Jul

13

I am sure there are people who treat what Goldman Sachs says as Gospel.

GEN AI: Too much spend, too little benefit?

Tech giants and beyond are set to spend over $1tn on AI capex in coming years, with so far little to show for it. So, will this large spend ever pay off? MIT’s Daron Acemoglu and GS’ Jim Covello are skeptical, with Acemoglu seeing only limited US economic upside from AI over the next decade and Covello arguing that the technology isn’t designed to solve the complex problems that would justify the costs, which may not decline as many expect.

Big Al writes:

The main critique seems to be one of cost/benefit: benefits are, as of now, too small, and costs too high. But it's apparent, even from the cheap seats, that lots of very smart people are devoting lots of resources to (1) improving existing models and creating new ones, thus increasing benefit; and (2) improving the efficiency of the hardware (and related software) to improve processing efficiency and reduce power consumption. Given how quickly things have developed over the last few years, it's hard to predict where we will be even 2-3 years from now.

Fwiw, this caught my attention:

GS US semiconductor analysts Toshiya Hari, Anmol Makkar, and David Balaban argue that chips will indeed constrain AI growth over the next few years, with demand for chips outstripping supply owing to shortages in High-Bandwidth Memory technology and Chip-on-Wafer-on-Substrate packaging—two critical chip components.

So I asked Gemini:

What companies produce High-Bandwidth Memory?

There are a few major companies that produce High-Bandwidth Memory
(HBM), but the top three are:

Micron Tech
Samsung
SK Hynix (https://en.wikipedia.org/wiki/SK_Hynix)
These companies have been producing HBM since its introduction in 2014, and they continue to develop new and improved versions of the technology. HBM is used in a variety of high-performance computing applications, including artificial intelligence (AI), graphics processing units (GPUs), and supercomputers.

What companies produce Chip-on-Wafer-on-Substrate packaging?

Chip-on-Wafer-on-Substrate (CoWoS) packaging is a type of advanced packaging technology that is becoming increasingly important for high-performance computing applications. Here are some of the leading companies that produce CoWoS packaging:
TSMC (Taiwan Semiconductor Manufacturing Company)
Intel
Samsung
ASE (Advanced Semiconductor Engineering)
Amkor Technology
JCET Group: JCET Group is a Chinese company that is a leading provider
of OSAT services in China.

K. K. Law comments:

Regarding the bandwidth constraints, the other one is communication bandwidth which still has yet to catch up with the computational speed. There are limits as to how close they can move their chips closer together even using 3-D integration to shorten the latency caused by propagation delay. In the near term, there are companies, including Intel, working to move the optical communication pipelines to the chip level. That in and of itself also soon presents a challenge as there are only so many optical fibers can be accommodated right next to the chip.

The generative AI appears to the biggest elephant for now because that gets the most attention from general public. However, AI/ML is penetrating into all sorts of applications that have not got much attention. The Goldman Sachs people, despite they are highly paid, they are surprisingly ignorant in this subject mater.

Mar

10

1999-2001 is an interesting period to study as there are many parallels to the present. Back then, the approach of the year 2000 caused a rush to buy new technology to fix all the old systems that used two digits to represent the year. The efforts were generally successful, and doomsday predictions failed to materialize, but the event pulled a lot of demand for technology from the future into the present. When that future arrived in 2001 and 2002, technology sales slumped because everybody already had shiny new systems–no need to buy more. A similar scenario occurred with the rush to enable remote work and communications in 2020, a bonanza for technology companies. What is different this time is that the technology bust resulting from the saturated market already occurred in 2022-23, but now there is a second wave.

K. K. Law notes:

Don't know if anybody ever mentioned about in the 2000 era many optical components manufacturers double-booked and triple-booked the actual TAM. I knew that first hand but too late. Then the CEO of a major component manufacturer called JDSU, now has become a much bigger company, due to multiple M&As, decided to resign in order to sell his company stocks.

Steve Ellison responds:

I vividly remember that JDSU reported blockbuster earnings in October 2000 that for a day or so reversed the Nasdaq rout that by then had been in progress for 7 months. Jim Cramer, who was still running his hedge fund and presumably still had an edge from calling his buddies at bulge bracket firms to find out what big orders they were working that day, pointed out that JDSU's sales were probably just piling up in inventory at downstream suppliers (that had all reported disappointing results); hence JDSU sales were about to go off a cliff. Since I still have the scars from owning JDSU, I developed a tracker of some important companies in Nvidia's ecosystem (chart below). If one of the other stocks turns down, maybe it will be an early warning.

K. K. Law comments:

The actual bookings of all the optical suppliers were over-inflated by many folds over the actual demands by the telecom system manufacturers and service providers . The market eventually realized that.

Steve Ellison adds:

Even in the absence of fraud, when there is a component shortage, buyers inflate their orders to try to gain priority, and executives at the supplier might misinterpret the orders as real demand. When the bottleneck is alleviated, and supply begins to actually flow, the buyers quickly cancel the excess orders in a phenomenon known as the "dreaded diamond", as I discussed in a Spec List thread in 2021.

Feb

7

It is not hard to see this is very late stages of speculative madness but I really would like to know how the risk management teams approve buying Nvidia stock here after adding $200 billion to market cap in 3 days?

Larry Williams offers:

Maybe my cycle forecast for NVDA would help:

Asindu Drileba writes:

I don't know why people are still buying Nvidia. But this is what I personally think of the stock. Nvidia has an 80% market share in the Graphics Card business. Their bread and butter used to be video gaming, 3d animation, video editing, later crypto mining, AI (computer vision), AI (Large Language Models), AI (Image generation) possible new advances may occur in Molecular Dynamics, Self driving cars etc. The CEO had an interesting interview where he talked about possible areas Nvidia may venture into.

But here is one strange thing about high performance computing (Nvidia's Niche): We would think that the better (higher performing) their products are, the less people would buy because people would do more with less right? It's actually the opposite.

— In gaming for example, when graphics cards improved people moved to less polygon looking characters and wanted more details like finer hair & plants. From there they even went to more computationally intensive algorithms like ray tracing that mimic real world scattering of light. Requiring even more compute in subsequent algorithmic advances.

— In Bitcoin, many people using Nvidia GPUs made it more difficult to earn money from crypto mining. Which requires people to have even more Nvidia GPUs just to continue earning the same income.

— In AI, when ever a new breakthrough was made, researchers often trained models with larger datasets, using more & more GPUs. Chat GPT for example was trained on 1 Trillion corpus of text.

So if they do maintain this 80% market share and these underlying industries continue to grow (and make new break throughs). It makes sense that Nvidia will be very valuable in the near or distant future. Buying now (at all time highs) is definitely dangerous but, even if the bubble pops, the underlying industries it facilitates will still be present. And if more breakthroughs in these industries are made, it makes sense that Nvidia still has some value left in it.

Cagdas Tuna responds:

Good fundamental points and there I have 2 counter outlook:

-Gaming industry; I almost everyday play an online game called Destiny 2, and their developer Bungie has reduced workforce around 10%. I know many other gaming companies are reducing/reduced workforce which doesn't give too much optimism in that area.

-Bitcoin mining; there is halving in a few weeks and this will require more powerful computers but it will also increase the cost which in the end will end up new miners losing money in most cases. Only way to maintain gains in mining is Bitcoin price to double or triple in a year.

Even on the best possible scenario it will not add 200 billion dollars worth growth in many many years.

Steve Ellison comments:

Words of wisdom from Rocky's Ghost, posted in the Spec List on April 4, 2017. And yes, I am long NVDA. I believe this is the study Rocky referred to.

Soros and I share very little. However, I have come to agree with him that the right position is to be long "bubble" (however defined). I used to subscribe to Anatoly's view and to be bearish during bubbled but I discovered that from a risk-adjusted-return perspective, it's better to be right "today" than right "tomorrow." Along this point, I read a study that shows a substantial percentage of stock returns occur during the last surge in a "bull market". If you miss this surge, it's very difficult to keep up with the indices in the short term. And in the long term, we're all dead.

Asindu Drileba replies:

Gaming Revenue was about $142B just in 2022. If cloud gaming, something Nvidia is planning todo is successful, I expect this to jump by several multipliers. I expect Cloud gaming to be a bigger business than say AWS. Gaming is really big, I believe you have heard about gaming being bigger than movies & music combined.

The Crypto market cap is $1.6T, a lot of these Crypto currencies use graphics cards to mine their currencies. So I don't think $200B is too much. For Nvidia which is well positioned in these industries, i.e., owning 80% of that market.

Humbert H. adds:

One fundamental point about predicting the future of NVIDIA. It's a complete accident (lucky for NVIDIA) that the hardware optimized matrix multiplication used for 3D graphics pipelines was also useful for AI.

K. K. Law riffs on The Great One:

Confirmation bias. And this is where the AI computation puck is at of course.

Cagdas Tuna realizes:

Now I see why everyone chasing this momentum with FOMO as all assumptions based on Nvidia will get all of the cake in the market!

Dec

31

Mr Fake Meats does not support is own research:

Health effects of dietary risks in 195 countries, 1990-2017: a systematic analysis for the Global Burden of Disease Study 2017

Findings: In 2017, 11 million (95% uncertainty interval [UI] 10-12) deaths and 255 million (234-274) DALYs were attributable to dietary risk factors. High intake of sodium (3 million [1-5] deaths and 70 million [34-118] DALYs), low intake of whole grains (3 million [2-4] deaths and 82 million [59-109] DALYs), and low intake of fruits (2 million [1-4] deaths and 65 million [41-92] DALYs) were the leading dietary risk factors for deaths and DALYs globally and in many countries. Dietary data were from mixed sources and were not available for all countries, increasing the statistical uncertainty of our estimates. [Funding: Bill & Melinda Gates Foundation.]

Note meat does not pop up in this data.

Jeffrey Hirsch writes:

Lot’s of meat works for me. Keto, exercise and sleep. I’m down 50lbs. Skipping the Booze was a big help.

Pamela Van Giessen comments:

Virtually all nutrition studies are pretty meaningless because it is almost impossible to confine study to one food to the exclusion of all else (do people who eat red meat also not drink and exercise regularly; do people who eat low grain diets also eat a lot of processed food and lack exercise, and so on).

Maybe you can hack your health and longevity with diet. Maybe not. I’d err on the maybe not side and get a lot of good exercise (mix of cardio and strength training), dial back the alcohol and soft drinks, drink a goodly amount of water, eat everything in moderation but be sure to get good protein, green veggies, and fruit, especially as you age. But know that your diet is meaningless without the exercise, good mental health, and purpose in life — whatever it may be for you.

Pretty much what my grandmother, born in 1901, used to say. Except I also drink a glass of athletic greens every morning. Can’t hurt. And stretch and do planks/core work. Both are super important to maintaining balance and agility. More ill health and deaths start with falls than anything else.

K. K. Law wonders:

No argument about the benefit of exercising. But a simple and cursory inspection of the regional maps of (a) and (b) show the people in the regions highlighted by red ellipses appear to have lowest death rates. Do they have something in common in their diets that lead to longer lives?

Pamela Van Giessen responds:

Shouldn’t the question be to first isolate commonalities in everything among the people in those regions as opposed to assuming it is solely a food such as fatty fish? Is it just omega 3 or do peoples in those areas also have lower obesity rates, for instance? If they have lower obesity rates (and where there are lower obesity rates, there are routinely lower premature death rates), how come? What are they doing? Is it all diet or are there other variables?

That said, I try to eat fish at least twice a week. Fortunately I have a neighbor who likes to fish but he doesn’t like to eat fish. So we have a steady stream of fresh Montana trout. And elk. Elk meat is fantastic.

Kim Zussman adds:

Genes are a big factor in longevity, likely the biggest factor (besides distance from windows in Moscow). Could explain regional performance since primates primarily mate locally. The best tactic is to choose your parents carefully.

H. Humbert writes:

The media story on how the 100 yr old lived that long because he had one shot of whiskey per day or ate French fries three times a week always crack me up. I’m not saying nutrition (and exercise) do not matter, but of course their longevity is most likely because they won the gene pool lotto and not because of whatever quirky dietary habit they had.

“Virtually all nutrition studies are pretty meaningless”. This comment always cracks me up. It is untrue. Of course epidemiological and observational studies (observation) have value, even if they are not double blind placebo. For example, if you observe four people eat strychnine and die, would you not conclude that it might be dangerous? Would you stay in line to be the fifth person, even though you have merely done an observational study, and strictly speaking causation is unproven by a scientific study? If your answer is “no” then you must believe that epidemiological and observational studies have some value. Otherwise, you would be “blinded by science” (and dead).

Humbert H. responds:

Of course simple studies, like is strychnine dangerous, are useful. However, studies of subtle effects are generally useless, because of the various biases involved. It is to this day not possible to know if Ivermectin helps fight Covid, or if so, to what degree. Partly is because people are invested in the outcome and the set up of the studies appears suspect, and partly is because the effect is seemingly not overwhelming. Hearing about various "Coffee is good/bad for your health" through the years is a more common example.

Big Al adds:

Another issue with broader studies is that we are learning more about how different individuals with different genetics respond differently to coffee or salt or red wine or a high-fat diet. It becomes more difficult to make conclusions like "coffee is good/bad for you".

Humbert H. replies:

I agree completely. Coffee, if I drink it for a week and than stop, gives me terrible, incapacitating headaches, and if I keep drinking it, eventually I will get the same headaches. I don't know anyone else who has the same side effects, but I can only drink it once in a while. So all the recent studies I've read about the positive effects of its consistent use are of no use to me.

H. Humbert agrees:

Yes, this is absolutely true. And the genes may respond differently to foods over time, as other lifestyle factors change. Epigenetics.

Big Al offers:

An interesting show to watch:

Live to 100: Secrets of the Blue Zones

Though thinking about the stats, you would assume there would be pockets of longevity around the world just by chance. Also stat-wise, he claims there is a correlation in Corsica between the longevity of people in towns with the steepness of the streets in the town (steeper = longer lived). Haven't seen the data, but that's an interesting one on an intuitive basis. Maybe you could compare NYC residents on the first floor vs those on the fourth floor of a walkup building. ;-)

Peter Saint-Andre is skeptical:

That Blue Zone hypothesis is somewhat questionable. Here's one critique.

My impression from previous reading is that in some of these remote and frankly somewhat backward areas (e.g., Sardinia, Ikaria), the original cohort of centenarians contained a large number of people who faked their ages (e.g., to obtain government benefits), which they could do because they were born before birth certificates were common. The centenarian numbers didn't hold up in cohorts born after documentation of birth dates kicked in.

Pamela Van Giessen maintains:

The comment is true. Nutrition studies are meaningless. It’s a backward science in crisis with a host of issues starting with what gets published (and then reported) to garbage analytical studies on the same data sets, most of which have null results (but don’t get published) done from a laptop in about an hour.

Until people spend some time learning how “science” gets funded and what gets published, and demanding change, our knowledge will remain more antiquated than my grandmother’s guidance which was at least practical and based on real world experience.

John McPhee wrote about the funding problem in geology in Annals of the Former World. His observations apply to most fields. In short, what gets funded is what is trendy until it is not and then the new trend gets funded. This process takes about 100 yrs. In nutrition it may be worse. Vinay Prasad does a nice recap of the problems.

Nov

20

Alston Mabry notes:

Some TLDR:
- The US will lead the post-covid recovery, not EM.
- Goods-producing jobs are back which will have a multiplier effect.
- Capex will lead in the US, and total capex is 4x stock buybacks.
- China unlikely ever to exceed US economy. China much more like
Japan since the 90s.
- She likes innovation and recommends ARKK.

Wealthtrack: U.S. MANUFACTURING RESURGENCE
November 12, 2021

Be prepared to question many of the negative assumptions you have been hearing and listen to some other data that shines a different light on the outlook. Our guest is a highly respected economist who is no pollyanna. She is just a top economist who looks at data many others miss.

Nancy Lazar is Partner and Chief Economist of Cornerstone Macro. Lazar and her team are challenging the assumptions that higher inflation is here to stay, that interest rates have to go higher and that emerging markets will be the driver of global growth post-pandemic.

K. K. Law comments:

China could be much worse than Japan in the '90s.

Sarah comments:

Is she assuming all manufacturers/categories operate the same way? As much as I would love for this to be the case, there appears to be an oversimplified view of manufacturing that stems from the en vogue ecommerce B2C who typically have less operational personnel, strategic planning, etc. Sales and marketing teams are out in full force to convince manufacturers to buy their products, but many B2B who are currently better positioned and quieter could be the slowest to change.

Archives

Resources & Links

Search