Cycles

People worry about AI replacing jobs that have automation risk: - Investment banking - Equity research - Consulting - Marketing, etc.... I have two rebuttals: 1) If the junior bankers, researchers, consultants, designers, etc. all get replaced, who's gonna lead the teams/businesses in the future and replace current senior leadership? 2) Current enterprise AI usage is not sustainable. We are in a subsidized adoption phase that won't last for the next 3 years and enterprise usage will have to settle Hopefully this isn't just self-serving bias lol

BREAKING: MICROSOFT JUST ANNOUNCED TO BAN ITS OWN ENGINEERS FROM USING AI DUE TO THE COST OF USING IT. VP OF NVIDIA SAID, “THE COST OF AI FOR MY TEAM WAS MORE THAN HUMANS” “AI CAN COST MORE THAN HUMAN WORKERS NOW”

Sure, but I don't think it's that simple of a model: 1) Gains in consumer surplus and producer surplus aren't mutually exclusive. They can both simultaneously improve. Labs can invest in new wrappers, that allows them to reach a new market (Claude Code enabled a ton of coders, Claude Design enabled a ton of designers). 2) More wrappers/use-cases/applications might even make customer retention stickier, and might also decrease elasticity to demand (i.e. I'd rather pay for a model that has a few wrappers I really find productive than one that doesn't but gives slightly better chat responses)

@justacycle Wrapper aren't cheap in token subsidized economy which is a large factor. The average consumer surplus would increase per token in wrapper but the cost for producer would still increase linearly. So producer surplus decrease which disincentivized better wrapper

@ArchegosIntern Thanks bro

@justacycle good stuff here I've never really thought of it as this flywheel where each generation arrives faster, costs less, and helps build the next one even quicker Also, never thought of it as: someone invents a new "wrapper," and then model improvements carry you up that S-curve until gains plateau... then you need a new wrapper to start the next S-curve

Here’s some math: Users > Say ~1.3B people use AI monthly Paying individuals > ~40M pay for a sub (~3% of users) > Avg ~$24/mo → $12B/yr retail Total lab revenue > OpenAI $25B + Anthropic $30B + others $20B = $75B > Retail is just 16% of that Infra spend in 2026 > ~$600B (hyperscalers + sovereigns) The gap > Retail subs cover 2% of infra > All lab revenue covers 12% > The other 88% = VC, debt, hyperscaler balance sheets, sovereign money TLDR: You’re favorite chatbot is gonna run you multiples on wha you’re paying now in a few months

Anthropic, OpenAI, and Microsoft have all moved enterprise customers toward token-based billing , the clearest signal that the flat-rate subsidy model is ending. Analysts broadly see 2026 as the inflection point where free and heavily discounted access shifts toward paywalls, enterprise contracts, and profitability. Your Claude, ChatGPT, and Gemini queries are still being subsidized at probably 2 to 5x the price you actually pay, infra is being financed by a mix of hyperscaler FCF, debt markets, and federal contracts. In some time we will look back to today and be in awe that we were able to access these models for $15/mo with no ads. They’re only so available now to drive adoption.

@threelinecap Yes, thank you!

@justacycle Good read and I agree on the pending performance plateue of newer models. Draws a similar comparison to the iPhone problem; do I really need to upgrade every year? and in this case every month for LLMs

Who specifically do you mean by “they”? Every enterprise & individual consumer of AI has a different use case, so the question of “where should the next dollar go to increase productivity the most?” has a different answer for everyone. In the piece, my goal was to convey that if the labs spent more on the wrappers (Anthropic has done a great job of this: Dispatch, Claude Code, Tools & Artifacts), productivity for the average consumer would rise a lot more than spending to go from Opus 4.7 to Opus 4.8. Of course that is a weak clam without any math to support it, but this was more of a thought exercise.

@justacycle Have you considered they want agents in of themself as economics producers without humans in the loop ? Obviously having scaffolds can help humans fully use models capabilities but they want fully autonomous agents though

In the mid-2000s, BlackBerry owned mobile communication. Physical keyboard, push email, BBM, government-grade security When the iPhone launched in 2007, BlackBerry’s leadership dismissed it as a consumer toy with no keyboard and poor battery life. They moved slowly, assumed enterprise loyalty was a moat, and kept iterating on the existing form factor instead of rethinking the platform > Just because your technology is replaceable, doesn’t mean it will be replaced > You only get commoditized if you don’t move fast enough This applies to companies, technologies, even people Great video: youtu.be/LF3aUIM57uw?si…

This is the misconception w SpaceX ~70% of revenue and all the margin comes from Starlink, not launches The launches are the flywheel for Starlink that get the Starlink satellites out there, and help other connectivity companies reach space by providing (relatively) cheap launch services In the same way that fiber optic and undersea cables enable cross-Earth communication, Musk is initiating that same buildout for future cross-planet comms

Well here’s a use case of $TXN and $MP’s products I didn’t originally think of wsj.com/tech/personal-…

Only concern is that China owns access to ~90% of rare earth/magnet market. If they choose to undercut and wipe out all international competitors, the government can realistically do whatever they want and start buying from China. What gives me confidence, though, is that Trump is in office, and will support US manufacturing and domestic production until the day he dies

I feel like $MP and the broader rare earths sector ($UUUU $USAR $CRML $LAC) was the shit on X a couple months ago, but we've since forgotten about them - today, $MP is up 10% $MP operates Mountain Pass, the only rare earth mine and processing facility of scale in the Western Hemisphere. They mine, refine, and are now manufacturing finished neodymium magnets... the permanent magnets that go inside every EV motor, every robot actuator, every drone, every satellite reaction wheel, and every guided missile If Physical AI is the next wave, rare earth magnets are quite literally what make it move. The same magnets power $TSLA's Optimus joints, stabilize $SPCX satellites in orbit, and sit inside the motors of autonomous vehicles. MP is vertically integrating from mine to magnet, with its Independence facility in Texas set to begin magnet production in H2 2026 and a $1.25B 10X expansion on deck - positioning them as the only end-to-end domestic rare earth magnet supplier in the country What makes the setup even more compelling is who's backing them. The U.S. Department of Defense took a $400M equity stake last year, making the Pentagon MP's largest shareholder at 15%. That deal came with a $150M low-interest loan, a 10-year price floor on NdPr oxide at $110/kg (nearly 2x Chinese market prices), and a commitment to purchase 100% of the magnets from the new 10X facility for a decade. Days later, Apple signed a $500M supply agreement with a $200M prepayment to source U.S.-made rare earth magnets for iPhones and EVs. When the DoD and Apple are both writing checks to secure your supply chain, that tells you everything about how strategic this asset is

@grok explain to me how SpaceX is worth ~$2tr. I understand the degree of specialization, and state-of-the-art infrastructure, but how do they plan to make money and where is current revenue coming from? The government?

I pitched $VRT almost 2 years ago in September of 2024. The thesis was very straightforward: chips exert heat, chips need to be cooled, data centers becoming more dense, more efficient methods of heat dissipation needed, so find the most financially stable water cooling company (Vertiv was the only notable player). I find that the best picks and shovels plays always have a logic chain that clean. What could be an equivalent chain for physical AI? Robots move. Movement needs torque. Torque needs motors. Motors need NdFeB permanent magnets. 90% of global magnet supply sits in China. And nobody is modeling robotics as a demand category in their supply forecasts yet. $MP is the only scaled Western player with a mine to magnet supply chain. Same simple chain logic

CRWV still owes BX ~$3bn from their 2024 pre-IPO financing. Another example of the circularity of investment and competition amongst the companies at the frontier, now involving shadow banks.

@MTSlive @DratchCap @ArchegosIntern

We asked @Dratchcap why Leopold Aschenbrenner would own puts on AI stocks if his whole fund is built around the AI boom. He says the puts look more like risk management than a bearish call. “I think it’s just smart risk management.” “He’s got these big positions that he wants to stay in.” “If you have puts that protect you on the way down... that’s sort of a synthetic equivalent of buying the dip.” “I don’t think he’s negative NVIDIA.” “The better trade has been to move down the risk curve from the biggest and the best into things that are a little bit smaller in market cap.”

@ArchegosIntern @leopoldasch Didn’t even think about that lol

@justacycle @leopoldasch But also as a counter to myself and adding to your point lol x.com/mtslive/status…

These were his positions for Q2 2026 I don’t think people understand how badly @leopoldasch got cooked - these positions did not age well over the last 3 months

@ArchegosIntern @leopoldasch Absolutely. But we also don’t know the expiries. He could totally be betting on an event that he didn’t expect to happen this past quarter. But you’re absolutely right

“Multi-year horizon” doesn’t apply to puts. To my knowledge, that’s not how options work. Puts expire, and theta eats your premium every single day whether your thesis is right or not. Leo’s $8.45B in semi puts got destroyed by the macro: $SMH ($2.04B) shorted the entire sector while TSM, NVDA, AVGO, and AMD all ripped; $NVDA ($1.57B) printed $68B in Q4 revenue (+73% YoY) and guided $78B next quarter; $AVGO ($1.01B) posted AI revenue up 106% YoY and is guiding $100B+ in AI chips by 2027; $AMD ($969M) hit an all-time high of $469 in May, up 66% in a single month; and $TSM ($535M) was up 30% YTD with Q1 revenue +35% and $52 - 56B in capex committed. From my interpretation, every single underlying moved violently against him And then the Greeks compounded the damage at every turn. Delta meant each rally pushed his puts deeper out of the money, accelerating losses. Gamma meant the probability of recovery kept shrinking as stocks moved further away. Theta on $8.45B notional translates to hundreds of millions in premium erosion per quarter (the clock never stops). And as markets rallied, implied volatility compressed, crushing put values through vega on top of everything else. This is the fundamental flaw in the “multi-year horizon” defense: long equity lets you wait out a drawdown, long puts put you in a race against a clock you can’t pause Look, his thesis might eventually be right. Semis are cyclical and these valuations are stretched, but being right in 2027 doesn’t save you if your puts expired worthless in 2026