows.
[38:30] if we look at today what we're seeing in terms of gentic growth even fast forwarding from the 24th of March March.
[38:41] This is just exploding.
[38:44] We're seeing this with SAS companies, whether it's Snowflake or Salesforce or Service Now who are developing all the agents relative to running in the backlane.
[38:57] The explosion of anthropic with claude code, codecs from OpenAI, all these agentic workloads are driving in more demand.
[39:06] And what that does in turn is drive a very very significant growth clicker is doing here in terms of where the CPUs go.
[39:14] So if you change the access on the Y
[39:18] side to units and you look forward in terms of what the growth rate looks like.
[39:23] CPUs are even growing faster than we had thought.
[39:26] And we are seeing this across the board.
[39:29] It's not just ARM.
[39:31] Of course, I'll be promoting ARM a little bit more later, but we're seeing this from everyone who's in the CPU business.
[39:37] The demand for these CPUs continues to explode because the agents beget agents beget agents.
[39:45] Now, is the number four one?
[39:48] Is the number six to one?
[39:50] Is the number 8 to one?
[39:52] I don't know.
[39:55] But what I do know is that it's getting bigger.
[39:58] that the agents continue to accelerate relative to the growth and with that CPU growth is also raising.
[40:06] We threw out a number back on March 24th around a CPU TAM in five years going to north of 100 $120 billion.
[40:17] And again, at the time when we did that event, we had
[40:18] A lot of questions from media investors, analysts saying that number seems a little too aggressive.
[40:26] Not sure how you got there.
[40:29] Fast forward, the numbers that people are talking about are almost twice that number, if not larger.
[40:34] What we do know is that AI at Gentic workloads because of the more tokens you generate, the more information that's being used, the more that they are at Gentic drives demand for compute.
[40:48] And of course, we have an answer for that.
[40:51] The ARM AGI CPU.
[41:00] Now, this CPU, as I mentioned before, 100% built in Taiwan.
[41:06] I'm going to show you a video that we showed on March 24th.
[41:08] Going to show it to you again for those that didn't see it.
[41:10] I want to show it again because frankly, I love it.
[41:14] It's a great video and says everything you want to know about the product, but also emphasizes the importance of the
[41:20] Taiwan ecosystem.
[41:36] Heat.
[41:58] Heat
[42:13] up
[42:15] here.
[42:23] Heat
[42:36] up
[43:08] I think I can watch that video every
[43:10] single day. Uh I I just get so
[43:12] motivated, enthused by by what I see
[43:14] there. So the ARM AJICPU
[43:18] built in Taiwan. TSMC, our partner, we
[43:21] are now in production of this product.
[43:24] One of the things that we emphasized
[43:25] early on when we talked about
[43:27] potentially delivering solutions into
[43:29] the marketplace was that we didn't want
[43:31] to talk about the product until we had
[43:34] customers, the product was shipping and
[43:36] equally as importantly that we had
[43:38] partners who could help deliver the
[43:40] product to market. We understand that in
[43:42] this world it's not just about
[43:44] delivering a chip, but it's delivering a
[43:46] full system with partners. And we've
[43:49] worked with some of the best on the
[43:52] planet all here in Taiwan. I understand
[43:55] there are actually some that are out
[43:57] outside there in the in the demo area. I
[43:59] think we may even have a full rack I've
[44:02] heard from Super Micro sitting out
[44:03] there. But whether it's ASRock or TSMC,
[44:06] our FAB partner, Quanta,
[44:09] Incycle, Super Micro, ASPeed, all
[44:12] fantastic partners who enabled our
[44:14] ecosystem to deliver amazing solutions.
[44:18] Now, this product comes in two flavors
[44:21] from a system standpoint
[44:23] and one of the things that we really
[44:25] emphasize with the RMA GI CPU is maximum
[44:28] performance, density, and efficiency. Of
[44:32] course, our hallmark is around energy
[44:34] efficiency. We were born from mobile
[44:37] phones. We designed a custom CPU way
[44:40] back in the day that had to fit into a
[44:42] plastic package and run off batteries.
[44:45] And that is a mindset that sits inside
[44:47] our engineers in everything that we do.
[44:49] And it translates to amazing solutions
[44:52] and products. An aircooled rack, 36 kW,
[44:56] 8,000 cores,
[44:59] and a liquid cooled rack that has over
[45:02] 45,000 cores, 200 kilowatts. So, two
[45:06] different solutions.
[45:08] But what's key about this product line
[45:11] is the performance per rack, performance
[45:15] per watt.
[45:17] Two times the performance per rack
[45:20] versus a comparable x86 system.
[45:22] Basically means same power envelope, two
[45:24] times the benefit in terms of
[45:26] performance.
[45:27] If you want a half the power, you still
[45:30] have equivalent performance. So it's
[45:33] incredibly
[45:34] efficient. But more importantly, when
[45:36] you think about what goes into these
[45:39] giant data centers,
[45:41] and we're seeing announcements literally
[45:43] daily. In fact, the parent company of
[45:45] ARM softbank just announced a
[45:47] partnership in France for a 5 gawatt
[45:50] data center. These data centers are
[45:53] incredibly capital intensive. The energy
[45:55] uh costs are huge. So having the benefit
[45:59] of performance per rack, more CPU in the
[46:01] same power envelope has huge huge
[46:04] benefits versus the competition. We
[46:06] estimate about 10 gawatt of capacity
[46:09] over 10 billion up to 10 billion of
[46:12] savings. But as we go forward and we
[46:16] have more and more CPUs inside the
[46:18] systems, you'll get even more benefit
[46:21] relative to using the ARM AGI CPU.
[46:25] Now, we were super proud back in March
[46:29] to talk about our partners, people who
[46:31] had embraced the solution, customers
[46:33] that we had signed up, Meta, Rebellions,
[46:36] SAP, Cerebrus, OpenAI, SK Telecom.
[46:41] And that was just on March 24th, and we
[46:43] talked about our customer base and who
[46:45] had adopted the product. I'm proud to
[46:48] say that since that time,
[46:51] even more companies have joined the
[46:52] family.
[46:54] Oracle huge partner with OCI. We have a
[46:57] long history with Oracle. They've now
[46:59] joined the ARGIC CPU family as well as
[47:02] Bite Dance. Two new partners part of the
[47:05] family validating that the Arma GIC CPU
[47:09] solves real world problems.
[47:12] Now,
[47:13] we talked about this back in March and I
[47:15] want to emphasize it again. We are now a
[47:18] full endtoend solution provider.
[47:22] So while we do have production silicon
[47:24] of the ARM AGI CPU, not everyone wants
[47:27] to buy the ARM AGI CPU and that's okay.
[47:30] We have compute subsystems,
[47:33] many partners who take that and we have
[47:35] just standalone IP in this space.
[47:38] Whether it's Google, whether it's
[47:39] Amazon, whether it's Nvidia, whether
[47:41] it's Microsoft, we have many, many
[47:43] customers who are on the lefth hand side
[47:45] of that slide and we intend to provide
[47:48] solutions to whatever the customers want
[47:50] to see. And that is very important
[47:53] because the momentum is really
[47:55] increasing for us now with Aentic AI and
[48:00] whether it's our own CPU or our
[48:02] partners. very significant announcement
[48:04] took place last month where Google
[48:08] announced for their TPU 8T and 8i that
[48:13] the head node, the CPU that interfaces
[48:16] into the accelerators, is going to move
[48:19] from x86 to Axion, which is their
[48:22] internal chip using ARM neoverse.
[48:26] 60%
[48:29] less power at the same performance.
[48:32] Andy Jasse had a great quote I think one
[48:36] of the earnings calls that basically
[48:38] said for Graviton.
[48:40] We had two customers say can we buy
[48:43] everything that you have?
[48:45] Graviton now has half more than half of
[48:48] their design starts are based on
[48:52] Graviton versus x86. From a few years
[48:55] ago that was zero.
[48:59] And of course, Nvidia
[49:01] who announced Vera,
[49:04] amazing partners. Vera is an amazing
[49:06] product. The list of partners is far
[49:09] larger than here. I didn't have a slide
[49:10] big enough for all of them, but Nvidia's
[49:12] had a tremendous momentum with uh with
[49:14] Vera.
[49:16] Now, our intentions are very clear for
[49:18] the ARM AGI CPU. We intend to be in this
[49:21] for the long term. It's
[49:23] multigenerational.
[49:24] ARM AGIC CPU 2 is already underway. And
[49:28] as you can imagine, it has more cores,
[49:30] more power efficient, better
[49:32] performance.
[49:34] An RMIG CPU 3 is on the way. But these
[49:37] are all based on the compute subsystems
[49:40] that we intend to deliver along with the
[49:42] chips and they'll be lined up roughly on
[49:45] the same cadence. So the CSS's that we
[49:48] deliver to our partners, those are what
[49:50] we use to enable our end devices.
[49:53] So that's ARM AGI CPU, which has had
[49:55] incredible momentum.
[49:58] Now I want to switch gears a little bit
[50:00] because Comp Computex to me always
[50:03] having come here 20 years ago for the
[50:06] very first one was always about the old
[50:09] exhibition hall, floppy disc
[50:12] controllers, USB cables,
[50:16] all kinds of things in terms of it
[50:19] malls.
[50:20] And you could go into these shops and
[50:23] buy almost anything under the sun. It
[50:25] was like a mini fries,
[50:28] 20 of them on a floor in a building that
[50:30] was 10 stories high. And that's how
[50:33] people bought PCs back in the day in
[50:36] terms of how they shop for them.
[50:40] And if you think about how these PCs
[50:45] uh were built
[50:47] and how we used to buy, it was very
[50:50] interesting. you you'd have literally
[50:53] every single price point you could think
[50:55] about, whether it was a base entry
[50:57] laptop. Raise your hand if you remember
[50:59] the netbook.
[51:02] I knew the Nvidia guys would remember
[51:03] that one. We have battle scars from that
[51:06] one.
[51:07] All the way up to high-end gaming
[51:10] machines. But literally, these units
[51:12] were priced at $50 price points. He had
[51:15] feeds and speeds,
[51:17] clock frequency, memory size, etc., etc.
[51:21] And everybody was trying to position for
[51:22] the slice of the pie.
[51:25] So much has changed obviously not only
[51:28] in how we buy PCs,
[51:32] but more importantly, how we use these
[51:34] products. Okay, how we use the products
[51:38] has really really evolved with obviously
[51:40] what the smartphone has done,
[51:43] what the web has done, what applications
[51:46] have done and what we see is that
[51:48] they've really started to bifurcate into
[51:53] kind of two two areas I would say. One
[51:57] is, and I think many of you can identify
[52:00] this on the bottom left, is I need a
[52:03] machine that is on the go. Battery life
[52:06] is really good, connects everywhere, and
[52:10] I need it to kind of look like a large
[52:12] phone with a keyboard where I can do
[52:15] work, but it maps very closely to what
[52:18] my phone does. And if I think of about
[52:20] myself personally, I have one of these
[52:22] flip phones which I use for reading
[52:25] documents and reviewing presentations.
[52:27] And I'm a CEO so I create very little
[52:29] these days. I I review many things. But
[52:33] what I find is I go back and forth a lot
[52:36] between that smartphone that flips like
[52:38] a tablet into the PC. But it's really
[52:41] super important that the PC and phone
[52:45] are synchronized and they can do things
[52:46] back and forth very very quickly.
[52:49] There's also an extreme performance
[52:51] workload and that is I'm either running
[52:54] agents,
[52:56] I'm either running models, I'm doing
[52:58] some development work, I need some very
[53:01] very extreme level of performance. So
[53:05] there's really two different components
[53:07] in two different areas in terms of how
[53:08] they all work.
[53:10] So
[53:12] only ARM
[53:14] really enables this for PCs. And I think
[53:17] that's a very very key distinction in
[53:19] terms of the way we used to think about
[53:21] this category back in the day where
[53:23] literally you had every single price
[53:25] point covered, every single feed and
[53:28] speed.
[53:29] Now you want two different ends of the
[53:31] spectrum.
[53:33] And whether it's long battery life,
[53:35] great AI experience, we're in that
[53:38] bottom category. But if you also want
[53:40] the agentic type of performance, we're
[53:43] there as well.
[53:45] Now, specifically when we look at the
[53:47] units that are there,
[53:50] you can see that you've got the Acer
[53:53] device, Mac Neo, pretty interesting
[53:56] product, the Google Book, Microsoft
[53:59] Surface,
[54:01] Mac Studio, of course, the Nvidia RTX
[54:04] Spark, which was just announced, which
[54:06] I'll talk about. But these two broad
[54:08] categories are very unique to ARM. And I
[54:11] get lots of questions, you know, over
[54:13] the years about Windows and ARM and when
[54:15] is ARM going to really take place to be
[54:17] a significant player in laptops and the
[54:19] comput space. I would argue that we are
[54:21] now actually there
[54:24] because when we look across the spectrum
[54:28] of the operating systems that are
[54:30] supported whether it's Linux whether
[54:32] it's MacOSS which is 100% on ARM today
[54:38] Chrome Windows only ARM can enable this
[54:45] across the board and this would not be
[54:48] done without huge huge huge cooperation
[54:50] from all of our partners who are listed
[54:52] there, the the folks on the operating
[54:54] side, operating system side that we work
[54:56] so closely with. We've worked for
[54:58] decades with Apple. We've worked for
[55:01] decades with with Google and Microsoft.
[55:03] This work does not happen overnight.
[55:06] There is a huge amount of effort to go
[55:07] off and make this happen. And I want to
[55:09] give an applause and thanks to all of
[55:10] our partners to uh to make this work.
[55:19] Now I want to talk about a product that
[55:22] we knew was being worked on and we are
[55:25] proud to be a partner with Nvidia on the
[55:29] RTX Spark powered by ARM.
[55:33] 20 cores ARMbased cores in the custom
[55:37] grace CPU.
[55:40] I believe that is the most CPU cores
[55:42] that you can find in a laptop anywhere.
[55:46] But when you pair it with Blackwell, the
[55:50] world's most powerful GPU for Agent, you
[55:54] have an incredibly special product. One
[55:56] pedlop of FP4,
[55:59] huge amount of memory, full Windows
[56:03] native on ARM,
[56:05] amazing product.
[56:15] And of course, as you'd suspect,
[56:18] partners who are there already, Acer,
[56:21] Asus, Dell, Gigabyte, HP, Lenovo,
[56:26] Microsoft, MSI, I think I saw a Surface
[56:30] Ultra uh that was announced. An amazing
[56:33] product. Congratulations again to the
[56:35] NVIDIA team for making all this happen.
[56:38] Now
[56:40] we're
[56:42] our role here was working very closely
[56:45] with Nvidia and with MediaTek
[56:48] using our CSS strategy.
[56:52] And again for those who are not familiar
[56:54] with what our compute subsystems do, the
[56:58] CSS is basically the building blocks
[57:00] that we use to put together everything
[57:03] to build a full end solution system. the
[57:06] CPUs, the GPUs, the system IP, the
[57:10] memory controllers, everything that goes
[57:12] into building a custom SOC. We provide
[57:16] these to our customers. We did this with
[57:17] MediaTek as either full solutions they
[57:21] can take or building blocks that they
[57:23] can start with. So we see a very
[57:27] significant opportunity again given the
[57:29] strategy we talked about with IP and
[57:31] compute subsystems around the armagentic
[57:34] uh CPU very very similar with what we're
[57:38] doing with the CPUs for the CSS's and I
[57:42] think the PC space is going to be a a
[57:44] very very interesting domain uh as I
[57:47] said going forward because with these
[57:49] use cases on the bottom left again the
[57:51] the kind of use that I am relative to
[57:53] using the the systems for uh for
[57:56] creation and things of that nature.
[58:00] The high-end systems when we start
[58:02] thinking about where agents can go and
[58:04] how agents interface with with us, it's
[58:07] going to be a very very different domain
[58:09] and I think this product from Avidia has
[58:11] really uh demonstrated its capability.
[58:16] So, I'm not sure if the systems are
[58:18] available yet, but we actually got
[58:22] access to uh some of the hardware and
[58:24] technology and we decided to uh to give
[58:28] it for a spin.
[58:31] Complete uh
[58:34] surgeons general warning here. This
[58:37] following video was AI generated. So,
[58:41] please don't uh have your legal teams
[58:43] contact us.
[58:46] But uh let's take a quick look.
[59:03] Back
[59:20] up.
[59:23] Back.
[59:42] Oh,
[59:53] heat, heat.
[01:00:18] We are ready for you.
[01:00:31] Now, I know you're probably saying,
[01:00:34] "I'm not sure that's AI because the dude
[01:00:36] always wears the same clothes." But on
[01:00:38] the other hand, uh those are events that
[01:00:40] I would not actually do myself, but I
[01:00:42] think that's just a small example of the
[01:00:44] kind of creation that can be done, you
[01:00:46] know, on these computers and uh where I
[01:00:48] think we're going to go with Agentic AI.
[01:00:52] Now, I want to be able to uh talk more
[01:00:55] about the product, but I'm kind of
[01:00:56] thinking that there's probably someone
[01:00:58] better than me to join me on stage to
[01:01:01] talk about the RTX Spark and everything
[01:01:03] that Nvidia does. So, I'm going to
[01:01:05] introduce a special guest here. My
[01:01:08] clicker behaves.
[01:01:25] That's a That's a pretty cool video of
[01:01:28] Renee. Superstar.
[01:01:31] >> Action hero.
[01:01:32] >> Do you buy
[01:01:33] >> Not just a superstar, he's an action
[01:01:35] hero.
[01:01:35] >> I think the night market was the part I
[01:01:37] thought was the coolest.
[01:01:38] >> Yep. Well, that's that's the most
[01:01:40] exciting part of your video.
[01:01:41] >> Yeah. Well, thank you for joining. I
[01:01:43] appreciate it.
[01:01:44] >> Yeah. So,
[01:01:46] >> so tell me Jensen, uh, congratulations
[01:01:49] on the RTX Spark. Amazing.
[01:01:51] >> Thank you.
[01:01:52] >> Uh, Windows on ARM is not a new thing.
[01:01:54] >> Yeah.
[01:01:55] >> Why is this one
[01:01:56] >> going to be different?
[01:01:57] >> Look at Look at his stock price. I
[01:01:59] announce
[01:02:01] >> I announce a product. Look at his stock
[01:02:03] price.
[01:02:05] Every product I announce, his stock
[01:02:07] price goes up.
[01:02:09] >> Nothing happens to mine.
[01:02:14] Let's also um let's also stay for the
[01:02:17] record
[01:02:18] >> that that's I'm very happy about that.
[01:02:20] >> Let's also stay for the record that you
[01:02:21] were a shareholder and you sold.
[01:02:23] >> Yeah. Yeah. Well,
[01:02:25] >> so
[01:02:25] >> I needed a cash.
[01:02:29] >> So, how
[01:02:30] >> what were we talking about?
[01:02:31] >> RTX Spark.
[01:02:32] >> RTX Spark.
[01:02:33] >> How was it going to be different this
[01:02:34] time?
[01:02:34] >> Well, we wanted to reinvent the
[01:02:37] computer. You know the PC has been here
[01:02:39] for 40 years and the operating system uh
[01:02:42] code written by hand is now going to be
[01:02:45] uh replaced with applications that are
[01:02:47] agentic. Now these agentic systems
[01:02:49] agentic agentic AIs will use the PC will
[01:02:53] use the tools in the PC. And so uh when
[01:02:57] we imagined this future uh we thought
[01:03:00] let's see what how would we change the
[01:03:01] architecture and how would we change the
[01:03:04] operating system and um uh reinvent the
[01:03:07] computer and um you know that that's
[01:03:09] kind of where we are and so one of the
[01:03:11] things that that we realized is that an
[01:03:14] agentic system really wants to have
[01:03:16] excellent CPUs which is the reason why
[01:03:18] we used ARM and it has a 20 core CPU has
[01:03:21] to have excellent single threaded
[01:03:22] performance the parameters the memory
[01:03:25] has to hold a lot of parameters and so
[01:03:27] we uh created a new uh numerical format
[01:03:30] called MVFP4 so that we can compress the
[01:03:34] large language models as much as
[01:03:36] possible and fit a very smart AI into
[01:03:39] the system memory. We also wanted to
[01:03:42] unite CUDA that is uh for accelerated
[01:03:46] computing and CUDA tiles our tensor core
[01:03:49] processing into one processor. And the
[01:03:51] reason for that is because when you're
[01:03:53] operating these agents and they're
[01:03:55] thinking and they're using the tools,
[01:03:57] the agents are fast. And when the agents
[01:04:00] are fast, they expect the tools to
[01:04:02] respond quickly. And so that's why we're
[01:04:04] accelerating all of the tools. We're
[01:04:06] accelerating Adobe. Adobe announced
[01:04:08] they're going to rearchitect Adobe
[01:04:10] Photoshop and Premiere so that it's CUDA
[01:04:12] accelerated and a aentically accessible
[01:04:16] and so we're accelerating applications.
[01:04:18] We accelerated Blender with RTX. We
[01:04:20] accelerated you know we're going to
[01:04:22] accelerate everything. We accelerate
[01:04:23] Adobe Autodesk the so seammens we're
[01:04:26] going to accelerate every tool. And once
[01:04:29] these tools are accelerated then they
[01:04:30] can respond to the agents very quickly.
[01:04:32] And so we now in order to build this
[01:04:35] computer this SOC unless you have the
[01:04:38] ability to integrate with the CPU and
[01:04:42] adapt the CPU to exactly the shape of
[01:04:44] the computer it's really quite
[01:04:45] impossible which is the reason why ARM
[01:04:47] is perfect.
[01:04:48] >> Well thank thank you and when you think
[01:04:49] about the agents running low
[01:04:52] >> and the key word there is ARM is
[01:04:53] perfect.
[01:05:00] The other key word is thank you. Um,
[01:05:03] a agents running locally.
[01:05:04] >> Naughty naughty.
[01:05:07] >> Yila,
[01:05:09] >> that's not a fair fight at this point.
[01:05:11] >> You're welcome.
[01:05:14] >> Uh, agents running locally versus agents
[01:05:17] running in the cloud. How do you think
[01:05:19] about that as a as a trade-off and where
[01:05:22] do you think that goes over time? Well,
[01:05:24] you know, when ultimately this this the
[01:05:27] computers, these personal computers are
[01:05:29] going to be becoming agents that are
[01:05:31] running all the time. They're
[01:05:32] autonomously use running all the time. I
[01:05:34] could imagine I today if I left my
[01:05:37] laptop at home or I left my laptop in
[01:05:40] the hotel, I won't use it again until I
[01:05:42] get there. But in the future, you just
[01:05:44] pick up your phone and you chat with
[01:05:46] your agent. you're chatting with your PC
[01:05:49] in the future and that you maybe there's
[01:05:52] something that you uh needed to have
[01:05:54] done and sent to you. Maybe uh maybe
[01:05:57] there's a a speech I need to have
[01:05:58] quickly written. And so, you know, I'll
[01:06:00] be working with my agent, working with
[01:06:03] my assistant, and that is now uh the ARM
[01:06:06] personal computer, right? And so,
[01:06:07] >> so so the PC is working in the back
[01:06:09] while you're not there. It's working.
[01:06:10] >> Yeah. And so, if I want to do something
[01:06:13] that requires a cloud API, of course,
[01:06:15] I'll I'll call it into the cloud API.
[01:06:17] But whatever I can do locally, we're
[01:06:19] going to continue to do on the PC, which
[01:06:20] is kind of the nature of PC.
[01:06:22] >> The nature of a personal computing
[01:06:24] device is that whatever you can do on
[01:06:26] the device, you do.
[01:06:27] >> You don't have to worry about metering.
[01:06:28] You don't have to worry about the time
[01:06:30] spent, but whatever you need to do in
[01:06:32] the cloud, you will.
[01:06:33] >> And when you think about the
[01:06:34] complexities of the models, do you think
[01:06:36] PC performance and architecture can
[01:06:38] scale? I mean, you guys are doing
[01:06:40] incredible work with where Blackwell and
[01:06:42] then Ruben, etc. How do how do you think
[01:06:44] that all maps together in terms of
[01:06:45] scaling the systems? Well, if you look
[01:06:47] at the RTX Spark PC, it's got 128
[01:06:50] gigabytes memory. If it was completely
[01:06:52] compressed into MVFP4,
[01:06:55] then you can have a 100 billion
[01:06:57] parameter model working on your PC all
[01:07:00] the time.
[01:07:00] >> Wow.
[01:07:00] >> And a 100 billion parameter open model,
[01:07:03] say Neotron 3 Super, say that's a really
[01:07:07] really good model. And so it could do a
[01:07:09] lot of the basic work and and whatever
[01:07:11] whatever deep thinking and frontier
[01:07:13] model that you need to use, it's just
[01:07:15] connected to cloud anyhow. And
[01:07:16] >> do you think that changes what happens
[01:07:18] in the cloud in terms of this classic
[01:07:21] client cloud model? I don't do I need as
[01:07:23] much compute in the cloud versus on the
[01:07:25] client or do you think there's just so
[01:07:27] much compute that needs to get done?
[01:07:28] What they just
[01:07:29] >> these agents are going to be they you're
[01:07:31] going to be you have agents and sub
[01:07:33] agents and teams of agents. Um they're
[01:07:35] going to be working uh in the cloud.
[01:07:37] they're going to be working on devices
[01:07:39] and so it's just like today in a lot of
[01:07:41] ways.
[01:07:41] >> Yeah.
[01:07:42] >> Mobile cloud is not cloud only, not
[01:07:44] mobile only, it's mobile and cloud
[01:07:47] >> and so it allows you to have have a
[01:07:49] really great personal computing
[01:07:50] experience, you know, your own
[01:07:52] experience, but whatever you need to
[01:07:54] connect to the cloud, you will. And so
[01:07:56] >> do you think and this may a bit of a
[01:07:57] provocative question, but as as these
[01:07:59] agents are running in the background and
[01:08:00] they're doing a lot of the work, does
[01:08:02] the operating system matter? Is the
[01:08:04] agent really the OS, if you will, and it
[01:08:07] does the work and isn't so reliant on on
[01:08:10] the hood? Where do you think that goes
[01:08:11] over time?
[01:08:12] >> Well, the operating systems get be just
[01:08:14] as important as ever before, if not more
[01:08:15] important. And the reason for that, and
[01:08:17] this is this is the controversial part
[01:08:19] that people say AI comes along, software
[01:08:22] is dead. You know, nothing is more
[01:08:24] nothing is further from the truth. And
[01:08:26] now people are starting to realize that
[01:08:28] >> when agents are here, they're going to
[01:08:30] use tools. And so those tools are more
[01:08:33] important than ever. And so they're
[01:08:35] going to use Adobe Photoshop, they're
[01:08:36] going to use Adobe Premiere, they're
[01:08:38] going to use Canva, they're going to use
[01:08:40] the, you know, they're going to use the
[01:08:41] SOS tool, Seammen's tools, they're going
[01:08:42] to use, you know, tools, whatever they
[01:08:44] have on the device. They're just, this
[01:08:46] is, this is the incredible part today.
[01:08:49] Most of us probably know 10, 15, 20% of
[01:08:53] the features of a tool. If you know how
[01:08:55] to use Photoshop, you know, use
[01:08:57] Lightroom, you're unless you're expert
[01:09:00] like my son, it's kind of hard for you
[01:09:02] to know all of the features. But now
[01:09:04] with your agent, you tell the agent what
[01:09:07] you're looking for and the agents know
[01:09:09] exactly how to use the tools because
[01:09:12] it's re read a skills file. It's
[01:09:13] essentially read the manual of that
[01:09:16] tool. Yeah.
[01:09:16] >> And so now it goes and uses the MCP or
[01:09:19] the CLI connected to that tool and it
[01:09:22] does everything you need to do.
[01:09:23] >> That's great. It's going to unlock.
[01:09:24] >> Yeah. So, it's going to unlock all these
[01:09:26] tools. These tools are going to be more
[01:09:27] useful, more valuable than ever. And
[01:09:29] that these tools run on the operating
[01:09:31] system. So, we're going to need we're
[01:09:32] going to need Windows. We're going to
[01:09:34] need, you know, all these APIs and all
[01:09:35] these tools for a long time.
[01:09:37] >> So, um, Nvidia is involved,
[01:09:39] understatement, in everything around AI.
[01:09:42] I mean, you guys do everything around
[01:09:44] the networking, the systems, you know,
[01:09:46] where all the bottlenecks are. When you
[01:09:48] think about over the next number of
[01:09:50] years, where where are the constraints
[01:09:52] to to growth? Where do you think they
[01:09:55] are?
[01:09:56] >> Well, it's probably going to be
[01:09:57] everywhere. This is this is um at this
[01:09:59] point, if you look at our evolution, uh
[01:10:01] first Hopper was designed for training.
[01:10:05] Then Grace Blackwell was of course great
[01:10:08] at training, but we also specialized
[01:10:10] MVLink72 for inference. And at first
[01:10:13] people thought, you know, inference was
[01:10:14] easy. And we explained to people that
[01:10:17] large language models and to be able to
[01:10:19] inference very quickly and generate
[01:10:21] these tokens as efficiently as possible,
[01:10:23] you're going to need a very complicated
[01:10:25] computer. And so GB or or Grace
[01:10:28] Blackwell MVLink72 is the most efficient
[01:10:32] and we produce the lowest cost tokens in
[01:10:35] the world. Okay. And so that was a big
[01:10:37] breakthrough and now people understand
[01:10:39] that that tok that you want very
[01:10:41] advanced systems to generate tokens at
[01:10:43] very low cost. Vera Rubin took, of
[01:10:47] course, all of that and we we extended
[01:10:50] it to run agents. At first, when I said
[01:10:53] that two years ago, most people had a
[01:10:55] hard time understanding what that meant.
[01:10:56] But now they realize that an agent is
[01:10:59] orchestrating thinking, is using tools,
[01:11:02] it's accessing long-term memory, it's
[01:11:04] dealing with short-term memory, you
[01:11:06] know, working memory, and it's
[01:11:07] compacting me, doing memory compaction
[01:11:10] to to remember to think about what
[01:11:12] should I remember for the future? How do
[01:11:14] I index uh SQL memory? How do I index
[01:11:17] structured memory? How do I index
[01:11:19] unstructured memory? And so how do I
[01:11:20] deal with all of that? That agentic
[01:11:23] system
[01:11:25] is what Vera Rubin is. And it's a large
[01:11:27] system. And and so people are now
[01:11:29] starting to understand that that when
[01:11:31] when we were thinking about agentic
[01:11:33] systems, we were really thinking about
[01:11:34] new computing application pattern and
[01:11:38] that it really requires a new new
[01:11:40] architecture. Well, now the big
[01:11:42] breakthrough of course these agents now
[01:11:45] are producing useful AI and that's the
[01:11:48] reason why all of our growth right your
[01:11:50] growth my growth it's just so incredible
[01:11:52] because when AI becomes useful then the
[01:11:56] tokens that are being generated are
[01:11:58] profitable and when token generation is
[01:12:01] profitable everybody wants to generate a
[01:12:03] trillion times more token. The other
[01:12:06] part is that the agent the application
[01:12:09] this agent compute pattern is a thousand
[01:12:13] times maybe a hundred thousand times and
[01:12:16] depending on the work it's a million
[01:12:18] times more than chatting
[01:12:21] and so you could see that the agents are
[01:12:23] working they're working for you know
[01:12:25] minutes
[01:12:27] sometimes days sometimes weeks and so
[01:12:30] instead of a chatbot which responds from
[01:12:33] one click now the AI I is thinking,
[01:12:36] using tools, reading, thinking some
[01:12:38] more, planning, trying and so the amount
[01:12:41] of tokens that we have to generate has
[01:12:43] increased tremendously. The profitable
[01:12:46] the profitability of the tokens
[01:12:48] obviously is driving demand. So the
[01:12:50] compound effect of need more compute
[01:12:54] with more demand that compounded effect
[01:12:56] is what you and I are experiencing. And
[01:12:58] so we're we're seeing you know
[01:12:59] constraints almost everywhere. In our
[01:13:01] case, you know, we were fortunate that
[01:13:03] we we planned, you know, one of the best
[01:13:06] things about ARM is that they don't have
[01:13:07] to worry about the supply chain.
[01:13:11] You know, you know, the supply chain of
[01:13:13] IP is electrons and you could use as
[01:13:15] many electrons as you need. Okay. And so
[01:13:18] I love his business model. I mean, I as
[01:13:20] you know, I know I tried to buy it.
[01:13:23] >> I tried to
[01:13:24] >> I tried to become ARM, you know.
[01:13:27] >> We we were willing. I was trying to
[01:13:30] become art. Were you really?
[01:13:32] Renee Renee and I used to work together
[01:13:35] and then we tried to work together
[01:13:36] again. But anyways, that was okay. I'm
[01:13:40] not I'm sad still. I'm a little sad but
[01:13:43] but this is a happy this is a happy
[01:13:45] meeting. So my my point is in our case
[01:13:48] we saw agents coming and we saw Bar
[01:13:50] Rubin coming. So we did a good job
[01:13:52] planning our supply chain and so our
[01:13:54] supply chain can support our very robust
[01:13:57] growth. We grew almost 100%
[01:13:59] year-over-year this year. We're going to
[01:14:00] grow very aggressively next year. And so
[01:14:02] we have our supply chain could support
[01:14:04] our growth. But the fact of the matter
[01:14:06] is demand is even higher than that.
[01:14:08] >> Yeah. I we was talking with uh with Cece
[01:14:10] and Kevin uh this week and they were
[01:14:13] saying, you know, at some point gravity
[01:14:15] has to take over. They've never seen
[01:14:16] four consecutive years of a
[01:14:17] semiconductor cycle that looks this
[01:14:19] good. But when you look at the things
[01:14:21] that you just described, there's no
[01:14:23] reason it can't continue in terms of the
[01:14:25] fundamentals.
[01:14:25] >> That's right. Take a step back. What's
[01:14:27] happening? Yeah.
[01:14:27] >> Take a step back and think what's
[01:14:28] happening. What's happening is the
[01:14:30] computer industry was limited by the
[01:14:33] number of people using the computers.
[01:14:35] >> Yep.
[01:14:37] >> And now we have agents that are
[01:14:39] autonomously using computers.
[01:14:41] >> And so we're going to have instead of 1
[01:14:43] billion humans using computers, we will
[01:14:45] have tens of billions, maybe more than
[01:14:49] that of agents and robots and
[01:14:51] self-driving cars using computers.
[01:14:54] And so the question is how large can the
[01:14:56] computer industry be?
[01:14:57] >> And so you know my sense is that at this
[01:14:59] point it's a foregone conclusion that
[01:15:01] what is a trillion dollar multi-t
[01:15:04] trillion dollar industry is likely 10
[01:15:05] times larger.
[01:15:06] >> Yeah.
[01:15:07] >> And so we're on our way to
[01:15:09] >> and that's why Nvidia is the, you know,
[01:15:10] the largest market cap company in the
[01:15:12] world. And if you combine the two
[01:15:13] companies, we'd be the largest in the
[01:15:14] world
[01:15:16] >> still.
[01:15:17] >> I I love that. I love that. That's such
[01:15:19] a great idea.
[01:15:21] So, you know, thank you.
[01:15:23] >> Congratulations on RGX Spark. Just
[01:15:25] amazing.
[01:15:26] >> Well, congratulations on everything you
[01:15:27] guys are doing.
[01:15:28] >> I have a small gift for you.
[01:15:30] >> Really?
[01:15:30] >> Someone's going to give here.
[01:15:32] >> Yeah.
[01:15:35] >> So, for those who may not recognize what
[01:15:37] this is, and I'm going to sign it, this
[01:15:39] is this is very very real, by the way.
[01:15:41] The very first this is the Jensen talks
[01:15:44] a lot about resiliency and sticking with
[01:15:47] things.
[01:15:48] Tegra 3 was the first Windows on ARM
[01:15:51] laptop that was announced.
[01:15:52] >> How come? How come when we were younger?
[01:16:01] >> I have to tell you, I think I aged
[01:16:04] better.
[01:16:08] >> Do you guys agree?
[01:16:13] >> I feel I aged pretty well.
[01:16:15] >> Come here. Here.
[01:16:16] >> You're You're my guest better. to you.
[01:16:19] >> If I sign it back to you, it's got
[01:16:21] treasure.
[01:16:22] >> No, you sign it back to me. There's a
[01:16:24] contract. There's invoices.
[01:16:27] >> We can't do that. We know that game. All
[01:16:29] right.
[01:16:30] >> Thank you very much. Thanks, guys.
[01:16:34] >> By arm. By arm. I tried.
[01:16:45] >> One of those things was real up there.
[01:16:47] That actually was a real system that we
[01:16:48] worked on and and and fish and cow those
[01:16:51] guys will remember on that. Um I think I
[01:16:53] aged a little bit better than he did by
[01:16:54] the way. Uh so to wrap up
[01:16:58] oneic platform cloud to edge showed
[01:17:02] these these products before it's the ARM
[01:17:04] AI compute platform that enables systems
[01:17:07] from the very very smallest to the very
[01:17:09] very largest. And we do this through a
[01:17:11] very consistent effort with software. 22
[01:17:14] million developers, the largest
[01:17:16] developer community across the planet
[01:17:19] for any comput platform. But as I said,
[01:17:22] none of this happens without incredible
[01:17:25] cooperation and dedication from our
[01:17:27] partners. And again, I just want to say
[01:17:29] thank you to Taiwan. ARM is nowhere
[01:17:32] without Taiwan, the ecosystem, the
[01:17:34] people, the engineers, the supply chain
[01:17:35] managers. Thank you so much for
[01:17:37] everything you've done. Thank you for
[01:17:39] attending today.