Weekend reading: One more time

Good links (as ever). The time taken to compile them is much appreciated.

Pretty sobering those NPR and Andrew Sullivan ones aren’t they? 🙁

The NASA business is basically a IRL version of the plot to Don’t Look Up.

Question now is are we 1/8th way through this madness with the Mad King Mrk 2.0, or will this all just go on and on, and get even worse in different guises, a la Putin’s playbook?

I don’t take much comfort from the fact here that the condition of humanity throughout most of history has been one of feudalism, servitude and ignorance. Let’s hope that the future isn’t ‘nasty, brutish and short’, as Thomas Hobbes characterised his contemporary world.

Good idea, but perhaps not so practical when typing on foot through open countryside en route to a village fête 😉 At least the typos mean you know it’s not ChatGPT speaking. I do have a Chromebook when at rest, but no classic laptop or desktop, and not for browsing, just for online banking, fund platforms and shopping, where security paramount.

#15,16,17 – 174,000 Terawatts of solar radiation are intersected by the Earth, of which, after 30% is reflected back by the atmosphere, 122,000 Terawatts reach the Earth’s surface.

Humans generate just 18 Terawatts overall (of which 9 Terawatts is actually productively used, allowing for losses in production and transmission), of which just over 3 Terawatts is for electricity.

All life on Earth uses somewhere between 100 and 500 Terawatts.

Given the Stefan-Boltzmann Law (total energy radiated by a black body is directly proportional to the fourth power of its absolute temperature), and bearing in mind Earth’s surface temperature is already about 300 Kelvin, then increasing human energy production from 18 Terawatts now to 1,000 Terawatts would result in Earth’s surface temperature rising by just ~1 degree Kelvin (or Centigrade) due to radiative forcing.

This compares rather favourably with the IPCC forecast of a 2.7 degrees Centigrade rise in global average temperature by 2100 due to the greenhouse effect.

If we crack deuterium to deuterium (DT-DT) fusion we can extract DT from seawater (~150 parts per million of hydrogen in seawater by atoms and one part in ~3,300 of hydrogen in seawater by atomic mass).

Limiting extraction to 50% of the available ~46 trillion tonnes of DT in seawater (to keep extraction costs to no more than double their starting cost); a 1,000 Terawatt civilisation (e.g. 100 kilowatts each for 10 billion people, compared to 25 kilowatts for each current Californian) could be supplied with energy from fusion alone for 250 million years.

As it happens, 1,000 Terawatts (a Petawatt) is the tolerable thermodynamic limit (of 1 degree Kelvin global warming from radiative forcing) which epoch.ai uses as the upper bound of practical energy availability (compatible with human activity at readily adaptable temperatures) for their projections for a near future ‘agentic ASI singularity’ data centre driven digital economy – see their Gate model here:

https://epoch.ai/gate

@TI #25: “consider the huge percentage of market valuations given to energy companies even 20 years ago compared to their relatively paltry weighting now”: A Reddit post from Feb 21 last year states Nvidia’s trailing twelve-month trailing earnings at the time were $19 billion, compared to the energy sector’s TTM net income of $145 billion—making Nvidia’s profits approximately 13.1% of the energy sector’s. This was framed in the context of their market capitalisations being roughly comparable around that period (with supporting articles from early Feb 24 noting Nvidia’s market cap was only about $100 billion higher than the entire S&P 500 energy sector).

https://www.reddit.com/r/NVDA_Stock/s/sl3NvRkMCi

Something’s up. You can get 19% yields on Columbian EcoPetrol and 13% on Brazilian Petrobas on 5 to 6 x GAAP 12 months trailing P/E ratios, or you can pay up 600-700x trailing P/E for AI bespoke software darling Palantir (which I’ve owned since last year and is up several hundred percent).

On any conventional view it’s insanity squared. But so was BTC, and look what happened?

I think the trick now is to avoid the middle, the average and the normal – It’s extremely cheap value plays, either by sector (like energy, especially legacy hydrocarbon EM producers), or by size cohort (SCV and even smaller deep value plays) – where you can get paid handsomely with divis just for turning up (so no price catalyst needed); or it’s the full bonkers territory, mad high valuations, picks and shovels AI plays in chips, data centres and software.

It’s a bit like casting your mind back to school.

Of 2,000 pupils take the top 1% and the bottom 10% of performers.

In future life, of the top 20, a disproportionate number will have managed great things. All as expected.

But, of the bottom 200, there’s also some later life super achievers in there too, and unexpected so because, paradoxically, being s**t academically meant that a few of them managed to get seriously entrepreneurial, weaponise that chip on their shoulder, and show the grit to beat the odds.

For the middle 1780 (89%) of students though, not so much.

They went on to lead – to one degree or another – regular boring schmuck lives from skilled manual to PMC.

Same with companies now.

Bessembinder’s winners aren’t going to be found in the modestly achieving, modesty priced but ultimately disappointing fat middle cohort.

They’ll be in the disruptors and the transformers or the unloved and left for dead but still highly profitable deep value legacy plays (e.g. Philip Morris is the best performing company for investors since 1925, up 2.75 mn x in a century in nominal terms with dividends reinvested, that’s 16.3% p.a. for 100 years. Yet the product is hated, is a health hazard to the extent that it kills its own consumers, has faced existential litigation and regulatory restriction, and is wholly uninnovative and undifferentiated).

The risk from picks and shovels AI plays is being too early to the party. As the Chinese proverb goes, the first bird to fly is the one that gets shot.

Massive over build out followed by disappointing early LLM offerings could follow the pattern in TMT with fibre optic over deployment where the build out (with cable data transmission capacity doubling every 100 days from 1994 to 2000 for a million fold capacity improvement) was then inevitably followed by the mother of all crashes between 2000 and 2002.

Mbps prices are now down from hundreds of dollars to a fraction of a dollar (i.e. 99.9% down), which on the back of massively slowed growth in data transmission, has seen telecom revenue halve since 2000 ($235 bn to $118 bn), and even more reduced in real terms (see also here my MV post earlier today on the weekend reading first they came for the call centres thread).

But it was off of this cheap, seemingly limitless, supply of fibre that YouTube was able to launch at initial data transmission costs of just $500k per month (and not an otherwise bankrupting $500 mn per month).

Limitless and inexpensive Machine Learning may similarly create new sectors for the economy which are currently not commercially viable.

But the beneficiaries – and investment opportunity – may not yet exist.

@ermine #27: “not a native”: yeah. Gen X rules. Nothing against Millennials, Z and Alphas mind, but so much of what’s wrong now is down to Boomers. 1988 may have been 1967 Mrk 2, but that earlier generation have really lost the plot. From Bob Dylan to Nigel Farage. Explain that one to me!

On the Big Bubble, it’s not going to be much different in the broad shape, although the precise details obviously vary from one speculative mania to the next.

But the bigger mistake is to see only the Bubble and not also the profound change.

Tech is always hugely overestimated in its short term impact and always massively underestimated in the long term effects.

It wasn’t until 2012/13 or so that the natural monopoly and cash machine characteristics of current internet Big Tech really became apparent, even though some of these businesses had been around from the mid to late 1990s.

And, yes, for every Amazon there were many Pets.coms, but even the idea behind Pets.com eventually had its day (i.e. Chewy).

Although, on a personality level, Taleb seems to me a colossal pretentious p***, one must never ignore the message just because of the messenger.

The world really does look to me a lot more like his Extremistan than it does his Mediocristan.

And Lars’ VWRL approach works here not because it’s a Mediocristanic embrace, but rather, and contrastingly, because it’s a bid at ensuring entry to Extremistan.

If you own everything by definition you will own Bessembinder’s winners somewhere in there, amongst all the mediocre c**p.

But the Barbell is a purer and, IMHO, more rewarding (per unit of risk and on CAGR) approach to winner capture. Eliminate the space where winners are least likely to be found. Stock up on risk off (precious metals, HQ gov bonds etc) to offset the increased volatility and return dispersion on the remaining (more concentrated) equity sleeve. And then pray for the best and hang in for the long haul.

I always circle back in my mind to those findings from Bessembinder’s two studies: (of 26,000 US & 62,000 Global companies continuously listed from 1926 and 1990): all excess returns (over T-Bills) from 2.5-4% of companies, 75% of excess returns from 0.5% of companies, 50% of excess returns from 0.25-0.3% and 25% from just 0.1% (equating to just 50-100 companies worldwide today, out of 60,000-100,000 listed).

In the next 30 or 40 years there are going to be some 1,000+ baggers coming out of ‘AI’, whatever that may mean IRL.

All agreed. The big Q now is “what time is it”? Is it 1994/5, with years (and hundreds of percent) of melt up to come, or 1998/9, almost at the precipice? Fortunes will be made or lost on getting that call right or wrong.

@ermine #32 & @Indexing, you say? #33: both your points agreed. There are different ways of thinking about the same data though.
– I’m overwhelmingly in cap weight global index trackers (although I’ve very recently switched en masse to the slightly more spicy WGEC ETFz which gives some synthetic global government bond exposure overlaying the global cap weight equities). I’ve got no problem with indexation per se. It is a very low cost and very practical solution. And a very good one at that. But it’s not the one and the only solution. And it should never become dogma.
– On absolute numbers v cap weight, the latter definitely would be better for an apples to apples comparison with index trackers. All agreed there. But that’s just the data Bessembinder had access to, and uses, and, tbf to him, the point which he’s making is not an anti indexing one, but rather to avoid typical performance individual shares because the typical (median) share performance actually loses you money, even though the equity market as a whole goes up in the long term, and by a lot – e.g.: US large cap stocks: ~+9.94% annual returns (1928–2024) and US Small caps: ~+11.74% p.a. (versus Gold: ~5.12% p.a., US Bonds: ~4.5% p.a, US Real estate: ~4.23% p.a.)
– The obvious shortcoming of being always 100% in passive cap weight index trackers, and nawt else ever, is that all the returns above and beyond the meagre ones from just holding short term government bonds (US T Bills to be exact) given by holding the many thousands of shares passively tracked by the ETF are actually coming from just a teeny tiny subset of the holdings, just 2.4% since 1990. The majority of the tracked shares subtract from the performance of just holding cash, actually in effect destroying wealth. That’s of course the necessary price of ensuring that, by covering every single share in the index, you capture the returns of the 2.4% who generate all the returns. But it’s very diluting. You hold the many duds to ensure that you get a tiny stake in each of the many winners.
– In contrast, if you go 90% passive cap weight index tracker and even just 10% into trying to select just one of the outsized winners what’s then the worst outcome? Mathematically, even in the unlikely eventuality that every single one of your active picks is a disaster and goes to zero, you still end up with 90% of what you would have had had you been all in (always) on indexing. 90% and 100% are, FAPP, virtually identical in terms of lived outcomes (will having £900k really make a difference to you compared to £1 mn?) But if say you start with £1 mn, put £900k into index trackers and then give it your best shot putting £5k each into 20 chances at selecting just one of Bessembinder’s super winners (the top 5) the results can be completely transformative. Nvidia’s up close to 5,000x since IPO in January 1999 and Indian tech behemoth Infosys’ total 30 year return with dividends reinvested is up over 10,000x. Bearing in mind the limited downside of a 10% allocation to a Pareto punt on picking the winners, and the hugely asymmetric payout if you manage to just get one of them amongst your (say) 20 best shots; it seems to me to be an alternative worth considering if your time horizon is long enough.

@ermine #38: US healthcare and biotech in epic drawdowns. The worst for former since 1993. More broadly, pharma ain’t looking great.

I’ve done a pitifully sized dabble into UnitedHealth and Novo Nordisk in the ‘just for fun’ T212 account.

Maybe it gets worse, but the amounts I’m punting on these 2 is so small that IJDC 😉

The biggest regret will be if it actually pans out – like Palantir did, at least until about 48 hours ago 😉 – and then you wish you’d put in 10x more.

But when the starting position is well below 1% of the portfolio (as with UH and Novo for me) then there’s next to no loss to feel any aversion from if it all screws up.

For LLMs, it looks like mundane utility. Even mundane utility could make money eventually, but I’ll play safe and wait for a crash before committing to picks and shovels at scale. There’s at most $35bn AI ‘badged’ revenues this year (including creative accounting with the $10bn OpenAI Azure MSFT cloud credit as revenue), for $550bn projected 2025 Big Tech Capex, now equalling that of the Oil and Gas sector.

Bonkers. Best to wait out for a better entry and to stay in broad index trackers until then.

The risk of going in too early is not so much getting burned because it turns out it’s 1999 – as you can DCA in to mitigate that one in £impact terms – but the losses being so extreme %wise that they then put you off backing up the truck on a 80% fall, like by 2002 with the Nasdaq.

As they say, a 90% drawdown is just a crash of 80% followed by another 50% down just after you took the plunge.

@ermine: “far fetched than AI”: far fetched but, at the same time, real in a peculiarly non-meatspace way, just like Case jacked into his cyberspace ‘deck’.

Just read and much appreciated your superb “The great AI bubble – I’m a believer” on SLIS. Excellent stuff.

Notwithstanding that @Curlew on the next thread considers discussions of arguably the greatest concern of this era (maybe a misplaced one, but no less prevalent if it is such) to be “life-draining bilge”; and in the hope that not everyone shares his sentiment to put their head in the sand, their fingers in their ears and their hands over their eyes concerning the impact of the reality of the largest Capex boom since the railways in the 1840s (leaving aside here the direct impacts of and wider significance to the underlying technology itself); I’ve dropped in a few thoughts on the ‘is this actually 1998/99?’ question into the ‘First they came for the Call Centres’ thread.

If it is 1998/99 already, then that MIT report this week saying that 95% of companies surveyed report basically 0% RoI from LLMs might be something akin to a latter-day version of the 20 March 2000 Barron’s cover article by Jack Willoughby (“Burning Up, Internet companies are running out of cash – fast”) which (either coincidentally or not) turned out to mark the absolute peak of the Nasdaq right through until May 2015.

The Telegraph seems to think as much:

https://www.telegraph.co.uk/business/2025/08/20/ai-report-triggering-panic-and-fear-on-wall-street/

[Accessible via archive.ph]