Leopold Aschenbrenner’s memory-heavy, leveraged AI-infrastructure book blew up, but the underlying thesis on memory, Neo Clouds, and AI capex is still being pressed by SK Hynix, SanDisk, Google, Amazon, and a broader stack trade into power, payments, and photonics. The cleanest money is now argued to sit lower in the stack: memory sold out through 2027, AI payments emerging, and power constraints forcing nuclear, solar, and even jet-engine data-center schemes. Across disclosed 13Fs, the loudest consensus is Alphabet/Google, with Amazon, TSMC, SpaceX, and then power/energy and memory as the next crowded winners. Berkshire’s new Greg Abel era is presented as an aggressive buyer of GOOGL and other names, while several large allocators hedge with QQQ puts or sit in payments and infrastructure.
Leopold Aschenbrenner’s top disclosed positions were 28.5% in SNDK and 28% in MU; together they were over half the book, with BE also at 9.5%, and leverage is what blew the portfolio up.
SK Hynix posted its best quarterly earnings ever, made more in one quarter than in the prior full year, then the stock fell about 20% the next week on the view that memory investors were too levered and that NVIDIA could no longer keep buying memory at the top.
SNDK’s investing day is used to argue the “GPU buyers-at-a-top” thesis is wrong: major investors doubled down at the lows, and its new high-bandwidth flash is already backlogged into end-2027 for inference.
Memory demand is described as unlike anything before, with the thesis still “very strong” and still “very alive and well,” even if Leopold was too over-levered.
A Bloomberg rumor put Leopold’s open equity position at roughly $500 million to $1 billion, with most of it still in an ANTH stake for now; a new “Moratorium” sheet was said to arise after a liquidation.
Greg Abel’s Berkshire has been deploying capital at high rates, with the quarter’s biggest add about $17 billion in GOOGL; it also added 44% stakes in DELL, MCD, and LEN.
GOOGL is framed as the only company owning the full AI stack: models, custom chips, distribution via Search/Gmail/G Suite/Android to hundreds of millions daily, plus infrastructure, power, and capex.
GOOGL is said to be on track to spend around $250 billion by year-end, with a data point claiming each $1 invested since inception produced about $2.8 in return; Berkshire’s buy is tied to that capital-allocation record.
The next layer of AI money is argued to flow into capex winners: the video says the two biggest winners this quarter are AMZN and GOOGL, with AMZN’s AWS AI capex around $200 billion this year and proof that revenue and margins are expanding.
Jensen Huang’s position is described as a $30 billion stake in INTC, after a $5 billion investment midway last year for about 5%; INTC is being bought for bleeding-edge CPUs to orchestrate hundreds of thousands of GPUs and handle AI agent tool calls efficiently.
Brad Gerstner’s top holdings are said to be NVDA at about $1.9 billion, then CERE, with META, TSM, CRWV, and ARM also in the book; Gavin Baker’s biggest position is SPXX? no — SpaceX at $4.7 billion, plus MU, CERE, ALAB, CIEN, and CRED.
Gavin Baker also has $2.3 billion of QQQ puts, framed as insurance that lets the portfolio survive a market drawdown; a similar downside hedge is said to exist in some Leopold-style portfolios.
Photonics and power are presented as the next buildout: GPUs need fast long-distance interconnects, copper becomes inefficient at scale, and light-based photonics solves it; COHR is a major manufacturer in that theme and is up 85% YTD, while GEV supplies materials/components that latch onto GPUs.
The money trail is said to continue into lower layers: power/energy, payments, and memory; Bill Ackman is “mostly in payments” with V, MA, SPGI, and NFLX, while Stripe bought Open Router and Ramp is building its own router.
Seth Klarman is said to run about $3.4 billion and have AMZN as his #1 holding, while Terry Smith is said to have dumped V and GOOGL; across the fund consensus chart, the most dots are on GOOGL, then AMZN, TSM, and SpaceX.
The hard constraint behind the AI buildout is now physical: insufficient plumbing and chip fabs, all memory supply for next year sold out, not enough power resources for future GPUs, and a scramble toward nuclear, solar, and even jet-engine-powered data centers as Elon Musk reportedly pursues by buying a company for that approach.
The final framing is that Leopold’s blowup may have been leverage rather than a peak in demand, but over the next couple of months the market still has to prove whether memory is simply over-owned or whether atoms and capacity limits keep the thesis intact.