news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

these two days, the "most wronged award in silicon valley" must be given to huang renxun.

after releasing an impressive second-quarter financial report, nvidia's stock price unexpectedly fell dramatically, with the maximum drop reaching 8.85%.

first, there are internal reasons. first, nvidia's gross profit margin has declined compared with the previous quarter. the official main reason is "inventory reserves for low-yield blackwell materials and an increase in the portfolio of new products in the data center", and it is said that it may continue to be affected in the future. second, nvidia's next-generation chip blackwell has a "design defect" and will be delayed in delivery. x survey | nvidia has entered a turbulent period, why are we still optimistic?

but more importantly, wall street no longer believes in nvidia. after all, in the face of the heated "ai bubble theory", nvidia, the "shovel seller", was the first to be affected and was even seen as "cisco 2.0".

however, after the financial report was released, huang renxun quickly responded to all doubts in an exclusive interview with bloomberg.the interview was a bit confusing, so shidao screened and summarized it, and the content does not affect the original meaning.

next year will be a bumper year. i never thought of becoming a cloud service provider.

regarding the blackwell incident, the host first put forward a conspiracy theory.

host:the market is looking forward to blackwell, especially people want to see more specific details. i am browsing all the core content.it seems pretty clear in this record that this was a production issue, not a design flaw.however, what is the reality? specifically, how is the timeline delayed? what is the impact on product revenue?

jen-hsun huang:well, let me think about it. we have made a large-scale adjustment to improve the yield of blackwell, and now we are testing blackwell systems around the world. we have given people a tour of the blackwell system that has been deployed and running. you can also find many pictures of blackwell systems online. we have started mass production, and the first products will be delivered in q4, and will gradually increase production thereafter.the demand for blackwell far exceeds the supply, especially in the early days, because the demand is so great. but we will provide a lot of supply and start to increase production in q4. we expect blackwell to bring in billions of dollars in revenue in q4 and next year, and next year will definitely be a good year.

apart from big customers, where can nvidia see the money?

jen-hsun huang:indeed, large-scale data center customers account for about 45% of our data center business, but we are relatively diversified today - covering internet service providers, sovereign ai projects, and various key areas such as industries and enterprises. these businesses account for the remaining 55%. accelerated computing is also widely used in many fields such as database management, data pre-processing and post-processing, video transcoding, scientific simulation, computer graphics and image processing.people use our products for a variety of purposes, and generative ai is just one of them.

regarding ai sovereignty, nvidia wants to cooperate with various countries.

jen-hsun huang: this is not necessarily (a deal with a single government), but involves regional ai infrastructure construction around the world, although they are sometimes government-funded. for example, countries such as japan, india, canada, the united kingdom, france, italy, singapore and malaysia are promoting the development of local ai infrastructure through government funding.because they realize that their country's knowledge and data resources are also a natural resource and national resource, and need to be transformed into national digital intelligence.this is called sovereign ai, and you can imagine that almost every country in the world will eventually realize this and start building their own ai infrastructure.

regarding transformation, nvidia said, "i actually don't want to be a cloud computing service provider."

host:do you plan to become a cloud computing vendor?

jen-hsun huang:(pause) no. we built the gpu cloud (dgx cloud) to be the best nvidia cloud service on every major cloud platform, and this strategy has been very successful.

first, we focus on our own and only do the best nvidia cloud. by building dgx cloud inside platforms such as azure, aws, and oci, we can ensure that our own gpu advantages are maximized to improve the efficiency of ai infrastructure.

secondly, the benefits should not be lost to outsiders. nvidia has carried out cutting-edge projects on dgx cloud, such as autonomous driving, general robots, and omniverse, to demonstrate how to use cloud services to promote technological progress. these internal applications not only verify the performance of dgx cloud, but also provide nvidia with valuable usage data to help further optimize products and services.

thirdly, nvidia aims to be the "tsmc" of ai model manufacturers. nvidia's ai foundry service provides professional ai model development support for small companies that lack resources, positioning nvidia as a "foundry" in the ai ​​field to meet the needs of various industries for customized ai solutions.

buffett's "successor": it's too early to sell nvidia

shidao once discussed "why is nvidia optimistic":

first, even if there is a bubble in ai, it is still a long way from bursting.

second, the current gpus are not fully utilized. for example, a16z plans to expand the scale of its gpu cluster to more than 20,000 gpus, based on its stockpiling of thousands of gpus. another example is that compared with the training market, the inference market is a larger market, and more importantly, it has not yet been truly opened.

third, nvidia's moat is very solid, and the era of general-purpose structures it represents is in the midst of disruptive demand.

of course, if nvidia’s strong bargaining power helps push the cost of ai computing power to remain high for a long time, thereby inhibiting large-scale innovation, backlash is also possible.

samantha mclemore, a well-known atypical value investor and one of the "next generation buffett", expressed her views.

she believes that the ai ​​boom will not repeat the mistakes of the internet bubble, and nvidia will not be the next cisco.

“we bought nvidia at about 25 times earnings, when everyone was worried about the sustainability of nvidia’s earnings and whether it was facing an ‘earnings cliff’, with nvidia being compared to cisco after the internet infrastructure boom in the late 1990s.as of the end of the second quarter, nvidia's p/e ratio was 47 times, and cisco's p/e ratio peaked at 150 times. i am not saying that nvidia's p/e ratio will also reach such levels.

the direction of nvidia's stock price will depend on whether its earnings can continue to grow.i don't see much risk in the short term: ai data center construction is still ongoing, meta, microsoft, and alphabet plan to continue increasing capital spending next year, and there is still a supply shortage of gpus, especially nvidia's blackwell chips due to be launched later this year.”

mclemore believes it's too early to sell nvidia.

“if we talk about selling, it is more from the perspective of valuation.we have been looking at a variety of possible scenarios, and nvidia has been beating expectations every quarter for a year.corporate ceos want to invest more money in ai because they don’t want to be left behind, and i don’t think this trend will end anytime soon.

we are concerned about nvidia's long-term growth rate. nvidia's current operating margin is 65%. when making forecasts, our assumption is not that nvidia's operating margin can always be maintained at this level, but around 40%, similar to microsoft's operating margin. huang renxun's vision is to create an accelerated computing platform, which is the future of computing. nvidia's hardware, software and developers are the company's competitive advantage, and it has the ability to maintain this advantage.if nvidia stock returns 10% per year, it will outperform the market, but it will be a wavy path.

so far, nvidia is at the forefront of the artificial intelligence revolution and has reaped most of the profits in the field.in the short term, there are still many fundamental positives in the technology and artificial intelligence sectors, and we need to continue to see what will happen in the end, but everyone is currently chasing companies in these fields.

taking all of the above into account, we believe it is too early to sell nvidia.”

chris miller: don’t just think like an economist

regarding nvidia and huang renxun, chris miller, author of chip wars, provides some fresh perspectives. (bonus: follow the shidao wechat public account and reply "chip wars" to get "chips war: the battle for the world's most crucial technology")

his underlying logic is: whether it is to understand company growth or international situation, you can't just think like an economist.

because economists view everything asutility function, but it has limitations. (the ps utility function is expressed as u=e(r)-0.5aσ^2, which is a tool for quantifying "risk & return".)

miller gave two examples respectively - "business tycoon" huang renxun; "political tycoon" putin.

ten years ago, after being warned by wall street, huang renxun still invested money in building the cuda ecosystem. today, the entire cloud ai demand depends on the cuda ecosystem, which has become nvidia's moat.

miller believes that for company founders, if you follow a purely economic mindset—focusing on return on equity and maximizing shareholder value—you will miss a lot of opportunities to truly drive competition and become a successful company.

for example, putin.

miller was shocked by putin's change: 20 years ago, when putin first came to power, he appeared as a relatively modern russian leader. he was reforming the tax system and doing what a political leader should do. at that time, when american scholars discussed putin's motivations, they often focused on the economy. some economists once said, "isn't money putin's main driving force?" there are many examples that prove that putin and his friends are corrupt and build luxurious palaces on the black sea coast. (狗头保命)

but miller gradually discovered that putin’s motivation was not all about money. when russia invaded ukraine in 2022, putin expanded in the name of peter the great and catherine the great. this showed that “modern people” are not always driven by modern impulses.the desire for power, glory, and control, the desire to be superior to others - for good or ill - occupies a central place in the utility function of many people.these impulses may seem primitive, but they exist in everyone to some degree, and we can ignore them at our peril.

miller argues that governments don't think like economists, either. they tend to maximize glory, territory, reputation or power. sometimes, they simply want to win by beating their rivals, just like great entrepreneurs.what drives nations and the individuals within them is often difficult to quantify, and these unquantifiable factors are often more important than strictly economic variables.

interestingly, huang renxun mentioned "ai sovereignty" in the interview; mclemore also raised concerns about geopolitics in the rest of the interview, and even considered it "the biggest investment risk at present"; miller also listed "great power conflict" as a risk that is currently seriously underestimated.

will these "nuclear bombs" actually explode?

miller said: as the saying goes: "history doesn't repeat itself, but it often rhymes."

the “most terrible rhyme” is the world before 1914, on the eve of the outbreak of world war i. it was also a period of geopolitical tension and an era of rapid technological development.for example, railroads facilitated the formation of european empires.

in 1800, european empires basically only ruled coastal areas or small islands. for example, britain only occupied the canadian coastline, some caribbean islands, and coastal areas of india. in 1900, with the rapid development of railways, every european empire went deep into the interior, and railway technology also made the colonization of africa a reality.

the "relatively positive rhyme" occurred in the 1980s and 1990s. at that time, computing technology began to be applied in various fields, which helped to greatly improve productivity in the 1990s and 2000s.

in miller's view, there are three key changes in our time.

first, the investment in and application of advanced computing technologies is accelerating. ai is not only a hot word this year, but also a long-term trend.

second, geopolitical competition is intensifying. miller is relatively pessimistic: "some people comfort themselves by saying, 'this is a high-impact, low-probability event with low expected costs.' if you say the risk is low this year, that may be true. but if you extend the time span to the next decade and take into account the compounding of risks every year, i think these assumptions are not valid."

finally, the politicization of international economic life. in the past 20 to 30 years, this area was not mainly driven by political factors, but now the situation is completely different.

everyone relax, maybe it's just sensational. after all, historians can be the best storytellers.