526: Microsoft Goes Nuclear, Qualcomm + Intel?, Nick Sleep Letters, OpenAI & Anthropic, Scale AI, T Cells vs Cancer, and SPQR

"The fallacy seems obvious to me"

It is wise to avoid militants of all plumage, to trust only the fanatically unfanatic.

–Frederick Franck

Made with Midjourney 💚 🥃

🛠️ ⚖️ Tools do not have a fixed utility.

(Here I mean “tool” in the broadest sense of the term)

Different people using the exact same things — whether a hammer, a book, Google, or ChatGPT — can get *incredibly* different amounts of value from them.

Some people may get almost no value while others may use them to totally change their lives, and other people’s lives, for the better.

As friend-of-the-show Jesse Livermore put it:

Can't be overstated how powerful ChatGPT is for a student trying to develop a deep understanding of something in a technical field. The student now has 24/7 access to a genius w/ infinite knowledge, infinite communication ability, and infinite patience, to "make it make sense."

This is also true of other things throughout history like public libraries and Google, to varying degrees.

So, how can you improve your ability to use tools?

That’s the high-leverage question, isn’t it?

Think about what you use every day. How much effort have you put into getting better at using these tools? Could you upgrade to better versions of these tools?

That was my argument to my wife when I suggested that it could be a good idea to improve her typing speed on a keyboard — it’s a major part of her day, and the effort required to get better is a one-time cost that will pay dividends forever after. 🫳 ⌨️

🛀💭🥧🏡🏚️ When I hear statements like:

“Billionaires shouldn’t exist”

The fallacy seems obvious to me. Those who think this are focusing on the wrong thing.

Centi-billionaires like Vladimir Putin shouldn’t exist. Pure negative sum parasites making the world much worse. They’re rich by making others poor and miserable and reducing their options and freedoms.

In contrast, a centi-billionaire like Jensen Huang only captured a small portion of the value he created and made our civilization better off. I’d love more of those please.

Never unthinkingly accept how others frame things, make sure to focus on the right things. Otherwise, we risk creating policies and incentives that prevent the things we want, while not doing much to prevent the things we don’t want…

🏪 🎲 🚗 💰☭ Speaking of wealth creation and incentives, ‘Communist Monopoly’ is pretty hilarious.

Check out from the 3 to 6 mins mark to get an idea of the rules:

---

🏦 💰 Business & Investing 💳 💴

Three Mile Island Unit 1’s steam, before being decommissioned in 2019.

🇺🇸🤖🔌⚛️ Microsoft Makes Deal with Constellation Energy to Restart Three Mile Island Reactor! 💰💰💰

Three Mile Island is best known for the partial meltdown of its Unit 2 reactor in 1979 (which caused no deaths or injuries and *may* have exposed people very close to the plant to the equivalent of 1 chest x-ray, which is what you get from an average airplane trip at 35,000 feet…).

It is a less known fact that the other reactor kept operating until 2019, and was shut down for economic reasons (fracking made natural gas cheap at the time — which seems very short-sighted to me to remove an asset that can produce clean baseload power for decades because nat gas is temporarily cheap…).

Anyway, this reactor may be coming back from the dead because GPUs are power hungry:

The restart effort has been in the works since early 2023, when Constellation began evaluating whether it made sense to switch the reactor back on. By early this year, the company had concluded that it wanted to pursue the project, and began talking to potential customers. Microsoft was immediately interested [...]

Constellation expects to spend around $1.6 billion to restart the reactor by early 2028. Microsoft has signed a 20-year power-purchase agreement with Constellation, the companies said Friday. The deal would help Microsoft pair its 24-7 electricity use with a matching source of nearby clean power generation.

“The most important energy commodity in the world today is a reliable and clean electric megawatt just because of the difficulty of replicating it and the need for it,” said Joe Dominguez, Constellation’s chief executive officer.

This unit has a nameplate capacity of 819MW and was operating at a capacity factor of 95.65% in 2017, which made it one of the best in the world.

AI data-centers are *incredibly* power-dense (and becoming more so, with water-cooling becoming more common to further increase density…).

We’re talking about hundreds of megawatts — soon gigawatts — going into *a single building*.

If the goal is to avoid coal and gas, it makes a ton of sense to match the 24/7 predictable demand to a nearby 24/7 source of similarly sized power, rather than trying to build transmission lines that are hundreds of miles long — for which there is already a multi-year backlog — and trying to cobble together enough variable, low-density renewable power from a gigantic area. That means you have the extra expense and complexity of overbuilding supply and then building massive amounts of batteries to store the excess power to have supply at night and during cloudy weeks.

Back to the Microsoft deal, friend-of-the-show Mark Nelson put it well:

However big you think this Microsoft nuclear deal is, I am telling you, it's bigger. 

One of the biggest power deals ever. 

~$800 million/yr for 20 years
$16 BILLION for one nuclear reactor

From a 40-year-old reactor, just to get to age 60.

We now have a market for newbuild.

More context from Mark:

I suspect Microsoft is paying a big premium to get this plant back as fast as possible.

This deal is revealing such interesting things about the limits of getting concentrated power from new renewables.

TMI nuclear power is being purchased for a price that appears to be multiples of what new wind and solar projects can get, demonstrating the high value of nuclear electricity and the low value of incorrectly located and erratically produced power.

Electricity is not, and never was, a commodity. It is a service with strict and uncompromising location and time-based constraints. Nuclear almost perfectly matches the specific service needs of the next generation of centralized AI datacenters in a way distributed weather-based energy cannot do while keeping costs reasonable.

Getting electricity to market when and where needed is a savage problem. The amount of constant power needed at individual hyperscale datacenters is unprecedented. Combine these two things and nuclear power from existing plants becomes extremely valuable because it's both constant and carbon-free.

This is so important. Finance people tend to think of electricity like they think of money — inflows and outflows that have to net out, and that’s it. But actual reliable infrastructure is a lot more complex than that.

Breaking down the numbers:

Today's operating nuclear plants make electricity for a cost to owners of $20-40 per MWh

Today's electricity markets have been paying nuclear plants $30-60 per MWh

Hyperscalers are willing to pay over $100 per MWh, on seemingly unlimited volume, if they can get it.

New renewables are being built for $40-100 per MWh, but their power is in the wrong spots at the wrong time, and thus does not solve the problem of constant power for datacenters.

I’ll be curious to see if this further changes the conversation around nuclear and whether other Big Tech make similar moves. AWS has already posted job openings for nuclear engineers and Oracle said it is looking at building Small Modular Reactors (SMRs)…

🏦 ‘14 Major Global Banks Express Their Support for Effort to Triple Nuclear’

Speaking of nuclear:

Today, nations endorsing the Declaration to Triple Nuclear Energy launched at COP28 in 2023 were joined by 14 financial institutions who expressed support for the call to action to triple global nuclear energy capacity by 2050.

The group of financial institutions in the convening include: Abu Dhabi Commercial Bank, Ares Management, Bank of America, Barclays, BNP Paribas, Brookfield, Citi, Credit Agricole CIB, Goldman Sachs, Guggenheim Securities LLC, Morgan Stanley, Rothschild & Co., Segra Capital Management, and Societe Generale.

These COP declarations rarely lead to much action… However, these banks clearly see a real opportunity to make good infrastructure loans backed by assets that can pay for themselves many times over during their useful lives.

I don’t think this is just about virtue signaling because nuclear power isn’t *that* popular. The virtue signaling was when they *didn’t* openly support it because they were afraid of backlash by NGOs...

🐜🐊 Qualcomm Approached Intel About a Takeover 👀

As I write, Intel has a market cap of $98bn and an enterprise value of $125.75bn. It’s a pretty big bite to swallow, especially for Qualcomm with a market cap of $186bn (they are *not* Big Tech):

A deal is far from certain, the people cautioned. Even if Intel is receptive, a deal of that size is all but certain to attract antitrust scrutiny, though it is also possible it could be seen as an opportunity to strengthen the U.S.’s competitive edge in chips. To get the deal done, Qualcomm could intend to sell assets or parts of Intel to other buyers. [...]

Qualcomm is a leading supplier of chips for smartphones, including ones that manage communications between phones and cell towers. It is one of the most critical suppliers for Apple’s iPhones, among a range of other devices.

A deal would significantly broaden Qualcomm’s horizons, complementing its mobile-phone chip business with chips from Intel that are ubiquitous in personal computers and servers.

Maybe there’s a way this could be a first step toward what I’ve been proposing lately — what if Qualcomm really only wants the chip design business and was interested in spinning off the fabs..?

Maybe they could be spun off to some kind of consortium backed by the Big Tech companies who would benefit most from an alternative to TSMC and the US Government which would most benefit from a national security standpoint?

I kind of wish it were Broadcom making this move instead of Qualcomm — they have a better reputation as operators — but I suppose that if Qualcomm sent a lot of its volume through Intel Foundry, it could help bootstrap it (though I don’t think it wouldn’t be enough). This may have the downside of making Intel Fabs dependent on internal customers rather than on third party ones, and preventing them from gaining the expertise they need to truly compete with TSMC.

📈 Nick Sleep and Qais Zakaria — Learning from the Best 🕵

My friend David Senra (🎙️📚) just released two excellent episodes:

• 🎧 #364 Nick & Zak's Excellent Adventure: How Nick Sleep and Qais Zaharia Built Their Investment Partnership

• 🎧 #365 Nick Sleep's Letters

I read the Nomad letters a few years ago. This made me want to re-read them! It’s a great overview of the principles and insights used by Nick and Qais to not only achieve great investment returns but also build what seems to be a great life for themselves.

It’s not just what you do but also how you do it.

I’d much rather achieve my goals will doing work that I love in a way that suits my wiring and rhythm than have to sprint a marathon and grind on uninteresting BS for years.

The destination isn’t worth *that* journey, IMO.

Of course, the specific insights in these letters may not work as well for investors going forward — partly because they popularized them — but their approach seems pretty timeless.

🤖🤖 OpenAI Raising Money at $150bn Valuation, Anthropic Wants Around $30-40bn 💰💰💰

Anthropic has started talking to investors about raising capital in a deal that could value the startup at $30 billion to $40 billion, roughly doubling its valuation from a funding that closed early this year [...]

in the wake of OpenAI’s pending deal to raise between $5 billion to $7 billion at a valuation of around $150 billion, nearly double the ChatGPT maker’s valuation from a share sale that closed around the first quarter of this year.

Then they both turn around and spend that money in Azure and AWS compute credits and Nvidia GPUs, right?

It’s like the carbon cycle in a forest! 🔄

✍️🤖 Scale AI Quadruples Revenue in 2024’s First Half 🔥

Scale nearly quadrupled its sales to almost $400 million in the first half of the year compared with the same period last year. Scale is among the privately held firms that have benefited the most from the AI boom by providing infrastructure or services—in Scale’s case, human contract workers—to companies building advanced AI. [...]

The company had about $980 million in cash

Interestingly, one of the biggest beneficiaries of the AI boom is a company that hires human contract workers…

Well, I mean they benefit when it comes to growth and valuation — Scale AI last raised money at a $13.8 billion valuation in May — but they are still unprofitable, with an operating margin reported to be around -20%.

I’m curious how durable the business is 🤔

How long will all these humans be needed to write and annotate data for LLM training and tuning..? At which point do we transition to a world where all this becomes more of a closed loop, with synthetic data and AI-annotated data replacing these humans?

Of course this has downsides, but it can also scale — ironic given this company’s name — much better and so the trade-offs may be worth it.

---

🧪🔬 Science & Technology 🧬 🔭

🔋🧫 ‘Giving T cells extra batteries supercharges them against cancer’ 👨🏻‍⚕️👩‍🔬

In a fair fight, the immune system would wipe out cancer every time – but this crafty disease doesn’t play fair. Along with its many tricks to avoid detection, tumors build microenvironments around themselves that are toxic to immune cells, draining them of energy. Sustained immune responses against a pathogen like cancer can lead to T cell exhaustion.

Finding ways to supercharge the immune system to continue the fight is the main goal of immunotherapy. Now, a team from Brigham and Women’s Hospital has demonstrated a new way to refresh T cells, by essentially replacing their batteries.

Mitochondria are organelles that produce chemical energy for cells, but during T cell exhaustion these little batteries can be lost or damaged. Previous studies have shown that cancer cells can use nanotubes like “tiny tentacles” to slurp up mitochondria from immune cells. For the new study, the researchers found that they could use the same mechanism to do the opposite – donate new mitochondria to T cells, from bone marrow stromal cells (BMSCs).

Power ups for immunte cells! 🍄🆙

The team cultured BMSCs and T cells together, and after 48 hours found that up to a quarter of the T cells had gained extra mitochondria. The researchers dubbed these juiced up immune cells Mito+.

In tests in mice, Mito+ cells were found to more easily penetrate tumors and launch a more robust attack on them. Tumors shrank drastically, and 75% of treated mice survived the full 60-day study period. In contrast, control mice saw their tumors continue to grow, and all had died by the 20-day mark.

Intriguingly, Mito+ cells could multiply quickly and pass their extra mitochondria to the new cells

With the caveat that a lot of things that work in mice don’t end up working as well in humans, this is very promising and the mechanism of action is simple and straightfoward enough that I hope we can test it out rapidly.

🚗📡🇨🇳 US Wants to Ban Connected Vehicles with Chinese Hardware or Software

The U.S. Commerce Department proposed prohibiting key Chinese software and hardware in connected vehicles on American roads due to national security concerns, a move that would effectively bar Chinese cars and trucks from the U.S. market. [...]

The proposed prohibitions would prevent testing of self-driving cars on U.S. roads by Chinese automakers and extend to vehicle software and hardware produced by Russia and could be extended to other U.S. adversaries.

The proposal would make software prohibitions effective in the 2027 model year. The hardware ban would take effect in the 2030 model year or January 2029.

I wonder if China will retaliate against US carmakers.

Elon Musk must be nervous about this…

🦠🤧🤒 FDA Approves At-Home Nasal Flu Shot

There are few things I hate more than being sick, so this is good news:

The agency on Friday gave the green light for people who have been screened to give themselves the FluMist nasal spray, which can be ordered directly from an online pharmacy, skipping the need to visit a doctor’s office.

It will still require a prescription from a doctor's office, however. It's expected to be available next year.

What’s kind of crazy is that there’s no technical breakthrough here. It’s purely a *regulatory* breakthrough after taking forever:

FluMist itself is not new — the live attenuated influenza vaccine has had FDA approval for more than two decades. But the ability for adults to order the vaccine at home to administer to people ages 2 to 49 is a breakthrough in convenience and access to preventative care.

I can’t wait until this comes to Canada — it may take another 20 years — and I can just order from a menu of shots online, take a sniff at home, and be done with it.

---

🎨 🎭 The Arts & History 👩‍🎨 🎥

🏛️ SPQR: A History of Ancient Rome🏺

I’m just a few chapters in, but I can already recommend this book.

There’s a lot of great stuff, like this bit about how historians can try to estimate the economic conditions thousands of years ago:

Roman coins can often be precisely dated, because at this period they were newly designed each year and 'signed' by the annual officials who were responsible for issuing them. They were minted using a series of individually hand-cut 'dies' (or stamps), whose minor differences in detail are still visible on the finished coins. 

We can calculate roughly how many coins an individual die could stamp (before it became too blunt to make a crisp image), and if we have a large enough sample of coins we can estimate roughly how many dies had been used altogether in minting a single issue. From that we can get a rough and ready idea of how many coins were produced each year: the more dies, the more coins, and vice versa.