AMD 与 Cerebras 推出低延迟 AI 推理合作方案，但直播内容并非新增可交易催化

【实时直播转录 · Bloomberg TV+】
来源:Bloomberg TV+(bloomberg 直播间 btv_us) · 标题:Bloomberg TV+
时段:2026-07-24 23:30–00:00(Asia/Shanghai),共 204 段
说明:以下为该时段该直播间的连续语音转录(直播 ASR,可能有同音/识别误差)。请从中识别与 AI算力/芯片/服务器/光模块/液冷/数据中心/大模型/机器人·具身智能/自动驾驶 等相关的、**可交易**的公司/产品/订单/产能/需求/合作/高管或官员表态,判断是否构成可交易事件;若通篇无可交易内容,判为不做。识别误差请结合上下文与常识修正,勿据孤立音近词臆断。

--- 转录正文 ---
[23:30] 🎼There has been a lot in the world of AI about security incidents, about the debate on closed versus.
[23:30] 🎼Open and a lot of industry moves as well, Capex in AI infrastructure has been a dominant theme, but it really is.
[23:30] On the AI and software side, where we focus, major tech leaders took to social media to tout the benefits of open weight.
[23:30] IIn a letter signed by NviIDdia, Meta, Microsoft and Dresson Horowitz and many others. This comes as Moonshots Kimmi K3.
[23:30] It's been shaking the AI industry just this week, spurring discourse on China, potentially using U technology to.
[23:30] head in the AI race, Michelle Geer, CEO of the Crack Institute for Tech Diplomacy at Purdue and also a former assistant secretary of State.
[23:30] For global public affairs under the first Trump administration writes, the U won't be able to win the AI race simply by limiting.
[23:31] China's technology and joins us now and frankly that's where the debate is, know you will have seen the headlines this morning.
[23:31] The open letter essentially on open weight, the timing of that, there must be a reason for it.
[23:31] I think you're actually seeing Ed is a really good example of Silicon Valley in Washington, DC. talking past each other because if you look at that.
[23:31] The real debate isn't about open weight models, it's about Chinese open weight models, which unfortunately.
[23:31] Are the most available, most price effective option that's on the table right now, and especially for a lot of new businesses and startup.
[23:31] That are trying to build their AI stacks, and so the real debate is on Chinese open weight models and the problem there right.
[23:31] We know that presents national security risk, we know it presents corporate risk, given all that we know about Chinese technology being untrustworthy, It's why we've banned.
[23:31] It's why we banned and then had to restructure TikTok, it's why there's legislation moving through the house right now about banning connected vehicles.
[23:32] that are coming from China, and so there's a pattern here and I think the focus on what do we ban versus what do we keep?
[23:32] Is actually misplaced focused by both the private sector and the U government. The focus should be on how do we turbocharge.
[23:32] Aica's open weight ecosystem, so we have a world class, robust, really price effective offering for the rest of the world.
[23:32] We can diuse American AI as fast as possible and not only overseas, but here at home, where businesses are trying to grow on the here at home bed.
[23:32] I've read the open letter as many times as I could before we came to air, and I'm thinking, what is the concern?
[23:32] them writing it and then putting it out. One take is the concern that Washington just overregulates over.
[23:32] So models right in a way that is detrimental to American interests, where do you sort of sit on that debate?
[23:33] Yeah, well, I think it's because open weight models cheaper, open weight models have become really core to a lot of how this.
[23:33] developing their AI stacks as you mentioned earlier, we just saw big tech wipeout $890 billion because AI is really expensive.
[23:33] Everybody's looking to see if we can keep up with this spending bubble, and so the cheaper, good enough versions are really important for our private site.
[23:33] To be building their AI stacks, the problem is, it comes from an adversary and there is no really robust US trusted.
[23:33] Alternative to the Chinese open weight models that we're seeing and then the Kimmi3 launch this week catalyzed our awareness of that. And so the debate now.
[23:33] Is not really just open weight, it's Chinese open weight and can we get an a American alternative out there fast.
[23:33] Did sort of an extended interview of Axios and basically said, the top lines are the Chinese models are excellent.
[23:33] And that the open source models that are excellent to should be used. And so you, based on your line of argument that.
[23:34] The risk is too great. If the open model comes from an economic adversary, it comes down to how influential is Jensen Wong.
[23:34] With this administration, well, I actually think it comes down to how fast can we turbochar inmer.
[23:34] Open weight ecosystem, look all the things that he's saying and other Silicon Valley leaders are saying, make sense if you're looking at this purely through.
[23:34] innovation and a commercial lens, It makes a world of sense. If you then factor in the communist adversary lens, the risk calculus looks a whole.
[23:34] lot different. And that's why these technologies are different. Look, we just spented the better part of a decade with businesses thinking about how they.
[23:34] Duple and they deris from China because of the supply chain risk, the financial risk, the corporate risk, the national security risk.
[23:34] Talking about entrenching Chinese technology into the very foundation of our American company's AI stacks.
[23:35] No sense, and so how do we focus instead on getting trusted American AI cheap, effective price.
[23:35] AI into the hands of American companies and as much of the world as possible, that's where the focus needs to go.
[23:35] the beginning, you know, your argument through the lens of policy research reflecting on your time in government is that at the end of the day, America.
[23:35] won't win the AI race by restricting China. So you indicate that America needs to be proactive in its own approach.
[23:35] Get its own house in order, how far does today's action go, what else needs to be done?
[23:35] Yeah, I think, look, the President has this great group on his PCAS to this presidential advisory board with a bunch of technology leaders I think they should get.
[23:35] As quickly as possible and figure out how they go on offense really quickly with American openweight AI and by the way.
[23:35] Get all of our allies on board because even if we ban and limit the use of Chinese technology, Chinese AI models here, if we're used.
[23:36] AI models from America here, but the rest of the world is running on a Chinese AI stack that's still a problem for us. and so I think they can rally around that.
[23:36] And look we have a lot of lessons when I was at the State Department, Huawei was a threat across the world and theres a lot of lessons to be learned.
[23:36] ThereAnd a big lesson is that we're not going to win on principle, we're going to win on price, so how do we get much more cost effective?
[23:36] 'Tmer AI open weight models out into the world now we're a post from.
[23:36] cratcios from a couple of days ago, we covered that story, the accusation that Kimmi K3 was distilled from an anthropic model.
[23:36] And used illegally in VDdia Blackwell systems. We've been over that. But just pointing that out that Mr. Crrattzys is also the.
[23:36] Chair of PCAS, along with David Schs Jensens on PCS, for example, Michelle Geer, CEO of the Crack Institute for Tech Diplomacy at Purdy Backumb.😊
[23:37] 🎼Thank you very so coming up the AI race also shifts to speed AMD and Crebbras andbel a new partnership.😊
[23:37] 🎼Rever CEO Andrew Feldman joins us next, Fast In, this is Bloomberg Tech.
[23:38] Is there anything that humans today?
[23:38] that will change the lives of the humans currently on the earth Well, first of all, people who switch over are getting the benefit.
[23:38] 🎼But coming back to the point about the fossil fuel industry and our debt to them, there's an old saying that came from the kingdom.
[23:38] Saudi Arabia years ago, the Stone Age didn't end because of a shortage of stones, it ended because something better came along.
[23:38] Some betters come along now to replace fossil fuels, which are dirty and polluting and threatening humanity's future.
[23:38] 🎼T in the history of the world is solar energy. Second chief is is wind energy. And now batteries that the only thing that's come down in price.
[23:38] Faster than solar is utility scale batteries doubling every single, why are we ending the wind farms that are now being built off the coast?
[23:39] The United States, we're ending them because Trump is insane on another subject.
[23:39] I to that let me add one point on the solar. If you look at all of the new.
[23:39] 🎼generation installed worldwide last year. I often asked people, what percentage of it was renewable. And people say, you know they're looking for.
[23:39] 🎼High number, say maybe 30%, 35%. correct answer is 93% because it is taken.
[23:39] 🎼the.
[23:39] InWhen and wherever it happens I'm mi mckee on the Mexicancan border and this is bloomberg.
[23:39] 🎼The globallobal economic playbook is being rewritten and the levers of power are moving in real time to see where the world is headed you have to look.
[23:39] 🎼the action I'm Ha Linda Ar in Singapore I'mhi in Mumb welcomecom to the new economy.
[23:40] 🎼each month we dive deep into how the world's most exciting rising economies are shaping the global future.
[23:40] 🎼Join the conversation, subscribe to imagine on your favorite podcast platform today to bloom this.
[23:40] I'm lis mateteay and I'm chrisine de Raffiti we're tracking breaking news today from wall street to washington.
[23:40] 🎼The bloomberg into your weekend routine what listen stream bringing you up to the minute space news when and where.
[23:40] 🎼it happens, I'm Ed Ludlow at NASA's Kennedy Space Center in Florida, and this is Bloomberg.
[23:40] Shares of Crebra systems.
[23:40] Down about 10% right now, they had jumped yesterday on some news that AMD is teaming up with Ru.
[23:41] Systems on a new server designed to slash response times, taking direct aim at NviDdia, joining us to explain the deal.
[23:41] The technology Andrew Feldman, cofounder and CEO of Crebra Systems. So this is how it's going to work. You basically have a helios server.
[23:41] Srebra server in combination, how is that going to work and?
[23:41] The sort of split on the workload Sure, good to be back, Ed, thanks for having me The way to think about it.
[23:41] Is that the inference problem is comprised of two parts, we call the first part processing the prompt.
[23:41] And we call the second part generating the answer, and those two parts of the problem have very different computational requirements.
[23:41] And that opens the door to address them with two different machines processing of the.
[23:41] The prompt, the analyzing the query, that part is a problem that can be parallelzed.
[23:42] And for that type of work, GPUs are very, very good, and helios will be the best.
[23:42] The second part of the work, which is generating the answer at extraordinary speeds.
[23:42] Sebs is the best in the world bar none, and so by bringing these two solutions together so that.
[23:42] Helios processes the prompt, Creus generates the answer. We create a single inference.
[23:42] That is the fastest in the world and has this extraordinary throughput.
[23:42] I'm not trying to make a scan or a negative out of it, but for you, how does the sales channel work, right?
[23:42] AMD has like this pipeline of projects and that infrastructure is like, okay, we're going to use helios systems but now.
[23:43] They can also get the cerebriris server alongside it. How have you guys agreed to split the revenues, How have you guys agreed to do this?
[23:43] In a way that is economically equitable, it's a really interesting deal.
[23:43] Speed makes markets bigger, it doesn't make markets smaller, this isn't about carving up something that's a fixed size.
[23:43] Right fast inference is productive inference, and where AI is productive, people are willing to spend more and more.
[23:43] And more and so the first application of this solution will be in the Crebers cloud.
[23:43] And that will be later this year and shortly thereafter it will be available more generally, I think we have an enormous backlog.
[23:43] They have an enormous backlog, I think we have customers around the world where who are demanding.
[23:43] Fast, fast inference at extraordinary scale, and the solution we're building is head show.
[23:44] There's above anything else. I find the technology approach like really interesting. You know, the parallel example is NviIDdia.
[23:44] And Goc with a queue and Goc 3 LPX, and like in that case, it's integration of the Goc.
[23:44] System into the same server design, right, you didn't take that approach. Did you talk about that like you know what what were the?
[23:44] OnWellWe don't quite know what NviDdia is doing.
[23:44] really delivered it to market yet, and so we are interested to see how that works out.
[23:44] Were able to do because we use open standards based technology for our IO.
[23:44] To build a solution with AMD quickly and easily.
[23:45] We sort of support open standards, remember we've also done this with trainingium from AWS.
[23:45] And so we adopted the same approach that by being open, by being standards based, right, half the leading.
[23:45] Chhiip makers, AMD, AWS with traininganium are now using R.
[23:45] In a disaggregated approach for high speed inference delivery. So I think the answer is.
[23:45] By being open and by being standards based, we can rapidly partner with other.
[23:45] Members of the you've also been a busy guy, we were talking off air about how busy everything is.
[23:45] You know, you just did a very big IPO kind of interesting to know like who approached who, whose idea was this?
[23:45] I've known Lisa for a long time, she was CEO when my last company was acquired by AMD,'ve watched.
[23:46] sortrt of her do extraordinary things with that company, the returns over her tenure as CEO.
[23:46] Our mind boggling and so we are in constant communication we are close with her CTtoO, Mark papermaster.
[23:46] People we've known for decades, and so we're in constant communication talking about how we might collaborate, what we might do.
[23:46] Together, and this was an idea that emerged from those discussions and made perfect sense.
[23:46] In the NviDdia Gck with a Q example, which I appreciate, not your company, like completely separate, but the mechanism was an aqua hirere.
[23:46] as you know, prior to your IPO, we reported at Bloomberg that some of the fabulouss chip names.
[23:46] Held talks or interests expressed interest in also acquiring cerebras prior to its listing.
[23:47] Did you discuss any of those kind of mechanisms with Lisa about the merits of her making an equity investment or some of the structures?
[23:47] The AMD has deployed in other of their arrangements Sure, AMD has been an equity investor for quite some.
[23:47] They were an investor in some of our mid and later stage rounds.
[23:47] And so we have been talking with them and engaged in discussions on how we might work together for a long.
[23:47] As we are with many of the players, and that's the advantage of being open, that's the advantage of not being a walled garden.
[23:47] ageOf not having a proprietary IO technology, but using standards based high speed.
[23:47] Ethernet technology, we can engage with AMD, we can engage with AWS, we could engage with Google.
[23:47] anybody who builds a part that would integrate into an extraordinary solution.
[23:48] Andrew, really quickly, I had a long conversation with SK group Chair Kt1 very recently, and he said basically the difference between China's approach.
[23:48] AI in the US is China's going for the lowest dollar per token, whereas America is still in a place where it's focused on the highest quality tokens.
[23:48] From the non HBMs standpoint, would you kind of weigh in on that. Yeah, right, So we don't use HBM and that's one of the.
[23:48] Real advantages we have, so we can generate tokens for for less because we're not dependent on this sort of.
[23:48] Sly chain constrained part I think it's unclear.
[23:48] Whether the Chinese model makers are actually developing for less or piggybacking on technology that.
[23:48] Other people have invented right I think that's an open question. What I don't have the answer to, but I know theres some.
[23:49] Sngly held views that they are stealing, distilling, I'm not sure, but what I do know.
[23:49] I don't understand their cost to build these models Open AI is one of our larger customers, we do have an understanding.
[23:49] of the amount of compute they need to build. And we have an understanding of the amount of compute that other frontier US labs have. We' really.
[23:49] Have an understanding of what the cost is, so the question is really are they making lower cost AI because they're borrowing?
[23:49] Other people's technology, or do they have some really interesting inventions that allow them to make AI for less?
[23:49] That's an open question and one that relates to your previous segment and what we should do about that if it is the case.
[23:49] Very very quickly, you've seen the letter this morning, leadersers in U technology emphasizing America's need to focus on open weight.
[23:49] You point on that Look I think right now the market's been made by open AI with a fast follow by.
[23:50] Anthropic, and they have invested an enormous amount of money, a mind boggling amount of money, and that they have made.
[23:50] This market, and I think the followers, both from China and others.
[23:50] OtherAre trying to capture this with a fast follow strategy. And I think it's unclear to me what the right.
[23:50] Approaches, I think obviously we want lower cost AI what we can control as cerebraus is.
[23:50] By building like we're doing with AMD, solutions that deliver more tokens per unit power.
[23:50] Deliver more tokens per dollar to deliver those tokens faster so they're more productive. so you can pay more, right.
[23:50] The problem isn't that tokens are expensive, the problem is that it's hard to measure how much productivity you're getting from tokens.
[23:51] And what we know is that when you make things faster, they drive up productivity, and so these are the dimensions we can work on Ed.
[23:51] We're urgently working on them every day A Feldman, Ru System cofoer and CEO, thank you very much coming up Viking G.
[23:51] 🎼B admit it's missing out on the AI trade, this is Bloomberg Tech.
[23:52] We should not deny.
[23:52] That change is coming and change is not coming just in data centers and AI. as a world, we are building infrastructure.
[23:52] We're doing energy transmission, we're building next gen manufacturing, we're ramping defense, and we're doing this thing called AI and data.
[23:52] We will see the largest need for capital ever, and that's what we've seen so far this year, and so now to your specific question.
[23:52] It's about underwriting, there are a number of very strong US tech companies who run diversified businesses.
[23:52] Amazon, Google, Microsoft, who are interested in seeing these centers built and are interested in not consolidating the debt on their bound.
[23:52] And to the extent they are prepared to lend their credit One can lend against those types of things and make sense.
[23:53] Underwriting decisions.
[23:53] 🎼Africa, a region rife with potential for growth and opportunities, but not without.
[23:53] It us as we get to the heart of African business, politics and economics and ask the question, what's next?
[23:53] On this episode, we're trying to build AI infrastructure so that it's close to where the users are and where the demand is going to be.
[23:53] 🎼ric dual goal is both connectivity and premier July24.
[23:53] 🎼Bloomberg training day is about to start and you're already looking for that edge opening trade brings you everything you need to know as markets.
[23:53] 🎼Open across Europe. I' Gu Johnson. I'm Anna Edwards. and I'm Tom Mc. This is your opening trade only on back.
[23:53] With AI stock soaring this year, Viking global investors told clients that its conservative position on the sector was a quote.
[23:54] Missed opportunity, the Fs flagship He fund gained just 2.6% in the first half of the year, trailing peers with greater AI exposure.😊
[23:54] Like2 and yet the firm doesn't plan to change course Bloomberg's He Palmer broke the story, so they're sticking to their guns before.😊
[23:54] Yes a really interesting stance from Viking global. This is a hedge fund that manages $56 billion.
[23:54] They're taking a divergent view from what we're seeing from many of their peers that are really jumping on this AI beta market.
[23:54] AndMeanwhile, Cou really a lot more cautious, not partaking in a lot of these stocks. And we're seeing this in their performance with their hedge fund.😊
[23:54] Up about 7.5 in the second quarter, but still only up 2.6 for the full year so far.
[23:54] That's a great degree less than many of their competitors. So if they're not budging and they acknowledge the kind of the.
[23:55] in a missed opportunity. Is it that they see some opportunity in their current positioning that's going to change their fortunes. So they're very concerned about.
[23:55] The risks that they see in the market. So they're concerned that we could see potentially a correction, some sort of sudden shock.
[23:55] That curbs buying of these stocks that are highly popular. They're worried about the valuations that they're seeing. So they do partake in some.
[23:55] IInvestmentsThey are a buyer of Samsung, which has worked out really well for them, but a lot of their portfolios and other things.
[23:55] Like consumer, financials, industrials, and some of those stocks, some of those trades really aren't working as well as they would hope.
[23:55] Some of them, the firm says to investors, is potentially being unfairly considered an AI loser.😊
[23:55] Very quick he, historically good performance of Viking or bad performance, you know, historically strong performance and because of their caution.
[23:55] A couple years ago in 2020 and 2021, that saved them from a lot of the troubles we saw amongst their peers in 2022.
[23:56] When we saw that stock market crash when it came to tech stocks and private valuations, and now it saved them then, investors may be hoping.
[23:56] 🎼There's a correction in the future it'll see them today Blue with a read you very much that does it for this edition of Blueloom.😊
[23:56] 🎼What a week more tech earnings coming up next week, this is what it looks like recap on the pod, I really recommend some of the conversations throughout today's show.😊
[23:56] This is Bloomberg Tech.
[23:57] In case you missed it on Bloomberg crypto.
[23:57] We've just launched the first regulated perpetual futures product here in the United States and the interest has been pretty astronomical.
[23:57] Get things right with our regulators at the CFTC on products before we list them. I think there's a lot of both retail and institutional interest.
[23:57] 🎼future we're likely to see that continue to How do distinguish between legitimate informed trading insider trading and there's.
[23:58] talkals with Congress like are you going to ban people who may have special knowledge that the public does not know so we prohibit members of con from.
[23:58] Trading on our platform, US law already prohibits members of the US government from insider trading on information that they have it has been.
[23:58] Exciting to watch Congress explore legislating in this space, and we're largely supportive of legislative efforts to target those contracts where they may have in.
[23:58] information and we do that work on the exchange to prohibit and attack that activity already don't miss bloomloberg crypto live tues.
[23:58] Mind being called an ideological lunatic or a bunch of left wing nut jobss, you know, I've been called worse things than that.
[23:58] 23 and me I think is so incredibly valuable, we're coming back.
[23:58] What magazines were you reading as a kid I don't think I was reading magazines until I was in them so is it paradise ahead or Gattica?
[23:59] Maybe somewhere in the middle.
[23:59] 🎼M this is We back to the opening trade. Itsberg money. This is the Asia trade. This is Wall Street read We to balance of.😊
[23:59] You're watching Blo welcome to B this weekend this is Bloberg bringing.
[23:59] Political news whenever and wherever it happens i'm jennifer aasja in the lasuttu highland and this is bloomberg.
[23:59] In case you missed it on Bberg grief, you have semi z in one corner, but another you know.