Lunchtime @_Lunchtime_

@edzitron You realize this is $36B a year, or less than 1GW of compute. AMZN will add >5GW of compute over the next 12 months, so this doesn’t even cover 20% of it.

@justjoshinyou13 x.com/_lunchtime_/st…

Lunchtime @_Lunchtime_

7 days ago

Despite the endless discussion on fintwit Anthropic is still very poorly understood. The reality is, contrary to popular opinion, Anthropic’s business model is elite. They have best-in-class unit economics and the genetics of a highly, highly profitable business. Let me prove it.

0 0 1 548 4

@FredaDuan Freda, what exactly is this gross vs. net difference? Can you explain what exactly OpenAI nets out that Anthropic keeps in as gross? I don’t think you can, because it doesn’t exist.

Despite the endless discussion on fintwit Anthropic is still very poorly understood. The reality is, contrary to popular opinion, Anthropic’s business model is elite. They have best-in-class unit economics and the genetics of a highly, highly profitable business. Let me prove it. According to FT reporting, Anthropic did $16.8B of revenue in Q3 with 85% GAAP gross margins and 11% operating margins excluding SBC. They were cashflow positive and generated $1.70 of revenue for every $1.00 spent on compute. First let’s get a couple things out of the way: 1. 85% gross margins are already best-in-class. That’s higher than most SaaS businesses, which just a few years ago were thought to be among the most efficient businesses in the world. 2. It’s pretty darn impressive for a business growing >10x YoY to be cashflow positive. I’d challenge anyone in private markets to find a business growing even half as fast that’s not burning serious cash. “But 11% operating margins aren’t that strong and if you include SBC they were probably negative. They’ll never make real money.” Let’s break that down. You can simplify Anthropic’s costs into three main buckets: i) inference compute COGS ii) training and internal compute expense iii) all other operating expenses. This is where the $1.70 of revenue per $1.00 of compute spend is an important disclosure. Take the $16.8B and round $17B to make the math easy. $17B of revenue with the 1.7:1.0 ratio implies $10B of compute expense. 85% gross margin on $17B revenue implies $2.55B spent on inference compute. That leaves $7.45B of training and internal compute expenses. Since we know operating margin was 11%, that means operating profit was $1.87B. Now we have enough to get the full picture of the P&L: $17B revenue (-) $2.55B inference compute expense (=) $14.45B gross profit (85% margin) (-) $7.45B training and internal compute expense (=) $7B profit after compute expense (41% margin) (-) $5.13B other operating expenses (=) $1.87B operating profit (11% margin) There’s a few things we can learn from looking at the P&L in this way: 1. Given the scale and growth rate of this business, non-compute operating expenses should quickly decline as a % of revenue. In just a few years you’d expect this line item to be 10% or less as the business deomonstrates operating leverage, which means operating margins should converge up close to the 41% “profit after compute expense” margins. Call it 35% as a realistic ceiling. 2. A credible path for 35% operating margins is already impressive, but it’s clear those margins could be much higher if Anthropic wanted them to be. How? Of the $10B they spent on compute only 25.5% went to generating revenue (inference). The rest was training and internal spend, which has no in-period revenue benefit. Of course their whole business is predicated on developing next-gen models so this expense can’t be taken to zero, but in the past both Sam and Dario have talked about a potential long-term compute split of 50% training and 50% inference. Assuming a steady-state 50/50 split at the same 85% gross margin removes $2.55B of cost from training and internal compute spend, which adds another 15pts of margin. All of a sudden true margin potential looks like 50%. Those are op margins even Hock Tan would be jealous of. 3. What about SBC? Without any publicly reported details let’s just take a stab. Anthropic has maybe 4,000 employees. How much does each get paid in SBC per year? Tough to know for certain but we know the average employee at META makes $300k a year in SBC. Anthropic has much higher research talent density, and researchers are expensive, so let’s assume Anthropic pays the average employee 5x more in SBC than META does. That’s $1.5M per employee per year. At 4,000 employees that’s $6B per year, or $1.5B per quarter. On $17B the $1.5B is 9% of revenue, which means GAAP operating income would be 2% (vs. non-GAAP at 11%). So they’re potentially already better than breakeven on a GAAP operating basis, and if we subtract the 9% from their 50% adjusted margin potential we still get >40% terminal margins. That’s MSFT-level profitability on GAAP margins, which is as good as it gets. “This is all backward-looking analysis. Anthropic locked in Q3’s compute capacity a long time ago. Compute prices have skyrocketed since - run current rates through their P&L and it blows up.” Let’s take a big step back and look at their economics all the way through the stack. It costs about $55B to build a 1 GW data center full of VR200s all in. SemiAnalysis estimates Anthropic ended Q3 with $74B in run-rate revenue, and they believe the business generated it on about 4GW of capacity exiting the quarter. From the P&L analysis above we know Anthropic only allocated 25% of their capacity to inference, which means only 1 of those 4 GW were generating revenue. So on 1 GW of inference compute Anthropic generated $74B of run-rate revenue, which is more than enough to cover the $55B of compute capex and roughly $3B of annual opex (e.g., electricity, water, on-site labor, etc.) required to keep the DC running. In fact, this means they’d have $16B left over to pay for the margin stacking of intermediaries they rent from. Functionally this means if Anthropic allocated all their capacity to inference they’d have <1 year paybacks on "look through" capex. Even at a 50/50 inference vs. training mix that’s <2 year paybacks on total capacity, or a >50% ROIC (ROIC not exactly the right calc here but you get the idea). I’d say the math works handsomely even if compute costs continue to rise, and encourage anyone to find me another business that can deploy hundreds of billions of dollars with a >50% ROIC… We don’t know everything about Anthropic’s financial picture, and of course competition could drive pricing pressure which compresses the margins calculated above, but from what we do know their business looks pretty fantastic. 85% gross margins are best-in-class, >40% terminal operating margins are excellent, and <2 year paybacks are unheard of when you’re talking about deploying hundreds of billions of dollars per year. You might expect competition to come in and compete away these margins, but Anthropic took some pretty good swings from OpenSource and OpenAI price cuts in Q3 and were still able to put up these results. With their anti-distillation efforts there’s reason to believe OpenSource lags from here, which if true means we’re looking at a ~2-5 horse race given the amount of capital required to compete at the frontier. In that case it’s quite likely that Anthropic is not only one of the largest businesses of all time but also among the most profitable and highest quality businesses we’ve ever seen. Can't wait to see the S-1 for more detail.

@KairosPraxis He’s talking about the cost to rent and/or factoring in depreciation on his own capacity. Well known that annual rental rates are ~$15B per GW if you’re a large customer like Meta renting from a neocloud. Cash operating costs are ~$2-3B.

@MetacriticCap They won’t. Anyone expecting guidance is going to be disappointed tomorrow. Position your portfolio accordingly.

If OpenAI’s 50B RRR is a “net” figure, what exactly would comprise gross? MSFT has access to all their IP through 2032 so there’s nothing flowing to OpenAI, and their presence on AWS is too small to matter now. “Gross” is a myth.

BI has historically been a pull function: You pull up the dashboard, notice a KPI has changed, then ask an Analyst to find out why. The future of BI in the AI-era is a push function: the AI knows which KPIs matter so it can track changes and identify root-causes automatically. Trends or inflections that require real-world attention are pushed to the relevant business leads. This is just one of many examples where AI will increase the scope of knowledge work and allow software to deliver more value to customers.

Salesforce @salesforce

2 weeks ago

We have signed a definitive agreement to acquire Loops, an AI-powered analytics platform. Loops’s AI goes beyond showing what happened — explaining why it happened and what to do next. Together, we’ll help customers better understand what drives business performance and move

15 23 137 39K 31

Can we please retire “ARR” as a metric to judge AI startups by? Two main reasons: 1. It’s not “ARR” - in most cases these aren’t annual contracts and it’s not recurring. Call it “RRR” for run-rate revenue. 2. RRR is functionally useless without understanding the margin profile. RRR from reselling tokens is more akin to GMV than it is revenue. Tell me run-rate gross profit and then we can have a discussion. ARR was useful in the SaaS era when customers signed annual contracts, churn was low, 70-85% gross margins were pretty much guaranteed. None of those apply to AI startups today.

@edzitron It’s almost as if AWS has a >$100bn business that has nothing to do with AI…

Is there any possible way OpenAI’s 50B RRR could credibly be grossed up to 70B? I’m sympathetic to the earlier argument of the MSFT rev share, but as of April MSFT’s rev share to OpenAI is over so that can’t be it. And even if it was the 20B gap is 30% of 70B, materially higher than the prior 20% rev share.

@danprimack Why is it either OpenAI or Anthropic’s responsibility to disclose their RRR or correct any “leaks” that are flat out wrong? They’re private companies. Many reasons for them to go public, this isn’t one.

The obvious answer is OpenAI and Anthropic, but I think COVID also played an important role. Peninsula = suburb life, which was fine for people when they got enough social interaction at work. During COVID everyone was devoid of personal interaction, which made them crave it more on the other side. Underrated reason why young founders overwhelmingly moved to the city and started companies there.

@value_invest12 Exhibit A for why P/E multiples are dumb. They completely ignore the balance sheet.

@BrianSozzi While I agree with the general point, it’s possible we’re still in a bubble. 20x consensus earnings could turn out to be 30x real earnings given the degree of over-earning we’re seeing in some industries. Just look at memory. Nothing like .com, but still.

I like to think about this in terms of earnings vs. multiple, and which you have to get right when making the investment. For example, you can probably estimate T+3yr AMZN earnings within a decent confidence interval. What matters for that investment is T+3 multiple. With NBIS, however, if you’re above consensus and right on T+3 earnings multiple will probably work out for you.

@LongGameEquity Why in earth would anyone pay 20x run-rate revenue for AWS? It’s a 33% margin business so that’s 60x EBIT for a highly capital intensive asset with eroding competitive position.

@geoffwoo Maybe even worse than jewelry? Jewelry will have higher margins and lower AI disruption risk. Soon Muse / Dot / etc. will be able to vibe code a hyper-personalized whoop competitor app in a day.

@babyfolio GLW. When scale up goes optical content will 9x. Their opportunity inside the box is even greater, and they’ve already been crowned by NVDA as the winner.

@BourbonCap Sure, but does that make any of them good? Cloud objectively a worst market than it was 3-5 years ago, Agentic shopping a new threat vector to the ads business (which drives all of retail’s profitability), and WMT increasingly closing the gap in ecommerce.