WEBVTT - Monologue: Anthropic Is Not Profitable Unless You Remove Its Costs 0:00:02.800 --> 0:00:07.000 Zon Media Hell and welcome to a very special better 0:00:07.040 --> 0:00:18.520 offline monologue. I'm your host ed Zron And. Wednesday, the 0:00:18.560 --> 0:00:21.239 Wall Street Journal ran a story about how Anthropic is 0:00:21.280 --> 0:00:25.000 about to have its first profitable quarter, specifically an operating 0:00:25.040 --> 0:00:28.600 profit or a bit duh profitability so earnings before interests, 0:00:28.800 --> 0:00:32.280 tax and bad stuff, really, saying that anthropics revenue was 0:00:32.320 --> 0:00:34.160 said to more than double to ten point nine billion 0:00:34.159 --> 0:00:36.960 dollars in the second quarter of twenty twenty six, which 0:00:37.000 --> 0:00:40.519 is when said profit would happen. But the journal Anthropic 0:00:40.600 --> 0:00:43.040 generated four point eight billion dollars in sales in the 0:00:43.040 --> 0:00:46.040 first quarter, and in Q two, that's that five hundred 0:00:46.040 --> 0:00:51.960 and fifty nine million dollar operating profits home interesting, there's 0:00:51.960 --> 0:00:54.760 a lot of certainty, considering we're barely through the first 0:00:54.760 --> 0:00:57.800 half of the second quarter, and quite a specific number 0:00:57.800 --> 0:01:01.400 given the fact that June hasn't start, and all of 0:01:01.480 --> 0:01:05.840 these numbers are mysteriously leaking exactly while Alanthropic is raising 0:01:05.840 --> 0:01:08.600 its funding round. And there's another important note from the 0:01:08.680 --> 0:01:11.200 story too. The journal adds at the very bottom of 0:01:11.240 --> 0:01:14.319 the article that it and I quote is unclear what 0:01:14.440 --> 0:01:17.360 accounting methods Anthropic has used to book revenue and costs. 0:01:17.520 --> 0:01:20.000 As the company isn't yet required to follow the financial 0:01:20.040 --> 0:01:23.840 reporting requirements of a public company. That's right, Anthropic is 0:01:23.920 --> 0:01:26.480 possibly going to be a bit to profitable for a 0:01:26.520 --> 0:01:30.440 single quarter on a non gap so generally is accepted 0:01:30.480 --> 0:01:34.040 accounting principles basis. Anyway, I wonder how they did it, 0:01:34.280 --> 0:01:38.120 because based on the unhelpfully labeled diagram from the Journal's article, 0:01:38.160 --> 0:01:40.960 it appears, as I said last year, the Anthropics costs 0:01:40.959 --> 0:01:45.120 scale linearly with its revenues, except they magically didn't in 0:01:45.160 --> 0:01:49.360 the second quarter. How and I wonder if Anthropic will 0:01:49.480 --> 0:01:53.480 somehow stay profitable. Let's see if the journal has a quote, Oh, 0:01:53.160 --> 0:01:56.600 there we go, Okay, the company might not remain profitable 0:01:56.640 --> 0:01:59.320 for the full year, as it plans spending increases due 0:01:59.320 --> 0:02:05.559 to its fasting needs. Interesting. That's very interesting. So Anthropic 0:02:05.760 --> 0:02:09.320 may be profitable very specifically in the second quarter of 0:02:09.320 --> 0:02:12.960 twenty twenty six, but might not be afterwards. It's almost 0:02:13.000 --> 0:02:15.399 as if it found a way to specifically cut its 0:02:15.400 --> 0:02:19.760 costs in May and June somehow, because it did. Remember 0:02:19.800 --> 0:02:23.679 that deal Anthropic sign with SpaceX to take over Colossus one, Well, 0:02:23.720 --> 0:02:27.120 it's also extending to Colossus two, and they're going to 0:02:27.120 --> 0:02:30.240 be paying SpaceX one point two five billion dollars a 0:02:30.240 --> 0:02:34.200 month starting in May and June, and in those two 0:02:34.320 --> 0:02:38.440 months they're going to have reduced fees. You know, the 0:02:38.560 --> 0:02:42.360 quarter in which Anthropic is magically profitable, the one with 0:02:42.440 --> 0:02:46.679 the lower costs ah per SpaceX is s one. Pursuant 0:02:46.680 --> 0:02:49.520 to its agreement, Anthropic will pay SpaceX one point two 0:02:49.560 --> 0:02:52.239 to five billion dollars a month through May twenty twenty nine, 0:02:52.240 --> 0:02:55.320 with capacity ramping up in May and June at a 0:02:55.440 --> 0:02:59.560 reduced fee as fifteen billion dollars a year in compute 0:02:59.560 --> 0:03:02.600 cost but due to an indeterminately discounted level for the 0:03:02.639 --> 0:03:05.600 precise months that Anthropic is using to tell its investors 0:03:05.639 --> 0:03:08.799 in the media that it has an operating profit. Five 0:03:08.880 --> 0:03:11.200 hundred and fifty nine million dollars in operating profit in 0:03:11.240 --> 0:03:13.800 a three month period is absolutely possible when you're not 0:03:13.840 --> 0:03:17.280 paying for all your costs. While I wouldn't say this 0:03:17.360 --> 0:03:19.600 is cooking the books, I would say it's definitely a 0:03:19.639 --> 0:03:23.600 Shiatsu grade massaging of the numbers. Anthropic is deliberately leaked 0:03:23.600 --> 0:03:26.240 a quarterly profit, and I'm doing air quotes where it 0:03:26.280 --> 0:03:29.480 knows it can suppress its costs. Specifically made sure that 0:03:29.520 --> 0:03:32.280 the journalist gave it an out of costs might increase 0:03:32.560 --> 0:03:34.600 and released it on the day of Nvidia's earnings as 0:03:34.600 --> 0:03:38.320 a means of keeping the AI bubble inflated. Nothing has changed. 0:03:39.000 --> 0:03:41.200 If Anthropic was paying the full rate for its compute 0:03:41.200 --> 0:03:44.120 in those two months, its economics would shift right back 0:03:44.120 --> 0:03:46.320 to where it's always been. Per my reporting from last 0:03:46.400 --> 0:03:49.040 year on its aws costs. A business that has its 0:03:49.040 --> 0:03:53.800 costs linearly increase with its revenues, I also severely doubt 0:03:53.800 --> 0:03:56.040 that Anthropic managed to make the cost of running their 0:03:56.120 --> 0:03:59.840 services profitable in the space of six months, per the information. 0:03:59.880 --> 0:04:03.000 In January, Anthropic missed on its gross margin projections to 0:04:03.080 --> 0:04:05.680 twenty twenty five, saying that its inference costs were twenty 0:04:05.720 --> 0:04:10.320 three percent higher than the company had anticipated. How did Anthropic, 0:04:10.440 --> 0:04:12.640 who faced a massive influx of new business to the 0:04:12.680 --> 0:04:14.640 point that it was forced to buy more compute from 0:04:14.640 --> 0:04:18.640 elon fucking Musk, suddenly have its revenue outpaced its costs. 0:04:19.560 --> 0:04:21.840 Have had a few guesses other than the obvious way 0:04:21.880 --> 0:04:26.080 in which they reduced their costs. Number one, for large enterprises, 0:04:26.080 --> 0:04:29.160 they're taking prepayment of tokens, so fifty million dollars intended 0:04:29.200 --> 0:04:31.600 to be spread over twelve months that it's taking us 0:04:31.680 --> 0:04:35.040 upfront revenue. This would both inflate revenue numbers and depressed 0:04:35.040 --> 0:04:38.680 costs because Anthropic wouldn't have actually provided the compute necessary 0:04:38.800 --> 0:04:41.480 to earn that revenue. Yet perfectly legal, because this isn't 0:04:41.520 --> 0:04:44.360 non gap. This is non gap accounting like you could 0:04:44.400 --> 0:04:47.280 just you could say whatever in a real company or 0:04:47.279 --> 0:04:51.440 a public company, that would be considered deferred revenue. Anthropic 0:04:51.520 --> 0:04:54.360 is already offering discounted tokens for Claude users too, through 0:04:54.400 --> 0:04:56.800 the by Extra Credits page on their accounts, with discounts 0:04:56.880 --> 0:05:00.400 ranging from ten percent to thirty percent. It again, maybe 0:05:00.400 --> 0:05:03.680 booking this upfront rather than saying, okay, let's book this 0:05:03.800 --> 0:05:06.599 as the person uses the compute. They could also be 0:05:06.600 --> 0:05:09.400 front loading annual commitments of basically any kind, subscriptions to 0:05:09.440 --> 0:05:12.480 Claude enterprise team agreements and so on. And they could 0:05:12.520 --> 0:05:15.400 have deliberately ratcheted downtraining to ease the burden on its 0:05:15.440 --> 0:05:20.880 infrastructure to provide inference. But if I'm honest, the revenue 0:05:20.920 --> 0:05:24.800 side is where the real problems lie. So Anthropic has 0:05:24.800 --> 0:05:27.200 said it brought in four point eight billion dollars in 0:05:27.240 --> 0:05:29.240 revenue in the first quarter of twenty twenty six and 0:05:29.240 --> 0:05:32.360 projects to ten point nine billion dollars in Q two 0:05:32.480 --> 0:05:36.680 twenty twenty six. This is pretty tough to reconcile with 0:05:36.760 --> 0:05:41.880 previous reporting. On February twelve, twenty twenty six, Anthropic claimed 0:05:41.880 --> 0:05:44.760 it had reached fourteen billion dollars in annual recurring revenue. 0:05:44.839 --> 0:05:47.520 As a reminder, ARR is the accounting tool largely used 0:05:47.520 --> 0:05:50.120 by startups but some public companies too, where a snapshot 0:05:50.160 --> 0:05:52.839 of a single month's income is taken and multiplied by twelve. 0:05:53.360 --> 0:05:55.440 This gives you an implied monthly revenue of like one 0:05:55.440 --> 0:05:59.160 point one six one point one seven billion dollars. On 0:05:59.200 --> 0:06:01.719 March third, twenty two, twenty six, Warrio Ama Day would 0:06:01.720 --> 0:06:05.080 claim Anthropic had reached nineteen billion dollars an ARR, so 0:06:05.120 --> 0:06:07.160 about one point five to eight billion dollars per month. 0:06:07.760 --> 0:06:11.240 Three days later, on March sixth, Krishna Rao, chief financial 0:06:11.240 --> 0:06:14.280 officer of Anthropic, would declare under oath in a court 0:06:14.320 --> 0:06:18.880 filing that Anthropic had brought in revenues exceeding five billion 0:06:18.960 --> 0:06:23.640 dollars to date, so just to be clear lifetime revenues. Also, 0:06:23.720 --> 0:06:26.080 keep in mind that the Information had previously reported that 0:06:26.120 --> 0:06:28.640 Anthropic had four point five billion dollars in revenue in 0:06:28.640 --> 0:06:31.480 twenty twenty five, which already puts US at the limits 0:06:31.480 --> 0:06:35.800 of credulity. While boosters may claim that exceeding could mean 0:06:35.880 --> 0:06:39.760 literally any number they want above five billion dollars. I 0:06:39.800 --> 0:06:43.919 found it doubtful that the CFO of Anthropic would, under oath, 0:06:44.240 --> 0:06:46.480 lead the court to believe its business was thirty to 0:06:46.480 --> 0:06:49.279 forty percent smaller than it was, especially when trying to 0:06:49.320 --> 0:06:51.719 convince it that the damage of being labeled a supply 0:06:51.839 --> 0:06:54.000 chain risk would ruin its business. They would want it 0:06:54.040 --> 0:06:57.320 to seem bigger than it is, not smaller. And at 0:06:57.320 --> 0:06:59.760 this point it's impossible to reconcile that twenty twenty five 0:07:00.160 --> 0:07:02.200 point with that five billion dollar number. No one wants 0:07:02.240 --> 0:07:04.280 to talk to me about it. None of the journalists involved, 0:07:04.320 --> 0:07:06.520 no no analysts. That no one wants to talk about 0:07:06.520 --> 0:07:10.200 this at all. It makes them uncomfortable and angry, I assume, 0:07:10.280 --> 0:07:23.840 probably because it's impossible to reconcile. If we assume that 0:07:23.880 --> 0:07:26.880 the arr claims made by Anthropica correct, we can presume 0:07:26.880 --> 0:07:29.840 that it made revenues of roughly two point four to 0:07:29.840 --> 0:07:32.080 two point five billion dollars in March, given that it 0:07:32.080 --> 0:07:34.600 claimed it had thirty billion dollars of ARR and April 0:07:34.640 --> 0:07:36.680 sixth if you get in confused with the numbers. By 0:07:36.680 --> 0:07:40.960 the way, there's a newsletter attached to this. Nevertheless, when 0:07:41.000 --> 0:07:42.840 you add up all those arrs, you get to about 0:07:42.840 --> 0:07:45.280 five point two five billion dollars, which is in excess 0:07:45.320 --> 0:07:47.240 of what the Wall Street Jurnal has said. And in 0:07:47.280 --> 0:07:50.560 some world maybe there's a way of cherry picking using 0:07:50.600 --> 0:07:53.160 particular periods to the point that the arrs makes sense 0:07:53.320 --> 0:07:55.080 and would actually be in the region of four point 0:07:55.080 --> 0:07:58.560 eight billion dollars. Fine, but they don't make a lick 0:07:58.560 --> 0:08:00.320 of sense when you bring up what Christian to Row 0:08:00.400 --> 0:08:03.800 their CFO said. If we believe anthropics leaks, putting inside 0:08:03.960 --> 0:08:06.760 all the R figures for a second, this means that 0:08:06.800 --> 0:08:10.000 Anthropic A made over ninety percent of its lifetime revenues 0:08:10.040 --> 0:08:12.920 in the first quarter of twenty twenty six, and B 0:08:13.400 --> 0:08:16.520 made virtually no revenue its previous years, and c leaked 0:08:16.560 --> 0:08:21.120 completely imaginary run rates of the media for years. While 0:08:21.160 --> 0:08:24.680 I acknowledge that anthropic abys to have grown significantly, that 0:08:24.760 --> 0:08:29.880 level of stratospheric growth does stretch the limits of their credibility. Moreover, 0:08:29.920 --> 0:08:32.720 the fact that the previous era. ARR figures are inconsistent with 0:08:32.760 --> 0:08:35.760 the leaked charts from Anthropic. Further raises questions about the 0:08:35.760 --> 0:08:39.760 credibility of well any numbers from the company. The only 0:08:39.880 --> 0:08:42.960 real defense that anybody has here is that Krishna Rau 0:08:43.320 --> 0:08:45.959 under oath low balled the US government and adjudged to 0:08:46.160 --> 0:08:49.600 such a dramatic extent that he hid an excess of 0:08:49.679 --> 0:08:53.000 four billion dollars in revenue. And as I've discussed before, 0:08:53.040 --> 0:08:55.960 and Flying Penguin helpfully collated, which I'll link in the notes. 0:08:56.160 --> 0:08:59.440 Adding up anthropics previously reported ARR from January twenty twenty 0:08:59.480 --> 0:09:01.880 five to my twenty twenty six already gets us to 0:09:01.920 --> 0:09:05.320 around six point six six billion dollars. Now, I know 0:09:05.520 --> 0:09:08.080 there are going to be some boosters who hear about 0:09:08.080 --> 0:09:09.440 this and they're going to be like, well, this is 0:09:09.480 --> 0:09:11.640 proof there's a business model emerging from AI and I'm 0:09:12.400 --> 0:09:17.320 I'm sorry, that's not what's happening. Dario Amadianelon Musk worked 0:09:17.320 --> 0:09:19.640 out a sweetheart deal framed as a ramp up that 0:09:19.679 --> 0:09:22.920 allowed Anthropic to artificially depress its costs. I also question 0:09:23.000 --> 0:09:24.880 how much of a ramp up there really was, or 0:09:25.080 --> 0:09:30.000 what Anthropic's actual compute constraints were because Anthropic immediately loosened 0:09:30.040 --> 0:09:32.959 rate limits for Claude subscribers on announcing the deal, meaning 0:09:32.960 --> 0:09:37.840 that it immediately started having higher inference costs, which somehow 0:09:37.960 --> 0:09:42.400 led to it making a higher profit or did. Musk 0:09:42.480 --> 0:09:46.000 has literally described in its s one have SpaceX charge 0:09:46.040 --> 0:09:49.280 Anthropic less for two specific months to make the numbers 0:09:49.320 --> 0:09:53.800 look better. In July, Anthropic will start paying SpaceX one 0:09:53.840 --> 0:09:56.480 point twenty five billion dollars a month or about fifteen 0:09:56.480 --> 0:09:58.520 billion dollars a year, on top of all its other 0:09:58.600 --> 0:10:02.000 compute deals with Google, Amazon, the Microsoft. If we assume 0:10:02.000 --> 0:10:05.200 that its spend is comparable on AWS and Google Cloud, 0:10:05.200 --> 0:10:08.480 and it's most assuredly more, that means Anthropic is spending 0:10:08.520 --> 0:10:10.920 around three point seventy five billion dollars a month in 0:10:11.000 --> 0:10:13.599 compute costs or eleven point twenty five billion dollars a 0:10:13.679 --> 0:10:16.800 quarter or forty five billion dollars a year. I'm one 0:10:16.840 --> 0:10:19.599 hundred percent sure it's more than that. There's also a 0:10:19.720 --> 0:10:22.680 very compelling argument that the Anthropics costs will increase and 0:10:22.720 --> 0:10:25.400 will eat up that profitability. To once again repeat the 0:10:25.480 --> 0:10:28.280 quote from the Journal, the company might not remain profitable 0:10:28.360 --> 0:10:31.120 for the full year as it plans spending increases due 0:10:31.160 --> 0:10:36.120 to its vast computing needs. Also have to wonder this 0:10:36.240 --> 0:10:39.400 company's profitable. Just humoring here for a second. If you're 0:10:39.440 --> 0:10:42.160 so profitable, why aren't you ipoing? Why not take this 0:10:42.200 --> 0:10:45.600 to the public markets, unless, of course, your only non 0:10:45.679 --> 0:10:47.760 gap a bit to profitable based on a two month 0:10:47.880 --> 0:10:50.720 long discount specifically covering the period in which you're profitable. 0:10:51.679 --> 0:10:55.120 I will give Diario Amida credit. Nobody does it better. 0:10:55.520 --> 0:10:59.720 Nobody does financial engineering and press lead information war better 0:10:59.720 --> 0:11:04.160 than our Anthropic. The utter willingness of the press to 0:11:04.160 --> 0:11:06.880 eat up incongruent numbers and the eagerness of many to 0:11:06.920 --> 0:11:09.480 jump up and find obtuse ways to explain away the 0:11:09.520 --> 0:11:12.640 obvious problems is only made possible when a company's perfected 0:11:12.640 --> 0:11:15.360 the art of manipulation and ingratiation of those who want 0:11:15.400 --> 0:11:18.360 to feel like they're first. If you take this as 0:11:18.400 --> 0:11:21.839 an incontrovertible proof that Anthropic is profitable, you are deliberately 0:11:21.880 --> 0:11:24.920 ignoring the blatantly obvious ways these numbers are being massaged. 0:11:25.320 --> 0:11:27.920 We've got its CFO saying numbers that don't match up 0:11:27.920 --> 0:11:30.560 with any of their leaks or their own marketing materials. 0:11:30.640 --> 0:11:32.840 In the aggressive and deluded way in which people ignore 0:11:32.880 --> 0:11:36.760 them is equal parts frustrating and depressing. So I want 0:11:36.760 --> 0:11:39.400 to speak to any AI boosters in the audience, and 0:11:39.440 --> 0:11:41.560 I want to speak with a little more empathy than usual. 0:11:42.559 --> 0:11:45.520 If you want Anthropic to win, you should be just 0:11:45.559 --> 0:11:48.360 as skeptical of these numbers as I am. You should 0:11:48.360 --> 0:11:50.360 want to smash my face in the tarmac with the 0:11:50.400 --> 0:11:53.760 most crystal clear impossible to argue with numbers, but reft 0:11:53.760 --> 0:11:58.679 of asterisks or discounts from suppliers or obfiscated accounting metrics. 0:11:59.240 --> 0:12:03.280 You should want better from your heroes. If you truly 0:12:03.280 --> 0:12:05.880 think this company is amazing, unstoppable and leading the tech 0:12:05.960 --> 0:12:09.199 industry to a glorious era of innovation, there shouldn't be 0:12:09.280 --> 0:12:11.800 this many questions and the metrics shouldn't be this murky. 0:12:12.559 --> 0:12:14.520 Every other time when a company has played this level 0:12:14.520 --> 0:12:18.040 of silly, weird bullshit, it's led to disaster. For example, 0:12:18.080 --> 0:12:20.280 we Work claimed to be profitable since the second month 0:12:20.360 --> 0:12:23.760 of its operations and repeated those claims of profitability throughout 0:12:23.760 --> 0:12:26.480 its existence until it turned out it was only profitable 0:12:26.520 --> 0:12:28.960 if you remove things like some of the costs of 0:12:29.040 --> 0:12:32.400 doing business. And I get where you're so defensive, and 0:12:32.440 --> 0:12:34.080 I get where you want this to work. A lot 0:12:34.080 --> 0:12:36.280 of you are very excited about generative AI, and being 0:12:36.320 --> 0:12:38.440 excited about it has given you a tremendous community of 0:12:38.480 --> 0:12:41.280 equally excited people. I get that you like these tools, 0:12:42.200 --> 0:12:43.880 and I need you to know that these companies are 0:12:43.920 --> 0:12:47.360 laughing at you. Anthropic timed this lead to focus on 0:12:47.400 --> 0:12:51.080 a specific quarter where it artificially suppressed costs and gave 0:12:51.080 --> 0:12:55.000 you the flimsiest proof imaginable, specifically crafted for you to 0:12:55.040 --> 0:12:57.160 share it as a triumph and spread the idea that 0:12:57.200 --> 0:13:01.280 AI labs are actually profitable when their core economics haven't changed. 0:13:01.800 --> 0:13:04.680 Costs increase linearly with revenue, and will continue to do 0:13:04.760 --> 0:13:08.160 so in perpetuity. I genuinely can't wait for Open AI 0:13:08.240 --> 0:13:10.439 and Anthropic to file that goddamn s once