H20s to China + 15% with Chris Miller and Lennart Heim

0:00:06 emergency podcast we’re bringing back leonard heim of rand and chris miller you guys know
0:00:15 chris miller is um to talk the new h20 drama uh we went from banning it in april to now selling it
0:00:22 with a 15 export fee all right leonard why don’t you start our narratives what is the h20 why should
0:00:27 you care about it and what were the first few months of the trump administration doing when it
0:00:33 came to this chip the h20 is this chip which nvidia designed as a response to export controls in 2023
0:00:38 right it’s a typical game you draw some lines and then new chips get created right below the lines
0:00:43 and the h20 is such an example but it did a neat trick it maxed out the specifications which are
0:00:49 not controlled memory bandwidth so putting on the best high bandwidth memory um world world currently
0:00:54 has on this chip and created a export control compliant chip which i think was introduced at
0:01:01 beginning of 2024 so a couple of months after the update and it was sold over 2024 lots of interest
0:01:09 in this chip and when then the uh term administration started in january they well we can say the bind
0:01:12 administration didn’t get to it i think it was definitely a way out of this problem many of the
0:01:16 officials came out many of them even favor but they just never got to banning to it because again
0:01:19 many stakeholders many different opinions and also just kind of running out of time
0:01:27 but trump then banned this chip as we saw reported in april 2025 not by the normal means how you do it
0:01:31 not via passing regulation via using a cool tool called is informed letters because they’re pretty
0:01:35 fast and you can just send a letter to the companies which produce these chips and tell them like hey
0:01:40 you can’t sell these chips anymore because we expect there’s an export control violation going on and
0:01:45 we actually think this chip is too good so that’s where we started so the chip was banned from my personal
0:01:50 point of view a big success i think this chip should not be solved um we need to reduce our
0:01:56 thresholds this is just simply too good of a chip and well then then this was the last latest status and
0:02:01 i think in the last few months last few weeks we saw some flip-flopping back and forth and every day
0:02:05 reveals more information and i don’t know while we talk probably more things will come out over time
0:02:08 all right so this was president trump on monday
0:02:15 two questions one about china one about russia if i could on china your administration agreed to send
0:02:23 uh the most advanced or advanced nvidia and amd chips no obsolete no obsolete chin well and then 15 percent
0:02:30 of the problem in the 20s no no that’s this is an old chip that china already has
0:02:41 jensen who’s a great guy and nvidia uh the chip that we’re talking about the h20 it’s uh it’s an old chip
0:02:49 uh china already has it in a different form different name but they have it uh or they have
0:02:55 a combination of two will make up for it and even then some now jensen also has jensen’s a very brilliant
0:03:00 guy and jensen also has a new chip the blackwell do you know what the blackwell is the blackwell is
0:03:09 super duper advanced i wouldn’t make a deal with that although uh it’s possible i’d make a deal uh
0:03:21 a somewhat uh enhanced in a negative way blackwell in other words take 30 to 50 off of it but that’s the
0:03:26 latest and the greatest in the world nobody has it they won’t have it for five years but the h20 is
0:03:31 obsolete you know it’s one of those things but it still has a market so i said listen i want 20 percent
0:03:36 if i’m going to prove this for you for the country for our country for the u.s i don’t want it myself
0:03:43 you know every time i say like uh like 747 i want i want yeah for the air force
0:03:50 that’s so i just wanted so when i say i want 20. i want for the country i only care about the country
0:03:56 i don’t care about myself and he said would you make it 15. so we negotiate a little deal
0:04:07 so he’s selling a essentially old chip that huawei has a similar chip a chip that does the same thing
0:04:12 and i said good if i’m going to give it to you because they have a uh you know they have a stopper
0:04:19 what we call a stopper not allowed to do it it restricted it’s really known as a restrictive covenant
0:04:24 and i said if i’m going to do that i want you to pay us as a country something because i’m giving
0:04:32 you a release i released them only from the h20 now on the blackwell i think he’s coming to see me again
0:04:41 about that but that will be a uh unenhanced version of the big one like i don’t know if you know it we
0:04:48 will sometimes sell fighter jets to a country and we’ll give them 20 percent less than we have do you
0:04:53 know what i mean it’s just yeah it’s crazy to see like this discussion also to the level of sophistication
0:04:58 here right knowing the h20s knowing the blackwells by name knowing what huawei it’s clearly a big topic
0:05:04 now where there’s this whole experiment of things in the spotlight so i think it’s a good moment to
0:05:10 kind of take a step back and look at the arguments uh in favor and against uh selling china ai chips
0:05:20 so against selling you have broadly the argument that uh selling uh ai chips helps upgrade the chinese
0:05:27 ai ecosystem that’s going to compete with america’s broadly and that there are specific applications of
0:05:34 the chips that we would be selling to china which we would be very uncomfortable with like um military
0:05:40 ones or intelligence ones or sort of broad human right by human rights violations that you wouldn’t
0:05:47 want sort of america and technology to be helping to further and then the arguments um in favor of
0:05:54 selling we have this idea that actually selling uh nvidia chips would retard domestic chip development
0:06:00 make it harder for uh smic and huawei and whoever else wants to try to build in a domestic ai chip to
0:06:08 find a marketplace and this idea that selling chips into china would maintain uh american you know chinese
0:06:14 dependency on the u.s stack so it keeps chinese developers using using qda building infrastructure
0:06:23 around u.s technology and you know some sort of like broad soft power um agenda setting advantages to be
0:06:31 determined that um china using nvidia hardware will um give to the u.s going forward so maybe we should
0:06:39 run through those maybe we should run through those systematically i guess um uh you know let’s start
0:06:46 with the biggest one which is you shouldn’t sell these chips to china because upgrading the chinese
0:06:53 uh ai ecosystem is like a strategic threat to the u.s and i think that’s a i think maybe we’ll throw this
0:06:59 one to chris i mean this is a this is like a like this is like a grand strategic question almost of like
0:07:07 how much of china’s rise is okay and how much of it isn’t because um you know the the military intelligence
0:07:13 human rights applications are almost secondary to like just how scary you see
0:07:23 a richer more flourishing more powerful china to be right yeah and i would even segment out the sort of
0:07:27 richer and more flourishing side and just talk about technological capabilities they’re obviously
0:07:34 interlinked but i i don’t think u.s strategy has been to try to make china uh poorer or less flourishing uh in
0:07:42 general i think the question is just who’s gonna lead in in ai uh the trend over the last five years
0:07:47 and really the last 50 years has been that if you want advanced ai you need lots of advanced computing
0:07:52 and there’s a small number of companies who produce the chips in question um and so if you think that
0:07:57 advanced technology has mattered in the past in geo street competition which i think is pretty hard to
0:08:01 argue with it’s probably gonna matter in the future and therefore who wins in ai you know it’s a
0:08:09 little bit reductive to ask but i think it also uh it matters um and just like we would be a less
0:08:14 happy if we were all using huawei phones and relying on alibaba cloud because i think there’d be pretty
0:08:20 significant political ramifications that were downstream of that so too if we find ourselves in a future where
0:08:29 either the u.s or third countries are relying on chinese ai providers whether for models or for applications or for
0:08:35 um ai cloud that implies less political influence for the u.s a weaker u.s and a stronger china so
0:08:42 that’s the stakes um and i think to to some degree both sides of this argument agree on that basic framing
0:08:48 question is how best do you get there uh you know one argument is that you restrict compute access and
0:08:54 thereby uh hobble the growth of chinese ai firms um a second argument is that you try to you know as
0:08:59 secretary lotnick has said get china addicted to the ai stack and the question to ask is you know how
0:09:03 addicted are they willing to become and how addicted could you actually make them and can you leverage that
0:09:08 addiction in the future uh or or not and i think these are where the the empirical questions are
0:09:15 are really focused and i think kind of one more on the in favor of selling argument this idea that
0:09:22 keeping china dependent on you know tsmc fab chips actually lowers the risk of a taiwan war which i have
0:09:26 some questions about but this is something that ben thompson has been pushing which i think has also
0:09:33 percolated into the uh into the administration and congress like my mom threw me that one the other
0:09:39 day um anyways i mean what do you think what do you make out of it jordan i’m like i’m like it was like
0:09:44 for me it’s like it doesn’t seem to be the main calculus behind it like i i buy it on the margin
0:09:51 i think maybe on the margin a little bit i think there’s two levels to the question first um at the
0:09:59 political calculus to go to war or not to go to war it seems very this is like this would be an extremely
0:10:07 kind of weighty decision um where you know the fates of nations would be at stake to do a serious blockade
0:10:16 or to you know do some strike or some lay of or some actual you know d-day style invasion and uh you
0:10:21 know whether or not the chips are there china’s sort of like you know gaining relatively or lose
0:10:30 relative losing relatively in uh ai hardware just strikes me as like the 12th thing that you would
0:10:37 be thinking about if you were a chinese premier i mean domestic political developments uh uh political
0:10:44 developments in beijing political developments in taipei um just how willing the u.s is seeming to fight
0:10:51 for taiwan how um excited japan is to let the u.s fight from uh right from its territory
0:10:58 all of those strike me as much more germane decisions and there’s a bit of sort of like
0:11:05 technological myopia of tech analysts thinking that like oh the chips are the one thing that’s like
0:11:12 this silicon shield stopping it it’s like no like we’re i’m sorry i i really as as as cool and important
0:11:18 and you know over a 50 year period potentially like world shaking as uh you know advanced semiconductors and
0:11:26 artificial intelligence may be um you know if you are a uh a head of state this is and making the
0:11:31 biggest decision of your life i don’t think it’s really going to come down to like oh well huawei
0:11:39 tells me they can only make 750 000 chips in 2028 so then it’s not going to work out but i do think sort of
0:11:47 one level down from that there is this very open question which we debated um on uh sunday’s edition
0:11:53 of uh second breakfast about to what extent the that the chips and the technology that are they’re going
0:11:59 to be enabling ends up sort of reshaping the military balance of power and i think that is still very much
0:12:07 an an an an open question that smart people can disagree on on whether or not um you know whatever
0:12:15 uh you can do with you know putting chips in your autonomous drone so they can like target without being
0:12:21 uh uh interfered with or what have you you can imagine a lot of a lot of different crazy futures where
0:12:28 where ai really matters and by the way like it could actually work in the other direction um of
0:12:34 lowering the risk of taiwan war if america has a big lead when it comes to uh semiconductors because then
0:12:40 uh a chinese or a leader in beijing would look at the military balance of power a balance of power
0:12:48 and the advantage that uh u.s and taiwanese forces get from um being more uh you know ai-ified that they
0:12:52 see oh you know there’s no chance of us winning why even why even try to play this game in the first
0:12:57 place i think the other key facet here is that if you look at sales of advanced chips from taiwan’s
0:13:02 ecosystem to china most of them are not ai chips it’s mostly smartphone chips and pc processors and
0:13:08 and ai chips are a portion but a a small portion i think it also gets back to the question i know you’ve
0:13:14 raised a lot of shows jordan which is how agi-pilled is xi jinping and the answer is you know
0:13:18 doesn’t seem that agi-pilled uh best evidence for which is that you know smic and its seven
0:13:23 nanometer production is still producing a whole lot of smartphone chips uh which you would not do if you
0:13:28 thought we were in a a race for uh for agi which will define the future so both of those facets again
0:13:35 point against the silicon shield as it relates to ai chips being absolutely central here yeah and also
0:13:40 just to clarify they’re not allowed to produce ai chips at tsmc they can produce everything else there
0:13:45 huawei not because an entity list you know they did some bad stuff but almost every other chinese
0:13:49 company can just go to tsmc and produce chips there right so there’s a significant flow of chips from
0:13:55 taiwan so as we’re speaking right now uh to china right just not ideally not ai chips you know we we
0:14:03 had some hiccups in the past where it was also ai chips right so i think a key question leonard is uh how
0:14:10 obsolete is the h20 relative to a what what black holes can do but probably more importantly be what
0:14:14 what huawei can do you want to walk us through the numbers what i would just say i don’t think
0:14:22 describing the h20 as an obsolete chip is this fair right um chips have many specifications right and
0:14:26 let me try to break it down two simple ones we should care about the computational plot power how many
0:14:32 flops it has how many operations per number per second it can crunch but then also memory bandwidth
0:14:36 this means basically you need to read and write memory and the memory capacity and the bandwidth how
0:14:41 fast you can read and write this memory is key and one of the key inventions we’ve seen over last few
0:14:47 years which actually amd did first is so called high bandwidth memory which is a pretty complex technology
0:14:52 we’ve got three companies in the world doing it right now uh sk hynix samsung and micron building this
0:15:01 hbm and the h20 is really bad on the flops 7x worse than h100 even more worse 14x worse than like the
0:15:06 upcoming chips in the b100s or more right so it’s definitely no competitive chip here um but on the
0:15:12 memory bandwidth side which is again key for deploying chips it’s pretty good it’s even better than h100
0:15:18 because the h100 uses five units of hbm whereas this one has six units of hbm so it gets a mind like
0:15:23 boggling four terabytes per second of high bandwidth memory and no chinese chip has such
0:15:28 good high bandwidth memory and more importantly even if they have right now like 910c which has some
0:15:34 hbm i think it sits at 3.2 terabytes per second they’re not allowed to buy it anymore it’s banned
0:15:40 since december 2024 right so right now china would be struggling to first of all getting their hands on
0:15:44 this hbm and they’re trying to produce it domestically but this will this will take a bit and even if they
0:15:49 produce it domestically it would initially be worse so i don’t i don’t think the h20 is an obsolete chip
0:15:52 it’s a pretty pretty competitive chip i think it’s definitely fair to say it’s a worse chip than many
0:15:57 others that’s fair to say but again if you look at this other dimension this dimension of deployment
0:16:04 it’s pretty good it’s really really good but i think that is one of the key uh uh axes of debate you know
0:16:08 some people say the goal is to stop china from training high-end and therefore you focus
0:16:13 on the flops if your goal is to constrain inference you focus more on memory bandwidth walk walk us
0:16:18 through uh the way these different chips are used yeah i think that’s a fair debate to we should be
0:16:22 having here right like we should think about with xbox controls what do we want them to achieve
0:16:27 and i think right now it’s fair to say the h20 is not an amazing chip for training ai systems
0:16:31 there are some things which numbers don’t always capture right like you still build on top of the
0:16:35 nvidia software stack if your company which used nvidia before there’s like a little bit of a pain to
0:16:40 switch there’s a bunch of problems with huawei chips which you don’t really see in the specifications
0:16:46 right the overheat you need more of them the software stack isn’t great yet do you can even get enough
0:16:52 right all of these things just mean that h20 is not a great training chip but beyond the numbers it’s
0:16:57 still you’re stuck on a software ecosystem so on the training goal i think that’s still being achieved
0:17:02 here um where the debate begins is what do we think about deployment and what at least i’ve learned
0:17:07 over time here and is if you want to be really precise right if your goal is to only stop them
0:17:11 from training but everything else is below or you only stop them from training big systems it’s really
0:17:16 hard to be precise on all of this and then ai is ever changing and i think the most biggest thing we’ve
0:17:22 seen over last year is like this this rise of test time compute of air models thinking right how do they
0:17:27 think will they produce tokens and that’s what the h20 is like really amazing at right so one could say
0:17:32 the usability and importance of the h20 only went up ever since because we got models which
0:17:37 do more thinking generating more tokens and also generating tokens to then train the next generation
0:17:41 of ai systems right and i think these are the arguments which people would say like well actually
0:17:45 this is a pretty good chip right for producing these new things which are like more important
0:17:50 than ai development life cycle so the other argument that the president made is that huawei already
0:17:55 makes these chips which is uh you know true to an extent but but walk us through the numbers there
0:18:00 as as you see them i know that there’s questions both about the quality of of huawei’s chips as well
0:18:06 as the numbers they can be produced i know secretary lutnik said they can produce 200 000 a year and i
0:18:11 suppose that’s right how does that compare with with what we’re going to see with h20s the key
0:18:15 dimensions here are just quality and quantity and i think many always talk about the quality argument
0:18:20 here i personally think the quantity argument is way more important you already mentioned the number
0:18:27 from i think latnik said and kessler also testified at 200 000 asin chips being produced in 2025 how does
0:18:32 this compare to the us i think we’re churning out around 10 million chips this year so significantly more
0:18:36 this means if we’re selling and i think there have been projections about nvidia selling
0:18:42 um a million h20s we sell them five times more what they can produce right and i think this is where the
0:18:48 debate starts the quantity argument is like really key here if you would only sell them a couple of
0:18:52 thousands or like 200 000 or something that’s a vastly different debate than selling them a million
0:18:58 or potentially even more and just the sign that china wants to buy them speaks to like their problems
0:19:05 with producing domestic chips right so on the quantity side china is simply not there yet they’re getting
0:19:10 better and they’re producing more chips the minute we speak that’s the case but they’re many difficulties
0:19:15 as long as even to produce more chips have enough high bandwidth memory how good is their smuggling
0:19:19 operation to get this memory how good is their packaging yield all of these things just add up
0:19:24 that you eventually just really can’t produce competitive chips then they at the end they get chips out
0:19:30 of it if you compare the asin 910c to nvidia’s best chip right now which is being sold to b200 it’s way
0:19:36 worse right it’s way worse on the high bandwidth memory part and it’s also way worse on the computation
0:19:41 performance and it’s also worse than h20 which you’re selling at least on the memory part um which we’re seeing
0:19:47 here and the point is if you’re selling the h20 i think what many missed there’s a chip or at least there were
0:19:50 rumors around it and i think there were like pretty good rumors there’s a chip called the h20e
0:19:56 what does it do it doesn’t use hbm3 it uses hbm3e so i previously said it has four terabytes per second
0:20:02 if you use hbm3e you can probably go up to five terabytes per second or even more right and which
0:20:06 indications do we have that this ship is not getting sold right the flops are still being capped but the
0:20:10 memory just continues going higher and higher and higher and i think that’s another thing to be
0:20:14 tracking here and as long as we don’t have updated regulations for it we just don’t know where where
0:20:19 the line is going to be drawn here in terms of quality of memory bandwidth but also most importantly
0:20:24 in terms of um quantity so if i could ask for one thing please reduce the quantity i think that’s the
0:20:30 key thing we should pay attention to here yeah i think uh you know one of nvidia’s lines uh that jensen has
0:20:35 been saying a lot is like they used to have 95 market share before the restrictions and now it’s
0:20:44 down to 50 first off they’ve never actually given numbers for that but second my guess is that they were
0:20:51 the only people making you know accelerators that people wanted so even if it did go down to 50 it’s not
0:20:57 like it was the same pie like the pie went down such that that five percent that it used to be
0:21:05 now turns into 50 of the whole pie so um the the idea that sort of huawei like that number does not
0:21:12 tell you that huawei necessarily has the capacity in order to fill it up and as and as leonard said like
0:21:19 the jensen cares about this because lots of chinese companies are willing to spend uh his projection is
0:21:26 what like 15 billion dollars a year in sales and to and to think that like i mean you know huawei and
0:21:32 baidu and tencent they are not dumb like they are going to spend billions and billions of dollars of
0:21:37 uh capex by the way this capex number um you know seems small if you’re talking about google and meta
0:21:44 but it’s actually pretty large um relative to the sort of total capex that you’re seeing from the chinese
0:21:48 hyperscalers so like they’re doing this because they think it is useful and important and relevant
0:21:56 to their sort of uh you know ai ambitions going forward not to like do jensen a favor or anything
0:22:02 could we talk about what we know in terms of who in china will be the large-scale buyers of
0:22:11 these ships jordan you mentioned uh tencent uh alibaba there’s ai firms like deep seek um there’s
0:22:17 byte dance a huge player in china’s a ecosystem i don’t know whether you have a sense of of numbers
0:22:21 if any of those are public or at least talk about kind of who are who are the buyers of these ships
0:22:25 inside of china i don’t think they have public reporting of exactly there’s definitely been some
0:22:32 reporting that the big hyperscalers the big cloud companies right tencent um by tens and others are
0:22:35 definitely interested in this i’m not sure i’m interested by tensers because they’re building
0:22:41 a tons of clusters in malaysia which by the way can buy whatever chips they want to buy um there and
0:22:47 just continue building um so i think just the normal hyperscalers go there um they will continue
0:22:50 buying these kinds of chips but they’re all hatched right they all also get ascent chips they’re not
0:22:54 stupid right we just see with the policy flip-flopping they don’t know when they’re going to get cut
0:22:59 off right so they’re all just hatching with like huawei’s and chips while they’re getting better
0:23:02 there’s something we would just subsidize their transition while we do this right and that’s
0:23:08 that’s like the thing which i’m worried about here it’s just it’s just the case that huawei will get
0:23:12 better they will produce better chips i think the chips will be significantly worse and significantly
0:23:16 less quality in the us but they will get better and that’s the thing we all need to acknowledge
0:23:21 and there was a policy at some point which was made which just told huawei they will need to
0:23:25 produce their own chips right and that’s that’s just a pathway going down here and i think there’s
0:23:30 like no going no going back here the question is like what do we do in the meanwhile right um and how
0:23:34 big will the gap potentially be and i’m i’m a firm believer that this will be quite a massive gap which
0:23:41 will have big impacts on the ai uh competition and i think that is one of the key uh lines of debate but
0:23:48 also you know empirical questions that’s hard to to really research or get hard data on which is
0:23:56 um the the decisions of the private tech firms in china the the alibabas the 10 cents and and others
0:23:59 because to the extent that you’re right that there’s a meaningful quality difference between
0:24:05 uh nvidia and huawei gpus for example they got a strong incentive to build as much as possible on
0:24:11 um on nvidia so you can see an argument that says well they’re gonna buy a sense but sort of put them
0:24:15 in the closet or not really take them seriously because they want to build their products um but
0:24:22 but you’re saying no that’s probably not the case because even those firms which don’t have a strong
0:24:28 incentive on their own to help out huawei uh do in the context of potential future export controls and
0:24:36 and loss of access to an nvidia chips and so the argument that um that controls sort of align the
0:24:42 incentives of tencent and alibaba with huawei the chinese state uh you think those incentives are already
0:24:48 at this point fully aligned yeah i think so i think more more importantly here we should just always
0:24:52 i mean let’s walk through the arguments for it there are arguments in favor of selling h20 and i think
0:24:56 that’s the same debate to be had here right on the other side i think it’s sometimes lacking some
0:25:00 technical details here i think the market share argument is a fair argument right just like you
0:25:05 want to maintain for nvidia bigger market share and reduce demand for huawei i just don’t think that’s
0:25:11 the case right it’s an existential priority for china to produce the semiconductor industry and importantly
0:25:15 it’s not like the semiconductor industry only gets better because of ai chips the majority of
0:25:20 chips the world produces are not ai chips who’s producing at the advanced and note at most advanced
0:25:26 node at smic but also tsmc it’s apple right usually we put those mobile phones first there
0:25:31 so they’re pushing it forward anyways for the newest huawei smartphone that produce probably soon
0:25:36 something like a six nanometer node which will then be leveraged to produce better ai chips right
0:25:42 so even if you reduce the market demand right now um semiconductor industry will get better and
0:25:47 these will lead to better ai chips eventually right if they then just like transition to to this and then
0:25:52 also what is the tech stack argument here right sure we keep them hooked on cuda right and it’s a pain
0:25:57 to go from cuda to eventually mind spot to the huawei ecosystem and i think we can model this as like a
0:26:02 one-time transition costs many American companies have done this google switched to tpus at some point
0:26:09 right um and fropic right now is using the trainium chips on aws i think they they pay a significant amount of
0:26:13 costs you to switching and like running these different hardware sticks but eventually they’re
0:26:18 doing it right and they will also eventually do it with huawei and it’s not like if you use cooter
0:26:22 your systems are not more aligned you know like if you would sell them ai systems who don’t spit
0:26:27 out ccp propaganda i’m in favor of that that’s like spreading American values liberal values right
0:26:34 that seems fine but if you would be just selling them chips there there are no no values no constraints
0:26:39 which come with selling chips you can just do whatever you want on it right and i think that’s again
0:26:43 where i’m just like it’s missing this tech tech component and we kind of got it right in the uae
0:26:47 sell them the cloud let microsoft build here versus here we just sell the underlying component they can
0:26:51 build whatever they want on top of it and i think that’s just missing in the debate you’ve made it 30
0:26:57 minutes into a podcast about age 20s do you want to do this full time do i have the opportunity for you
0:27:02 the horizon fellowship provides a direct path into the world of emerging tech policy a fully funded
0:27:08 placement at a federal agency congressional office or think tank in dc for up to two years for its
0:27:15 2026 cohort horizon is actively seeking candidates with expertise on china’s technology ecosystem its
0:27:20 policy landscape and broader strategic goals prior technical or policy experience is not required join
0:27:25 a community of over 80 fellows working on some of today’s most pressing issues that i can personally
0:27:30 vouch for having met a good chunk of them are some of the smartest minds working on these topics in dc
0:27:35 the application deadline is august 28th check the show notes for a link to learn more and apply i think
0:27:40 this is a key aspect to the export control debate which which is fascinating a lot of people don’t
0:27:47 get which is that if you if you restrict sales of tools then you hurt the tool makers but you help
0:27:52 the users of those tools and so in the chip industry if you sell fewer lithography tools it’s bad news
0:27:57 for asml but it’s actually probably good in the long run for dsmc and other companies that face less
0:28:04 chinese competition similarly if you uh sell gpus to china it’s bad news for gpu sellers or it’s good
0:28:09 news for gpu sellers but bad news for usai firms who face stronger competition and so one of the kind of
0:28:16 strategic questions is which level do you try to cut off and the us has i guess decided until recently cut
0:28:21 off at multiple levels and is now shifting well i guess we’ll see where we are next week but this week
0:28:26 seems like it’s shifted towards a policy of sell the gpus but keep the controls on the on the chip
0:28:32 making tools which makes sense right like if we would like reverse selling them uh extreme ultraviolet
0:28:38 photography machines from asml um i would be way more on a rampage than selling them ai chips right
0:28:41 and i would sort of complain more if we start selling them blackwells over h20s
0:28:46 right so i think i think that’s a fair debate to we should be having here people can fall on different
0:28:51 types of positions here and again we can disagree on some arguments here and like again yeah and then
0:28:55 you have like these different types of controls which stack with each other and the eye chips are
0:29:00 the first ones to fall i think that makes sense so yeah so it’s one of the arguments is that if you
0:29:06 make china addicted to ai chips you you gain long-term leverage and i think like the mental model that
0:29:12 people think of here as well if you get them using uv lithography tools they don’t have their own
0:29:18 ecosystem and it takes a decade to uh try to replicate your own tools so maybe this is going
0:29:25 for leonard uh does the same dynamic hold here or is uh or if not why what are the differences yeah i
0:29:31 guess there are many different facets to being being addicted to something right i mean in the ideal case
0:29:36 it just means all chinese firms are like really reluctant to adopt chinese chips right and therefore
0:29:40 they have like less revenue smic is wondering nobody wants to buy their chips and instead all the chinese
0:29:45 isn’t just buy us chips i already talked upon how just like smic and semiconductor gets anyways
0:29:51 independent of ai right but it’s a fair thing to say like the less people use like huawei’s ai
0:29:57 software ecosystem the worse it is i think that’s that’s a fair argument to be made i just think they
0:30:01 know they want to produce it anyways they just know we need our own ai chips at some point they’re not
0:30:05 full steam on this i think they could go strong if they wanted to maybe they’re full steam on it but i
0:30:11 just don’t do better for many reasons um so like them using the us tech right now i think like maybe
0:30:16 delays it to some way even subsidize it let’s just think about i don’t know volkswagen you know my
0:30:21 german heritage they love to showcase to china how’s this going right now did this stop byd
0:30:26 not really right and i expect like the share of volksvang is being sold to china right and the future
0:30:30 will be low the argument to be made here they made a ton of money in the meantime and i think that’s a
0:30:35 fair argument to be made the reason why i just feel nervous about ai chips is well you supplement
0:30:41 you’d like increase the total compute deployment training capacity in the time if again let’s just
0:30:46 say agi is the singular point ai is just not going to matter the nice five years then all we discuss
0:30:50 here doesn’t matter that much right because the good thing about ai chips is they get exponentially
0:30:55 better we’re not going to talk about age 20s in five six years from now because we have
0:30:59 exponentially better chips already here right so i think that’s an argument we can just say like
0:31:03 don’t worry man we just sell them we make some money they get a little bit better ai but ai is
0:31:09 not going to be decisive in the next four to five years but then later ideally then later we stop it
0:31:14 right um and we don’t sell them we have like better chips which are like exponentially better again it
0:31:20 goes back where do we draw the threshold and when and how does ai matter which is a diffuse question
0:31:25 right and i i have a pretty uncertain view here i’m just like man ai could be a really big deal in the
0:31:30 next three to four years it seems likely it’s going to be a big deal um bigger or less big right
0:31:34 depending on how it goes from just like like transformative economic growth being like
0:31:39 determined the future of the military up to just going to fizzle out right i think we should address
0:31:44 this uncertainty here um i just work on national security risk and i’m trying to minimize downside
0:31:49 risks right and i don’t see see the benefits here in the long run that why we should sell them so fair
0:31:53 argument i think there’s some good arguments here but overall i think it doesn’t cut it at least for me
0:32:00 when you look at some companies like it’s a really big deal having chinese market access
0:32:08 like nvidia excuse me like intel like 35 percent of their revenue is selling cpus into china um this
0:32:13 was a big deal for the tool manufacturers like you know in some years it was 30 40 percent over the past
0:32:20 few years um nvidia is a four trillion dollar company like they will be just fine um and still be able to
0:32:28 deliver you that exponential curve of uh you know rapidly improving ai chips even without um uh you
0:32:33 know the extra 10 billion dollars of sales um i mean i guess you know there’s like the maximalist version
0:32:40 of this question of like okay like you are a hundred percent sure that ai is not matter does not matter and
0:32:46 it’s not a sort of like strategic technology then yeah sell it you know go crazy do whatever you want
0:32:54 with it but it is it is a it’s like a it’s a it’s a tricky line of thought we’re like a you know we’re
0:32:58 writing an ai action plan where we want to make ai dominance we think this is going to usher in a new
0:33:09 golden age but we’re willing to like take some of this downside risk um that we’re making it easier
0:33:17 on a china which we’ve identified as a like like major strategic threat in order to um
0:33:25 i don’t know what i mean make some money we’re still yeah well well it’s like it’s like okay so
0:33:31 if we if we do think this is a big chip like there is a broader context of the relationship
0:33:39 that you can try to trade things in right and say you know we wanted them to scuttle some submarines
0:33:47 or um you know uh stop uh messing with the philippines or uh i i don’t know there’s just there
0:33:54 there are lots of other asks you can make from a sort of like balance of power regional dynamics
0:34:02 perspective that you could have put on this and it’s kind of wild that it didn’t even seem to be in
0:34:11 the context of the um of the debate of the sort of discussion between the the us uh china trade deal
0:34:16 but was just like a decision that uh trump made independently because jensen got to him and we
0:34:23 wanted to have like good vibes in the relationship and the 15 tax we’re putting on it i mean is it just
0:34:29 going to like pay off the national debt i mean what if it went to like buy uh drones for taiwan
0:34:35 or to sort of shore up you know go into funding bis so they could do a better job of uh tracking down
0:34:41 all the chips that are getting leaked out into china uh you know there are some sort of lines of this
0:34:48 where if you really are going to follow the like okay like we’re we’re all right with the premise that
0:34:55 like china is a strategic threat and we gotta kind of watch uh and and hopefully shape just how much
0:35:00 they’re going to gain on the us from a sort of relative technological competition perspective
0:35:09 like there are other moves you can do to make it more to to sort of like use this card more in your
0:35:15 favor than just like letting the other side pocket it i think i think john that’s an excellent point
0:35:21 right and if you would just have like more insight here because what we’ve seen so far is the h20 got
0:35:27 sold again then postdoc somebody said it was part of the trade talks others denied it then the chinese
0:35:33 came out denying it was part of the trade talks right and eventually what i don’t like about it
0:35:38 expo controls were always a national security thing right when the diffusion framework came out and there
0:35:42 were certain countries like poland switzerland in this tier two many were complaining like you’re
0:35:47 dividing a european trade union but it’s a national security thing it’s not a trade deal we’re doing
0:35:52 here this is at least where expo controls originally came from right and now we’re mixing them with
0:35:57 the straight things and now we get 50 of the revenue share and amazing let’s pay off the debt let’s
0:36:02 do other great things i don’t think national security is for sale here right if the if we
0:36:07 would get like other national security concessions here in return that’d be amazing right it would
0:36:11 just be nice to hear more and communicate about this there’s definitely things where like people
0:36:15 like me and others are willing to walk back hell yeah let’s sell in the h20 because we got a beautiful
0:36:21 deal out of it i just don’t think 15 of the sales is is makes the cut here just money money doesn’t help
0:36:27 you they get the computing power you make money you might even make more money because there is nvidia
0:36:32 making money and maybe microsoft or your favorite hyperscaler in between right why you still have more
0:36:36 control and more leverage right so i still generally think i think many just miss the point you don’t
0:36:41 need chips like in your basement to run them you just access them remotely and so they could literally
0:36:47 dial in they could dial in into our beautiful new uae 5 gigawatt cluster or dial in in the us and like
0:36:51 existing cloud providers and then in the future in case they go rogue or you really want to make sure
0:36:56 it doesn’t go to certain military linked entities you just have way more leverage right and i think like
0:37:00 if we do the concessions we talked about the different things you want to walk back before you sale sell
0:37:05 chips just tell them you can use the cloud which is by the way perfectly legal as we’re speaking right now
0:37:10 so if they really want to the computing power use our cloud it’s all legal you can go for it we still
0:37:16 make money so um i think there’s a political economy dynamic here which which leonard i think you’re
0:37:22 you’re referencing which is you’re getting back to if you sell the tools you enable the chip maker if
0:37:28 you enable chip makers that type of competitive dynamic and and what we’ve seen is um gpu sellers nvidia
0:37:34 i think most prominent be very vocal on this issue we haven’t seen hyperscalers be vocal at all even
0:37:40 though i think one should conclude this implies more competition for them um and then we’ve seen
0:37:46 mixed responses from ai model companies i think anthropic has been pretty vocal opposed i haven’t
0:37:52 seen open ai um and so it strikes me that companies that have a lot of stake have been taking very
0:37:58 different strategies uh some being vocal some not um i don’t know what exactly explains that uh more
0:38:03 observation i got an explanation go for it you know which cpu still are using and videos and if you
0:38:09 speak out against them jensen is going to get you right so if you look at anthropic who’s like
0:38:16 slowly migrating to like more google tpus and amazon trainium you can see the deals they can speak out
0:38:21 against it where everybody else is like um reliant on jensen and i can at least confirm from many
0:38:26 conversations with many people in his companies the only competitor here is amd right and i think it’s
0:38:29 only going to be the market share for my video is only going to go downhill from here
0:38:34 right the total market will go up ai is a big deal but amd is getting better google gpu is getting
0:38:39 better microsoft chips getting better amazon chips getting better we have more and more startups
0:38:44 getting better right we just have way way more ai chip competition i think like nvidia is also like
0:38:48 slightly nervous on all of these issues and i think i would love to live in a world where nvidia
0:38:51 had a smaller market share and see what the hyperscalers and the ai companies would say and i think
0:38:56 many of them would actually come out and open ai at least came out in favor of export controls
0:39:00 historically um when they talked about energy dominance and more and i think right now they’re
0:39:04 all just all quiet because somebody else might then knock on the door it’s it’s fascinating in the
0:39:09 context of i’ve gotten lots of questions of what does industry think and of course you know what you’re
0:39:14 saying is well which part of industry are you looking at which which segment which specific companies
0:39:18 uh yeah fascinating political economy there jordan back to you why don’t you do that why don’t you do
0:39:23 the the the hbm political economy this has apparently been reported that the chinese government is asking
0:39:32 for high bandwidth memory as part of a concession number two um maybe like what would that what is
0:39:38 that what does that tell you leonard that ask um if i’d be be running china i would ask for high
0:39:43 bandwidth memory over asking for h20s personally right because i’ve got my sovereign drive anyways i
0:39:47 want to build better and better ai chips and if i look at my current ai chip industry i would like to
0:39:52 want euv but maybe this is too much to ask for right because we did this early on trump did it back in
0:39:57 the days but what is the thing we’ve only recently done is banning high bandwidth memory units right so
0:40:02 we got our chip and next to the chip we put like the memory and these memory units being produced by
0:40:08 samsung and sk hynix and micron they’re not allowed to go to china anymore right and we’ve seen reporting
0:40:13 that at least the chinese again the chinese put forward could we trade hbm is that like something we
0:40:19 can do here i hope the u.s government will draw like a clear red line here right we talked about how you would
0:40:23 walk back things there are arguments in favor of setting them chips we talked about them what we do
0:40:29 here is not selling them our chips what we do here is enabling them building better chips the best way
0:40:33 how the 910c or the 910d whatever the next best next chip is they produce will get better is by having
0:40:41 higher bandwidth memory and right now china does not have the capacity to abuse even hbm3 there’s reporting
0:40:47 about first trial production of hbm3 in contrast nvidia and others are starting to equip hbm4 and using
0:40:52 hbm3e right now so again don’t get wrong china will get better they will eventually produce high bandwidth
0:40:57 memory there’s lots more to be done we could stop them from producing better memory but in the meantime
0:41:01 while they’re like scaling up this production and trying to get better at least we should probably
0:41:06 not sell them our high bandwidth memory to make the air chips more competitive because we might regret
0:41:11 this in many years when we’re then competing of course emerging markets and uh huawei has a better
0:41:16 chip which can better compete with our chip and the interesting dynamic the memory space is that the
0:41:21 two of the three producers are are not us but korean so indeed that’s also why we see probably tons of
0:41:26 smuggling here because um it’s pretty close by and there’s like certain tricks to get more hbm so don’t
0:41:31 get me wrong i think china smuggling hbm right now we just throw sand in the gears but again i’m in
0:41:37 favor of throwing sand in the gears and ideally we get better enforcement so um yeah they get less hbm
0:41:46 eventually should we switch to the political economy of this in china so july 15th we get the news that
0:41:54 the trump administration is letting nvidia start to uh sell h20s a week later mss publishes a notice to the
0:42:02 public saying to beware of digital spying via ford and produce chips um uh ten days later uh the uh
0:42:07 cac the cyberspace administration of china summons nvidia representatives over risks of being able to
0:42:14 control a chinese remotely and accuses them of having a kill switch in them then we have a private uh
0:42:21 leading cyber security research firm in china tianxin um they publish a report which goes viral talking about
0:42:29 all the ways that um there could be back doors and uh ten days after that so just three days ago august
0:42:36 9th we have uh state television doing a whole big report talking about how there might be back doors
0:42:44 already in these h20s and they cite uh former china talk guest tim fist c uh cnass report on this topic
0:42:51 which he came on to talk about on china talk a few years ago so um you’re welcome uh i don’t know chris
0:42:58 what’s your uh what’s your read on this like kind of interesting brushback pitch we’ve gotten from uh
0:43:07 the central organs about h20s in china so i think there are three potential explanations not mutually exclusive
0:43:14 one is that the chinese security services are paranoid um and the discussion in washington of the chip
0:43:21 security act which would mandate uh geolocation um uh verification which has been happening
0:43:28 simultaneously to the h20 debate has intensified those uh concerns that that’s explanation one
0:43:35 explanation two is that it’s part of an effort to discourage the private chinese tech firms from using h20s
0:43:40 um and that there are people around huawei or in the government who are afraid that h20s will
0:43:44 actually take into the market share and so this is a way to say buy more huawei chips as well
0:43:50 uh and then i think the third explanation is that this is actually pressure on u.s firms like nvidia
0:43:56 to say actually we need you to do more um or else we’re not going to let you back in uh the market we’ve
0:44:01 seen this in other uh segments of the tech sector where china will ramp up pressure on a private u.s firm
0:44:08 with the aim of having that firm then try to use its resources to shift the debate in washington you
0:44:15 could maybe envision the hbm um debate being part of what china is looking for the broader trade negotiations
0:44:21 that are uh underway but it certainly wouldn’t be um a very attractive endpoint for nvidia if they got
0:44:26 approval from the us side then didn’t get approval from the chinese side to sell um so perhaps china
0:44:34 thinks it has uh some some leverage there how exactly to kind of uh tribute uh uh these three
0:44:38 causes i’m not exactly sure what the what the shares i would put on each of them but they both
0:44:43 seem or all three of them seem potentially relevant do we do we talk about also put out guidance a while
0:44:50 ago on energy efficiency this was actually shortly in in april or may when the h20 was no sorry even
0:44:56 before so when the h20 was sold before it got banned initially they put out guidance that the h20 is
0:45:00 famously pretty energy inefficient if you look at flops right because of the export control bandwidths
0:45:05 and basically i don’t know exactly what this guidance means but basically discourages companies from using
0:45:09 it i guess nobody’s been following it because now they’re buying it in the well up in the single
0:45:14 millions um this chip but i think it feeds into the same narrative here right just like you try to push
0:45:20 certain companies um or you create artificial demand for like some huawei chips like slowly tell them
0:45:24 like hey guys at some point we want to do our own ai chips right i think as chris was saying i think
0:45:28 all of these stories are simultaneously um true right i think it all just makes sense and there’s like
0:45:32 no big downside for them to to do these kinds of things only benefits but actually there’s a state
0:45:37 media source i don’t know if you’re referencing jordan but but one of its criticisms of the h20 was
0:45:44 is that it was environmentally unfriendly yeah no they cite this exact uh ndrc line that uh leonard
0:45:52 talked about how the goal is um five teraflops uh or half a teraflop per watt and and the h20 can only
0:45:58 give you 0.37 so it’s pretty bad pretty environmentally friendly for training but pretty damn environmentally
0:46:04 friendly for deployment of ai chips way better than any huawei chip i can tell you that i mean i think
0:46:09 like here’s a moment where some mirroring might be in order we’ve just had an hour-long conversation
0:46:16 about how messy and convoluted um and sort of like how many conflicting priorities we have in american
0:46:21 policy towards artificial intelligence and like the same thing is happening in all these different
0:46:28 ministries in china and um you know this is big news this is like a um a a a change in the landscape and
0:46:34 people want to um you know people want to have their say and make their their stamp on it so you
0:46:40 you don’t necessarily need to sort of like attribute some like 4d chess of like oh now’s
0:46:46 the time to squeeze them but like i’m sure the people in the mss read tim’s thing and are like oh
0:46:53 this would be really stupid if we bought all these chips only for them to turn into um you know bricks or
0:46:58 like spy on us or like have bombs in them that are gonna blow up like beepers and and lebanon or
0:47:06 something um i’m sure uh folks in cac feel the same way and then you know the same debate that we’ve been
0:47:12 having for the past hour of oh is it sort of like net positive or net negative for um domestic self
0:47:20 self-sufficiency like uh to have a a sort of competitor to um uh to huawei potentially take
0:47:25 a big chunk of uh of of the market domestically is something that’s that’s being played out in china
0:47:33 so i agree with you that like at a broad level now is the right time to ask for more stuff from nvidia
0:47:40 that now that they’ve gotten the green light and there is this um kind of you know 10 10 15 billion
0:47:45 dollars of demand and these chips sitting on a lot somewhere in taiwan that they’re really excited to
0:47:52 ship out to say hey you know um you know you better step it up or cut the price or um you know do do an
0:47:57 extra screen to make sure there aren’t any kill switches on it or whatever you know the way this is
0:48:02 playing out on twitter is oh china’s saying they don’t want them that means we should definitely sell them
0:48:09 i don’t think it is necessarily dispositive just reading that chinese state media or state organs
0:48:15 are saying something means it is true i mean there is it’s not that hard to play the let’s not even give
0:48:20 this credit for 40s chess this is just 2d chess of saying oh no we’re worried about the chips we don’t
0:48:26 even want them chips that changes the political economy of the debate in washington where it makes
0:48:32 selling these chips potentially easier to go down so that’s also something to watch out for as
0:48:37 we see the chinese government saying ah no we didn’t really want these all that much this
0:48:42 isn’t actually a big concession we’re kind of worried about the knock the second order effects of this
0:48:50 but uh the fact is people like the demand is not going anywhere like it’s not like alibaba is not going
0:48:56 to buy these chips because uh because of these sort of warnings i think is the is the main i think alibaba
0:49:01 i would be pretty sad if they suddenly only need to rely on other inferior chips where they can’t
0:49:06 produce enough of them right like ideally again if i would run the chinese government i would like put
0:49:12 out so many regulations that i can sell all of the huawei chips i can produce and then fill the rest with
0:49:16 like some nice nvidia chips here but i think also what’s interesting i think there’s just some
0:49:20 misunderstanding of what the chip security act is supposed to do and location verification the idea
0:49:24 is not to check if a chip is well the idea is to check if a chip is in china and then we have a problem
0:49:29 right so idea is like we put this tracker on a chip in malaysia singapore wherever you think
0:49:34 that we smuggled and then check they don’t end up in china right so this was never supposed to be go on
0:49:38 chips which go to china because ideally we don’t have any chips going to china at least not the advanced ones
0:49:43 right so this is an interesting confusion right like this whole debate of this hardware name and
0:49:49 mechanism location verification was big in the uae and saudi arabia and singapore and malaysia for all
0:49:55 of these smuggling hot spots which people were worried about um and again i think some people have been
0:50:00 pushing and i think if you think about if we now stop selling chips i’m arguing we should sell them
0:50:05 cloud that’s one line before people got say we can sell them chips but put something on the chip
0:50:09 right but like just knowing a chip is now in china and we know it’s in this location and this
0:50:14 city over the how how does this help us again here right again like and then you everybody can dial in
0:50:19 remotely even if it sits at tencent who says that the pla is not using it right you can just dial in
0:50:24 remotely so i i don’t know what’s going on there if there’s some misinterpretation of documents just
0:50:30 some vibes it’s it’s a confusing status jordan you previously made a point about intel which i think
0:50:35 is an interesting one we’re like saying like intel made a lot of money in china right and intel is
0:50:40 still allowed to sell their cpus and intel cpu share in china is only going down right and i think
0:50:45 we will see the same with with nvidia and ai chips basically even if you’re allowed to sell your share
0:50:49 will potentially going down why is this the case there is similar guidance for example for all government
0:50:56 computers to go to homegrown domestic produced chips we can’t trust intel anymore on this right we will
0:51:01 see the same on ai chips so yes intel made a lot of money in china but the share is going down over
0:51:06 time and they’re pushing on the self-reliance basically to produce their own ai chips and they
0:51:11 also named security concerns here right that’s why the government is coming first i don’t know the exact
0:51:16 numbers of intel’s sell sales right now in china and how much money they’re making there but i’m pretty
0:51:20 confident it’s been going down and the government is not buying any more intel chips because they just
0:51:24 put out this guidance here so we’ve seen this playbook playing out before right the only difference is now
0:51:28 we have this confusion with which ships allowed to be sold which ones are not to be sold and how
0:51:33 good are they actually right but the story is nothing new could we talk about what we know in terms of
0:51:38 the big buyers of ai chips in china and their relationship with the state so you got the private
0:51:46 you know quote-unquote private tech firms baba and tencent um you’ve got the ai labs you know deep seek most
0:51:54 prominently um i don’t know what leonard jordan would you have a have a view here but one of the key
0:51:57 questions seems to be how what is the relationship with the state today and how is it changing
0:52:03 and to what extent should we see them as you know arm of the state is is certainly not accurate totally
0:52:08 independent is certainly not accurate there’s a spectrum um and so to what extent are these political
0:52:18 priorities shaping their procurement decisions look uh there was um reporting which was clearly sourced by the
0:52:26 uh intelligence community over the past few years that after the mss hack of the sf 86 so that’s the
0:52:33 form you um submit to the us government uh when you want a security clearance which like talks about you
0:52:39 which basically like you try to you know it’s your like confession of sorts to the catholic church where
0:52:44 you talk about all your divorces and all your debt and everything you know that a foreign intelligence
0:52:51 community might want to know about you um that that data was actually uh that the mss tapped alibaba and
0:53:02 bike dance engineers to sort of like put into a more useful format so um like we’ve just seen over the
0:53:08 the past few weeks reporting from business insider about a public tender uh from some corner of the
0:53:15 pla that wanted uh h you know eight h20 is to like do whatever they wanted to do with it i used to be
0:53:23 more sanguine on this type of thing um but i think there’s again this is like the most this is the dual use
0:53:30 technology to like beat out other dual use technology so it just seems to me to be like a little preposterous
0:53:35 that like insofar as this is a strategic resource that the chinese government would not be able to
0:53:43 leverage um you know data centers that are living in china um that the u.s did not have any sort of
0:53:50 kill switches or on-chip governance on um in order to do whatever they want to do with it whether that’s
0:53:56 um you know build a surveillance system or help with uh weapons manufacturing and to be clear like um
0:54:03 you know the pentagon has now signed like i think it’s like a 200 million dollar contract
0:54:10 with open ai and that this is just the beginning right so like clearly this stuff is useful um that
0:54:16 we’re willing to pay a lot of money to like get it into the pentagon uh and in in in one form or another
0:54:27 so um if what we’re sort of convinced with uh if if what you find uh what leonard says is convincing is
0:54:34 that you’re selling a lot of h20s like materially raises the amount of sort of like usable functional
0:54:41 compute that can be put into anything in china it would be really surprising for me um if you didn’t
0:54:46 have the chinese government want to take out these new toys for it if you if you didn’t have this sort
0:54:53 of uh chinese like military police complex want to take these new tools out for a spin so i think there’s
0:55:00 kind of two two uh two points you get on right so one is if ai tools exist will the military use them
0:55:08 and i think obviously the answer is yes there but on on the prick if you’re a data center procurement
0:55:16 official uh or uh executive at alibaba cloud to what extent is your decision making there shaped by
0:55:22 what you read in state media versus what your boss tells you to build an effective cloud in which case
0:55:27 maybe h20s are your best option versus ascends again how do we think about the because those
0:55:31 are the people who are going to decide how many ascends to buy right um unless they’re getting a
0:55:36 dictate where they are from the top now i guess like the the sort of counter example i’m thinking to
0:55:40 myself is you know there was a time when parts and maybe still i don’t know parts of the u.s military
0:55:45 were using chinese drones um not because there was a policy chinese drones because they didn’t have
0:55:50 any u.s drones and so i is is there a scenario in which like your your procurement executive at alibaba is
0:55:57 is is is just going to ignore the sentence because they were told to build a good data center yeah i
0:56:03 think i think at some level yes i mean these are these are companies that report quarterly earnings
0:56:09 that pay their employees based on how well the company performs right and you know people get stock
0:56:18 options so um i think by and large um uh the the the incentives of the people who are buying these chips
0:56:27 is to you know drive the most revenue um for the money you’re spending on your capex but um you know
0:56:35 it only goes so far uh and i do think that there is this sort of broader strategic realization um which
0:56:40 you don’t even which you don’t need beijing to tell you right is that this is like a this is this
0:56:46 you know this door could be closed at any time um and you need to ignore beijing to close it already
0:56:52 so you know yeah i mean i think i think maybe now is an interesting moment to sort of talk about
0:57:01 the sorts of things that could change the dynamic we’re on now um on chips and sort of on the broader
0:57:08 u.s china relationship i mean we have congress as a variable um there have been a number of uh senators and
0:57:13 congress people who’ve been like wait what are we doing selling these ships to china i thought we banned
0:57:20 and said we’re you know um uh you know our our our our our golden ticket to the 21st century
0:57:29 and then i think just because trump is doing this at such a personal level like we’ve seen him turn on
0:57:36 putin right and i mean we’ve seen him gone from gone from like all putin to like we’re gonna ask some
0:57:42 questions about this guy and you know we’ll see what the hell happens in in alaska but i do think that
0:57:50 there is like jensen saying the wrong thing taking too much of a victory lap or you know she doing
0:57:55 something really obnoxious i mean there’s there are a lot of sort of like personal interpersonal
0:58:02 dynamics that could change what the um uh what the trump what the trump administration ends up doing
0:58:08 which is like probably the more relevant variable than like whether or not leonard can convince you
0:58:15 that huawei can only make x amount of chips yep yep it’s it’s an interesting moment in time because
0:58:20 we just have all of the trade negotiations right so everything is like volatile and it’s just like
0:58:24 certain things are just on the table and they’ll be willing to discuss them right and we see the
0:58:30 chinese bringing forward at least according to reporting the idea of hbm right and it will just
0:58:33 be interesting to see what the government is going to say it’s going to draw a red line we had statements
0:58:40 before the trade negotiations in london that h20 is above the red line right so they wouldn’t negotiate it
0:58:45 and again we can all try to put together the story what happened here what not if it’s part of it we
0:58:49 we won’t know for sure right um but it will be more discussions about these kinds of things the chinese
0:58:54 companies can bring it up but i’m also more interested in the semiconductor manufacturing
0:58:58 equipment companies if nvidia got this beautiful deal i know what be doing right they’re all trying
0:59:03 to give the president elsewhere ring um and it just seems like it’s a handful of people are making
0:59:07 these these decisions and i just hope they’re like well informed of which things are more important
0:59:13 right um if if i see any news about euv machines being sold to china i’m probably going to get a heart
0:59:18 attack and i just really don’t want this to happen i mean i think just from a personal transaction
0:59:27 perspective like there isn’t someone in the semi-cap equipment ecosystem that trump’s going to give the
0:59:33 time of day like he felt like he had to with jensen because you know this is like america’s most important
0:59:41 ceo um i don’t think any of those folks have this sort of like panache and skill to make it work and
0:59:46 even you know ben thompson who i gave some i gave a hard time for earlier in this podcast
0:59:53 understands very clearly that uh there’s a lot of downside risk in settling um more tools to china
0:59:58 than we already have i would go even so far it wouldn’t be good for jensen if huawei is not good
1:00:03 at producing ai chips right so wouldn’t wouldn’t be in their interest to say like hey yeah let’s make
1:00:08 sure we sell them our chips but really let’s make sure to hit them on every single dimension we can
1:00:14 to make sure huawei is just less competitive right um i would love to see that this would be at least a
1:00:21 good part of the story here i think going back to congress i think congress will be interesting to watch
1:00:28 on on this issue because you the trend in congress has been congress has vocally pushed for tougher
1:00:33 controls both in the first administration and under biden um not universally but i think that’s been
1:00:38 the the predominant push and so now i think we’ll be interesting to watch uh senator cotton for example
1:00:44 and and what he does or does not say publicly on on this issue chris do you want to sort of tie the
1:00:52 you know tease out the russia comparison a little bit i mean congress like really not happy um they
1:00:57 ended up putting some sanctions on the table um what was the what are the dynamics been uh there
1:01:03 over the past six months well i guess the last six months in russia have seen congress
1:01:09 officially not play much role at all they put sanctions legislation on the table and then pulled
1:01:13 it back actually after trump requested it but i i think i would say there’s been a number of
1:01:17 republican senators who have been influential in shaping trump’s thinking lindsey graham for example
1:01:23 seems to play a role in shaping trump’s thinking on uh putin over the last six months and the way that
1:01:28 putin is stringing along now you know we’re going to alaska later this week and so maybe all that will
1:01:33 prove irrelevant if trump changes his mind but it does seem like you could argue that that even though
1:01:39 congress has done nothing on russia in fact it it has helped change thinking in the white house i
1:01:43 wonder if the same will be true here but but this seems like a place where trump’s going to make more of
1:01:48 his own decisions especially as far as it intersects with the china trade negotiations which it seems like
1:01:54 it may well yeah and it’s it’s kind of a less salient thing than a land war there’s no domestic uh
1:02:02 constituency yeah yeah yeah just weirdos tech national security podcasts um
1:02:06 other stuff we should get to
1:02:12 planet do you have anything to say about the um because before before this week it was reported
1:02:19 that nvidia is coming out with a downgraded version of some new downgraded chip post h20 but i guess
1:02:26 that’s now the b40 or b30 i think yeah yeah that’s now irrelevant because of the h20 um it’s unclear
1:02:32 right if people we we flip-flopped the decision on the h20 but notably there is still a license requirement
1:02:36 right so nvidia got the license granted so if they wanted to go all the way back they could
1:02:44 have removed the license requirement right so from october 2023 to april 2025 there was no license
1:02:49 requirement then they introduced the license requirement which is still intact the only thing
1:02:55 which happened as of speaking last friday is they granted the licenses according to reporting right so if
1:03:01 they still want to set a chip which is not subject to export controls they would produce a new chip called
1:03:08 b30 b40 it needs to be lower the computational power threshold so the same as the h20 and also have lower
1:03:12 memory bandwidth and according to the reporting i think ft leaked what is in the isn’t formed later
1:03:19 it needs to be less than 1.4 terabyte per second memory bandwidth again the h20 is at four terabyte per
1:03:24 second right so the b40 would then probably not use hbm anymore it would probably be used to like
1:03:29 an inferior memory technology but significantly cheaper because why use hbm if you can’t have that many memory
1:03:35 bandwidth anyways uh so-called gddr technology which we usually use for our graphic gpus and
1:03:41 again if people talk about this is only the fourth best chip i don’t think the h20 is the fourth best
1:03:48 chip i think the b30 b40 that’s a more fair description of a fourth best chip and i would still
1:03:53 not call it obsolete chip but it’s definitely a worse chip it’s only like you know um with a chip where like
1:03:57 again the us government at least decided here’s where we draw the new lines this chip is fine to be
1:04:01 exported before the license so it could still be coming i have not heard they’re stopping the
1:04:06 production yet i guess nvidia is making a calculus right now on how much demand there is but it’s
1:04:10 clear the case the h20 is better the question is will all the licenses be granted going forward right
1:04:17 and trump has said or he said at the press conference a couple days ago that
1:04:26 he’ll consider a downgraded uh blackwell are there ways we should think about what that might look like
1:04:31 if in fact material material materializes of course with huge questions over whether or not that’s actually
1:04:38 real one thing which stood out he he said like 30 or 15 to 50 percent less performance and i think what
1:04:44 many people are just missing on on ai chips and computing chips get exponentially better if your
1:04:50 chip is 15 less that’s nothing that’s still the same generation right so if you really want to sell worse
1:04:56 chips you need to go back a few generations and then the chip needs to be like seven times worse not only
1:05:02 50 or 15 right so there’s an argument to be made that you want to sell worse chips but it’s not a little
1:05:07 bit of a downgrade we really need to take the exponentially into account right if we trim down
1:05:14 a blackwell chip for example a b200 by 15 to 50 percent still like twice or three times as good as
1:05:20 the huawei chip right and again we can produce millions of them but huawei struggles according to
1:05:26 reporting producing 200 000 this year so again that’s just a key thing to get right here um and yeah people
1:05:32 keep in mind the exponentials here right like chips get exponentially better 15 to 50 percent trim
1:05:38 is is nothing in the grand scheme of things and i would um at least yeah make my voice heard to say
1:05:42 this is this is probably not a good idea of what you’ll be doing here the government two lines before
1:05:44 and the lines are way lower and i think that’s where they should be
1:05:57 all right well it’s been a long road since october of 2022 thanks for sticking with us everyone chris
1:06:07 leonard always a pleasure um yeah never a dull moment on the china ai semiconductor export control
1:06:11 i’m sure we’ll speak again let’s see
1:06:18 this information is for educational purposes only and is not a recommendation to buy hold or sell any
1:06:23 investment or financial product this podcast has been produced by a third party and may include paid
1:06:29 promotional advertisements other company references in individuals unaffiliated with a16z such advertisements
1:06:36 companies and individuals are not endorsed by ah capital management llc a16z or any of its affiliates
1:06:44 information is from sources deemed reliable on the date of publication but a16z does not guarantee its accuracy