Señor Pete

[Señor Pete @Pete_Analysis]

INTC trades like it is geopolitically strategic with US gov’t backing, has the potential to become a legit competitor to TSMC over time, and is experiencing strong AI demand from end markets with both more competitive differentiation and less risk of a pricing cycle. MU trades like current DRAM prices are unsustainable and will roll over within the next couple years. These may both be wrong but are not really contradictory

[Señor Pete @Pete_Analysis]

The model progress up to this point has been incredible, so I agree this is possible. That said it is quite the leap of faith to gamble our entire economy on. ROI looks good for spenders and lenders now, but we will see how it looks as we add 20+ GW per year with higher throughput hardware & software efficiencies, massively increasing supply. The main issues I have with some investors in the space is they seem to fail to consider scenarios where models get better, adoption keeps growing, but there is still a ton of malinvestment as it simply takes time for enterprises to figure out what use cases they are willing to spend trillions in aggregate on, especially given the most obvious opportunity is labor substitution which will bring with it significant potential political issues

[Señor Pete @Pete_Analysis]

@gregory_FTA They all are, MU and SKHY are just more cyclical because they are in a massive shortage driven pricing cycle on top of a capex cycle

[Señor Pete @Pete_Analysis]

I mean customers are worried about getting allocation today, so they give better SCA terms. Not sure it has much signal for where price will be two years from now, nor am I sure that customers willing to bid prices up in a commodity shortage have ever been good predictors of future prices

[Señor Pete @Pete_Analysis]

@HyperTechInvest You should calculate operating income at the new gross margin holding COGS constant, not revenues/ASP, as the ASPs are what will come down to meet the price floor. It is a lot uglier than you think

[Señor Pete @Pete_Analysis]

Everyone who is behind in the race is unsurprisingly ok with distillation. Regardless of morality, hard to see how distillation is better for US competitiveness if it prevents frontier labs from capturing profits that will justify exponentially larger training rounds and more exploratory research

[Señor Pete @Pete_Analysis]

@P_Remarks MU said LTAs have GM floors above past cycle peaks. Even if the LTAs are honored and floors are in 70% range, EPS gets crushed if those floors come into play. Feel like people haven’t done the math on what a drop from 85% to 70% GMs means for ASPs when holding cost constant

[Señor Pete @Pete_Analysis]

@SemiAnalysis_ Takes a lot of pretty aggressive unspecified assumptions on the demand side to get to numbers like this

[Señor Pete @Pete_Analysis]

If everyone is bringing on a ton of MW, and better hardware drives more throughput per MW, the constraint could pretty quickly become how much $ businesses and consumers can spend on AI. We are probably within shouting distance of a range where enterprises will need mass layoffs to keep scaling spend

[Señor Pete @Pete_Analysis]

@GavinSBaker Are LTAs enough protection in your view even if they are honored? The way Micron described theirs suggested earnings would still get crushed if the floors came into play

[Señor Pete @Pete_Analysis]

@lokoyacap How does he pay for that

[Señor Pete @Pete_Analysis]

@HedgeyeComm Generally believe the ROIC will be there, but wouldn't it be fair enough to argue that the incremental revs per GW on Anthropic's 10th GW may be lower than their first few?

[Señor Pete @Pete_Analysis]

As fair as all of these anti-Anthropic arguments are, it is naive to think that the US gov't wants potentially superintelligent models to be open weight. A healthy US open-source ecosystem is fine as long as it trails the frontier, but depending on how much smarter these models get, open-source at the frontier would be a national security threat. Regulatory capture in this scenario is rational even if you don't like it

[Señor Pete @Pete_Analysis]

@amir Pretty sure there are a few inputs more scarce than ODM capacity

[Señor Pete @Pete_Analysis]

@chamath I don't think the US gov't cares as much ab token costs as they do having the most powerful models in the world in the US and run by people loyal to the US

[Señor Pete @Pete_Analysis]

@AtlasShrug1 @JonahLupton They were definitely pumping SPCX but also came off as more cautious on BG2 in early June. I believe Gavin even called the end of the "bottleneck bros"

[Señor Pete @Pete_Analysis]

@TBU12345678 Does this mean CRWV doesn't even need to hedge but they want to gamble on some options anyway

[Señor Pete @Pete_Analysis]

I worry that large enterprises may just not be labor constrained at this point, so will be hard for AI investment to lead to more revenue. Enterprises will more likely have to realize ROI through cost savings/layoffs which bring their own issues. Smaller more innovative businesses should be seeing ROI already with the available tools

[Señor Pete @Pete_Analysis]

@TXMCtrades Efficiency gains similar to WDM are happening in AI every year, but demand is more than keeping pace and we are in a compute shortage. Not obvious to me that this can't continue, models get smarter as you provide them more compute, there is no good dotcom analogy for that