The easy money in AI infrastructure has been made. Ask the Koreans.

For the last three years, the investment framework was almost embarrassingly simple:

Find the bottleneck. GPUs. HBM. CoWoS. Networking. Optics. Power. Memory. Jensen’s five-layer cake with rum and whipped cream.

Demand exceeded supply. Capital followed scarcity. Scarcity earned the rents.

The hyperscalers and frontier labs were (and still are) willing to spend big to build ahead of the adoption curve (same as every historical analog). Initially with operating cash. Increasingly with debt, equity and ever more creative financing structures.

For a while, the only serious question was whether anyone would earn an adequate return on the capital being feverishly deployed. This largely manifested via widening credit spreads of the ‘spenders’ (i.e. the hyperscalers + neoclouds who tapped the debt markets with abandon) and via relatively prosaic equity performance. Encouragingly though, recent cloud growth has made that question less urgent. Not for Michael Burry, admittedly.

Yet, a harder question continues to nag. To sustain this euphoric kumbaya, investors require a bridge from carte blanche token consumption to economically useful (and monetizable) intelligence.

In my last post: ‘Are the Kids Alright?’, I made the distinction between capability and intelligence; and while the former doesn’t need much of the latter (so long as we mean human-like or mammalian when we say ‘intelligence’), there are deficiencies in reliably executing long-horizon tasks, however banal they might be. We can debate endlessly how to fix those deficiencies with better context, but we cannot ‘harness-wash’ (I am trademarking that) the truth that the models need to get better to seriously penetrate the $5-8 trillion Knowledge economy in the US.

My base case remains that the vast, vast majority of inference tokens will ultimately be serviced by open-weight (and/or distilled) models, small models and closed-weight non-SOTA models (many of which the frontier labs will offer). Incommensurately, a tiny minority of tokens shall be serviced by the frontier SOTA models. Said differently, token economics will likely become extremely Pareto distributed.

Chart below is just an illustration. The y-axis is not how intelligence will be denominated in the future, but it makes the point. That the frontier labs will service a fraction (< 5%?) of tokens with their SOTA models and capture >80% of the revenue.

For the next leg of the cycle, I think the entire AI capital cycle concerns can increasingly be reduced to three macro questions:

1. Can model deployment itself become the next scaling axis?

2. If someone crosses that continual-learning threshold first, can anyone catch them?

3. Won’t better intelligence ultimately need less hardware?

The first question flirts with how models can get significantly better given Chinchilla scaling limits. Current models may appear to have reached asymptotic competence, but they possess little native intuition for state management across time (and task adherence over long-horizons), especially in less highly-verifiable domains. Do we need a breakthrough?

The second question pokes at: what is the point of being the frontier when Chinese models have reduced the half-life of SOTA down to months? Why even bother with the investment and heartache?

The third question asks why hardware bottlenecks would exist in a world where models learn continually and hyper-optimize architecture? Won’t the thing we’re pining for kill the bottleneck trade?

Let’s explore each one in the simplest terms. As always, I am a taker of feedback, and as usual I question more than I am capable of answering.

1. Can deployment itself become the next scaling axis?

Today’s models learn principally before they work. Pre-train. Post-train. Freeze the weights. Deploy. (Alas, much of the post-training and fine-tuning today bottlenecks on human expert annotations).

Everything after deployment — RAG, retrieved memory, tool use, long context, agentic scaffolding, inference-time compute — gives the model access to more information or more computation. But the underlying parameter space remains largely static. It feels like the model can adapt but it doesn’t really learn from that experience, at all.

That distinction may become the next scaling frontier. The post-Chinchilla era has already pushed capability gains away from simply scaling pre-training FLOPs. Instead: model improvements have come from better post-training. RLVR, inference-time search, test-time compute, and tool use. Longer horizons through better harnesses.

All of these are, in different ways, attempts to extract more intelligence from a static set of weights. However, continual learning would collapse the distinction between training and deployment.

Conceptually here’s how it could work: the model generates tokens; it ‘acts’. Observes the consequence. Updates weights (with some batch cadence of course, not on every interaction – as 99% of interactions are pure noise). It acts again. Deployment itself becomes part of the training loop. Experience stops being ephemeral context and becomes permanent capability.

Alternately, some ML experts frame continual learning as a heavily scaled version of in-context learning and runtime exploration, rather than this ambitious idea of continuous weight updates during deployment. My own sense is they are probably right but in highly containerized, verifiable environments. For a model to seem highly capable in high-entropy / less-verifiable domains, it probably needs to learn from interactions, and convert that trajectory into training data to teach the next iteration.

Further, human experts verifying short-horizon tasks is straightforward. However, deciding which multimillion-dollar frontier run to commit to is high-stakes and difficult to verify without actually running it. Totally unscalable.

The scale of learning through deployment is unfathomably large. So much so that even if only 0.001% of interaction trajectories in deployment (production environment) contain novel, generalizable learning signals, it would still imply 1000x speed of current evolution.

At that point, distribution is no longer merely a monetization advantage. It becomes a potential learning-rate advantage.

But do we need different architecture to get there? Modern AI infrastructure has been optimized around a relatively clean separation between training and inference. Of course, GPUs and TPUs can perform both inference and backpropagation. But they don’t continuously interleave inference, memory, weight updates and serving at production scale without destroying latency, stability or economics.

That may require much deeper algorithmic and hardware co-design. Not necessarily a successor to the GPU (that sounds too dramatic). Potentially something more subtle but highly consequential. To achieve decoupled (or asynchronous) feedback loops? I don’t have the engineering chops to suggest an elegant solution but on first principles I know that inference is largely memory-capacity bound. While backpropagation & optimization are compute and state-capacity bound. Merging both on the same accelerator would lead to some gnarly outcomes.

And herein lies the paradox. Before machines can recursively improve their own architecture, human ingenuity may first need to come up with the design of the physical substrate capable of recursive improvement.

2. If someone crosses that threshold first, can anyone catch them?

Today, frontier intelligence has surprisingly little half-life. SOTA arrives. It looks discontinuously better. A few months later, someone else is close. A few more months later, someone else is SOTA. Distillation accelerates the process. Architectural innovations diffuse. Post-training recipes diffuse. Synthetic outputs themselves become training data for competitors.

So, does being first only matter if you can pull the ladder up behind you? A deployed model begins accumulating proprietary experience. More deployment creates more trajectories. More trajectories improve the model. A better model attracts more users and agents. Those agents generate more experience. And so the flywheel goes. This is the recursive self-improvement argument in its bullish form – i.e. a small lead compounds.

However, the learning has to remain proprietary. If the breakthrough is primarily architectural, it will diffuse to competitors. The supply chain is too vast and complex to keep it secret. In a scenario where a small oligopoly crosses the rubicon within months of each other, distribution becomes the moat. Because the moat would be accumulated experience.

That distinction determines whether continual learning creates a genuine intelligence moat or simply accelerates another round of commoditization.

Also, there may not be one ladder, there may be many. The winner-takes-most argument depends heavily on there being a single recursive path, which is far from obvious.

3. Won’t better intelligence ultimately need much less hardware?

Today we compensate for static weights with enormous amounts of machinery. We are, in a sense, substituting compute and memory movement for persistent learning. If a model could actually internalize experience, much of that architecture might become unnecessary.

Why retrieve the same knowledge repeatedly if the model has learned it? Why carry enormous context windows if the system can compress experience into its weights or another persistent memory substrate? Why spend 100 units of inference-time compute searching over solutions if the model has already internalized the strategy from millions of prior attempts?

The relevant metric therefore may not remain cost per token. The more important denominator becomes total cost of compute per unit of economically useful intelligence. Under the conventional Jevons argument, lower costs stimulate enough demand that aggregate infrastructure demand continues rising. But it is possible that recursively improved model architecture, memory systems, compilers, interconnect and silicon collectively generate 10^x improvement in useful intelligence per unit of hardware. If something becomes a multi-year bottleneck, AI designs around that constraint. No?

That outcome is entirely possible.

So is the opposite. Continual learning itself might be extraordinarily compute-intensive. Agents may generate effectively unlimited demand for inference. Scientific discovery, software production, robotics and simulation may absorb every unit of compute made available. We may truly go back to the original and final moat – compute (fwiw – Google will win this).

We simply do not know which path. All I know is that in this next leg of the cycle, hardware efficiency won’t be exogenous to the intelligence being produced. The intelligence being produced can begin attacking the cost of producing intelligence.

The bottleneck trade has always been about how much infrastructure intelligence required. The next cycle may be about how little.

My view where that cocktail shakes out: That super-exponential demand likely overwhelms exponential upgrades in hardware efficiency.

The distributions are so very wide.

The Last Ladder

Lately, I’ve been venturing down Lewis Carroll-inspired rabbit holes, wondering which ‘futures’ lie ahead of us, as a society, if ‘AI’ (the whole soup of it) continues to march exponentially forward. As a dad, I reflexively think about likely states of the world that await my kids and how best to calibrate my approach to guiding them. I am certain I am not alone in saying that the distribution of possibilities feels uncomfortably wide. If you’re under 30 years old and reading this – No, the 90s didn’t feel like this.

Maybe because the AI oligarchy’s liquidity releases loom large (read: IPOs allowing VC’s to finally distribute cash), I have decided to burrow deeper. Or perhaps it was my wife lamenting that the song stuck in my head that I can’t stop humming, from a Janis Joplin-meets-Brittany Howard-meets-Amy Winehouse sounding artist called IngaRose, is in fact, AI. Or maybe it was noticing the latent despair inside two recent epistles: one from the moral and intelligent Rahm Emmanuel, offering a grand bargain with Universities that felt utterly antiquated given what’s ahead for higher education. The other: a recent paper by the Anthropic Institute suggesting we’re almost at the ‘recursive self-improvement’ precipice, that felt utterly Orwellian.

Come to think of it, what really persuaded me was the despondency in the rampant booing of college commencement speakers recently on the mere mention of “AI.” Whatever the motivation, this note is an attempt to build some causal connections between what AI is inevitably transforming and our potential ‘futures’ (stealing the term from Peter Frase).

Every prior technological transition had a cognitive higher ground to retreat to: farmer → factory → office. The escape valve for the low-productivity-class was always “move up the skill ladder.” But AI automates the ladder itself. There is no higher ground when the thing being displaced is general cognition.

You don’t need Intelligence for Capability

Progress in AI capability continues to surprise me (a reasonably heavy adopter). This next-token prediction business was meant to be a stand-in for intelligence, and not intelligence itself. Yet somehow that student who aces every possible multiple-choice practice test is getting exponentially and scarily close to doing original research, with the help of some tutoring, of course – from Mr. Reinforcement Learning (RL) and Ms. Fine Tuning (FT) to help it align better to what we humans (verifiably) want.

Now here is the thing about an exponential, which is a word we have all agreed to use casually the way we say literally, meaning we have stopped feeling it: it does not announce the bend. Frankly, the progress has been dizzying. METR’s task-time horizons for AI models double every few months — the machine that did four-minute jobs last summer now does twelve-hour ones, and will, if the line holds, do twelve-day ones by the time you finish shaking your martini — and at no point does the curve send a courtesy text warning ‘arriving now’. ‘SOTA’ (state of the art) is simply another Tuesday. And meanwhile, distillation steals the big model’s soul and pours it into a small cheap one.

It is genuinely hard to keep up with all the frontiers, the use-cases, and the domains being conquered each week.

We know that with models, as with all primates, capabilities arrive in jumps. Skills ‘switch on’ in our prefrontal cortex at thresholds rather than fading in gradually; think of it like heating ice – nothing happens from -20 to -1 deg C, and suddenly near / at 0 deg C the hydrogen bonds holding the crystal lattice snap and the molecules collapse into a denser fluid: water. Same with AI models: intelligence increases (read: loss function / error rate falls) > passes some threshold (different for each domain) > capability (“skill”) emerges.

This (now well-understood) flywheel of scaling compute > eventually more capability > task automation > displacement gets pretty dystopian once you graduate from textual modality and introduce autonomous machines / robots (or “embodied AI”) that demonstrate capability in the physical world.

I’ve always been skeptical of intelligence emerging in the real world using the existing model architecture. Because real world data has materially higher entropy and lacks the structure inherent in all human-created content (text, audio, video) available on the internet that LLMs exploit so well. This is also why solving autonomous driving is so much harder than coding agents (the literal two ends of entropy and verifiability).

Though once again, my skepticism seems misplaced. I am not precisely sure why I am wrong, but it appears I am. These vision-language-action models (VLAs) used by autonomous robots are literally just vision-language models (VLM) fine-tuned so that, given an image of its surroundings plus a text instruction, the model directly outputs (at a very high frequency) low-level robot commands that actuate dexterous actions. Picking up a slippery deformable object in a messy kitchen demands orders of magnitude more complexity and computation than re-writing Hamlet in the style of Tupac. Watch videos of Optimus Gen3 or Figure 03 on YouTube and you’ll realize how close they are to mimetic learning and impressive fine-motor skills.

Here I am, someone who always assumed that causal reasoning — something categorically different from next-token prediction — was the hill I’d die on. Real cognition, I held, requires a model of causes, not a glorified correlation engine, and the math seemed to be on my side: give a correlation machine 98% per-step accuracy, hand it a fifty-step, multi-week task, and 0.98^50 leaves you a 36% chance of success. The errors compound precisely because the thing has no idea why any step works. That’s a deliberate oversimplification, of course — I’m ignoring the scaffolding that bolts verification and error correction onto each step. But the deeper point stands: human-like intelligence is the capacity to few-shot generalize and to reason outside the distribution, the exact opposite of what an LLM or VLM does. And yet every month — most recently watching Mythos and GPT-5.5 — I’m forced to swallow the same humbling thing: you don’t need AGI, or human-like intelligence, or causal anything, to displace the bulk of human cognitive work. My priors are both correct and irrelevant. I’m right about what intelligence actually is. I’m wrong that most human work was ever using it.

Capital Becomes Labor

Anthropic told us that 80% of their code base has now been authored by Claude. Further, they “expect that embodied intelligence (i.e., robotics) might quickly follow recursive intelligence, and follow a similar path of increasing returns at decreasing cost.”

The combination of digital AI and embodied AI will disrupt and, depending on the task, displace both white-collar cognitive work and blue-collar labor at a velocity faster than society’s ability to recalibrate and adapt (which, eventually, it will).

This isn’t a controversial statement. And it isn’t novel either. Notwithstanding it being a PR problem for the hegemonic frontier labs, I believe it is a near-incontrovertible conclusion. I am not even discussing alignment. I assume the self-replicating embodied AI economy will be fully aligned with humans (a separate, worthy debate for another time).

The bigger point is that since the dawn of our species, Capital has needed Labor (and vice versa). Now Capital is Labor. That’s an upper cut that knocks out both Adam Smith and Karl Marx. So, what becomes of the Proletariat?

At this point, I am going to fast forward down the rabbit hole. The middle steps are worthy of debate another time – for example, ‘human in the loop’ which will be the dominant model for cognitive displacement for the next few years will mostly affect the younger, inexperienced workers – the twenty somethings. They haven’t acquired good judgement because they haven’t made or observed enough mistakes yet. Their work isn’t novel cognition, rather its domain specific pattern matching. Same with autonomous machine capability – few will care when it’s mostly contained in industrial warehouses and factories. Their come-to-Jesus moment occurs when humanoid machines inhabit our homes.

Skipping turns and going directly to GO, let’s reason some ‘end states’ for the proletariat.

Capital may have fired its workers, but it cannot fire its customers — a machine that produces everything and sells to no one is just an expensive way to make heat. Something has to manufacture demand, and the only relevant question is the plumbing: do you hand down a flow (income) or a stake (ownership)?

Another axis is power. Embodiment retires the proletariat’s last lever — you cannot storm a Bastille defended by an autonomous drone swarm. So the question becomes whether the social contract gets rewritten while the mob still has a vote? This isn’t prophesizing, it’s deduction.

Picture society is an A380 careening through the fog in descent. Where do our messianic pilots – Dario, Sam, Demis, and Elon – intend to land the plane?

Three Landing Strips

#1: Rentism

Everything reproducible is free (or marginally at the cost of energy to produce it) and everything that matters is not. The robots print abundance and cannot, it turns out, print Malibu or Manhattan, or a Nurse who knows you and sits and holds your hand, or a seat at the table where the table is, or John Mayer playing the blues live. AI cannot print status. You get a check. A check that is the fruit of a negative income tax system or something more daft only a politician can devise like universal basic income. Whatever the legislative mechanism, that check buys a magnificent television and not the room to put it in. This is not poverty. It is something newer and more humiliating: provision without standing. This is where abundance and hierarchy meet. The proletariat doesn’t take over because the aristocracy have private robot armies.

#2: The Shareholder Republic

Capital, having automated its workers, notices with mild alarm that it has also automated its customers, and that a billionaire cannot personally eat 8,000 salt-laden Big Macs. Demand, and not labor, was load-bearing. Therefore, ownership gets handed down — basic capital, not just basic income — because the cabal in Silicon Valley and Beijing who own the robots need someone to sell to, and the only fix that survives the arithmetic is making everyone a tiny owner. Income can only buy the cheap basket; ownership buys the scarce one. An income post AI singularity is a merely claim on reproducible goods — the basket deflating to zero — so the check buys ever more TVs and never the house. The majority realize it is a grant, handed down by someone who has stopped needing you (labor, white or blue) and finished building the means to ignore you.

But why capital? Surely income suffices? Because Income can be taken away by political whims, and the sheep have no lever. Ownership can’t be cut without expropriation, a higher bar. Capital redistribution arrives not because the Bernie bros won the argument but because the spreadsheet demanded it. Here the mob has a vote before everything gets re-written. Nobody is happy. This is the good ending.

#3: Mind the Gap

A title, a deed, or a tokenized-share of an AI factory is ultimately just paper; guns can always expropriate. #2 above is open only if institutions don’t collapse. In 2018, on a plane back to London from New York I read Walter Scheidel’s The Great Leveler. An engrossing yet upsetting read. He summarized that the only forces that ever actually compressed inequality were mass-mobilization war, revolution, state collapse, and catastrophic plague. Three of the four run on mass human violence. Retire the need for human bodies and you retire the only leveling engine that has ever worked. This landing strip is more the movie Snowpiercer than Elysium.

The layoffs (unnoticeable at first, then all at once) land years before the abundance does. The cognitive economy gets distilled by 2029; the cheap-robot dividend meant to pay for it shows up in 2034-ish. In between is a flat dark stretch of runway full of twenty-six-year-olds who were told to learn to code, then told to learn to weld, and are now being told, gently that they are nearly out of comparative advantage to burn.

The Brace Position

So — are the kids alright? The honest answer is that the distribution is too wide for me to say, which is itself the truest sentence in this entire note. But “I don’t know” is not the same as “I can’t act,” and here the day job is useful. What capability does human intelligence (running on 20 watts, not 20,000) have that AI, in its current form and trajectory, will not? That is a question deserving of more than one sentence, but I’ll stop at: Epistemic Self-Awareness (i.e. meta-cognition), Agency, Taste, and Generalization (from first principles). All of them, sans one (Taste), is teachable.

The path ahead for risk assets is a function of SoH reopening. And SoH reopening is a simple function of ‘intent to reopen’ x ‘speed of reopening’ x ‘extent of reopening’. All narratives (and there are plenty on both sides) have to be distilled down to what it means for flow through the strait.

Several ways to frame what's ahead. Like many, I have no particular edge in assigning subjective probabilities. But frameworks help you trade, and this is mine.

The Scenarios

A) Escalation: Boots on the Ground. Full kinetic escalation, ground deployment (amphibious and/or parachuted). Very bad outcome (protracted closure), and very unlikely — the hostage rescue operation told you everything about this administration’s sensitivity to American casualties. I assume < 10% probability.

B) Escalation: Critical Infrastructure attacked. US strikes Iranian civilian infrastructure; Iran retaliates against GCC oil (UAE, Saudi), gas (Qatar), and desalination (Saudi) infrastructure. This is the scenario where the conflict metastasizes beyond the Strait. Very bad. I assume <15% probability.

C) Kumbayah. A deal that includes SoH reopening conditions, sanctions relief, and a sizeable OPEC quota increase for Iran (they want this more than enriched Uranium). Russia and Israel unhappy (but US will tell Bibi to desist); China happy the US is out of the Gulf. Good to very good outcome, depending on the speed and conditions of reopening. I doubt this is the ‘next’ outcome, but I assign < 20% probability — however, conditional on Scenario B occurring first, the probability of C as the subsequent resolution rises to < 30%.

D) Full Taco. Trump abdicates the GCC protection commitment. SoH reopens under Iranian control — effectively a toll system. This is the base case. I assume >50% probability.

The Problem With Full TACO

D is my modal outcome (for now). But, D contains a wide distribution of outcomes, and the variable that matters is shut-in volume:

D.1 — Manageable (2–3 mbd shut in). Mildly stagflationary. Risk assets rally, gold rips, rates probably rally too. The world absorbs 2–3 million barrels offline with friction but without structural damage. Asia doesn’t like it. Europe pays up. US fine.

D.2 — Painful (3–4 mbd shut in). Bad. Rates don’t know what to do, gold probably sells off, dollar strengthens, Asia in real trouble, EU in trouble, US relatively insulated. Value over growth, probably. Who knows.

D.3 — Ugly (5+ mbd shut in). This is the scenario where the IRGC toll system effectively stalls Bahrain, Kuwait, Qatar, and the UAE (they hosted US bases > retribution). That’s the ugly number. Go to cash — everything does poorly except the dollar. (Bahrain is 70% Shia, so perhaps I’m wrong about them being on the wrong side of the toll.)

I am optimistic (hopeful) the outcome is D.1 while also believing that's not necessarily the next outcome.

On a more philosophical note: how should one view bartering US hegemony for political survival?

Iranians on local broadcast are calling this — in Arabic — “the final war.” I wonder what happens in Scenario D, not to markets, but to the architecture of American power. The Suez Crisis finished British and French hegemony in 1956. They were never viewed the same way again.

Hope for a quick resolution.

Although open to being convinced otherwise, I believe the idea of the US occupying Kharg island and operating Iran’s oil infrastructure is a terrible one. A naval blockade as an alternative is only marginally less bad.

1. The IRGC is there now. And they’d rather destroy it than hand it over.

Kharg Island receives crude through subsea and overland pipelines from Iran’s major producing fields on the mainland. It handles north of 90% of Iran’s oil exports. If a US amphibious force approaches, the IRGC will likely destroy the infra, the storage tanks, and the jetties before a single MEB sets foot on the coral. Think scorched earth. History is littered with precedent of regimes destroying their own infrastructure rather than ceding control to the enemy.

2. Dynamite and duct tape.

You don’t need ballistics to blow up the infra; just dynamite and duct tape (Saddam style). The easier option is to simply shut in the wells on the mainland. Close the valves. Nothing flows through the pipes to Kharg.

3. Even in the good case scenario, this doesn’t solve anything.

Assume the US somehow occupies Kharg with oil infrastructure intact and Iran, bizarrely, continues sending crude through the pipes. The plan, presumably, would be for the US to allow China and others to load tankers and pay into an escrow account — make it RMB-denominated to keep Beijing happy — that Trump declares is being held “for the Iranian people.” Set aside how fanciful this is. Even in the best case, it doesn’t solve the actual problem.

4. Kharg occupation doesn’t open Hormuz. And probably sends Oil higher.

Nothing in this plan restores the flow of GCC oil through the Strait of Hormuz. It only shuts off Iran’s 1.7 million barrels per day. Now analyze this from a market perspective. Right now, there is some non-zero probability that this de-escalates in weeks — because, frankly, it needs to. Putting boots on the ground at Kharg, with zero change in the arithmetic of oil flowing through Hormuz, signals to the market that this is a longer campaign. That probability of quick resolution decreases. Brent moves higher. Ergo, own goal.

5. China has been passive so far. Kharg may change that calculus.

Beijing has quietly provided satellite imagery and some intelligence support, but has otherwise been passive. They want this over. If Kharg gets destroyed, China probably pays to have it rebuilt — they’re Iran’s primary customer and they think in decades. But the US physically taking over oil shipments to China is a qualitative escalation. China has roughly six months of strategic reserves, gas flowing in from Russia via pipeline, and no domestic political pressure on a timeline. Trump has way fewer reserves and far less time, with political pressures mounting weekly. In a staring contest with China, we will blink first. We’ve already blinked with Iran.

6. Iran (mostly) ran out of ballistic missiles. They did not run out of drones.

Iran’s ballistic inventory is largely spent but the Shaheds keep coming, because Russia is resupplying them. Putin is deliriously happy with Hormuz closed. China isn’t happy, but if the US intervenes at Kharg, the dynamic shifts: Iran won’t ever run out of ballistics either courtesy China.

I understand the motivation. The IRGC holds an asymmetric edge in controlling Hormuz, and Kharg looks like leverage to rebalance that. But seizing it with marines (which looks like it may happen) doesn’t rebalance anything — it prolongs this conflict.

And this isn’t just about headline crude prices. Even if Brent and WTI stall at $110/$100 levels, give it a few more weeks and the downstream effects — diesel, sulfur, fertilizer, jet fuel — are going to be significantly worse. The refining margin squeeze is coming, and it will hit consumers. Even that doesn’t really matter — the pain is through the wealth channel, a.k.a S&P 500.

The off-ramp is really the US walking away (“all military objectives achieved”) with some ceasefire deal that guarantees Hormuz reopens and keeps the cumulative supply deficit under 5 million barrels per day. Trump needs to offer Iran something tangible for this (doesn’t need to share what he offered). He’s caught in an awkward position: he can’t finish the job militarily without sending oil to $200+ and he doesn’t have the political rope to inflict the kind of pain that disrupts 10% of global consumption of Oil. I suspect Trump has already decided he wants to walk away. He just needs guarantee that Hormuz reopens.

Jeff Currie said in a recent interview that this conflict has permanently changed realpolitik no matter how it resolves. He’s right.

Perhaps I’ve missed some 4D chess move, but I don’t see occupying Kharg helping us achieve what’s in America’s interest, which is Hormuz open and quickly.

Disclaimer: These are personal opinions. I have positions across markets long and short. Do your own research. This is not investment advice.

Read more

On Oct 29th I wrote the “Journey from Stupid to ‘Max Stupid’” (link) arguing that fears of an imminent “AI bubble” were premature—that you need the fragility of leverage to pop a bubble, and that disappointment (if it arrived at all) would more likely manifest as a plateau in productive capabilities than as a sudden capex retreat.

Roughly a quarter later, the SPX and NDX are both (roughly) flat. Beneath the surface, the inner concentric circles of the AI trade tell a story of divergent outcomes:
Software has retrenched another 15%.
Memory has gone parabolic (the cyclical is now rebranded, ahem, secular).
SemiCap behemoths have re-rated an entire weight class (AMAT, ASML, and LAM all +35%).
Companies that enable scaling—up, out, and across GPU clusters—have compounded handsomely.
The fabless chip designer “AI darlings” have stalled or faded.
The neo-clouds (CoreWeave & Oracle) are down bigly (-35%).

Bottom line over the last three months: the suppliers to the shovel outperformed the shovel-makers, who outperformed those attempting to dig.

The march to ‘Max Stupid’ continues, albeit unevenly. As I said before, nobody knows what summiting that mountain looks like—but I can try and imagine it:
Intel at $100 on 14A euphoria; Micron hovering at $600 because, you know, “2028” (not 2027) is now peak earnings; SemiCap buzz is all about words like “epic” and “historic” and AMAT flirting with a $15 EPS bogey; and it ain’t late-cycle until Bloom Energy is guiding to $5bn of revenue (imagine even the Jefferies analyst turns bullish!).

It cannot even be close to a bubble until you wake up to Blackstone underwriting a fresh $30bn CoreWeave bond issuance, and Elon redefining physics by launching football-field-sized radiators on Starship to perpendicularly plug into giant solar panels powering xAI’s first orbital data center. NVDA at $180 and Google at $300 is not what bubbles are made of, folks.

Sure, meanwhile we get rolling bubbles—animal spirits hopping from memory to semiCap to interconnect to foundry and perhaps, perhaps even… (closes eyes) …software.

You don’t have to be early. You’ll get a few bites at the apple.

As noted in October, the next endogenous node is the release of the first frontier models fully trained on Blackwells (Grok 5 coming soon). Improvements so far have been mostly along predictable axes—longer context, better tool use, lower inference costs—rather than step-function capability jumps. Even absent “jumps,” another swing factor is whether enterprise adoption evolves from tightly scoped, human-in-the-loop deployments toward something meaningfully more autonomous.

Without that FOMO catalyst, expect more of what the last three months delivered: messy dispersion, zero-sum math (Google takes from Broadcom, AMD eats where Intel slips, blah blah), and a fertile environment for chasing short-term momentum—long and short.

My bias remains the same. To observe (rather than always predict and time) these ‘jumps’ within various pockets and chase the strength or weakness that follows. We’re equally likely to materially de-rate the entire datacenter power-infrastructure (GEV, ENR, CEG, IREN, BE the whole brood) on increased NIMBY-ism around datacenter projects in the US as we are to re-rate semicaps another 50-100% on a renewed demand surge + broadening / redistribution of cowos capacity. Don’t discount this probability either (it could happen easily with the OpenAI narrative flipping – you just don’t know).

Either direction, the market will give you plenty of bites at the apple – you can miss the first 15% turnaround rally in Salesforce (CRM) if the market turns around on some parts of SaaS being a valuable ‘system of context’; there will be 2x more to capture after the first 15%. Heck, you could’ve missed the whole NAND chronic shortage trade 3x times and still made +50% on your fourth try (see SNDK chart).

Broad Macro Check.

The market is broader than software and semis, of course. And on that front, the Trump-Bessent run-it-hot campaign is on time for its pre-midterm crescendo. That said, tax refunds included, I am less buoyant on overall fiscal impulse in 2026 mostly because it is all Q1/Q2 front loaded and then a coin toss after.

While consensus can be dangerous this year (see my summary of where baseline consensus is on macro: here), I wouldn’t lean short beta in the US in H1. We’re going to see more volatility out of the WH with commensurate TACO-ing. I don’t think spending ~120 bps for SPX April upside (say 3% OTM / 35 delta calls) at 12.5 implied volatility is the worst way to rent beta (while you work on which granular pockets underneath you want to size). Again, I believe themes within the AI space will auto-correlate and run further (both directions) than you anticipate. Once it broadens to all parts then we can start discussing how far from ‘Stupid’ we’ve gotten. Early days.

Europe, a quick perspective check post Davos.

A short quip on Europe. While it pains me—as a European living in America—to watch the unnecessary Greenland fiasco unfold ahead of Davos, what really struck me was the subsequent European self-congratulation. The rah-rah around Carney’s speech, followed by headlines like the New York Times declaring that “Europe and China are, in some ways, natural allies,” was… jarring.

As much as I miss living in London, I struggle to see a European renaissance. I respect the pushback this view will get, and yes, the rules-based order is changing—and that change is exposing not just Europe’s past mistakes, but the ones it continues to make. I see very few credible paths out of the contradictions Europe has deliberately engineered for itself. Some perspective (directional—don’t quibble on technicalities):

A) Europe (incl UK) has a structural energy deficit that is insurmountable over the next two decades. The UK and EU produce roughly 10% of the primary hydrocarbons they consume. The US supplies ~70% of UK LNG imports and ~60% of the EU’s. Over two-thirds of EU “renewables” are biomass—largely imported US wood pellets shipped on LNG-powered vessels and labeled “green”—while less than 8% is solar. Europe runs a 35–40 exajoule energy deficit; the US and Canada run a surplus. This dependence on the US for energy is existential (in a ‘new world order’).

B) Layer on demographics and pensions. The UK and EU carry unfunded pay-as-you-go liabilities of roughly 300–400% of GDP, requiring 10–16% of GDP annually just to service. Sustaining primary balance demands growth and higher taxes—yet growth increasingly comes from abroad. UK and EU pension funds hold between 45% (German) to 65% (UK) of US Equity exposure which has (relative to European equity market) has delivered +450% in excess return over last 15 years (FX hedged). Strip out US innovation and term premium, and Europe’s fiscal arithmetic deteriorates rapidly.

C) Trade offers no escape. The UK and EU run a ~$200bn goods surplus with the US (offset by a $50–70bn services deficit). To sustain a current-account surplus (given deficit with China), Europe needs the US consumer. It cannot pivot to deficit-financed domestic investment because the energy constraint caps productive reinvestment.

D) The AI race only highlights this reality – EU accounts for just 15% of Datacenter capacity globally, but remove Ireland and its down to 8%. This is a joke. The relatively recent CRMA (Critical Raw Materials Act) plans to have 40% processing of critical minerals by 2030 – but any of rare earth processing (acid leaching, extraction, etc) violates almost every EU environmental law. This is classic Brussels bureaucracy at its finest.

Ten years from now, Europe will still face a deep energy deficit, marginal participation in AI value creation (notwithstanding the token Mistral and Loveable entries) across sectors (biotech especially), and little vertical integration across autonomous or robotic stacks. It is unlikely to lead in space (with respect to Miura and Isar) once launch costs fall below $500/kg across US players. I want Europe to turn this around. I just don’t see—credibly—how it avoids getting in its own way every time a good idea appears. I hope I turn out to be dead wrong.

I don’t believe I’m alone in this, but: digesting dozens of 4000 word ‘outlooks’ gives me a headache. While I enjoy infrequently reading a credible, long-form deep-dive on a particular sector’s outlook, keeping up with 2000+ word posts on ‘everything 2026’ aches my brain. And now add AI slop (which, if nothing else, reliably increases wordcount) we all truly drown in ‘infobesity’.

Today, is a short post about correlations, what it says about risk concentration, and to keep an open mind about why the path of risk in 2026 may not be ‘more of the same’. I’ll keep it to under a 3-4 minute read.

Let me first set the dinner table, and revisit baseline consensus for 2026. I’ll do it in punchlines. You’re very welcome.

Consensus ‘Baseline’ (as I interpret it) in 60 seconds:

Real GDP growth in the US will be positive but modest; dispersion across forecasters is the narrowest in a decade (!) Everyone independently arrived at the same answer; definitely not ominous.

Inflation glides lower and is a total snooze fest; math is against higher inflation; stagflation addicts remain a lost cause, still doom-scrolling 1970s charts.

Fed to cut twice next year; three cuts more likely than one, and four cuts about as likely as none. The Goldilocks distribution, lovingly curated.

Fiscal impulse turns positive in 2026, but most of it is front-loaded; growth bump early, fiscal vacuum after. Eat dessert first, then maybe salad for dinner.

Jobs will cool, duh. Nobody knows what break-even NFP keeps the UER calm. Hiring freezes > layoffs (for markets). White-collar slack builds quietly (“nothing to see here”).

Productivity is awesome — yay. The rGDP gains usually lag by years, but that’s fine. Meanwhile, just be happy it’s supportive of corporate margins as Rev/Employee ratio keeps climbing.

AI capex continues. 2026 is about “adoption,” not “spend.” Fewer ‘existential’ substack posts about capex, sans Burry – he shall keep drumming.

OpenAI funding issues are misplaced FUD, but even the bulls remain uneasy about the $1.4T (and growing) commitment. Enterprise penetration matters; WAUs growth, less so.

SpaceX to IPO pre-midterms, along with a few other darlings; ECM gets busier. Pitch decks include 10x more adjectives than in ’25. Pension funds (and endowments) to finally get some distributions; that money likely finds its way to public hedge funds, and not back to VC/PE.

Ok, that should’ve saved you 40 hours of reading.

Now, very briefly, let’s look at what assets are saying. All charts below use weekly returns and correlations are measured on a trailing 26 and 52 week basis.

Two noteworthy observations that speak to concentration:

1) Nasdaq vs. US Treasury Long Bond correlation has become even more unstable. It moved from ‘Inflationary Shock’ regime (+ve corr) to ‘Growth Panic’ regime (-ve corr) but it has oscillated (+ve and -ve) more in the last year (if viewed on a trailing 60day basis) than it ever has since 2017. A fragile relationship here suggests more exit velocity on var shocks.

2) Energy has de-coupled from the dollar (usually -ve corr, now +ve). Which, outside of uncommon supply shock regimes, is usually a sign of global demand cooling. But Metals (thanks to ‘AI’) have diverged from this relationship. The Metals complex vs. equal-weighted S&P500 (i.e. SPW) correlation is at 91%ile. Further, correlation across pro-cyclical assets (Metals, Nasdaq, Bitcoin, SPW) is near its highest in 20 years!

Concentration alone doesn’t imply negative expected returns. Concentration and Crowding are two different measures, in my opinion. In the spirit of brevity, here are two dumbed-down definitions:

Think of concentration as the # of net-effective positions your capital is allocated across, or how many positions explain a majority of you P&L variance.

In contrast, think of crowding as the amount of $ deployed, by cohorts of investors with a similar utility function, into assets relative to that asset’s exit liquidity. If investors who behave similarly in response to price moves (likely those with similar duration of capital) are piled into the same assets, that may create crowding.

Counter-intuitively, being long Crowding has positive alpha (especially in certain sectors like Healthcare, less so in, say, Financials), but also regularly has nasty left-tails (the distribution is very non-normal). Concentration doesn’t tell you anything about the distribution of forward returns outside of large var shocks. There’s a lot here to unpack, but that’s for another note/post at another time.

Bottom line — Cross-Asset correlations are telling me that 2026 is unlikely to be as easy a ‘dip buying’ tape as 2025. Fragility in correlations typically means difficulty in portfolio construction. And difficulty in portfolio construction typically means higher velocity of rebalancing. And more rebalancing typically means more volatility surges clustered around the same events / nodes.

Last (quick) observation, take the Nasdaq v. Bitcoin example – despite the underperformance, Bitcoin remains a mainstream risk asset. Far from ‘digital gold’ narratives. Last year, Gold (+64%) and Nasdaq (+20%) exhibited a weekly correlation of -4%. Nasdaq and Bitcoin? +51% corr.

I maintain three beliefs concurrently:

1. The direction of progress in ‘LLM intelligence’ (currently drunk on Reinforcement Learning) is actually moving further away from ‘Human/Animal intelligence’ (i.e. ability to generalize).

2. The use-cases for these “not-very-intelligent-but-useful” models are vast, and their productivity amplification is unhindered by the endless, circular AGI debates.

3. Some innovation will emerge that takes a hard turn from how we build AI models today (i.e., maximizing reward vs. minimizing ‘surprise’).

This note deals with belief #2: That these un-intelligent models are prepping us for a massive, useful Agent Economy.

While the headlines obsess over the datacenter capex required, less ink is spilled about the opex of this future economy. If we imagine a world 3-5 years out where inference dominates >90% of compute cycles, does the math even work?

To find out, I built a first-principles toy model of the inference economy — starting with the ‘Jevons’ demand layer, then the Silicon efficiency layer, and finally the implied Revenue hurdle needed (a.k.a. ‘Productivity breakeven’ number).

I’ve made some seemingly smart and evidently dumb assumptions. I encourage you to tear them apart. I am neither Dylan Patel, nor Sam Altman. Like many of you in the ‘early AI adopter’ cohort, I fall somewhere between that wide spectrum.

Here’s the bottom-up view:

1. The Demand Due: Agents and Video

I begin by dividing the demand vector amongst humans and agents. No embodied AI Robots (yet). Just humans and their digital swarms.

If you can see the images of my tables below, you can also read how I’ve contextualized my assumptions in the comments column. Here are the questions that matter:

· The Ratio: I’ve assumed 10 agents per human (20B agents + 2B humans). You can choose 100:1, but you might need to build 100 Hoover Dams to service that!

· Frequency: Humans ping 10 queries/day on average (median ~ average). Agents ping every 5 minutes on average, constantly (median << average).

· The “Thinking” Tax: DeepSeek R2 showed us that “thinking” tokens (CoT) act as a massive multiplier on compute load. I chose 25x.

· Video: Even if video is just 2% of queries, at 300 tokens per second, it dominates the load. My timid assumption already implies 10x the total hours uploaded daily.

Punchline: Using these inputs, I get a total global demand of 6,000 trillion tokens per day from inference alone.

Table 1: The Demand Layer

2. The Silicon Reality: 30 GW of Inference

To service 6,000 trillion tokens/day, physics gets in the way.

I’ve assumed material hardware efficiency improvements—modeled at a ~50% efficiency gain over current GB200 (FP8) specs and dropping energy costs to 0.27 Joules per token (i.e. ½ the cost today).

With 30% utilization headroom and a 1.15 PUE, I arrive at 30 GW of estimated power.

Now 30 GW might look small compared to the “100 GW” numbers thrown around as if OpenAI were embarking on spice farming in Arrakis. But 30 GW is still more than the total capacity for AI workloads deployed today. It’s roughly 260 TWh of electricity per year—or what the entire United Kingdom consumes—just for inference.

Table 2. The Silicon Reality: 30 GW of Inference

You can dial up or down the Joules/token or Joules/TFLOP assumptions or squeeze the PUE, but ballpark, 25-45 GW is the range where you’re going to land.

3. The Economic Bridge: From Physics to Finance

I start by assuming a blended power cost of $60 per MWh. Change it to whatever angers you least – perhaps you’re a solar disciple, or instead, like me believe in pragmatic inevitability of Nat Gas being the primary fuel source. The bill for 30 GW or 270 TWh of electricity is $16bn per year (in the US, a fraction of that in China and the middle east obviously). The full Opex Layer Cake: Add in leasing ($120/kW/month), maintenance, fiber, and cooling, and I estimate the base cost to run these centers (in aggregate) is ~$65-70B. And then add depreciation (a different discussion for another time).

The Hardware: 30 GW implies ~20 million GPUs (assuming 600k GPUs/GW in a post-GB200 world. Let’s not quibble with this one).

The Rental Check: Assuming a wholesale GPU rental rate of $3.25/hour, the AI Labs (OpenAI, Anthropic, x.AI, etc.) will pay the Cloud Providers roughly $540 Billion a year to service their human + agent customers.

I’ve chosen $3.25/hr. I understand the neoclouds are currently cheaper with H100s, but we’re talking a few years down the line and I assume bulk pricing will hover between $2.5-$4. You can toggle as you wish.

4. The $2T Productivity Breakeven

This $540B is just the “Cost of Goods Sold” for inference. It doesn’t include Pre/Post Training / Tuning, R&D, or the expensive humans building the models.

If we assume Inference is only 50% of the AI Labs’ total costs, their total annual spend is $1.1 Trillion.

Now, the final question: What Gross Margin is defensible? I have to fight my personal bias here (I am bearish on value capture at the model layer). But let’s assume a standard 50% GM (this is a ‘Dario level’ estimate, well below what Sama as alluded ‘eventually’). Truth be told, once model layer capabilities converge, I just think 50% is too high.

At 50% gross margins, the “Model Layer” needs to generate $2.2 Trillion in annual revenue to justify the build-out. Said differently, the amount of productivity gains at the services layer to justify consuming multi-modal LLMs *starts* at the $2T number in a few years.

Is $2+ Trillion too optimistic?

It sounded irresponsible at first given a 3-5y timeline (after all, we’re pretty much at zero right now, none of this has shown up in economic barometers yet, its all just stock market EPS for now, i.e. revenue vs. depreciation cycle accounting).

But if you view $2T in the context of 10% of US’ Services GDP (3% of Global Services GDP) and that much of these gains will occur inside large ad business (Google, Meta, others), it doesn’t appear outside the distribution of probable outcomes. This exercise oversimplifies the revenue model; after all, the labs themselves will ‘subsume’ much of the enterprise services layer as it exists today. Nevertheless the incremental value has to be created in aggregate, so my assumptions very much hold up.

Many factors not considered above can swing assumptions materially (vertical integration - see x.AI; neoclouds margin compression; larger hardware efficiency gains; NIMBYism in DC buildouts that slows deployment, CoWos capacity shock, et al).

Depending on the extent of my errors, whether the actual number is $1.5T or $4T, I do feel differently about the agentic economy. I am still bearish on the model layer value capture thesis, but this exercise leaves me more open minded about the productivity S curve in the future. These models won’t necessarily become more ‘intelligent’ as they inevitably become more ‘capable’ (two aren’t synonymous – see 1^st belief above). Until, of course, there’s a step change in approach (see belief #3 above).

I’m a taker of feedback, as always.

A few “macro” AI narratives have recently converged, intersecting in a peculiar way that is causing a “quiet panic” among public and private market investors. If I abstract these narratives to their essence, we are looking at a three-part structural shift:

A) OpenAI’s model dominance has eroded, signaling slower ChatGPT hyper-adoption. Slower growth requires a bigger discount factor applied to their massive “unfunded capex” commitments. This triggers a cascading impact on the entire incestuous ecosystem of vendors and cloud providers.

B) Google’s vertical integration is a cost advantage hiding in plain sight. While their hardware cost advantage (TPUs) is well-known, their data advantage is under-appreciated, and their distribution advantage is only now blooming as they finally pull their head out of the sand.

C) NVDA’s gross margins at its scale resemble that of a temporary monopoly and the inference explosion ahead of us removes CUDA’s moat as workloads emerge that don’t require it. Ergo, the ‘spot’ looks amazing, the ‘forward’ appears suspect.

There are intuitive causal relationships here. For example, if Narrative A (commoditization) becomes true, Narrative B (Google’s advantage) is amplified.

Here is the thing about vertical integration: nobody cares about it—there are plenty of insanely profitable businesses that aren’t vertically integrated—until the product becomes a commodity. That is Narrative A posits. Whatever “jagged edge” of intelligence LLMs ultimately exhibit, if those lines converge with low dispersion, then she who has a 40-60% cost advantage (e.g., Google’s TPUv7 “Ironwood” vs. Nvidia’s Rubin) wins the war.

In this context, OpenAI is a price-taker. If such a re-pricing occurs, nobody in the AI ecosystem—sans Google—escapes without a flesh wound (see image):

The $100 Billion “Napkin Math”

While momentum is not a moat, distribution often is. Before we write ‘OAI: AI’s NetScape’ or ‘AI’s Myspace’ obituaries, let’s attempt some napkin math (high level) on Sama’s claim that they get to $100b in Revs by 2028 (or “maybe sooner”). Don’t nitpick the assumptions, we can take a wide bid-ask on them as the idea is just to ballpark it.

Think of the end-state, for a moment, as individual consumers + enterprise seats + API/usage fees + Advertising-tier. Let’s handicap the first three with some very bullish assumptions.

First table: Maps each region by non-PPP adjusted incomes (obviously), addressable population with generous assumptions on paid conversions and wallet spend per region. Assumes by 2028/29, OAI has 2.2 billion users globally with ~150 million paying. Est. revenue range: $25-30bn (this is bull case).

Second table: Maps white collar work with generous penetration rates (higher than MSFT365 fwiw). OAI captres 200 million seats (MSFT365 has ~340 million globally). It doesn’t keep all the licensing fees so there’s a wide bid-ask here ($24-$38 billion).

Third table: API/Dev Usage based model – one you can take the biggest issue with as median vs average usage is skewed by concentration of customers. Est Rev here was $4-10billion.

All together that’s a napkin-math estimate of $75billion ARR in 2028/29, with inarguably generous (and erroneous?) assumptions. That’s an impressive bluesky number but still deficient of the $100bn target and ambitions of a $1trn IPO.

FWIW, this category breakdown (above) goes against the ‘one comprehensive AI subscription’ statement Sama made; but I don’t know what the heck that even means when its consumer + enterprise. Co-Pilot has not been an inspiration.

Ad-monetization is… complicated

Which leaves the Ad-monetization business. (I am going to ignore Johnny Ive product biz which isn’t going to be ready until mid-2027 anyway).

You would think hiring Fidji Simo was a clear signal they are heading in this direction. But this journey is non-trivial. Unlike search, there is a scaling marginal cost per query. Furthermore, they face a lethal decision on ad targeting: Contextual vs. Personalized.

The revenue model is going to have to be genuinely novel. This isn’t a simple shift from CPC to Price-per-Token. Speaking in trading terms, the Loss-Given-Default of users perceiving ad optimization to be embedded in the core model’s loss function is catastrophic. If people think the core model is nudging answers for ad yield, trust collapses.

Also, is the share of input tokens with commercial intent similar to that in keyword Search or structurally lower? I don’t know, and I don’t think anyone does (yet). My bet is that Google comes up (slowly) with the initial Ad-monetization model and OpenAI and others adopt it. They have the most to lose (‘blue links’) and the infrastructure to experiment at scale.

Ultimately, if model performance does converge (as we surely march towards a world where >90% of compute is inference) then a hardware TCO advantage in the 30-50% range is a death sentence. And the fallout of a slow-motion implosion of OpenAI is very difficult, in my experience, for the market to efficiently handicap & price in a matter of weeks. Even the obvious casualties—Microsoft, for instance, if OpenAI is currently driving >50% of Azure’s incremental revenue growth—are difficult to taxonomize. Microsoft also owns OpenAI’s IP, and I’ve yet to hear a serious take on what that’s worth in a downside scenario beyond the knee-jerk “they’ll just internalize it.”

Mathematically, if Narrative B (Google’s vertically integrated advantage) fully realizes, it forces investors to confront Narrative C (Nvidia’s forward earnings power). De-rating Nvidia on a gross-margin compression story isn’t complicated, even if it is slightly ironic—Jensen just reiterated that they expect to maintain mid-70s gross margins into 2026/27.

Notwithstanding, despite beating insane expectations, the market has de-rated NVDA ~30% over the last 3 months. Make of that what you will doomers.

The NVDA gross margins ‘F.U.D’

My bias is to take the other side of Narrative C, for the moment. This is not about broad price reductions for GPUs and ASICs. Imagine for a second NVDA had to drop prices (ludicrous in this environment of undersupply) – what do you think AMD would have to do? What do you think that does to Hyperscalers pursuing their own silicon – the ‘FUD-du jour’?

I am likely more persuaded by the NVDA bear thesis premised on narrative A than one premised on gross-margin compression / price cuts. And specifically for the AI-tourists, note: many of the Hyperscalers (NVDA’s customers) already have their own silicon. Amazon’s recommendation engines and Alexa don’t run on NVDA’s GPUs. Neither does MSFT’s internal application layer intelligence. Obviously GCP hosts NVDA GPUs but none of Google’s internal ecosystems use them. This is all obvious, yet it seems lost on many people in my conversations. Tesla’s AI6 is nearest silicon competitor to imminently make an appearance and that’s going to compete with other ASICs (including TPUs) and not NVDA (!)

Verdict

Three fulcrums: whether OpenAI can sustain enough x-segment distribution and pricing power to justify its unfunded capex; whether Google’s integrated stack hardens into a durable TCO moat as models converge; and whether Nvidia’s monopoly on training gives way to a broader, ASIC-heavy, more deflationary world of inference. My bias, for now, is that the market is over-pricing near-term fears on GPUs and under-pricing how messy the transition will be if model commoditization actually arrives. That path is unlikely to be smooth— and while it is fashionable to price long-duration structural shifts in short spasmodic windows – rest assured it will be noisy and in that very noise the best risk-adjusted bets in the AI stack are likely to appear.

Both are unserious questions — not because levels of AI optimism aren’t worthy of inquiry, but because benchmarking to 1999/2000, and the framing itself, is intellectually lazy.

The Internet wasn’t some monolithic bubble. The real blow-up of 2000 came from Telcos, not tech (this is conveniently forgotten). Debt-soaked carriers tearing up the earth to lay fiber, collapsing under leverage, not exuberance. Pets.com didn’t tank the economy — WorldCom did. Though, I’ll admit, the former running SuperBowl ads 30 days before it all went ‘poof’ is a sexier story.

Directionally speaking, with usual caveats, when hyper-speculation is equity-financed, it fizzles. When it’s debt-financed, it detonates. That distinction matters when reasoning by analogy: 1999 vs 2005.

I remember reading Alasdair Nairn’s ‘Engines That Move Markets’ in college. His dual-S-curve idea is helpful here as a toy-mental-model: capital inflows surge well before adoption S-curves catch up (i.e. fast->faster->collapse vs. slow->fast->plateau). The danger lies in the gap (which was the implicit premise of David Cahn’s “AIs $600B Question” paper in summer’24). It is obvious that this prisoner’s dilemma of capital intensity has capital running way ahead of adoption (Jensen’s $500b guide tonight implies $1.25T (!) capex on datacenters alone) therein creating an extraordinary time horizon mis-match, similar to the fiber-optic binge of 2000.

But that’s where the similarities end. The vast majority of 5-5.5% of nGDP that got lit on fire in the late 90s was borrowed (i.e debt). This AI capex cycle, also expected to approach 5% of nGDP, has – so far – been primarily funded through cash (read: retained earnings + operating cash flow) and equity. Rough estimates: hyperscalers ~80%, venture & other private capital ~10%, sovereign funds ~10%. Sure, in the last 6 months we’ve witnessed some off-balance sheet shenanigans and new players step in (a la “the neo-clouds”) through debt-financed investments (Oracle the most notable spender & borrower). Notwithstanding, whenever you wish to start the clock for this AI cycle, the point remains that most of the money that’s been spent has been unlevered.

Now, most of the money that’s been committed-to-be-spent in the next 36-48 months might not be. And that’s where the journey to ‘Max Stupid’ must begin, else there won’t be a ‘pop’ – Bubbles don’t pop without leverage. They need a self-liquidating mechanism. Otherwise, it’s a fizzle.

I define ‘Max Stupid’ as the Mt. Everest peak atop a convex price curve. You can consider base-camp to be some cocktail of: ‘this feels stupid, this is a bubble, clearly euphoric overspending, how can this turd trade at 40x, this is equivalent to funding an Apollo program every ten months, c’mon 100% of GDP growth was AI capex, all the PE/VC checks are concentrated in a few players, did you even read the MIT study?, Circular deals.. etc’

After 20 years in the markets, I believe the most money is made in the journey from ‘Stupid’ to ‘Max Stupid’. In the very least it is the most exciting phase. Markets rarely listen to Robert Downey Jr’s advice in Tropic Thunder. Once there’s sufficient momentum, prices will summit Max Stupid.

And each speculative cycle summits differently and benchmarking to 1999/2000 is self-defeating (and a little silly) because anything deemed short of those levels of hysteria and irrationality will somehow be treated as simply enroute an even higher peak – dangerous. This infrastructure overinvestment cycle can peak well before, especially if there’s insufficient leverage and disappointment comes, not from a sudden stop in capex intentions but, from something more fundamental – which is the more likely scenario in my mind.

Hyperscalers and neo-clouds most likely won’t pare back from spending in the near future. But other revelations might test our faith – such as model disappointment: For ex, what if Grok 5, the first one trained on Blackwells, isn’t a ‘wow’ moment? That could bring the marginal buyers’ attention to what the leaders of AI, the ones that’ve been at this the longest (Hinton, Karpathy, LeCun, Demis), have been recently suggesting – that new architectures are required for breakthrough to AGI. No matter how ‘jagged edged’ AI capabilities may turn out, as Josh Wolfe’s team at Lux put it: “The frontier of agentic AI isn’t defined by model size, but between what we can verify and what we can’t.” More GPUs don’t bridge that gap — cognition does. Until then, it’s more jagged-y edge-y with sufficient slop in the midst.

Until such disappointment surfaces, if it does at all, the rise to Max Stupid is certain. The FED has chosen to run it hot. Trump-Bessent duo wants mr. market pumped (certainly until close to the mid-terms), and deteriorating jobs only gives FOMC more rope to cut (cherry-topped with a new and likely more dovish Chair, coming in May). I won’t know precisely when we’ve summited, but I am pretty sure we’re far away right now.

So whenever I get asked by fellow parents to share my perspective on this I usually point them towards data and countless studies on this topic – especially this one from Sapien Labs circa 2023, because it spans 28,000 young adults across 41 countries, has a strong dose-response gradient, low temporal anchoring, and measures across dimensions rather than single outcomes.

Whatever verifiable data you choose to consume on the subject (if you are so inclined to value data over narratives), my punchline is that I believe this is ‘settled science’. Arguing whether TikTok and Instagram damage your twelve-year-old’s dopamine pathways is like arguing whether cigarettes damage their lungs. The correlation–causation quibbles are identical, but the evidence has reached similar levels of undeniability.

That said, what this means for when parents choose to introduce smartphones is a personal choice influenced by several other factors (many uncontrollable). And all of us, including the ‘Wait until Eighth’ brigade, should hold their judgments lightly.

My own stance: arm parents with data and then be empathetic towards their response function. If you find yourself scowling at an eleven year old face deep in their phone, take a moment to steelman the other side of the debate – that confounds are real (family dynamics / older siblings), that benefits while contextual still exist (community, information, media literacy), that skepticism due to our long history of moral panic is understandable (video games < > violence, rock music < > delinquency, internet < > dystopia). But weigh it all up, and the evidence is clear: the harms dominate. Nevertheless, I understand the varied responses to the same cocktail of facts.

More recently, I’ve come to another realization. That I, along with the vast majority of highly-educated and thoughtful parents in my community, underappreciate how short-form video (SFV) algorithms broadly work (even at a high level). I am not surrounded by luddites, nor disciples of Aaron Stupple who endorse abdicating agency. I think most of us have a working model that, in plain speak, goes something like this: As you swipe videos, like and dislike content, the algorithm learns your preferences and calibrates, through machine learning, to feed you what it knows you like (‘echo chamberism’). And this can cause a chronic lack of focus among other things.

That mental model while directionally correct is incomplete in a material way. Over the years, I’ve gone down rabbit holes with software engineers from Bytedance and Google (and one from Meta) that once worked inside, or adjacent to, the core tech that drives content recommendation.

SFV algorithms used to infer interests (and still largely do) through behavioral signals (watch time / rewatch rates) and interaction patterns (likes, shares, comments et al.). But since 2021, biometric inference is (opaquely) a major input. TikTok, to its credit, reveals a lot about tracking and data collection including voiceprints and fingerprints (scary), including knowing how you captioned a video you shared (through iMessage or WhatsApp for example). But they still conceal (and in some cases deny) much around vision / gaze detection / eye tracking (they never admitted so in a settlement when explicitly accused of facial recognition to determine age x ethnicity). I’ve been personally told it knows when you linger vs. when you press but swipe with delay vs. when you swipe rapidly. And where you touch the screen (right thumb vs left thumb) and how tired you are (touch dynamics).

This is where parental intuition often falls short. We know the tech is addictive, but we don’t grasp how invasive and adaptive it has become. These systems are designed to hack directly into a child’s prefrontal cortex—suppressing inhibitory control, overstimulating reward circuits, and reshaping neural pathways. MRI and spectroscopy studies show reduced grey matter in impulse-control regions from chronic SFV exposure—patterns disturbingly similar to substance abuse. The average teen already spends 2.5 hours a day inside these loops.

Wherever parents land on the smartphone introduction debate, my quiet plea would be this: keep them away from ‘Shorts’ (whatever the platform) for as looong as possible (yes, well past 14!). Everything else (Insta, FB, toxic ‘group chats’ etc) is trivial in comparison. I hope parents viscerally absorb this and then do the best they can given whatever circumstances prevail.

My humble tuppence for the moment. Good luck.

Dad of two precious kids with a wicked sense of humor. Husband to my own rock of Gibraltar.

Here I occasionally write about markets, industry, politics, and parenting. None of it should be taken seriously. I am, if nothing else, forever curious.