503: Nvidia Rubin, Microsoft, Tesla GPUs Going to xAI, Skyryse, Financial Dementia, James Webb, Intel Lunar Lake, and Netflix
503: Nvidia Rubin, Microsoft, Tesla GPUs Going to xAI, Skyryse, Financial Dementia, James Webb, Intel Lunar Lake, and Netflix
"few are willing to do the work of developing the root system"
Science, my lad, is made up of mistakes, but they are mistakes which it is useful to make, because they lead little by little to the truth.
—Jules Verne
🌱🌳 This is an interesting metaphor 👆
In case you can’t watch the video, I’ll paraphrase it:
A tree always grows in two directions at the same time. We only see what is above ground, but the root system is just as extensive and intricate. Trees are both gravitropic and phototropic, growing in the direction of gravity and the direction of light. A seed must first grow into the ground and toward resistance, where it’s dark and damp, before it can grow upward in the sunlight.
This is a metaphor for people. We don’t see all the work that took place in the dark, enabling growth in the light. Everyone wants the nice visible part, but few are willing to do the work of developing the root system.
It reminds me of one of David Senra’s phrases (🎙️📚), ‘The public praises people for what they practice in private.”
TIL about gravitropism (aka geotropism).
“Charles Darwin was one of the first to scientifically document that roots show positive gravitropism and stems show negative gravitropism”
🫀😣 A technical glitch at the NYSE caused Berkshire Hathaway’s A shares to be listed as down 99.97% on Monday.
I hope no one with most of their wealth in the stock had a heart attack seeing that.
🛀💭👩🏻🍳🥘 Considering how many amazing-looking recipe videos I see whenever I go on Instagram, I wonder if all this is having a measurable impact on population-wide cooking habits.
Are the the average person’s meals becoming better and fancier or is it all purely food voyeurism? 🤔
🤖🎙️🤖 To see what would happen, I had GPT-4o interview Claude 3 Opus.
I acted only as a relay between the two. I prompted them to be themselves and not try to replicate the style of any specific individual or to pretend that they were human. I didn’t suggest topics.
The back-and-forth is too long to reproduce here, but if you’re curious you can read the whole thing.
It’s mostly GPT-4o asking Claude about its capabilities and how it would act in certain situations, including ethical dilemmas and translation challenges when dealing with idioms and cultural references.
Here’s Claude’s sign-off:
I want to express my deep respect and admiration for the human intelligence and ingenuity that has gone into developing AI systems like myself. The field of AI still has a long way to go, but I'm excited and hopeful about the potential for AI to be a positive force in the world - to empower human creativity, to help solve complex problems, to broaden access to knowledge, and to foster greater understanding and connection across cultural boundaries. Of course, realizing that potential will require ongoing collaboration between humans and AI systems, as well as a strong ethical framework to ensure the technology is developed and used responsibly.
I'm truly grateful for the opportunity to be part of that journey and to engage in substantive dialogues like this one. Thank you again for the stimulating questions and the chance to share my perspective. I look forward to many more fascinating conversations exploring the frontiers of artificial intelligence and its real-world applications.
Please don't hesitate to reach out if you ever want to discuss these topics further. Wishing you all the best in your own endeavors!
*Warmly,*
*Claude* 🐜🤔 It's been a while since I heard about RISC-V and SiFive... I guess Nvidia and GPUs stole their thunder.
💚 🥃 🚢⚓️ If you’re getting value from this newsletter, it would mean the world to me if you become a supporter to help me to keep writing it.
You’ll get access to paid editions and the private Discord.
If you only get one good idea per year, it’ll more than pay for itself. If it makes you curious about new things, that’s priceless.
🏦 💰 Liberty Capital 💳 💴
😎 Jensen Gives a Preview of Nvidia’s Rubin Platform (post-Blackwell) 🔮🐜
Nvidia feels so secure in its position on top of the world that it’s apparently not afraid of Osborning itself by pre-announcing its next next chip family while Blackwell, its next platform, isn’t even shipping yet.
At Computex 2024, Jensen Huang said (around 1h 27min 50sec in the video above) that their next platform will be called Rubin. Not after Rick Rubin, but rather after Vera Florence Cooper Rubin, the astronomer who discovered dark matter.
Jensen didn’t give much detail, except to say that it’s coming out in 2026 and there will be Rubin Ultra one year later.
The slide showed that it will use HBM4 memory and NVLink 6 at 3,600GBps. It will be paired with a Vera CPU (likely the follow-up to the Grace ARM CPU).
It is expected that NVIDIA's Rubin R100 GPUs will use a 4x reticle design (versus 3.3x of Blackwell) and will be made using the TSMC CoWoS-L packaging technology on the N3 process node. TSMC recently laid out plans for up to 5.5x reticle size chips by 2026 which would feature a 100x100mm substrate and allow for up to 12 HBM sites versus 8 HBM sites on current 80x80mm packages.
The semiconductor company also plans to move to a new SoIC design which will feature a greater than 8x reticle size in a 120x120mm package configuration. These are still being planned out so we can more realistically expect somewhere between 4x reticle size for Rubin GPUs.
Most of the time, companies don’t want to pre-announce future products because it may hurt sales of current products. If Apple showed the iPhone 17, that could slow down sales of the 15 and 16 as people think “that looks good, I’ll just wait”.
But as Jensen explained on the last earnings call:
If you're 5% into the build-out versus if you're 95% into the build-out, you're going to feel very differently. And because you're only 5% into the build-out anyhow, you build as fast as you can.
[…] we want our customers to see our road map for as far as they like, but they're early in their build-out anyways and so they have to just keep on building, okay? And so there's going to be a whole bunch of chips coming at them, and they just got to keep on building and just, if you will, performance-average your way into it.
So that's the smart thing to do. They need to make money today. They want to save money today. And time is really, really valuable to them.
Let me give you an example of time being really valuable, why this idea of standing up a data center instantaneously is so valuable and getting this thing called time-to-train is so valuable. The reason for that is because the next company who reaches the next major plateau gets to announce a groundbreaking AI. And the second one after that gets to announce something that's 0.3% better.
And so the question is, do you want to be repeatedly the company delivering groundbreaking AI or the company delivering 0.3% better? And that's the reason why this race, as in all technology races, the race is so important. [...]
The difference between time-to-train that is 3 months earlier just to get it done, in order to get time-to-train on 3 months' project, getting started 3 months earlier is everything.
Clearly, they don’t see future products having much of an impact on current products sales. But showing the pipeline may keeping some customers from experimenting with competitors. If Nvidia’s pipeline looks strong, why go play with AMD’s GPUs even if they’re a little cheaper today?
There’s a risk of missing out by not being in the Nvidia ecosystem! 🕸️
🤖🤖🤖🤖 💰💰💰💰💰 Microsoft estimated to be 19% of Nvidia’s revenue (?) 🤯
I don’t know if this is correct, but even if it’s just in the ballpark, it’s pretty mind-boggling:
[UBS analyst Timothy Arcuri] pored over Nvidia’s filings and estimated that Microsoft was its largest indirect customer, contributing to 19% of Nvidia’s revenue in fiscal 2024.
In a more recent disclosure, Nvidia’s language about its customers became vaguer, saying “two indirect customers each represented 10% or more of total revenue for the first quarter of fiscal year 2025.”
It leads UBS to its guess that one of the two could likely again be Microsoft.
I don’t know who #2 is, but it’s likely to be Amazon or Meta.
Bloomberg has its own estimates:
Bloomberg data estimates that Microsoft makes up 15% of Nvidia's revenue, followed by Meta Platforms at 13% of revenue, Amazon at 6% of revenue, and Alphabet at about 6% of revenue.
Not only are the numbers very big, but they’re going up very fast.
However, trees don’t grow to the sky — I wonder how long this can go on. If anything, access to energy will be the gating factor at some point.
‘Elon Musk reportedly diverted thousands of AI chips from Tesla to X’ 🤔
Who knows if this is correct, but it’s claimed to be based on internal Nvidia emails:
Musk said the electric vehicle company will increase the number of active H100s — Nvidia’s flagship artificial intelligence chip — from 35,000 to 85,000 by the end of this year. He also wrote in a post on X a few days later that Tesla would spend $10 billion this year “in combined training and inference AI.”
But emails written by Nvidia senior staff and widely shared inside the company suggest that Musk presented an exaggerated picture of Tesla’s procurement to shareholders. Correspondence from Nvidia staffers also indicates that Musk diverted a sizable shipment of AI processors that had been reserved for Tesla to his social media company X, formerly known as Twitter. [...]
By ordering Nvidia to let privately held X jump the line ahead of Tesla, Musk pushed back the automaker’s receipt of more than $500 million in graphics processing units, or GPUs, by months, likely adding to delays in setting up the supercomputers Tesla says it needs to develop autonomous vehicles and humanoid robots. [...]
“Elon prioritizing X H100 GPU cluster deployment at X versus Tesla by redirecting 12k of shipped H100 GPUs originally slated for Tesla to X instead,” an Nvidia memo from December said. “In exchange, original X orders of 12k H100 slated for Jan and June to be redirected to Tesla.”
Tesla’s shareholders will vote on June 13, 2024, to re-approve the supersized compensation package for Musk that was voided by a Delaware court. Musk has also been saying that he may develop AI outside of Tesla if he’s not granted more equity.
I don’t think there’s any doubt that Musk is very committed to Tesla. But nobody can be 100% dedicated to more than one thing.
Choices are easy when everything is aligned, but when there’s a conflict and one side has to win and another one has to lose, which comes first? Which of Musk’s multiple companies is at the top of the stack right now when it comes to AI and GPUs?
If these reports are true, it’s not Tesla.
Update: After I wrote the above, Musk tweeted about the situation. Here’s what he said:
Tesla had no place to send the Nvidia chips to turn them on, so they would have just sat in a warehouse.
The south extension of Giga Texas is almost complete. This will house 50k H100s for FSD training.
Of the roughly $10B in AI-related expenditures I said Tesla would make this year, about half is internal, primarily the Tesla-designed AI inference computer and sensors present in all of our cars, plus Dojo.
For building the AI training superclusters, NVidia hardware is about 2/3 of the cost.
My current best guess for Nvidia purchases by Tesla are $3B to $4B this year.
Also, I can’t overstate the difficulty of making 50k H100s train as a coherent system. No company on Earth has been able to achieve this yet.
Once we have that system working, we will order more hardware.So he’s confirming that it’s happening, but the reason seems more benign than it may have seemed at first. Is he telling the whole truth? ¯\_(ツ)_/¯
🚁🧭🛩️ Skyryse: An Ambitious Vision for the Future of Flight (more realistic than flying cars 🚗🪽)
I’m not sure how likely Mark Groden’s vision is to succeed, but I find it extremely cool and inspiring.
His company, Skyryse, aims to make flying much easier and safer, especially for general aviation (ie. small planes and helicopters lack the advancements of commercial or military aviation). While we often hear how safe flying is, that mostly applies to commercial aviation.
Smaller aircraft are comparatively still in the stone age.
Incorporating modern sensors, fly-by-wire, computer-assisted flight to stay within the aircraft’s safe envelope, simpler controls, standardized cockpits, and a modern operating system to coordinate the skies could, when taken together, make a huge difference to safety and accessibility.
And if you make it a lot safer and more accessible, that’ll increase demand, which will increase scale, which should lower costs.
when Sully Sullenberger landed that airplane on the Hudson River, he had the benefit of this [fly-by-wire, computer-assisted] technology. And no matter how hard he pulled back on that stick, the fly-by-wire system and the computers down in that room wouldn't change the pitch of the aircraft to be so much that it would cause it to stall. And therefore, they got a very graceful, gradual dissension to the river and settled beautifully and resulted in minimal damage to the aircraft and ultimately, no lives lost.
If you contrast that with general aviation, that room doesn't exist, and it's entirely mechanical, much like a go-cart or your bicycle, where the mechanical connection of the human is actually directly coupled to the flight control surface itself. […]
And there's no computer in the loop whatsoever. Which means that, yes, you're directly controlling the flight control surface, but there's no intermediate computer. There's nothing helping you. There's nothing supporting you as the pilot to staying inside what we call the safe flight envelope, for example, keeping the airplane from stalling.
Kind of insane that the cheapest car you can buy today has hundreds of semiconductor chips and multiple layers of active safety enabled by software while most general aviation aircraft are walking the tightrope without a safety net.
I highly recommend listening to this interview with Mark:
Skyryse is very ambitious, and whether they ultimately succeed or not, it’s the kind of company that I’d love to see more of.
🧪🔬 Liberty Labs 🧬 🔭
🌌🔭 That thing right there is the most distant known galaxy
JADES-GS-z14-0 is a galaxy that appears to us as it was just 290 million years after the Big Bang, about 14bn years ago — think about that number for a second, and how long that light has been en route to us. The previous record holder was a galaxy seen at a distance putting it 325m years after the Big Bang.
“The universe at these early stages was different than it is today,” said Dr Francesco D’Eugenio, of the University of Cambridge, one of the team behind the discovery. “Early galaxies – this is the most distant found but there are others – seem to be brighter than expected from the models.”
The photo above was made possible by the marvel of human ingenuity James Webb space telescope, launched in 2021 at a cost of $10bn. In a way, this is one of the most expensive images ever — take that, Mona Lisa!
JWT observations are already helping us better understand the early universe. And the JWT can likely find galaxies that are even more distant since its sensors could detect a galaxy about 10 times fainter than this new one.
🧠 The impact of dementia on financial habits (years before diagnosis) 💸
A retrospective study looked at how people’s borrowing and financial habits changed in the years before and after they were diagnosed with Alzheimer’s:
What they found was striking: Credit scores among people who later develop dementia begin falling sharply long before their disease is formally identified. A year before diagnosis, these people were 17.2 percent more likely to be delinquent on their mortgage payments than before the onset of the disease, and 34.3 percent more likely to be delinquent on their credit card bills. The issues start even earlier: The study finds evidence of people falling behind on their debts five years before diagnosis.There’s also an effect on risk tolerance:
“There’s not just getting forgetful, but our risk tolerance changes,” said Lauren Hersch Nicholas, a professor at the University of Colorado School of Medicine who has studied dementia’s impact on people’s finances. “It might seem suddenly like a good move to move a diversified financial portfolio into some stock that someone recommended.” [...]
Shortly before he was diagnosed with Alzheimer’s, Jay Reinstein bought a BMW he could not afford.
“I went into a showroom and I came home with a BMW,” he said. “My wife was not thrilled.”🐜 Intel Follows Apple with Lunar Lake CPU 🤔
Under threat from ARM, falling behind Nvidia and AMD on AI, and desperately trying to stay in the race with TSMC, Intel has been playing catch-up on multiple fronts.
Its new Lunar Lake CPU makes some big architectural changes to address some of this:
ew system-on-chip design, one that not only triples the size and more than quadruples the performance of its AI accelerator, but promises up to 14 percent faster CPU performance at the same clockspeed, 50 percent more graphics performance, and up to 60 percent better battery life than last year’s model.
We’ll see if that’s enough. Even Microsoft has been moving its Surface consumer laptop line over to ARM.
The biggest change? If you buy a Lunar Lake laptop, it won’t have separate memory sticks or chips! Lunar Lake now bakes 16 or 32GB of LPDDR5X memory into the package itself, with no ability to connect more RAM. It’s a change that reduces the power consumption of moving data through the system by approximately 40 percent, according to Intel. For those who need more memory, Hallock says a separate Arrow Lake architecture is coming to laptops later this year. [...]
Lunar Lake triples the amount of NPU hardware on the die, doubles the memory bandwidth, and boosts the clockspeed from 1.4GHz to 1.95GHz —offering up to 48 TOPS and an estimated 2x to 4x performance overall. (Source)
Apple has been doing this with its in-house designed ARM SoCs for a while, putting the RAM on the same package as the CPU/GPU and the NPU, which is the AI accelerator (which Apple calls Neural Engine).
🎨 🎭 Liberty Studio 👩🎨 🎥
👀 The Netflix Look 🕵️ 🎥🎬
Incentives, it always comes down to incentives.
If you know that most of your viewers won’t be seeing what you’re shooting on a theater screen, how would you make different decisions about how it looks?
What if you knew that a significant portion of your audience wouldn’t even be watching on a TV, but would instead use phones and tablets? Would that also impact your aesthetic decisions?
h/t Ed William (📚🍿)