Tom Demaecker

@Midnight_Captl @PythiaR @TMTLongShort 2.25T: Out of which budget, sir?

@PythiaR @TMTLongShort CY26: $700B CY27: $1T CY28: $1.4T CY29: $1.8T CY30: $2.25T

Maybe I've spent too much time in Claude, but pulling hyperscaler capex ests for N5Y I feel like they are materially too low. Have seen enough sell side models that are so lazily done (capex as a simple % of sales, zero buildups, not thinking about rev drivers etc).

@pernasresearch According to inferencemax of @SemiAnalysis_ the improvement was even 65-100x, thanks to algorithmic improvements and more memory. Because for trillion parameter models, inferencing is memory bound and 72 cluster is more memory efficient than just 9x 8 cluster.

@pernasresearch Blackwell was 9x better due to more GPUs per rack. 2x better due to 2 dies in 1 chip, and 2x better due to number representation FP8 to FP4, for a total of 36x. Except for number representation, which is coming to an end (we come from FP64), improvement per watt was very low

Why is there only 2x improvement from Blackwell to Rubin when there was 35x improvement from Hopper to Blackwell? Is system scaling slowing down that much?

@fabknowledge Positive sum!

I modestly raise you fabricatedknowledge.com/p/gpt-3-and-th…

@laurenbalik In that case, pretty strange they have a 4.8 star review on app store.

Pharma and med device spending behavior. Lack of an alternative. Concern about a potential DTC ad ban looms, so this is marketing to doctors to market to end consumers, not DTC. It's a rickety pass-through of pharma ad spend. 🤷‍♀️ There are multiple ways for $DOCS to fall apart.

@patrickc Pretty misleading, as this only captures retail drugs dispensed at pharmacies. Expensive oncology drugs, biologics, or specialty therapies given under supervision, are included in hospital costs

Remarkable that US spending on pharmaceuticals is just ~1.6% of GDP. Amazingly good value for money considering the welfare gains. (And especially given the extent to which the US subsidizes the rest of the world.)

@MartinShkreli @peter_mantas Pretty misleading, as this only captures retail drugs dispensed at pharmacies. Expensive oncology drugs, biologics, or specialty therapies given under supervision, are included in hospital costs

i've been SAYING

@firstadopter Price per GPU would drop. The more you buy the more you save!

HSBC's Nvidia AI Server pricing assumptions for coming years. $2.6 million going to $8.8 million!

@fabknowledge CRCL shareholders. Apparently stablecoins will expand into cross border advertising

Who is liquidating GOOG nonstop right now?

@SouthernValue95 That means they’re in the hands of retail (or strategic investors like Berkshire). Not sure that’s what you want either

Narrative violation: big cap tech is under-owned (by active managers). This is somewhat structural but is below normal levels. Only $META is an active mger overweight among the Mag6. $MSFT $AMZN $GOOG $AAPL $TSLA are underweights

@TechFundies But might impact brick and mortar dollar stores more

Not sure I see de minimis duty change impacting $PDD Temu / Shein. For example, a shopping excursion for a submersible watermelon pool toy: $AMZN shows the high-priced, original products up top ($25 for 2). Now, it also shows lower-cost versions below (it used to not have as much of this stuff so good for them to broaden selection). Lowest-cost version is $9. $PDD is $3.27. Even adding 30% duty makes it $4.25. Not sure value prop that banged up. Also worth noting Temu pushing more inventory to local warehouses to avoid this change.

@long_equity Any split on where performance comes from? Rerating or better EPS growth?

Over the last 15 years, quality investing has crushed value investing. • The MSCI's quality index has an average outperformance of 2.71%, compared to -2.32% for their value index. • Quality has outperformed in 11/15 years, while value only managed 5/1 • Quality delivered the lowest volatility (SD=4.36%). Critics may argue that 15 years is a short time frame, but the consistency and lower volatility speaks for itself.

@fundam_inv @fabknowledge If only it was only about the tools… and not about the thousands of materials PhDs

Just like how BA went up on a horrible print, INTC going up on a horrible print tells you that anyone who could sell, sold lmao

@BLyons151 Hi Brad, can I still sign up?

@convequity But depending on the location of the datacenter the latency could be a limiting factor, if they can’t peer at colo’s?

Fortinet $FTNT: If FTNT pursues its proprietary data center strategy it will cut out the gross margin of the colocation providers. As $EQIX's gross margin is ~48%, FTNT can save 48 cents on the dollar by running its proprietary PoPs and make it far more competitive versus SASE vendors running on colo DCs and those running on hyperscaler infrastructure (e.g., $PANW on GCP ( $GOOGL)). For example, if FTNT's ASICs save 10x in compute/energy costs and the proprietary PoPs save 3x in hyperscaler IaaS (assuming 66% hyperscaler gross margin), then all else being equal, we calculate FTNT's raw COGS would be 10x less than PANW's. This is based on an estimation that the compute/energy component of total COGS is 70% and the remaining 30% is made up of networking and storage. Then, we add together 3x and 7x (10x * 70%) to arrive at a 10x reduction in total COGS versus PANW. This will allow FTNT to further increase its competitiveness and increase cash flow and shareholder returns simultaneously.

@PrescientCapInc @StockSavvyShay 😂

@StockSavvyShay Nope. Sentinel One $S will steal Crowdstrike $CRWD clients soon

I'm going to 'speak this into existence' -- $CRWD ($75B MC) will acquire $ZS ($27B MC) to officially cement themselves as the go-to all-in-one cybersecurity platform for enterprises 🤩

@techfund1 Makes more sense :) thanks

@DemaeckerTom These are servers, not GPUs

Projections on the AI server market from Dell Oro, a 30% CAGR looks to me very doable: - inferencing will be the far larger market compared to training, I'm estimating 5x to 10x larger - training infrastructure should upgrade to the latest $NVDA GPUs, you want to be able to retrain models as fast as possible to give you an edge - probably also inferencing infrastructure will upgrade to the latest NVDA GPUs - if you can cut the time down it takes to answer a query by more than a second, it would be compelling to make the upgrade. Users will flock to the generalized AI which produces the best and fastests results So, $NVDA's ASPs are likely to rise at a 20% CAGR and we get a 30% unit growth rate or so, while consensus is only looking for 20% top line growth next year and we have the shares trading at a very reasonable 36x forward PE. Shortages in CoWoS could remain problematic for this bull thesis, but overall, still think the upside is looking interesting here.

@fabknowledge Amazing note, man. Kudo's!

Jensen's keynote highlights something universally true in computing since the beginning: it's much cheaper to simulate reality than to observe it. AI is the final rendition. The whole keynote is just permutations of that truth. Read more here: fabricatedknowledge.com/p/jensens-worl…

@Nick09348805 Actually the real length is 24nm+ It’s been a while since node names don’t track the actual size anymore.

Cutting edge chips have transistor gate lengths measuring just 3 nanometers. For comparison, this is ~1 million times smaller than the size of a typical garden ant. And only 6x wider than an atom. A human hair is 70,000 nm wide. Our fingernails grow about 1 nanometer every second