Weekend reading: First they came for the call centres

Telegraph today back onto crash fears:

https://www.telegraph.co.uk/money/investing/stocks-shares/how-to-protect-your-money-from-a-market-crash/

Rereading the excellent Monevator 2010 Peter Lynch piece, I was struck that one of his 13 principles was to invest in firms that use tech, not make it.

I think there may be a lesson in there for the hyperscalers & semiconductor firms.

Interesting angle today on improvement to memory (the straw) being more important than getting more ‘compute’ (the liquid in the glass).

http://uk.investing.com/analysis/ai-chip-war-just-shifted-why-memory-may-matter-more-than-compute-200619722

The scope for improvement in this area could be substantial and low hanging. Looks good for Qualcomm.

And a very deep dive into why it’s not a Ponzi:

https://open.substack.com/pub/artificialintelligencemadesimple/p/the-ai-is-a-bubble-narrative-is-stupid

Sure vendor circular financing may be no Ponzi but simply saying OpenAI has millions of paying customers doesn’t cut it.

Amazon survived the Dot.com disaster of 2000-02, and trades at north of $200 now; but, split adjusted, it still fell from a intraday peak of $5.60 in 2000 to 30 cents at one point in 2002, nearly 95% down, and it was right that it traded down so far then given how far valuation had gotten ahead of earnings.

The financing may well secure CUDA lock in for Nvidia, but so what?

Show me the money.

Where are the hundreds of billions of revenue to go with the product (and to frank the Opex and Capex)?

Of course it’s not a literal Ponzi, and, yes, Ponzi as a term does get bandied about too liberally.

But it does look distinctly like there’s going to be an epic revenue gap ahead, even if massive end user revenue does eventually arrive (as, in my example, it ultimately did for Amazon).

From “What GPU pricing can tell us about how the AI bubble will pop” (Bryce Elder):

“One odd thing about AI equipment is that it’s very expensive to buy and very cheap to rent. Want an Nvidia B200 GPU accelerator? Buying one on its release in late 2024 would’ve probably cost around $50,000, which is before all the costs associated with plugging it in and switching it on. Yet by early 2025, the same hardware could be rented for around $3.20 an hour. By last month, the B200’s floor price had fallen to $2.80 per hour. Nvidia upgrades its chip architecture every other year, so there’s an opportunity for the best-funded data centre operators to lock in customers with knockdown prices on anything that’s not cutting edge. From the outside, the steady decline in GPU rental rates resembles the kind of predatory price war the tech industry relies upon: burn money until all your competitors are dead. The evidence, however, is more complicated. … among the hyperscalers (Amazon’s AWS, Microsoft’s Azure, Google and Oracle) prices have hardly budged. The result is an ever-widening gap between rates charged by the big-four and a growing number of smaller rivals.”

I ask myself, does the low rental cost but high(er relative) residual values for slightly older GPUs situation, support, or become disjunctive with, the analysis in the Asymmetric Opportunities substack post linked to immediately above?

Also, I’d understood/ read that Nvidia had moved to annual GPU release cadence, not every other year as before.

Nvidia margin= everyone elses’ opportunities, or is it instead Jevons’ paradox redux?:

https://open.substack.com/pub/davefriedman/p/when-gpu-demand-peaks

Real time AI boom or bubble monitor dashboard now live, and registering 1 out of 5 red flags:

https://boomorbubble.ai/

Anatomy of a bubble:

https://open.substack.com/pub/netinterest/p/bubble-trouble

MIT said 95% of businesses seeing no RoI on Gen AI but now Wharton says that in fact 74% are!:

https://open.substack.com/pub/bigtechnology/p/wait-are-74-of-businesses-actually

Meanwhile, MSFT’s Sep quarter had hidden ‘bombshell’: “a $4.1 billion loss from its equity-method investment in OpenAI”:

https://open.substack.com/pub/appeconomyinsights/p/microsoft-openais-wild-ride

And Boomer Bill Bonner takes the longer view:

https://open.substack.com/pub/bonnerprivateresearch/p/the-astonishing-ai-boom

With yearly release cadence for GPUs / TPUs / NPUs and a ~2.5x FLOP improvement/ watt per generation, with effective FLOPs (‘eFLOPs’) per FLOP up ~3x p.a. from algorithmic improvement and at least (maybe much more than) 1.35x p.a. from system architect optimisations (‘unhobbling’), and with inference/ eFLOP pegged at a ~2.5x increase p.a.; were currently looking at up to 25x inference/ watt/ p.a. increases.

With power usage doing 1.6x-2x p.a., overall that’s a 40x-50x inference p.a. increase.

The 40 fold figure is likely the source of Sam Altman’s claim of a 97.5% annual cost reduction for inference.

If demand for intelligence (or at least inference) is infinite and insatiable, then Jevon’s paradox reigns supreme, and picks and shovels plays will boom.

If it doesn’t though, then IMO the scenario looks similar to the Canadian cannabis industry post legalisation, where an irrational bubble in listed grower stocks met a massive oversupply in production, crashing wholesale prices, and resulting in escalating and ruinous losses:

https://open.substack.com/pub/stocksandstones/p/my-most-comprehensive-cannabis-update

If the second situation (i.e. no Jevons’ paradox) applies, then sitting out is the only winning move.

Otherwise, if it’s Jevons’ paradox all the way down on inference costs, then there’s no alternative but to stay invested in the AI game.

Erratum: the Epoch AI GATE study by Erdil et al is 12 March 2025, not April 2025:

https://arxiv.org/abs/2503.04941

Korinek’s research is here:

https://www.korinek.com/research

As Michael Burry issues yet another bubble klaxon (checks notes, predicted 7 of the last 2 recessions) How Money Works plays Devil’s Advocate today:

https://youtu.be/tAXKxKTGWFQ?si=SPSzAtXngufbh8yC

The ever brilliant David Friedman now has this to say:

https://open.substack.com/pub/davefriedman/p/why-the-gpu-boom-wont-burst-yet

https://substack.com/@davefriedman/note/c-172518526

Calling this one right is hard! 🙁 🙁

The $64 trillion question (or quadrillion, if the AGI / ASI accelerationists are right)…..

I still can’t work out where the 2 or 3 year GPU cycle is coming from. Nvidia is an annual releases and 2.5x FLOP/watt cycle improvements. I feel like the Captain of the Nostromo Arthur Dallas in Alien in the botched landing on LV-426 exclaiming “will someone give me a straight answer!’

AI Capex now = Metaverse in 2022?

Is this “discretionary spending”?

And, if it ever stopped due to low / no RoI, then there’d be a big boost to the distributable FCF (and, therefore, to buybacks / divis) of the hyperscalers, which, in turn, *should* help to support their share prices.

A Capex “dilemma” but with (some) upside on the downside?:

https://open.substack.com/pub/bestanchorstocks/p/a-nuanced-view-on-the-aicapex-dilemma

@Chris Samp (on Substack comments) has this (very thoughtful) counter argument to Dave Friedman’s excellent point (made in David’s own Substack notes, that, “[t]he fiber analogy only half works, because fiber is durable infrastructure while GPUs are perishable infrastructure. GPUs are consumable instruments of a rapidly moving technological frontier. Their utility decays with every model update, quantization improvement, and ASIC iteration”):

“…..as a counter factual. Current GPUs obsolete in 18 months as the AI infra players push hard on returns to scale. (Ofc they have to to justify their valuations, all the way around.) And the hard metrics of gigaflops or whatever is used today will respond to Moore’s law. But will utility of LLM? ChatGPT 3.5 was a zero to one event. ChatGPT 4 was a 1 to 2 event. Now we are getting to 4 to 5 events, not 4 to 8 events. Utility looks arithmetic not logarithmic. Which means the lifecycle of a GPU is much longer. Look at actual gaming. For a while you needed to upgrade your gpu to run the latest AAA games and have them look their best. But now you can stay a few generations back, for years, and have a fine experience. Even the Xbox, which is competitive to PC gaming now unlike console gaming used to be, has a several-year lifecycle. Which means the current GPU install base will remain useful for much longer, rescuing the owners of that capital. Abandoning return to scale would destroy the valuation of just about every named player currently in the space, while preserving the parallel to dark fiber.”

God knows what I think on this anymore. The market seems to think Nvidia is the real deal. At least it’s team up with Palantir is giving the latter some momentum!

Sooooo….

It’s certainly not a bailout

It’s not a bail in

And it’s not even a backstop

https://open.substack.com/pub/sophiecapital/p/inevitability-weekly-3

Ahhh. Enlightenment. It’s all so much clearer now (ROFLMAO 😉 )

Altman’s posse must take Uncle Sam for a fool.

If you are TBTF then you pay a toll for a government guarantee.

No money now to pay up for the toll – then you have to cough up some non dilutable equity instead. And a few board seats to go with it.

Why is this different to the Rare Earth miners the US Government has just taken stakes in?

Beyond (standard) LLMs:

https://open.substack.com/pub/sebastianraschka/p/beyond-standard-llms

Then again, maybe the doomers are not just wrong but, from a ‘Western perspective’, dangerously even counter productive, helping, unwittingly, to embed PRC / CP of China values into current models and, therefore, into any future AGI:

https://open.substack.com/pub/campbellramble/p/the-doomers-paradox

A doomer paradox alongside Jevons’ paradox.

Ed Z today on the case of OpenAI’s revenues again (the man, the legend, he’s published over a hundred thousand words on the finances of the model providers and the hyperscalers in just the last three months – that’s commitment!).

Quelle Surprise, it could be even worse than we thought and feared:

https://www.wheresyoured.at/oai_docs/

Ed’s doing a sterling job of fighting the good fight for the haters but this is perhaps the most intractable and the most important problem I’ve ever seen.

The more I read around (and on) it, then the less I know of it, save for gaining knowledge of our collective ignorance about it.

On the one hand Copilot is real. GPT5 Pro/Ultimate and what not is real.

For the business level (£200-£250 pcm) paid products, there’s something there already.

This isn’t crypto and SBF’s ’empty boxes’. Or just asking it to plan a holiday. It’s not even just chatbots.

Indeed, just 3 years in from the launch of GPT3, and already, from the enterprise grade transformers, I’ve now seeing outputs (admittedly after several iterations, and a day’s manipulation/ checking) with unique and (more or less) flawless (if a bit dry and over structured) compellingly professional answers to highly complex, wholly novel questions (i.e. which have never been asked, less still answered, before) on technical issues, to which there are definitionally no answers to look up (nor any other deep research fix), and which require the exercise of professional judgment, and genuinely deep reasoning.

Tasks it would take a 20 to 30 year Post-Qualification professional at least a few full days to do manually from scratch (so at least a 3x speed up, even allowing for iteration and quality control).

Obviously, the second / third rate free product tier offerings for Gemini/ ChatGPT/ Grok/ Perplexity/ Claude are no comparison with full fat fully priced enterprise level.

But the free to use junk versions are not what’s in the frame to be replacing white collar roles here (including mine 🙁 , then again might not be a bad outcome, at least if I can get early retirement, or, failing that, a nice redundancy payment out of it eventually 😉 ).

On the other hand, none of this AGI.

Not
Even
Close

It’s not even clear if it’s on the track to AGI.

You canna’ change the laws of physics, Mr Altman.

Pure compute scaling only reduces entropy loss at the derisory rate of the one twentieth power of the increase in effective Floating- Point Operations: A ten million fold observed increase in effective compute applied to LLMs from 2017-23 (10x p.a.) only halved the entropy loss rate: so 10x more compute p.a. reduces entropy loss by 10% p.a.

And parameter increase seems to scale at the one tenth power.

An improvement to be sure on compute scaling, but is it realistic to envisage going from a trillion or so parameter models to models with many septillions of parameters, as we’d need for fully autonomous AGI?

If we can get to real AGI or ASI and it doesn’t kill us then we’re under investing massively, as Pueyo observes.

If that turns out to be the case, and we are on track for imminent AGI, the only rationale thing (if we actually want AGI) is to bet everything on it, and then borrow to bet more.

But progress seems to be slowing just as scaling laws predict. It’s just not, and never will be, an economical proposition (as opposed to a proof of concept) to put exponentially more resources in to get sub linear improvements in output.

And if we’re not on the road to AGI then the existing Capex (yet alone the expenditure in the pipeline) is never going to make a return. It’ll be the mother of all AI winters. Dark fibre in 2000 won’t even touch the sides for a comparison. I just hope a hard crash in the Mag 7 and the unlisted profitless AI companies doesn’t end up tanking the whole economy. We’ve had more than enough macro, geopolitical and exogenous economic shocks since Monevator started just over 18 years ago!

Data centre Capex assumptions in the 58 page JPM dissected here:

https://open.substack.com/pub/tscsw/p/jp-morgans-7-trillion-ai-warning

Electricity and credit as primary bottlenecks. However, can low latency (delay in getting grid connections in the US) not be addressed by just always building all the power infra at the data centre site itself (i.e. Solar PV/thermal or (SMR?) nuclear directly into the data centre)?

Also, $650 bn p.a. perpetual revenue isn’t actually that much, barely 0.5% of Gross World Product.

If we eventually see the white collar information economy related jobs (the PMC) fall from, say, 2 bn to 1 bn high quality roles globally (thereby saving say $30k p.a. per role on average globally, based on a $60 T p.a. global salary / payroll bill), all for an inclusive fee to the agentic model provider per remaining role of $300 pcm, then that makes $3.6T p.a. recurring enterprise revenue before any consumer subscriptions.

Re: “cut and paste machine”: are you using a full premium/ultra (whatever label) enterprise package at £200+ pcm?

The enterprise versions give you (for example) ~12.5 million tokens pcm (so 375,000 day with a 10% buffer), against just a few thousand per day/ deep research query on rubbish free versions.

Even with enterprise level, you still really have to iterate, wiping the memory of the query and then inputting it’s output as amended by you (and adapting the prompt between cycles) to get it to expand and improve over output cycles.

The free stuff on my own phone is basically no more, at best, than turbo search.

This, I think, is perhaps why retired MV readers rate LLMs as crap.

The paid up enterprise level at work though is a quite different beast – a significant (albeit still mundane) utility (at least until agentic autonomous automated work flowing becomes possible, at which point – when/if – it may become more than that).

I see mid hundreds of billions through to low single digit trillions of dollars of business revenues from this (IaaS, Intelligence as a Service) annually (on maybe 50% margins eventually – this is too ‘inputs intensive’ to get upto the traditional SaaS 80% gross margins) by some point in 2030s.

We’re looking at several orders of magnitude improvements to each of algorithms and hardware possible (in effective FLOP/watt/$).

It all depends if businesses can / will get rid of huge numbers of relatively higher remuneration professionals.

Sorry about the delay in replying – hectic at work
🙁

Well, LRMs and LLMs are unreliable generally, and especially for basic maths.

Strangely, frontier models optimised for advanced maths are able to get into top tier positions in Mathematics Olympiads!

But, for day to day stuff, you’re better off with a calculator and Excel or, for more heavy duty work, with Wolfram Mathematica:

https://www.wolfram.com/mathematica/

LLMs/LRMs are revealing one aspect of consciousness with respect to concept space in natural language, and appropriate models do seem to handle formal symbolic languages like mathematics.

But, for most basic maths operations, simple logic gates in a calculator will beat the uncertain backward propagation weights adjustment in a deep layer neural net.

It’s horses for courses.

The brain has specialised areas for particular functions.

AGI, if it’s possible, may require several radically different architectures and approaches to work together seamlessly within the same system.

That means, in turn, that the different schools of thought on ML/AI have to come together, uniting behind a shared goal in spite of their personal, philosophical and professional divisions.

Sadly, given the falling out between Stuart Russell and Geoffrey Hilton on the one hand, and Gary Marcus, on the other, that’s not a given.

Perhaps a figure like Richard Sutton could be a future unifier, but just at the moment the symbolic language reasoning and neural nets approaches are poles apart, seemingly irreconcilably opposed despite a shared utopian goal, like the split between Trotsky and Stalin divided world communism in the 1930s.

The neural nets have all the money, power and prestige right now, as the hyperscalers embark on own their mad dash 5 year plan for a stakhanovite data centre build out based on LLMs; with the formal symbolic approach faction being exiled to the periphery, with no money or powerful backers, and besieged by naysayers.

But biological intelligence, shaped by millions of years of natural selection, converges upon the same solution of integration, with different parts of the brain undertaking different roles and with System 1 and System 2 modes of thought.

Language models and image generators are essentially superhuman System 1 engines. They are masters of pattern recognition and prediction, trained on massive datasets. But they have yet to show any System 2 attributes.

As you say, they still don’t consciously reason and can’t. Something more is needed.

Evolution did not just scale our primate System 1 (a perceptual and intuitive network) until a System 2 magically emerged. Instead, evolution’s solution was an architectural one.

It developed a new, physically distinct, and specialised modul with the prefrontal cortex.

The PFC is not just more visual cortex. It has a different cellular structure, different connectivity, and a different job (working memory, rule based manipulation, abstract goals).

Evolution then formally integrated this new “System 2 module” with the preexisting System 1 modules.

If scaling a System 1 network was all it took to get a System 2 response, the prefrontal cortex would have been unnecessary.

Evolution’s own “solution” to generality was hybridisation.

For investors I think you have to look beyond the (legitimate and important, but not fundamental) second order questions of whether, for example, OpenAI can raise $1.4T by 2030 (Moneyweek reckons they need $12T by 2032 to meet Altman’s 250 Gigawatts goal, about the same as India’s current electricity usage!)

The fundamental question is (and always was and will be) “does scaling continue to work”.

Without it continuing working, I just can’t see how we get AGI anytime soon, given that that’s the only approach the money and the bulk of the talent is getting behind.

#242: ‘”run out of data for scaling anyway”: will the sequence be ‘data dries up’, ‘electricity taps out’ and then ‘the money runs out’, all leaving huge improvements left on the table in effective computation (eFLOP) through algorithmic and hardware progress?

The scaling wall looks fundamental to LLMs (10x compute = only ~10% less entropy loss in the model outputs, and ~10 orders of magnitude (10 bn x) more parameters = just a 10 fold reduction)?*

But there’s no rule out there to say that we even manage to get to the point of us crashing into those fundamental constraints.

These guys think electricity is the biggie:

https://open.substack.com/pub/nicholasvardy/p/the-real-ai-arms-race-isnt-about

It’s looking like fully ‘all in’ and ready to go / installed 1 Gw capacity for cutting edge GPU compute (with the frontier models in the US) is now coming in at around $50-$60 bn.
Hence Moneyweek’s estimate of $12T for OpenAI to meet its 2032 ambition of 250 Gw of compute (up 1,200x on what they had in 2024).

Of that, the power plant (assuming nuclear, to ensure a constant level of electricity supply) comes to about $35 bn.

But in China they can do a 1 Gw nuclear plant for as little as $3 bn / Gw.

Add on Chinese DeekSeek levels of focus on doing more with less both algorithmically and architecturally, their laser like focus on deployment of ML into immediate RoI industrial uses, their 4 x and rising no’s of computer science and engineering grad’s (of likely higher quality) than the US, and the canny Chinese ability to use dirt cheap last gen chips to get more cost effective raw FLOP /watt (with the cheapest power pricing in the world): and China looks better placed than most expect to ‘win’ an ‘AI’ ‘arms’ race’. The tortoise and the hare (or should that be the dragon and the eagle) all over again. I’ve a feeling that Deng Xiaoping’s vision will end up being realised long before Sam Altman’s.

*I’ve seen research paper assumptions that the amount of algorithmic improvement (eFLOP / raw FLOP) could be anything from a very pessimistic 50 x to as much as 100 mn x (i.e. 8 orders of magnitude) from here onwards; with hardware having another 4 to 6 orders of magnitude (10,000 x to 1mn x) potential raw FLOP/$ improvement to go.

More on the man behind the shorting legend:

https://open.substack.com/pub/shanakaanslemperera/p/the-cassandra-protocol-why-michael

He’s right on the depreciation. If anything his figures seem far too cautious.

I’m seeing estimates of up to $700 bn Capex in 2026, maybe $800 bn in 2027, some $900 bn in 2028 and implied totals of around $1.1 tn and $1.2 tn respectively in 2029 and in 2030 (composite from multiple sources – with the last 2 years derived from estimates of cumulative expenditure over several years).

I’m also looking at an annual release new Nvidia GPU cycle with a 2.5x improvement in FLOP/W per cycle now such that the fully economically realistic depreciation is:

Year 1 – $100
Year 2 – $40
Year 3 – $16
Year 4 – $6.40
Average years 1 through 4 = $40.60

If you were to flat line depreciate over 4 years then, in contrast, you get
Year 1 – $100
Year 2 – $75
Year 3 – $50
Year 4 – $25
And nothing in year 5. Average years 1 through 4 = $62.50

Where I think Burry may be erring is on timing.

Directionally he’s likely to turn out to be spot on, but this bubble can stay irrational longer than he, or any of us, can stay solvent.

Maybe I should sell my (unfortunately, despite a nearly several hundred percent gain in not much more than 18 months) still tiny Palantir holding 😉

Then again the tiny position sizing to begin with was my own hedge.

There’s no way I can defend it’s 449x forward PE (or 136x trailing 12 month sales) at the peak price of just over $200. Growth is accelerating fast, but that still can’t justify the market valuation. It just can’t – whatever one thinks of Alex Karp etc.

And this today from the DT, showing that now the Worrying Classes of the Tory Shires are also…. worried:

https://www.telegraph.co.uk/business/2025/11/16/the-hole-lurking-big-techs-trillion-dollar-ai-blitz/

But at least the VC community can’t see a bubble when there is one (or do they have it right after all ??!):

https://open.substack.com/pub/thevccorner/p/coatue-ai-report-18-charts

Just spotted an epic maths fail in the OP X/Twitter screenshot in Dave’s piece linked to above.

It’s $30 per million tokens via US providers, like OpenAI, and a mere $0.14 (yes 14 cents) per million tokens from Chinese ones, like DeepSeek.

So, that means that we’re assuming a startup uses around 100 billion tokens per month, and not 100 million as quoted, with the cost of using US models being $300,000 pcm on that basis, versus just fourteen hundred bucks in practically chump change using the Chinese models.

A token is 0.7-0.8 words, so 100 billion tokens per month makes much more sense for a startup, especially given that LLMs are crunching their way through literally quadrillions of tokens every single month now globally.

In other news, Buffett buy Alphabet:

https://open.substack.com/pub/jamesfoord/p/ai-bubble-buffett-disagrees

I’m intrigued. I’ve also seen Amazon recently tipped as a hidden (in plain site) relative value play versus the rest of US mega cap tech.

What an absolutely superb analysis of brownfield, low margin industry uptake potential for small reasoning/inference and small language models in edge computing/ offline settings:

https://open.substack.com/pub/artificialintelligencemadesimple/p/the-low-tech-revolution-why-ai-will

Great alternative name for a Substack too (Chocolate Milk Cult).

Again, I think China will be massively in the lead on this, given the preference over there for practicality over the SV crowd’s grandiose (VC bait) narratives.

Yep. It’s either an non-arbitrageurable value gap compared to the US, or Chinese AI related tech is a potential trade of the decade:

https://substack.com/@poezhao/note/c-178784343?r=2kxl2k

A more optimistic take, and a counter punch to that August 2025 MIT study finding 95% of firms reporting no RoI:

https://www.pymnts.com/artificial-intelligence-2/2025/how-leading-enterprises-really-measure-gen-ai-roi/

Highlights:

“Nearly all (96%) executives at the enterprise level in our studies report favorable positive results, up significantly from even this time a year ago. That’s even though the technology is still in its very earliest innings of potential. And by everyone’s admission still has a long way to go”

“we find a 3x increase in enterprise executives saying Gen AI is highly effective for product development and improvement and a 40% increase in the share who find it highly effective for improving workflow management and internal processes”

“with 28% reporting a very strong business case to support their Gen AI deployments. In just the last six months, the share who will increase their investments in the technology has more than tripled”

Anecdotally this matches my own experience of using enterprise level (£200 pcm per user license) models. It’s a different world from free turo search. I can at least 3x my core professional written output even allowing for multiple iterations and extensive editing / QA.

But will the business subscriptions come in fast enough to save the hyperscalers and model makers? Too late might be almost as bad as not at all. In any event, does it even matter if China does it nearly as well but 10x – 100x cheaper??

A ‘hot take’ on the (relatively) subdued reaction to the Nvidia earnings beat yesterday (it’s not back over $200 last I looked):

https://open.substack.com/pub/shanakaanslemperera/p/the-nvidia-paradox-how-a-57-billion

The leccy’s gone out! 🙁

Big takeaways: “For investors, the signal is clear: Nvidia’s earnings beat validates near-term demand. The margin compression, circular financing structures, and infrastructure constraints validate long-term concerns” and “The AI boom is real. The question is not whether it will transform the economy, but whether current valuations reflect the timeline and obstacles that transformation will actually face”.

It’s a race against time to secure visible RoI and electricity. Eventually the bet will probably pay off, but not necessarily before one or both of the money and/or the power runs out. The creative funding carousel / vendor financing is just an attempt to try to shorten the odds before the clock ticks down.

Oh…

And again agreement that China’s well in the lead on reliable power availability and for future installations.

A decisive advantage.

Hilarious but true, the crash is all down to Google’s TPUs. Weep ‘Masaponzi Son’:

https://open.substack.com/pub/shrubbery/p/aibspart-vi-kiss

Energy again – thoughts on the electricity input play into the AI ‘no moat’ paradox:

https://open.substack.com/pub/marketsentiment/p/the-great-ai-trade-energy

On moats, if CUDA was a moat for Nvidia then it is a software moat for GPUs, not for TPUs. If Google’s TPUs topple the GPU paradigm then CUDA will be no asset or help for Nvidia.

The questions are, will Google keep TPUs for themselves, and, regardless, if they rent their capabilities out then will they undercut the competition?

If either or both questions answer “yes” then Nvidia will find out that the only truly durable moat is the ability to adapt, change and to reinvent.

Further to #237 above – now with more on that 58 pager from JPM on the $5-7 tn buildout: every bubble starts with a spreadsheet:

https://open.substack.com/pub/tscsw/p/jp-morgans-7-trillion-ai-warning

AI 2027 officially postponed to after 2030:
https://open.substack.com/pub/garymarcus/p/breaking-the-ai-2027-doomsday-scenario

We’re not getting to AGI (let alone ASI) via LLMs anytime soon, that’s for sure – not with error (entropy loss) reduction scaling at the one twentieth power of compute increases (10x compute = 10% error loss improvement).

I share Gary’s earlier stated estimate for p(doom) resulting from LLMs (by 2030, IIRC) at <1%.

Of course, there could in the (even near) future be a groundbreaking conceptual breakthrough leading to exponentially more intelligence (or at least inference type reasoning) from the same, less or only linearly more computation.

That would carry a much higher p(doom), but is, of course, necessarily conditionalised upon the breakthrough first happening, which (prima facie) is itself quite unlikely.

The more likely scenario for doom in 2027 is not existential risk but the prospect of data centre finance (circular vendor finance and private credit) collapsing by then, or of the power supply just not being there (given connectivity issues). No money and/or no leccy for data centres seems much more likely risks than no human life left tbh.

It might only be a mundane utility technology these LLMs, and not (and maybe never) AGI, but, at it's very best, it's actually quite impressive mundane utility.

Although I seriously doubt that we'll see accelerating growth, or even a big one off step change up in growth, I do think that we will see a bit more than just very marginal or even just negligible improvement to growth rates.

Using the tech of enterprise level reasoning models feels a bigger deal than access to the internet and email did back in the 1990s (from one who remembers when both appeared into our world).

For what I do, you need to interate, iterate. iterate. Save each version between cycles. It’ll become part of the input for the next one.

For me Copilot (Microsoft) works best with more material, but is (at least for me in my work environment) currently limited to 3 uploads per prompt.

So, use PDF Pro to merge docs into a meta doc for uploading which can be hundreds/thousands pages long.

If you have a decent length (tens pages) client tech sub that helps. Prompt. First draft advice output. Delete Chat memory.

Tell it to do a vigorous and critical review of its draft (on the basis its your first draft) identifying any and all areas of inaccurate, incomplete or underdeveloped text.

Feed it (upload) the original attachment(s) together with draft 1. Repeat cycle deleting chat memory between each.

Then ask it to come up with counterpoints and alternative views.

Then delete memory and upload those with previous drafts and attachments (merging docs into one attachment PDF to keep within 3 item upload limit), and ask to weight up different contentions.

Delete memory and reintegrate its last assessment with everything else.

Then ask it to evaluate the evaluation and, if it thinks counterpoints have merit, to integrate them into advice.

Delete memory and get it to check for accuracy and rigour and to turn anything overly ‘bullet pointy-ish’ sounding and styled into fully developed natural language English sentences.

And so it goes on.

After half a day’s to a day’s work, and a dozen iterations, you’ve got the best, most complete, most rigorous, accurate and precise advice that could possibly be written on the most mind warping complex and difficult subjects which, writing from scratch with no AI resources, would take not less than a full week, and probably slightly more.

The ‘trick’ here I think is to get it to think deeply and to fully reason and reflect before answering.

If it spits out its answer in just a few seconds it’s going to be surface level. Minimal utility. Avoid that.

If your prompt gets it to ‘think’ for 100, 150, or even over 200 seconds (it counts up its reflection time on screen) before answering then it’s likely going to be an impressive output. Some of the responses I’ve gotten out of it at 200+ seconds inference time have been borderline genius level. You almost have to trick it to think for that long though.

So, in truth, for this type of extreme complexity and amount of materials work, it’s probably a 5x productivity boost by output over unit time.

But, with more straightforward stuff, the advantages are correspondingly less. So I don’t really bother it with simple stuff.

Given that, overall I’d say its perhaps a 3x increase to productivity across the board.

On the memory – I can only really speak for work GPT-5 on Copilot for enterprise level paid product delivered via a desktop (although I’ve got free Gemini, ChatGPT, Perplexity, Claude and Grok on my own phone). It doesn’t seem to have great memory permanence between chats or between uploads (i.e. you have to re-upload between iterations).

It’s also frankly s**t at numbering and at general document formatting (it will try if you ask it to, but, by and large, just don’t bother).

Somewhat counterintuitively for a machine, it excels (or at least can excel, albeit not always reliably repeatably) on higher / highest order reasoning (and with having the benefits of having north of thirty trillion words in the training run set); but the simpler, ‘mechanical’ type tasks are occasionally epic fails.

So, the harder, the more abstract and the more complex the task; then the easier it crunches through it; and the more straightforward and brain light editorial type the chore (like formatting), then the higher the risk of it falling over.

I delete chat history / memory between iterations as I want to prompt it to look at each iteration completely afresh, as if it were someone else’s work that it were considering for the first time. This also stops it trying to ‘defend’ previous versions, which isn’t conducive to progression, both in LLM output building and with much else in life.

At each and every iterative improvement I give a different (or at least slightly new) role to the model, as compared to before, e.g.:
– a critical but fair and impartial reviewer, told to carefully and exhaustively check accuracy and completeness;
– an enthusiastic rewriter told to identify areas lacking clarity and fully developed sentences (otherwise it often reverts to, and your left with, very logical and grammatically correct sentence structure and language choices but ones which can be overly direct, a bit too structured and rather lacking in natural flow); and
– an argumentative counterweight, told to try and unpick the piece, find fault with it (i.e. regardless of merit) and to find any and alo alternative arguable views (i.e. to pressure and stress test the previous iteration output, and to generate possible new and alternative avenues of analysis for exploration in subsequent iterations).

Prompts between iterations which use deliberately exaggerated criticism of the previous (in reality perfectly good and accurate) iteration like, for example, “this lacks sufficient depth and detail. It needs to be improved and developed. Suggest appropriate text” or “this hasn’t nearly begun to bring out all the arguments and relevant points. Identify where and how it’s gone wrong, and supply additional paragraphs to remedy this by reference to the preexisting numbering” etc can all help push up the overall levels of quality, reflection and depth of analysis to ones that are as good as, or even better than, any human author.

Sometimes it doesn’t work though. Sometimes the model just ‘collapses’ in on itself and gives out an inferior reiteration.

Just delete that attempt and press on.

It never responds in the same way twice. Like a ghost in the machine. A digital Alibaba which has to be coaxed out of the cave to grant your wish.

And each time it needs a different set of prompts and inputs. Open Sesame doesn’t always do it!

But when it works, and invariably it does eventually work (with enough iterating), then it’s worth it.

Needless to say this typically uses up an enormously large number of tokens.

If task length can be improved in future models then all of the above iterating will sooner or later be done by agentic models taking different roles (like the DeepSeek Mix of Experts approach, but on steroids, and run both concurrently and then sequenced consecutively). At that point I’ll likely be out of a job. Hopefully, if that happens, then it takes at least 5 years – which is all the time that I need (I currently hope) to FIRE.

Turning to new links, this is good on the forthcoming Google ascendancy:

https://open.substack.com/pub/exponentialview/p/can-ai-escape-googles-gravity-well

As the piece says, its got: “(1) research labs rivalling OpenAI’s talent density, (2) infrastructure investment approaching $93 billion annually, and – most crucially – (3) distribution channels no startup can replicate.”

IMO the winning move, at this point, would be for Google with its TPUs and Nvidia with its GPUs to merge. OpenAI is yesterday’s story. Seen and gone.

That’s a very (very) interesting proposition @TI (and thank you 🙂 ) I’m going to have a think about it, and then DM you, if I may, via email.

The existential dread of a cautiously sceptical AI optimist again.

If you needed a reason to pour a stiff drink and stare blankly at your Vanguard LifeStrategy allocation, this reading list from the last day or so is certainly it.

A central tension exists between the “Nvidia Monopoly” narrative and the resurgence of hyperscalers owning their own stack. 

Money Machine and The Dutch Investors highlight that Google’s Gemini 3 (running on internal TPUs) proves one can bypass the “Nvidia tax.”

https://open.substack.com/pub/moneymachinenewsletter/p/what-gpu-shortage

https://open.substack.com/pub/thedutchinvestors/p/why-alphabet-will-win-in-the-age

The trade is shifting from “buy the shovel seller” (Nvidia) to “buy the sovereign utility” (Google/Alphabet).

Tech Fund and Sam Ro debate whether this CAPEX spend is a bubble or the new industrial base, but Shanaka’s work suggests that owning the compute and the energy source is the only moat that matters.

https://open.substack.com/pub/techfund/p/ai-outlook-and-nvidia

https://open.substack.com/pub/samro/p/ai-capex-buildout-vs-past-market-bubbles

Shanaka Anslem Perera’s posts (“Stellar Imperative,” “Monetary Energy Singularity”) argue that economic sovereignty is now just physics.

We are transitioning from a Type 0 to a Type I civilization (Kardashev scale), where the only currency that matters is harnessed joules (solar/fusion) converted into inference (intelligence).

https://open.substack.com/pub/shanakaanslemperera/p/the-stellar-imperative-why-planetary

https://open.substack.com/pub/shanakaanslemperera/p/the-monetary-energy-singularity-how

The “GPU Shortage” is not just a a supply chain glitch, but  a species level bottleneck. Exponential View reinforces this by tracking the “thermal wall” and the exponential demand for energy to power these “sovereign minds.”

https://open.substack.com/pub/exponentialview/p/ev-551

Just as nations risk becoming vassal states to AI clouds, traditional investors risk becoming irrelevant observers of a machine driven market. 

https://open.substack.com/pub/shanakaanslemperera/p/the-inference-sovereignty-thesis

The “China Tech Profitability Trap” is the geopolitical echo of this where the old models of profit don’t work when the structural rules change.

https://open.substack.com/pub/hellochinatech/p/china-tech-profitability-trap

There is a delicious, if terrifying, irony running through these links. 

On one hand, we have the techno optimists (Shanaka, The Dutch Investors) essentially arguing that we are mere rounding errors on the Sun, frantically trying to build a Dyson sphere out of Nvidia H100s and Google TPUs before the lights go out. The “Stellar Imperative” isn’t just an investment thesis; it’s a physics lecture with a ticker symbol attached.

On the other, we have the creeping realization that the “easy” trade buying the shovel sellers might be over. 

https://open.substack.com/pub/shanakaanslemperera/p/the-structural-collapse-how-googles

https://open.substack.com/pub/shanakaanslemperera/p/the-great-ai-depreciation-disconnect

The Money Machine note on Google bypassing the “Nvidia tax” with Gemini 3 is the sort of vertical integration pivot that makes active managers wake up in a cold sweat. 

It seems the only thing more dangerous than betting against the bubble is betting on the wrong monopoly.

And then there’s Michael Green, appearing like the ghost of active management past, whispering that “his life is a lie.” 

https://open.substack.com/pub/michaelwgreen/p/part-1-my-life-is-a-lie

One suspects he is articulating what many in the City won’t say aloud: that in a world of “Inference Sovereignty” and “Monetary Energy Singularities,” the old skill of reading a balance sheet is about as useful as a dowsing rod in a desalination plant.

It all reinforces a boring, stubborn adherence to the global haystack. 

Let the titans fight over who owns the “sovereign compute” stack. 

Let the macrogurus argue if we are in 1999.

Own a sliver of all of them. 

If humanity does manage to ascend the Kardashev scale and harvest the sun, I presume the index will capture at least some of the upside 😉

ROFLMAO – but you know I’m not going to settle for the index. No one wants to be average even if that’s not a bad thing to aspire to.

So I’ll keep turning over those rocks. And maybe it will work. Maybe it won’t. And if not, then, well, my SWR will be the least of my worries.

An interesting take this evening:

https://open.substack.com/pub/neuralfoundry/p/the-ai-scaling-laws-hold

For a couple of months, as momentum in the SPY/Nasdaq lost steam (coming off the epic rally from the tariff tantrum crash in April), the market whispered that the AI scaling party was over. The narrative was simple. Costs exploding, but capability gains shrinking.

This piece thinks that the latest frontier models confirm this fear was a mirage and that scaling laws are not broken; they’ve just entered a more sophisticated phase.

The key, it thinks, is sparse scaling, proven by models like Gemini 3’s rumored 5 trillion parameter Mixture of Experts architecture. This explodes total capacity, while keeping inference costs more rational. And surprise, surprise Google, not (washout) Open AI are the (again, only Western) innovators.

But does this in and of itself fundamentally derisks the long term investment thesis for AI infrastructure?

And will value accrue so heavily in the future to the foundation builders like NVIDIA (as supero omnia hardware enabler) and to the hyperscalers (Google Cloud, Azure, AWS), who, respectively, provision and manage the massive, tightly coupled compute clusters?

I just don’t know.

However, I think that I do know that one correct (or at least pertinent) question to ask now is whether (or not) bottlenecks are any longer primarily just chip centric (including GPU power requirements), or, instead, now operate at the level of both the coherent compute clusters and the software required to run these gigantic MoE systems?

Noone can say if the infrastructure trade is now/still cemented, or if, instead, the only wall that matters will be algorithmic efficiency.

Well, my take on his approach is that, as a modern Cassandra, he thinks that we’re a bit screwed because of the hedonic trap of participation, which means that we need far more income than we’ve actually got as a median in order to feel like valid members of a consumption economy ($140k/household).

Given that, the only realistic option (for Millennials, Gens Zs and Alphas) is to play the market casino, and hope it pans out.

With everything else he complains about (with some cause, I think, TBF) on market structure / no real price discovery (passive flow/constant bid, for now), this makes fundamentals a mere passenger, with momentum and flow/illiquidity the real drivers of returns, and, therefore, also the determinant of economic ‘health’, given that, as Market Cap / GDP (the Buffett indicator) expands, the economy becomes the market rather than the market merely reflecting/ following the economy.

That’s the thrust to me of his writings.

The AI spin here is that this (eventual AGI) might be the last, best hope as a way out of this, but only if AGI both actually significantly increases general wealth and also markedly decreases inequality (whereas, presently, it’s set up fundamentally to do the very opposite, i.e. to get rid of jobs in the squeezed middle 25%-75%, the PMC, and to reduce most waged income for those remaining).

BTW, further to all of the above links referencing Gemini 3, here’s another ecstatic review re Google’s future:

https://open.substack.com/pub/nexteconomy/p/the-one-company-economy

Gary highlighting today a swift, ironic pivot in SV’s economic philosophy. He contrasts tech VC David Sacks’ recent laissez faire stance that companies should be left to sink or swim with his sudden shift toward “safety net socialism” for AI giants like OpenAI. Marcus wryly notes that if OpenAI stumbles, Google is ready to step in without a bailout, exposing the convenient inconsistency of tech leaders when their own investments are at risk.

https://open.substack.com/pub/garymarcus/p/a-tale-of-two-ai-capitalisms

Shanaka argues that while Nvidia dominates training, it suffers from a “structural blind spot” in inference that Google is aggressively exploiting. Google’s custom TPUs create an “inference arbitrage” delivering far superior cost-efficiency for running models compared to Nvidia’s general purpose GPUs. This hardware divergence allows Google to undercut competitors on price and scalability, turning an architectural difference into a massive strategic moat against the rest of the market.

https://open.substack.com/pub/shanakaanslemperera/p/the-inference-arbitrage-how-googles

Missed this one off yesterday. Uncover Alpha delivers a comprehensive deep dive into the Google TPU, tracing its evolution from a 2013 internal necessity to a primary strategic weapon. The analysis details how TPUs use “systolic arrays” to strip away GPU architectural baggage, achieving superior FLOP/watt. It argues that this vertical integration allows Google to escape Nvidia’s margin trap, offering a sustainable path to cloud profitability that GPU dependent competitors simply cannot match.

https://open.substack.com/pub/uncoveralpha/p/the-chip-made-for-the-ai-inference

You could always hive it off as Moguls’ side project perhaps.

Maybe 95-99% of investment substack is, at least I’d guess from my own inbox’s content here (I’m soon going to have to pay Google to upgrade to 100Gb!), basically US focussed; and I reckon 50-80% of the rest is geared towards Europe, Japan and Emerging Markets.

The number of UK orientated sites like yours, is basically you and a couple of others (BoW, FvL – but only occasionally posts – and David Stevenson now).

There are some UK small cap bloggers (Paul Scott), but noone with your Peter Lynch crossed with Phil Fisher with a dash of Anthony Bolton active style (I’m guessing – but that’s what I see in your Moghuls’ write ups). I’m style agnostic (There’s no right way. Just your way. Then again, there are plenty of wrong ways).

So, I think that you have space to explore new themes, AI included.

On AI engagement. Yeah it’s a mystery to me.

Even with just mundane utility falling way (way) short of AGI (and even with China winning the AI race from where I’m sat, with only Google and maybe Nvidia looking pretty at this point, and OpenAI and Anthropic in deep trouble I think), I am still staggered at the lack of interest about ML/AI (whether here or more generally in society).

Everyone is planning as though precisely nothing has changed. Nothing to see here!

The know unknown.

Comforting but untrue/ invalidated certainties rather than deeply uncomfortable unknowningness and upheaval.

What is it to be a person? To think. To be conscious. This is what separates us (whether by degree or qualitatively, or both?) from a rock.

This is the realm we are exploring now.

If intelligence has a long term future, and if it eventually explores its future light cone, then it seems pretty unlikely that it will be the done by unaided creatures of slender flesh and meagre bone which evolution has so awkwardly and haphazardly produced in its flawed, localised and incremental way.

Maybe it was like this for the Wright Brothers.

For literally years the wider World either refused to believe or simply ignored that they’d actually made a heavier than air flying machine, even though many thousands, including reporters with cameras, had seen them fly it!

Or perhaps it’s just the age of the audience now.

I never really grew up you see. Like F. Scott Fitzgerald writing on his characters’ childlike wonder and imagination, and stubborn insistence on their own realities, (and whatnot) 😉

Today’s quota of new AI posts from around and about Substackopia.

The legendary investor Drukenmiller has swapped Microsoft for Amazon, Meta, and Alphabet/Google.

He reckons the easy money in AI infrastructure is gone; the real lucre now lies with the ‘monetisers’ who can turn all that spend into cash.

A pivot from expensive dominance to undervalued acceleration, apparently.

https://open.substack.com/pub/nicholasvardy/p/why-druckenmiller-ditched-microsoft

A bit of creative accounting at the tech giants, it seems.

Microsoft and Google are extending the ‘useful life’ of servers to flatter earnings, just as AI makes that hardware obsolete faster than ever. It boosts profits on paper now but risks a nasty writedown cliff later.

Buyer beware.

https://open.substack.com/pub/henryzhang/p/the-depreciation-mirage-how-big-tech

OpenAI has quietly slipped out a new coding model that purportedly thrashes the competition.

While the internal benchmarks look tasty, some external tests are less convinced.

Is it a genuine leap forward or just another incremental update to keep the venture capitalists purring?

One for the tech heads, but just check out how GPT 5.1 ‘breaks the graph’ on the exponential curve of stable task length plot, suggesting to me that AGI is a long way off.

https://open.substack.com/pub/thezvi/p/chatgpt-51-codex-max

A cheat sheet for your next pub debate. It covers twelve key arguments for and against the bubble from terrifying Capex numbers and ‘infinite money glitches’ to productivity doubts.

Whether it’s a trillion dollar ouroboros or the next industrial revolution, you’ll at least sound like you know the score.

https://open.substack.com/pub/derekthompson/p/how-to-sound-like-an-expert-in-any

Nvidia posted blowout numbers again, yet the share price wobbled.

Why?

Was the market is fretting about customer concentration and the inevitable ‘digestion’ phase after such a binge?

Or had Mr Market heard about Google’s seventh generation TPU breakthrough with Ironwood?

The author sees it as a healthy consolidation, but the valuation leaves precious little room for a stumble.

https://open.substack.com/pub/qualitystocks/p/stock-of-the-week-nvidia-why-is-the

Rebound Capital is feeling chuffed about their Google call, up 35% in two months.

They argue Gemini 3 has leapfrogged OpenAI, and Google’s custom TPUs let them sidestep the ‘Nvidia tax’.

And with billions of users and its own hardware, the search giant looks surprisingly sprightly.

https://open.substack.com/pub/reboundcapital/p/notes-on-google-goog

And on the same theme, while everyone fights over Nvidia’s training chips, Google might have a trump card in inference.

Its custom TPUs are purpose built for running models cheaply, undercutting rivals stuck with pricier, general purpose GPUs.

It’s a structural edge that could turn a technical difference into a massive economic moat.

https://open.substack.com/pub/shanakaanslemperera/p/the-inference-arbitrage-how-googles

The narrative has shifted from euphoria to alarm. 

We face a pincer movement of financial skepticism with fears of wasted CapEx and bursting bubbles and existential dread regarding job displacement and loss of control. 

Whether it is the wrong architecture, the wrong economic model, or simply too much hype, the consensus is that the current path is unsustainable.

Friedman argues that if Ilya Sutskever is right, AGI won’t come from ever larger LLMs but from “continual learners” that bootstrap their own intelligence. 

This pivot would render the current trillion dollar infrastructure of massive GPU training clusters obsolete overnight, turning today’s prized assets into tomorrow’s scrap metal.

https://open.substack.com/pub/davefriedman/p/what-if-were-building-the-wrong-kind

This note likely underscores the widening disconnect between soaring AI stock prices and the sobering reality of revenue generation.

It captures the shifting mood in the market, where blind faith in the “AI Supercycle” is beginning to curdle into doubt about return on investment.

https://substack.com/@marketsentiment/note/c-181726653

The monopoly is cracking. Google is aggressively pitching its own TPUs to cloud clients and high-frequency traders, directly targeting Nvidia’s core data center dominance. 

By keeping workloads on its own silicon, Google threatens to siphon off a significant chunk of Nvidia’s revenue, initiating a true chip war.

https://open.substack.com/pub/archiveinvest/p/google-vs-nvidia

A visual reality check on the AI trade. This note points to the stark divergence between capital expenditure and actual software earnings. 

The math is becoming increasingly difficult to justify, suggesting we are nearing the peak of the hype cycle before reality sets in.

https://substack.com/@quantitativo/note/c-181383454

Zvi reviews the arrival of Gemini 3 and GPT-5.1. While these models offer distinct upgrades and new safety frameworks, they arguably fail to deliver the transformational “mundane utility” promised.

We are seeing better tools, yet the leap to true general agency remains elusive and the safety measures unproven.

https://open.substack.com/pub/thezvi/p/ai-144-thanks-for-the-models

This piece reflects on how the integration of synthetic intelligence is subtly reprogramming human biology and behavior. 

In a world increasingly mediated by algorithms, maintaining a “curious mind” becomes a radical act of preservation against a technology that seeks to automate our very thought processes.

https://open.substack.com/pub/afewthings/p/the-curious-mind-reprogramming-life

Coding as a profession is entering its twilight. With agents like Cursor and Google Antigravity, software engineering is shifting from writing syntax to supervising autonomous bots. 

The barrier to entry has collapsed, and with it, the premium wages that defined the sector for decades are likely to vanish.

https://open.substack.com/pub/aisupremacy/p/is-ai-eating-code

A sobering look at the “race” to replace ourselves: Billionaires are building a future without humans.

This video from the Future of Life Institute is particularly relevant as it ties together the economic incentives driving the AI arms race with the existential risks of losing control.

The warning here is stark: tech titans are not building tools to help us, but agents to replace us. 

The competitive pressure between nations and firms makes safety a secondary concern. 

Without coordination or hardware level controls, the race to AGI becomes a suicide pact where the only winning move might have been not to play.

To prevent a loss of control, the Institute proposes hardware based “dead man switches” on chips, creating a physical choke point to shut down rogue superintelligence before it becomes ‘sovereign’.

https://youtu.be/cuGWq82QHnM?si=IchA8QOnf2vNsL_X

We are witnessing the decoupling of corporate growth from employment. 

Companies are boosting profits while cutting headcount, risking a “consumer death spiral” where the displaced workforce can no longer afford the goods AI produces.

The economy is sleepwalking into a crisis of demand, not supply.

https://youtu.be/MGa2SQ_eRq8?si=Qjal5ZmMLfDD0q_3

If Sutskever’s “cortex” architecture works, it invalidates the brute force scaling laws that currently underpin the entire industry’s valuation.

https://open.substack.com/pub/davefriedman/p/does-ilya-sutskevers-safe-superintelligence

Marcus contends the machine learning community is finally waking up to the “madness” of the LLM detour. 

We have incinerated a trillion dollars chasing a technology that cannot generalise like the human mind, wasting years on a dead end while neglecting the neurosymbolic approaches that might actually work.

https://open.substack.com/pub/garymarcus/p/a-trillion-dollars-is-a-terrible

The AI boom is a classic bubble, mirroring the Dot.com era’s overcapacity and delusion.

Katsenelson warns that not every gold rush makes the miners rich; in fact, the massive infrastructure spending is likely to end in a glut. 

Smart money should avoid the hype and wait for the dust to settle.

https://open.substack.com/pub/vitaliy/p/why-smart-investors-should-sit-out

Taken together, this half dozen serve as a potent antidote to the feverish techno optimism currently gripping the markets.

When the structural skeptics, the value investors, and the engineers all point to the same fissures in the narrative, it may be time to reassess one’s exposure to the hype.

The tech is undeniably impressive, but, as I am fond of repeating, a great product does not always make a great investment especially when the price assumes perfection.

Burry sees the current AI frenzy as less of a technological singularity and more of a Dot.com replay, where circular revenue streams prop up valuations that defy gravity.

https://open.substack.com/pub/aisupremacy/p/going-short-on-generative-ai-michael-burry-nvidia

Friedman recounts observing two “frontier” AIs debating a strategy question, only to watch them hit a cognitive glass ceiling.

The experiment suggests that for all the probabilistic mimicry, these models remain brittle when forced off script, lacking the genuine reasoning required to solve novel problems.

https://open.substack.com/pub/davefriedman/p/where-large-language-models-break

A scathing autopsy of the OpenAI investment thesis, stripping away the ‘AGI’ mysticism to reveal a cash incinerating utility.

The critique centres on the uncomfortable truth that ChatGPT has no defensible moat; it is trapped between open source competitors driving prices to zero and a copyright minefield that could turn its training data into a toxic asset.

Investors in this hypothetical float wouldn’t be buying a software monopoly, but rather subsidising Microsoft’s cloud revenue while holding the bag for inevitable regulatory fines.

A necessary corrective to the valley’s distortion field and a sharp reminder that in an era where “vision” can be generated by the token load, the only defensible moat is execution.

HMG argue that while the market chases the next shiny model, the real value will accrue to those doing the unglamorous work of actually bolting the machinery together.

https://open.substack.com/pub/sbeautiful/p/never-buying-chatgpts-ipo-and-the

https://substack.com/@lxpwsk1/note/c-182246243?r=2kxl2k

Brock takes a sledgehammer to the temple of AGI, arguing that we are confusing a very sophisticated autocomplete with a sentient mind.

He posits that the current architecture is a statistical dead end for true intelligence, suggesting the “god in the machine” is merely a parrot with a thesaurus.

https://open.substack.com/pub/mikebrock/p/why-im-betting-against-the-agi-hype

Gary Marcus delivers a grim pre-mortem on the generative revolution.

3 years in from the launch of ChatGPT on 30th November 2022, the shortcomings (hallucinations first then, lack of uniqueness and genuine and valuable discovery now) aren’t a bug; they are a feature.

If the technology cannot progress its scope of competencies (and generalise) substantively, and becomes fully agentic (able to do task after task both simultaneously and sequentially, all maintained reliably over time) after many, many hundreds of billions in investment, then investors might well wonder if it ever will be suitable for critical work at scale.

https://open.substack.com/pub/garymarcus/p/three-years-on-chatgpt-still-isnt

Less a case of creative accounting and more one of valuation gymnastics re: TSLA / X v xAI:

https://youtu.be/NNCqzfrZq9k?si=q3nA5VGejMvfzlhv

xAI: Currently ~$500 mn p.a. of revenue, $1 bn p.c.m of expenses, and a $200 bn private market valuation: Musk wealth in 2022 $200 bn, in 2025 $500 bn.

Dave Friedman today: “We’re now in an AI regime where the decisive variables are land, pipes, and defaults, not parameter counts. In that world, Google and Microsoft look like what they are: landlords. OpenAI is not a landlord. OpenAI is a very important tenant. The right way to think about this isn’t OpenAI vs Google. It’s:

Landlords vs tenants in the AI stack, and

Which layers actually get to keep the profits.”

So true. Revenue is vanity, profit is sanity.

Also today, Uncle Sam (the nation not Altman) muscling in on next gen EUV:

https://open.substack.com/pub/marklapedus/p/xlight-signs-letter-of-intent-to

Fallen behind on the links re: ‘AI’ (those two letters should be TradeMarked as a narrative framer, use the language you chose to create the perception you desire).

Anyways, here they are for the last few days:

The maths behind the magic to justify a $100bn valuation.

OpenAI must execute a flawless, narrow path to dominance.

With revenue needing to compound at eye watering rates, one wonders if the “valuation” is pricing in a monopoly that regulators and competitors won’t permit:

https://open.substack.com/pub/exponentialview/p/openais-narrow-100-billion-path

Some chilling parallels for those who remember 2008.

Friedman suggests the current AI financing structures where cloud giants invest in startups that immediately spend the cash back on cloud credits look suspiciously like the round tripping revenue schemes of the subprime era:

https://open.substack.com/pub/davefriedman/p/the-ghost-of-cdos-is-haunting-the

Financial engineering masquerading as organic growth? Surely not 🙁

The moat mirage is drying up. OpenAI’s “Code Red” isn’t just internal panic but a signal that their dominance is eroding faster than expected.

With open source models catching up and competitors like DeepMind lurking, the first mover advantage is looking rather perilous:

https://open.substack.com/pub/aisupremacy/p/openai-code-red-2025-the-fall-of-chatgpt

In a similar vein, perennial skeptic and LLM permabear Gary Marcus argues that OpenAI’s pivot to a for-profit structure is a sign of desperation, not strength.

He suggests the “Code Red” is less about safety and more about a business model that burns cash faster than it generates utility:

https://open.substack.com/pub/garymarcus/p/openais-code-red

A chart heavy inspection of two distortions in the current tech bubble: Private Credit and the financial sorcerers’ apprentice, Stock Based Compensation.

The former is fuelling risky corporate debt away from prying eyes, while the latter flatters earnings reports by pretending employee pay isn’t a real expense:

https://open.substack.com/pub/offthecharts/p/pc-and-sbc

Accounting magic at its finest.

While the West employs lawyers, China employs engineers.

Noah Smith argues China is reimagining the state apparatus to treat governance as an engineering problem.

It’s a formidable, if dystopian, model of efficiency that could give them the edge in the hardware heavy phase of the ‘AI’ ‘race’:

https://open.substack.com/pub/noahpinion/p/china-has-invented-a-whole-new-way

Are we entering a 5–15 year “Capex Supercycle” where spending on energy and chips becomes non-negotiable for nations.

And the twist if that’s so?

This inflationary spending might make boring old Treasuries a dangerous place to park your cash. Infrastructure as the new gold?:

https://open.substack.com/pub/aryadeniz/p/the-ai-capex-supercycle-and-why-treasuries

A deep dive into the philosophy of Ilya Sutskever. If you’re worried about the robot apocalypse, this won’t settle your nerves.

The discussion covers the path to Superintelligence and the near religious conviction required to build it safely. Essential reading for those betting on or against humanity:

https://open.substack.com/pub/thezvi/p/on-dwarkesh-patels-second-interview

A timeline for the near future. We are moving from “wow, a chatbot” to “agents that do the work”.

However, the gap between the exponential capability of the tech and the glacial adaptation of our institutions is about to cause some serious friction.

Buckle up:

https://open.substack.com/pub/exponentialview/p/the-next-24-months-in-ai

You can’t build data centres out of code. This piece highlights the looming physical shortage of copper required to wire up the ‘AI revolution’.

While everyone chases Nvidia, the smart money might be looking at the digging equipment.

A classic pick and shovel play for the commodities cycle:

https://open.substack.com/pub/nicholasvardy/p/the-coming-copper-crisis-why-smart

Like the Trojan priestess, Michael Burry is cursed to speak the truth and never be believed.

This revisits his bearish signals on the passive investing bubble and the tech over-valuation. If he’s right (again), the index trackers might be in for a bumpy ride:

https://open.substack.com/pub/shanakaanslemperera/p/the-cassandra-paradox-michael-burrys

Sanctions were meant to kill China’s AI dreams; instead, they forced efficiency.

DeepSeek’s latest model rivals the best of the West but runs on a fraction of the compute. It turns out that when you can’t brute-force progress with unlimited chips, you get very good at writing efficient code:

https://open.substack.com/pub/hellochinatech/p/deepseek-v3-2-china-ai-sanctions-paradox

A look under the bonnet of the DeepSeek architecture. Raschka explains how Mixture of Experts (MoE) and other optimisations allow this model to punch well above its weight class.

For the technically inclined, it’s proof that architectural innovation is just as important as raw GPU grunts:

https://open.substack.com/pub/sebastianraschka/p/technical-deepseek

China doesn’t just want to build the best model; they want to embed it in everything.

This piece highlights how Chinese tech giants are leveraging their “Super Apps” (like WeChat) to distribute so called AI instantly to billions, a speed of deployment the fragmented Western app ecosystem can’t match:

https://open.substack.com/pub/hellochinatech/p/china-ai-distribution-advantage

Prediction season, with bets that 2026 is the year the infrastructure wars truly begin. Data centres becoming the primary geopolitical asset?

If you thought 2024/25 was loud, wait until the physical constraints of energy and land start biting:

https://open.substack.com/pub/aisupremacy/p/will-ai-eat-the-world-in-2026

A philosophical look at capital vs. labour.

If AGI arrives, the value of human labour could trend towards zero, inverting the traditional economic model.

Capital (the robots) becomes the only game in town.

A terrifying thought for wage earners, a potential utopia for the owners of the machines:

https://open.substack.com/pub/shanakaanslemperera/p/the-great-inversion-what-the-agi

While the world worships the GPU, Google has been quietly building its own tensor processing units (TPUs). This piece argues Google’s vertical integration gives it a cost advantage that the market is ignoring. Nvidia takes the headlines; Google might take the margin:

https://open.substack.com/pub/philoinvestor/p/what-the-tpu-is-going-on

Stop looking at the chatbot wars. Friedman argues the real battle isn’t who has the best LLM, but who controls the distribution and the underlying rails of the internet. Google’s entrenched monopoly is harder to kill than a mere “better answer” engine suggests:

https://open.substack.com/pub/davefriedman/p/openai-vs-google-is-the-wrong-game

A scattered but insightful collection of thoughts, ranging from the energy costs of inference to the sociological impact of AI companions. The recurring theme: we are vastly underestimating the second order effects of this technology on our daily lives:

https://open.substack.com/pub/exponentialview/p/ten-things-im-thinking-about-ai-part-4

The weekly roundup from Azeem Azhar. It reinforces the theme that while the hype may be frothy, the underlying “compute economy” is hardening. Key takeaways include the shift from training to inference and the continued divergence between tech valuation and broader market reality:

https://open.substack.com/pub/exponentialview/p/ev-552

Fair few to catch up with on the link front/ info tsunami:

https://open.substack.com/pub/nourielroubini/p/americas-economic-tailwinds-will

Roubini suggests the long-term tailwinds of the US economy are so powerful they’ll overcome the noise of tariffs and political lunacy. Glad someone is looking past the next five minutes, even if they are known as ‘Dr. Doom.’ 

https://substack.com/@aisupremacy/note/c-184484769?r=2kxl2k

China’s home grown chip industry is booming, suggesting that U.S. sanctions might just force them into self-sufficiency sooner. Nvidia’s loss is Baidu’s gain, proving geopolitical risk trumps quarterly earnings, as ever. 

https://open.substack.com/pub/aisupremacy/p/going-short-on-generative-ai-michael-burry-nvidia

The Generative AI rally is showing all the classic bubble signs that even passive investors spot. Citing Burry, the author suggests a short on high flying Nvidia might be sensible, assuming you enjoy heart stopping volatility. 

https://substack.com/@marketsentiment/note/c-181726653?r=2kxl2k

HSBC forecasts a cool half a trillion in operating losses for OpenAI, begging the question: is this a technology company or just a very expensive charity? Apparently, funding the future is costlier than owning a server farm. 

https://substack.com/@ftav/note/c-184141268?r=2kxl2k

FT Alphaville’s simple “We’re not even mad, that’s amazing” note perfectly captures the current market’s absurd valuations and peculiar trading. It’s comforting to know even the financial press are just slack jawed onlookers sometimes. 

https://open.substack.com/pub/coughlincapital/p/china-is-still-the-most-asymmetric

Unfashionable China stocks, with their depressed valuations and state support, offer a rare ‘asymmetric’ trade. While US tech prices for AI perfection, China is quietly paying you decent dividends to wait for a long overdue rebound. 

https://substack.com/@capitalistletters/note/c-184124261?r=2kxl2k

AI model costs are collapsing, threatening to turn foundational providers into glorified, low-margin ISPs with giant CapEx bills. The real profit, naturally, will be found downstream in clever application layer businesses. 

https://substack.com/@michaeljburry/note/c-184120488?r=2kxl2k

Michael Burry observes that the S&P 500’s bottom 250 firms are now a measly 8.5% of the index. This extreme concentration, fueled by AI hype, signals correlation is overdue for a return, or as we call it, a proper rout.

https://open.substack.com/pub/marketsentiment/p/the-state-of-ai

The AI boom, despite its calculus roots, is a modern South Sea Bubble. Valuations disregard the prohibitive inference costs, confirming Sir Isaac Newton’s old rule: never underestimate the market’s capacity for irrational madness. 

https://substack.com/@marketsentiment/note/c-183795486?r=2kxl2k

Investors are valuing AI like a high margin software business, ignoring the messy industrial reality. Massive capital expenditure combined with fast depreciation means the unit economics simply fail basic common sense tests. 

In the grand cycle of technological revolutions, AI’s ascent mirrors the Dot.com frenzy: venture capital floods in, valuations soar on promise alone, only for maturation to expose the frail.

Yet does this bubble’s deeper virtue lie in its excess pumping capital to forge breakthroughs that sober times cannot?

For investors, the lesson is eternal: ride the tide with discipline, for optimism, not skepticism, reaps the harvest, even as realism tempers the fall.

History whispers that such manias, painful as they prove, propel humanity forward, linking fleeting hype to enduring progress.

https://open.substack.com/pub/capitalistletters/p/the-inevitable-ai-bubble

The “Productive Bubble” hypothesis (often associated with economist Carlota Perez, that excess pumping capital does really forge breakthroughs) is a standard defense of tech bubbles found in this type of macroanalysis though. So, familiarity with it maybe breeds a certain degree of contempt for its professed certainties.

Echoing Netscape’s 1990s downfall, OpenAI’s dominance in large language models rests precariously on Microsoft’s Azure, Windows, and Office ecosystem.

This reliance invites replication, as the tech giant clones capabilities, suffocating the innovator it once nurtured.

The theme resonates profoundly: in tech’s Darwinian arena, platform control trumps invention, underscoring fragility where symbiosis sours to rivalry.

It links to broader AI narratives, warning that infrastructure lords, not model pioneers, may dictate the future’s architecture.

https://substack.com/@davefriedman/note/c-184853817?r=2kxl2k

However Dave Friedman recently published a full article titled “OpenAI isn’t Netscape,” where he argues the opposite: that the comparison is flawed and OpenAI has leverage that Netscape lacked.

Unless this specific “Note” was a “Devil’s Advocate” thought experiment, this describes the fear the market has, whereas Friedman’s prior analysis likely refutes it.

The reflection that Microsoft will “suffocate the innovator” is the bear case Friedman is likely debunking.

Neolabs emerge as agile crucibles for AI innovation, blending academia’s rigor with venture’s velocity, unburdened by legacy labs’ inertia.

From bio-AI hybrids to scalable inference, these nimble entities promise accelerated discovery, democratizing tools once gated by giants.

Their significance unfolds in decentralizsation’s quiet revolution: fostering diverse, rapid experimentation that counters monopolies, weaving links to open-source ecosystems.

In this shift, we glimpse technology’s true ethos: collective ascent over siloed dominance, heralding a more resilient innovation tapestry.

https://open.substack.com/pub/theaiopportunity/p/the-rise-of-neolabs-where-the-next

*That* (infamous/ prescient/ flawed/ overtaken???) MIT probe (95% AI roll outs fail IRL) into bold AI wagers reveals leaders who staked fortunes not on prediction, but conviction, yielding outsized returns amid volatility.

The study’s core illuminates risk’s alchemy: all in bets, tempered by foresight, transcend gambling to strategic mastery.

Deeper, it probes leadership’s soul, courage as the bridge from uncertainty to legacy, linking personal fortitude to organisational renewal.

In AI’s whirlwind, this underscores a timeless truth: profound change demands vulnerability, forging paths where caution merely observes.

https://open.substack.com/pub/leadershiplighthouse/p/i-went-all-in-on-ai-the-mit-study

The author (Josh Anderson) explicitly validates the high failure rate (“The MIT Study Is Right”) but uses it to frame a narrative about leadership conviction.

The article is not about the failure itself, but the fortitude required to be in the successful minority.

Regarding the “MIT Probe” itself:

Infamous: Yes. The “95% failure” (or similar stats like Gartner’s 85%) is widely cited to scare executives.

Prescient: Yes. It accurately predicted the “Pilot Purgatory” we see today, where companies build chat demos but fail to deploy them to production due to reliability/governance issues.

Flawed: Arguably. Critics say these studies count “experiments that taught us something” as “failures,” ignoring the value of iteration.

Overtaken: No. It is still highly relevant as companies struggle to move from “Chat” to “Autonomous Agent.”

Ruben Hassid often writes about the tactical decline of “lazy” AI usage (e.g., like ChatGPT is declining because users need better prompting or different models”).

In this vein, amidst apparent AI euphoria, “Declining” charts a sobering arc: hype’s peak masks eroding returns on compute, talent saturation, and regulatory headwinds curbing exponential gains.

Ruben dissects this inflection, revealing not collapse but maturation where unbridled growth yields to sustainable scaling.

Technology’s promise hinges on equilibrium, linking overinvestment’s perils to wiser allocation.

It invites reflection on cycles’ universality, urging investors to discern signal from noise in the machine’s hum.

https://open.substack.com/pub/ruben/p/declining

This is Dave Friedman’s bread and butter. He consistently argues that energy (Power) and Capex (Capital) are the hard constraints of AI.

This dispatch unravels AI’s voracious appetite for power and capital, tracing grid strains and billion dollar data center bets that redefine geopolitics.

Friedman’s lens exposes the nexus: energy scarcity as innovation’s silent governor, channeling flows from speculative froth to infrastructural steel.

Deeper themes emerge in asymmetry vast rewards for the resourced few, echoing industrial dawns.

It binds to global stakes, where capital’s might not only fuels models but redraws power maps, demanding stewardship beyond profit.

https://open.substack.com/pub/davefriedman/p/sunday-ai-dispatch-power-capital

The “first attempt constraint” unveils innovation’s hidden toll: nascent tech’s raw edges demand iterative refinement, stalling adoption until friction fades.

Anslemperera probes this barrier, from buggy interfaces to unscalable proofs, illuminating why breakthroughs languish in prototype purgatory.

Its significance lies in humility’s mandate, embracing failure as scaffold, not setback; linking to AI’s maturation rite.

In this, we confront creation’s essence: progress as patient forging, where constraint births resilience, transforming vision into viable reality.

https://open.substack.com/pub/shanakaanslemperera/p/the-first-attempt-constraint

Azeem Azhar’s dispatch dissects 100 trillion tokens’ odyssey, unveiling the “glass slipper” dynamic: models lock loyalty by solving intractable pains first, not leading benchmarks.

Amid superhuman persuasion’s ethical shadows and Waymo’s moral crossroads, it grapples with AI’s dual blade: efficiency’s boon laced with manipulation’s risk.

Deeper, it links fragmenting chips to extinction’s receding specter, affirming exponentials’ layered dance.

Here, significance crystallises: technology reshapes not just tools, but trust, urging vigilant navigation of abundance’s unforeseen contours.

https://open.substack.com/pub/exponentialview/p/ev-553

A nice summary of the importance of the above linked Anslemperera authored, 1st shot only, AI alignment piece:

https://substack.com/@shanakaanslemperera/note/c-185125721?r=62vrvp

1000+ to 1 ratio of data centre to alignment spend. Good job it’s only the future survival of the human species in the balance….