Pinned Tweet
I was early in and right about Nvidia / AI.
I was early in and (hopefully) will be right about IonQ / quantum computing.
Sensei retweeted
Si vous êtes investi dans l’IA, vous voulez peut-être vous asseoir avant de lire ça
Hier a été une journée absolument horrible pour les actifs de croissance
L’IA change de phase; quelques explications :
1. Depuis le début des années 2010, le cash-flow des hyperscalers suffisait à financer le capex
Désormais, il faut emprunter pour construire et deux nouvelles à contre-courant sont tombées cette semaine :
A/ SpaceX a publié des résultats énormes et rentabilise ses datacenters en un an environ
B/ $META, à l’inverse, doit diluer ses actionnaires pour continuer à financer son capex, ce qui montre très clairement que le cash-flow ne suffit plus.
2. Ce changement se heurte à deux forces
A/ L’argent cesse d’être bon marché
L'argent pour le buildout c'est le nerf de la guerre, l'argent coute cher au moment précis où l’on en a besoin. Le 10 ans US est à environ 5,1 %, la Fed est en cycle de hausse et environ 70 pb de hausses supplémentaires sont pricées sur 12 mois. Quiconque a étudié l’histoire sait que construire à crédit pendant que le coût de l’emprunt monte a mené, à maintes reprises, à une période classique de fragilité
B/ Les revenus de l’IA font douter
Anthropic ajoutait environ 8 Md$ de revenus annualisés par mois au premier semestre, contre 2 à 3 Md$ en août et septembre. Beaucoup comptent sur les agents IA pour prendre le relais, mais côté grand public, Muse ne retient qu’environ 22 % de ses utilisateurs à 14 jours, contre environ 27 % il me semble pour ChatGPT
Le marché vend donc l’incertitude née du décalage entre un capex certain, financé de plus en plus cher, et une monétisation encore floue. Ce décalage devra se résorber d’une manière ou d’une autre
Je vous encourage donc à étudier 2000 de près, car les différences sont aussi instructives que les ressemblances
En 2000, on finançait à crédit une demande qui n’existait pas encore, et seulement 2 % de l'infrastructure était « allumé ». Aujourd’hui, la capacité est vendue avant d’être construite, les capex sont backés par des contrats solides, et ce qui manque, ce sont les puces et l’électricité, PAS les clients.
L'infrastructure IA n’est pas la fibre, par contre je peux le concéder, le parallèle tient sur le point le plus important en top down, le bilan car si les revenus ne suivent pas la dette, ça casse, et c’est ce que ce bon vieux marché a testé hier
Je suis tenté d’y voir un énorme discount sur les actions liées à l’IA qui connaissent leur première vraie phase de repli, mais j’ai aussi besoin d’être rassuré par les earnings et par les midterms. $GOOGL publiera mercredi 28, le jour de la Fed, et le S-1 d’Anthropic semble déjà pricé
Plus d'update à venir
$NVDA $AMZN
The AI skeptics taking a victory lap over Nvidia GPUs sitting in Amazon warehouses are either clueless or missing the point.
These GPUs aren’t idle because demand collapsed. Amazon is deliberately stockpiling billions in GB300 racks early because every hyperscaler is compute-starved. Amazon can’t hold off on buying until the supply crunch eases. If it does, it falls behind on the buildout while competitors lock in capacity.
The racks are staged in warehouses mainly because sites lack the power to turn them on. That is the simple answer. If you dig deeper, the bottleneck goes well beyond power. You can’t actually light up and fully utilize the systems w/o aligned relief across power, memory ($MU, $SKHY), and optics/networking ($LITE, $ALAB, $GLW, $COHR, etc.). Progress on one layer alone doesn’t unlock them.
This is also old news. Almost a year ago, Satya already said Microsoft had chips sitting in inventory that it couldn't plug in because of power and infrastructure limits. The “dark GPU” narrative keeps getting recycled.
Sensei retweeted
$IONQ NVIDIA just set a record in fault-tolerant quantum computing. And it was developed in the very center where IonQ's Superion 256 is headed.
On Monday, NVIDIA announced the largest comprehensive simulation of a fault-tolerant quantum computing system to date: 104 qubits, run on 8,192 GPUs on JUPITER, Europe's first exascale supercomputer. It validates magic state cultivation, an approach that requires 90% fewer qubits than previous methods.
The work did not start in Germany. An NVIDIA GB200 NVL72 system at the NVIDIA Accelerated Quantum Research Center (NVAQC) served as the development platform, scaling tests from four GPUs to hundreds before moving to JUPITER.
That is the same place IonQ is going. In September, IonQ announced it will be the first on-premise quantum computing deployment at NVAQC: a Superion 256 system directly linked with an NVIDIA GB200 NVL72 via NVQLink, with workloads orchestrated by CUDA-Q. Installation is scheduled for next year.
Jensen Huang: "Useful quantum computing will require fantastically complex, fault-tolerant systems - and understanding how they behave at scale is essential to building them."
NVIDIA is building the classical side of fault tolerance. IonQ will be the quantum computer sitting right next to it.
nvidianews.nvidia.com/news/n…
#IonQ #NVIDIA #QuantumComputing
$NVDA just ran a record 104-qubit fault-tolerant quantum simulation using 8,192 GPUs at Europe's JUPITER supercomputer validating a technique that could require 90% fewer physical qubits.
This comes as $GOOGL advances same approach on actual quantum hardware and Anthropic uses AI to design more efficient quantum circuits bringing hardware development, AI and classical simulation closer together.
Sensei retweeted
Everyone's watching GPUs. Jensen's watching substrate.
"Substrates are getting larger and there are more chips... it's going to be hard to close that gap for a few years."
Morgan Stanley's Oct 14 ABF note: undersupply hits 9.3% in CY27 and 13.7% in CY28, prices rise 25%+, and new capacity doesn't arrive until 2028 to 2030.
Aunty Lisa took her trips to Taiwan too late. The real bottleneck is the thing the chips sit on anon
$NVDA $AMD
"Sight unseen, who can you place a $50, $100 billion order on? We're literally the only one." - Jensen
Hyperscalers order racks a year or two before they see them. That's the bar Helios has to clear.
I was at Moscone for AMD Advancing AI in July. Helios looked credible on paper, but it's AMD's first rack-scale system and still unproven at volume manufacturing. My call was meaningful volume in 2H27, two quarters past AMD's own window.
Meanwhile AMD is getting squeezed from the other side by the custom inference silicon showing up everywhere.
Nobody writes a $50B check on a slide.
newsletter.bepresearch.com/p…
$NVDA $AMD
$TSLA $SPCX
It is looking increasingly likely that the merger announcement will happen soon. As anticipated, SpaceX will absorb Tesla in an all-stock transaction and emerge as the surviving entity. Based on my analysis, the base case conversion rate could be 3 SpaceX shares for each Tesla share owned, implying a 20–25% premium.
I am obviously rooting for a more favorable conversion rate, but for anyone who thinks a 3 to 1 conversion rate will never get approved, think again. The merger only needs a majority vote to pass. Elon alone owns close to 20% of all Tesla shares while all retail investors combined make up less than 30%. Most loyal Tesla investors who likely hold more shares than regular investors will approve the merger regardless, mainly due to timing-before any uncertainties heading into the next presidential term.
Let’s see if Elon meant what he said to retail investors: “I won’t let you down.”
Sensei retweeted
The S&P 500 is now in bear-market territory..
..the valuation, that is. P/E has contracted from 23.2x a year ago to 18.7x today.. .that's just about a 20% de-rating @dualityresearch
Multiple is the cheapest since Liberation Day April 2025
Sensei retweeted
NVDA, AVGO, MU — BofA's Arya: chip stocks already trade like post crash Cisco
+ BofA analyst Vivek Arya says Nvidia, Broadcom and Micron are already priced like Cisco after the dot com crash. "Investors are paying post crash multiples for a cycle that hasn't crashed," he wrote in a note dated Friday, per Benzinga.
+ By his math the three trade at about 13 times forward earnings on average, inside the 10 to 20 times range Cisco sat in after the bubble burst. The catch: those multiples only hold if earnings forecasts do. Lower estimates would make the stocks look less cheap at the same prices.
+ The note landed with chip stocks staring at a weekly drop of about 10%. As of Friday morning SOXX was down 9.8% for the week after eight straight daily declines, with Micron off 12.5%, Broadcom 7.1% and Nvidia 5.4%. Memory names slid after CXMT said it plans server memory on a smaller cell design by year end, and Wednesday's hot inflation print added fuel.
The irony is the crash may never come. TrendForce projects Q4 DRAM contract prices up 10 to 15% and NAND up 15 to 20%, which is the earnings backdrop that would keep those 13x multiples honest. If memory pricing holds through Q4, the "post crash" label is just a discount on growth.
Sensei retweeted
$MU $SKHY I think Samsung, SK Hynix, Micron & CXMT will all be making a lot of DRAM in 2028, not just China.
The difference: the memory trio will be shipping HBM5, DDR6 and next gen 3D DRAM architectures (2029+) into AI data centers. CXMT is still generations behind, selling mostly into China and mostly to consumers.
And not everyone will buy Chinese DRAM. Just like Americans aren’t buying Chinese EVs. 🙃
Section 301. 50% tariff on Chinese chips today. Another one scheduled for June 2027. CXMT already on the Pentagon’s list.
































