AI is improving knowledge work faster than most of Wall St appreciates imho. From my own use, it keeps gettin better and faster. Things particularly improve when you make models / agents argue with each other in front of a judge w/ verifier agents. This is all anecdotal, but the gap versus a few months ago is notable to me. The catch, and I’ve harped on this in the past, is the tools are as good as the curated context (including what services / research you connect) “Give me the bull case on MU” gets you a meh answer on the chat bot. Point the same model with a harness, curated “source of truth” (curated sources), persistent memory of key debates / work you’ve already done and the output is miles better. As an example of recent workflow, I had Claude build a “ Refresher”, a living html doc per name / theme for everything I own. It has the basic thesis, a change log since it was last read, estimate changes, valuation, and my leans on the top few debates on the name based on prior work / IBs / emails etc. And where I have no recorded view the model proposes a lean and labels it as its own until I sign off. The build itself was agents all the way down. 3 agents debated the format, a four-analyst council reviewed each initial doc, all in front of a judge. Most impressively, much of this was done in parallel (ultracode!) in a few hours. It’s simply never been easier to curious and wild to think this wasn’t possible a year ago.

There is a lot of Artificial Anxiety these days, from war / oil / stop outs in macro to lab safety self-owns. But one worthy debate we’ll continue to have is on the revenue needed to pay for all the hyperscaler capex. Goldman's recent framework puts the hurdle at roughly $475bn a year in 2028-30 for the ’26-27 investment phase. The numbers seem insane to many, but we’re living in insane times if you hadn’t noticed! I do think the dollars become easier to understand when you break it down between already contracted biz, returns inside the existing core businesses, and then “new demand”. Start w/ the contracted stuff. AMZN / MSFT / GOOGL had $1.7T of combined backlog in 2Q. Now, that spans multiple years and includes some non-AI business, so it needs to be matched correctly. But part of this buildout already has a customer and a payment commitment attached to it. For that contracted capacity, the provider's job is to deliver, do it efficiently,and get paid the owed rate. These payments can support attractive infra returns across a range of outcomes for the lab / customer’s own margins. The lab doesn’t have to become an extraordinary business for the provider to earn its return. Instead, each hyperscaler’s concern is whether they can deliver compute w/ good per-token economics relative to their competition. Next there is the compute used inside existing businesses. Better ad conversion, lower opex etc are all part of the equation. Some of meta's return can come through its ad biz; some of Google's can come through search / YouTube. These are also huge existing businesses. For scale, Meta and Google Services already do ~$620bn. Growth of say 10–12% adds ~$60–75bn in 1yr, and maybe $55-65 after some ad related costs. That’s a meaningful chunk of the $475bn. God forbid they do better than that btw. For example, is Zuck a Double Winner in all this? Perhaps the ad fly wheel cooks AND instagram eyeballs become more valuable in a world where consumers aren't on websites as much. Of couse some of all that may turn out to be what I'd call “shadow ROI.” If a business would have lost a step without the investment, maintaining its growth path can be a meaningful return. Some of the payoff may well be profit preserved, which never appears in a separate line called ‘AI Revenue’. Which brings us to the gazillion (?) dollar question: new spending. This depends a lot on how much each layer keeps. If compute absorbs 40% of customer spending instead of 30%, the same infrastructure revenue requires less spending at the top. There is no requirement that every layer earns wonderful margins. For illustration, let’s say ~$175bn of the ~$475bn from GS is supported by internal use. That leaves ~$300bn of external compute revenue, including the portion already contracted. At a 40% compute share, that implies ~$750bn of annual end-customer AI spending. 50% compute share lowers the requirement to $600bn. Don’t crucify me for these SWAGs, the point is to show how much the headline depends on the biz mix and margin. So where the hell do we get $600bn? Well, annual software spending is ~$1.5T, IT services is similar. The knowledge work pool is ~$6T. Yes, these budgets partially overlap somewhere I'm sure, but they give you some context for the $$ involved. Some spending shifts to AI; some gets justified by more productive white collar work, better logistics etc. Oh, add consumer subs and some activity that becomes economical because of AI on top. Use your imagination. The labs have to tap those same budgets and new use cases too, of course. That’s where their revenue comes from, and part of that revenue pays the hyperscalers’ compute bills The point is that there are a lot of permutations for how the pie is sliced. Customers can keep a substantial share of the gains while infra providers earn… traditional infra returns. Perhaps labs maintain sick margins and end-customer spending is higher. OR, dare I say it, useful lives turn out to be longer. Depreciation accounts for ~60% of GS’s revenue hurdle after all…

But markets are *very* imaginative when it comes to Doom Swarms / ways for everything to fail! This is due to the intersection of how hard it is to conceive of exponentials (in a biz based on compounding! lol) and how much smarter you sound as a skeptic. Happy Friday :)

Well, we’ll go back to ZIRP one way or another… 😂🤦🏻‍♂️

It's been so long I almost forgot what "overbought" feels like! But turns out if it hasn't happened in a while it's a positive signal in the shmedium term :). 🫣

Market found its muse 😜

I find things like Jev very bullish. Cheap, fast classification can identify more work worth calling an LLM for AND lower the cost and latency of chained workflows (TAM expander) Could be ultimately be an unlock for diffusion too by being a bridge between deterministic workflows and generative intelligence.

Remarkable how quickly everyone forgot @finkd is the OG machine learning monetizer. @alexandr_wang also looking like a $$ hire

Refreshing to see an account avoid FUD for clicks and pursue truth instead. Good sober overview. 🫡

Markets do love irony! Plus, we love the slop and, most importantly, our agent can’t watch it for us 😝

So the one where 5 banks did a $22bn loan to Blackstone 3 days ago was the end? 😂

This person posted patently false things about CRWV last week and deleted his tweet after I called him out. Now he's doing it again with more made up stuff. This is gross. You can put his claims into the LLM of your choice and have it compare against public filings. Do it 1x i assume an accident, do it 2x you are 1) a deceptive person, or 2) a moron. And should have zero followers. Horrible.

Musings about Muse: The Death of Dot-Com? I sense the June rip and subsequent unwind in stocks have thrown people off the scent. Investors have been fitting growth and profitability narratives to price action (shocking, I know). But step away from the chop and charts… what’s actually changed in demand, products and outlooks since June? The clouds keep delivering. We’re getting more detail on ROI, firmer pricing and older GPUs renting for longer. Competition is heating up. OAI is back. Products keep improving. And now the ‘Muse Moment’. The vibe feels like when Uber and Tinder became habits. Suddenly it was hard to remember life before. Personal / Social AI feels early in that transition. Consumers can finally feel the magic coders and early adopters felt. There’s also something uniquely satisfying about bossing around an agent, and I suspect control is part of it. Meta’s distribution could bring that to many people, fast. We’re still of course in the messy middle. Computer use lets agents use today’s internet, clicking through sites built for human eyes. As services reorganize around agents, we’ll spend less time navigating apps and websites ourselves. Things will shift to ‘agent first’. That’s the ‘death of dot-com’ I’m imagining. Useful delegation also becomes habitual. People, especially company leaders, bring those expectations to work. How long does a CEO who delegates to an agent at home tolerate a company that can’t? Then a competitor gets it working and waiting starts to cost you. I’ve also run into a lot of ‘diffusion must be slow’ takes. Roadblocks, data issues, bureaucracy all get in the way. Skepticism is always so cosmopolitan. But businesses will move mountains when lagging gets expensive enough. That’s how AI diffusion starts to look like “reverse bankruptcy”… slowly and then suddenly. “Your agent can’t talk to my agent? Better get cooking…” Most importantly, AI can work on its own adoption bottlenecks. Better agents can build integrations and help refactor the systems they’re being deployed into. Innovation raises the cost of waiting while making adoption easier. Tough opponent to be running against! So where does that leave the market? My sense is sentiment is down and positioning is cleaner. If oil eases, investors likely refocus on AI into year end with many winners well off their highs. I also suspect the next lab ARR updates will look better than the mood suggests, which likely helps a lot. Two parting reminders. Price => narrative works on the upside too (and asymmetry is way better) Greed has a very short memory (a certain call buyer is already back…). Godspeed.

And so begins yellow (web) pages 2.0. Agent economy. Ready set go.

There is literally no stability to be found in the AI stack. In the last few months: 1.) I moved all my intense recurring workloads from OpenAI to Anthropic and now back to OpenAI 2.) Used GrokBot as an over lay on top of OpenAI and Anthropic but am now switching this to Muse 3.) liked Instinct for a week until Muse came out 4.) Was using Exa and TinyFish only to now discover this new model called Jev which can save me money on all of the tools listed above In the matter of days, something that was an “oh shit moment” becomes underwhelming. As a consumer of these tools what a fun time to be alive

I can see it now: Full Claude Co-PM.

Dawn of the robot scaling law: more pretraining data predictably improves transfer to robot actions. Next thing you know we'll need lots of compute for robot safety ;)