1) Inflection will likely be the first of many VC-backed Foundation Model companies to fail.

Foundation models without proprietary, real-time data AND massive distribution for RLHF are the fastest depreciating assets in history.

2) Irony is that while models are commodities today, the ultimate future is likely one where there are only a few truly massive models with proprietary real time data and vast distribution.

Only a few will make it. And they will be super valuable.

Analysts often assume they are *the* audience for conference calls.

There are many audiences (regulators, employees, competitors) and incentives (sometimes want the stock lower i.e. if in the midst of a big buy-back or heading into option/RSU pricing).

Creates opportunities For example:

When a company is speaking directly to competitors by saying they are willing to incur losses to defend share.

A public conference call is really the only way to legally convey this intention and help rationalize competition.

1) GPU utilization rate is the new ROIC for any company working on AI.

“In deep learning, nothing is ever just about the equations. It’s how you put them on the hardware, it’s a giant bag of black magic tricks that only very few people have truly mastered.” 2) Cloud/virtualization/containers have abstracted hardware away for a generation of software engineers. Now back with a vengeance.

The MLPerf storage benchmarks will be a catalyst for storage by showing the massive impact on GPU utilization from faster storage.

1) Over the short-term, stocks trade on the 2nd derivative of Rev/EPS/FCF growth and ROIC changes along with beats/misses vs. expectations.

Valuation determines the magnitude of the move based on ☝️and 👇

Over the long-term, ROIC and growth in FCF per share drive performance. 2) I only say this as a reminder because there were so many replies to the tweet linked below pointing out that YoY earnings growth was weak as if that was a relevant or important fact.

The 2nd derivative and surprises vs. expectations are what matter. Always.

https://twitter.com/gavinsbaker/status/1653052376494817283

1) "Most people in high-stress, decision-making industries are always operating at this kind of simmering 6, or 4, as opposed to the undulation between deep relaxation and being at a 10. Being at a 10 is millions of times better than being at a 6."

tim.blog/2019/07/03/the… 2) "train your intuition, your somatic introspection, to feel when your quality of presence, your quality of energy is slipping from like a 10 to a 9. When I start working with top mental performers, very often, it can go from a 10 to a 2 before they even feel the slip."

1) Reread Jassy letter. Surprised that he hasn’t at least delayed Kuiper, their Starlink “competitor.”

Zero chance Kuiper succeeds until someone other than SpaceX can reliably land rockets.

In the unlikely event Blue Origin succeeds, then Kuiper will be a conflict of interest. 2) Starlink gets SpaceX’s internal launch cost (15m marginal launch cost for reused Falcon9 per an Aviation Week interview that is going to $2m post Starship) while Kuiper gets the Atlas V launch cost of $100m for a payload 20% lower than Falcon9.

1) Curious to hear what the large banks say about the consumer tomorrow.

I think it is now abundantly clear that the US consumer slowed dramatically in the last two weeks of March post SVB.

But cc data (multiple datasets) has begun to reaccelerate and normalize post Easter. 2) To cover the most obvious replies, well aware that the credit creation slowdown will impact the economy in several months, etc.

And oil going up is problematic as am becoming more convinced that the “consumer” is equivalently/more sensitive to the price of gasoline vs rates.

1) The strong perf of megacap tech in Q1 (up 31% per GS) is strange to me.

They were all in their own monopolistic swim lanes for 10+ yrs.

Now they are all facing an existential threat/opp from Generative AI/LLMs. 2) More importantly, competitive intensity is going up significantly as soon to be seen agentic AI forces them all into the same swim lane. Competition between them was previously either contained (iOS/Android), overstated (search vs. social) or oligopolistic (cloud).

1) If *only* there were another datacenter GPU coming to market in a few months. One that benefitted from OpenAI’s Triton such that almost any algorithm in PyTorch could easily be run on it.

theinformation.com/articles/ai-de… 2) Perspective on Triton:

openai.com/research/triton

1) Mkt seems tricky right now.

+’s: Futures pointing to a Fed pause and potentially outright cuts while inflation implodes in real-time.

-‘s: US consumer has rolled over since SVB per cc data. First real decel since Covid. And commodities might ruin all of the +’s above. 2) Tech companies sounded good at MS, but this was *before* SVB and different world since then. Although they all have religion on cost-cutting, margins and SBC.

As an aside and to reprise a 2022 debate, turns out that most SaaS co’s *can* just flip a switch on margins.

Glad this is being investigated.

CDS are a great product, and more predictive of future risk than credit spreads AFAIK, but illiquidity *might* make them vulnerable to manipulation over short timeframes.

Short DB, short indices, buy DB CDS, profit.

bloomberg.com/news/articles/… And I am obviously far from an expert here, so happy to be educated if I am off base. Just seems like a lot of this went on during the GFC with no consequences. If I’m wrong, ok.

1) Away from stocks/tech and on @destiny2

I no-lifed Trials of Osiris last weekend and got the god-roll Adept Immortal (Rangefinder/Target Lock).

First time I’ve had a “broken” gun before the nerf.

But my K/D in IB this weekend is *only* 0.04 higher than last few season avg. 2) This is a much lower improvement than I would’ve expected.

Part of this is due to replacing my almost perfect Messenger with a Stars and Shadow and being more of a ranged player.

SMGs just aren’t natural for me.

1) Anyone confidently saying there is zero risk of contagion from SVB should read the “Panic of 1907.”

Especially with Treasury saying they are carefully monitoring “several other banks.”

Contagion is a real risk. 2) Separately, risks to the VC ecosystem are reasonably contained.

Even absent a depositor bailout and/or acquisition, recoveries for uninsured deposits should be comfortably above 80% with some % paid out promptly and good VCs likely to supply any necessary interim cash.

1) The economy is accelerating hard in 2023, which cannot be “gratifying” to Powell. They were hoping for steady deceleration, which just isn’t happening.

No landing = hard landing is a good take from Michael Hartnett.

Bearish charts, followed by more positive ones next tweet. 2) Most positive fact for the market is that we are going through the most rapid decline in job openings without a recession in modern American history. Job switchers were driving inflation, so super positive.

Charts on this:

1) Grateful for the kind words from Jeff Green, CEO of Trade Desk and top 20 CEO IMO.

Superb execution in a very difficult space.

Reminds me of the way Nvidia came out of a scrum of ~10ish "3d graphics" companies in the late 90s and emerged as dominant.

joincolossus.com/episodes/83995… 2) My memory is slightly different btw.

It was a late stage round circa 14/15 and we extensively discussed the dominant incumbent.

I ultimately decided I couldn't invest *entirely* due to a conflict with an existing venture investment, so I sent it to a friend/competitor.

1) Amused that after the overwhelming conviction in material datacenter spending reductions:

META will still grow. GOOGL will grow their “technical infrastructure spending.” MSFT guided next q up. AMZN ahead on the q and no guide. Capex only commentary. 2) Interesting that the effective sum of the above decisions in terms of YoY growth is ahead of consensus data center revenue growth for the x86 share gainer stripping out El Capitan.

Also: Pensando/Mi300 impacts are not well understood.

Nothing to see here, move along.

1) Stripe doing a clean down round is healthy for ecosystem.

Many companies that raised at absurd multiples (avg SaaS deal in 21 was done at 114x ARR per IVP) will need to raise again soon.

Majority will be down rounds or heavily structured. Both options should be considered. 2) There is a lot of pressure from board members to raise structured rounds so marks can stay flat (i.e. artificially high).

Sometimes this is the right answer - generally when a company is doing extremely well - but often it's not.

It's hard to outrun structure.

1) Great @stratechery interview w/ @danielgross and @natfriedman

“the transformer and Stable Diffusion have given them this massive gift, which is we are finally CPU-bound. Software has not been CPU or GPU-bound for a long time.” 2) “It’s mostly been memory bound, and now that’s flipping again. And I think it’s a fantastic thing for anyone who makes hardware for TSMC, for Nvidia, for Apple. And so..I imagine that if they did Stable Diffusion that quickly, I can imagine they have a lot more than that.”

1) The FinTwit consensus dismissal of LLMs as a potential threat to search feels like the consensus dismissal of IDFA/ATT as a risk to social networks in 2021.

Difference is the dominant search company has wisely invested in a phone OS which helps mitigate risks. 2) When someone like Sridhar, who ran search ads, says that LLMs are a threat maybe don’t just dismiss the idea. Obviously he has incentives to say this

The largest social network was also quite cheap right before it imploded on IDFA/ATT & then the stupidity of FRL as a response

1) We will see more and more of this across industries.

AI can be “superhuman,” but the way it delivers these results is often unknowable.

i.e. AI can determine a persons sex by looking at a retina, something no human can do. But no one knows how.

adexchanger.com/commerce/meet-… 2) AI is better than any human at Chess, Go, lots of video games, making content recommendations (TikTok) so I’m not surprised it is better than humans at dynamically optimizing advertising spend across a variety of formats.

Advantage Plus also apparently off to a good start.

1) It is really strange that no image/video focused social network has released a native version of Lensa or otherwise incorporated generative AI.

Instagram *started* as a photo filter app. 2) TikTok represented a mortal threat to the incumbent and was ignored until almost too late in favor of trying to spend the Metaverse into existence.

Generative AI represents a clear opportunity/adjacency.

All the new filters are way behind Lensa. Yes, aware of Blenderbot.