0:02 Hey guys, it's Sasha. The hype around AI
0:05 exploded this week as OpenAI released
0:07 Sora 2, its latest video generation
0:10 tool. Tech bros and AI fans began
0:12 dribbling all over their keyboards as
0:14 they sang praises on Twitter. I tried to
0:16 download and try the app out, but it is
0:18 apparently not available outside the
0:19 United States. So, I had to rely on the
0:22 experiences of people who have appeared
0:24 to be able to actually try it out. And
0:26 distilling through the social media, it
0:28 appears that although there are definite
0:31 improvements, the story seems to be much
0:33 the same as it was before. To get one
0:36 short usableish clip, you typically have
0:39 to go through at least dozens and dozens
0:42 of completely unusable clips with major
0:44 errors in them. Here you can see a
0:46 couple of typical examples that I saw on
0:48 Twitter. First, there's a skateboarder
0:50 who seems to grow arms while performing
0:53 about a dozen shoulder dislocations in
0:55 midair. And then a boxing video with the
0:58 typical signs of AI slop with breaks and
1:00 continuity, weird morphing just doesn't
1:04 look at all right. And as per usual,
1:07 there is no way to edit existing clips
1:09 that is not coming for some time. For
1:11 example, making a very small alteration
1:14 to an already generated video. Maybe
1:15 there's one little detail that you want
1:17 to change. No, instead you have to
1:20 change the prompt that you put in and
1:22 hope that the next video, which will be
1:24 completely different to the previous
1:26 attempt, is good enough and doesn't have
1:28 issues. It appears that the videos that
1:31 come out are a lot better if the type of
1:34 video that you want can rely on very
1:37 similar pre-existing videos being out
1:40 there already that were used by OpenAI
1:42 in the training data set for the model.
1:44 So, if you just want to blatantly
1:46 plagiarize or copy something that's
1:48 already been done and just with slight
1:51 tweaks, then your chances of the video
1:53 not looking completely broken increase.
1:56 And naturally enough, this opened up the
1:58 usual contest on Twitter between people
2:00 who want to defend the rights of
2:02 copyright holders, you know, people who
2:04 have produced the original content for
2:06 which OpenAI does not pay them money but
2:08 then uses, people whose work has been
2:11 outright stolen in order to train the AI
2:14 model so that it can then spit out near
2:16 identical output for commercial gain.
2:17 And by the way, I was just reading just
2:20 before recording that apparently in the
2:22 first few days, OpenAI is already
2:24 restricting your ability to generate
2:27 videos with some characters, especially
2:29 Nintendo based characters because
2:31 apparently Nintendo really likes suing
2:33 people. But obviously for smaller
2:35 creators, you can just go and get
2:37 [ __ ] right? Nobody cares about you.
2:39 They'll just steal your stuff and let
2:41 the AM model train on it because you
2:43 will never have enough dollars to sue
2:46 them. Tech bros on the other side are
2:48 busy shouting very loudly that copyright
2:51 law is for boomers. AI companies are
2:53 allowed to do whatever they want, of
2:55 course, and completely ignore the law
2:57 because it's the future, bro. Although a
2:59 number of lawsuits have already been
3:01 launched against Open AI and other AI
3:03 companies by the likes of New York
3:05 Times, a bunch of other ones. The
3:07 progress of these lawsuits seems
3:11 extremely unusually slow and there seems
3:13 to be a complete lack of interest in any
3:15 of this from the government including
3:17 the Congress and the Senate. They don't
3:20 want to do anything about this even
3:22 though it is so brazen which of course
3:24 is very important to remember. Very
3:27 important has nothing absolutely nothing
3:29 to do with the fact that all of these
3:31 politicians stand to benefit financially
3:33 personally from these AI companies
3:35 seeing the valuations go to the moon
3:38 because they were lobbyed by those
3:39 companies. The politicians were lobbyed
3:41 to look the other way. Anyway, a few
3:44 days ago, we found out that OpenAI made
3:46 $4.3 billion worth of revenue in the
3:49 first half of 2025. That's a roughly $130%
3:50 $130%
3:52 increase on the revenue run rate that
3:55 they had in 2024, which is very
3:58 impressive. The growth in revenues is
4:00 doing really well. New people are
4:03 signing up, but at the same time, they
4:05 are still continuing to lose eye
4:07 watering amounts of money. While they
4:10 made the $4.3 billion worth of revenue,
4:13 they then also spent $6.7 billion on
4:15 research and development, another $2.5
4:17 billion on inference apparently, and
4:20 another $2 billion on sales and
4:22 marketing, which means that overall they
4:26 lost almost $7 billion while taking in
4:28 $4.3 billion from its customers. And
4:31 that's before accounting for capex
4:34 spend, which is probably dwarfing all of
4:36 the other costs put together because the
4:38 billions upon billions of dollars that
4:41 are being spent on building out the data
4:43 centers and the Nvidia graphics cards
4:45 that go into those gra data centers
4:49 comes on top of these costs. But the
4:51 cost of buying graphics cards, building
4:53 all of these giant data centers just
4:56 gets amatized over long periods of time.
4:58 Even though the likelihood is that these
5:00 graphics cards will get replaced over a
5:02 much shorter period of time as new ones
5:05 come out and as they just break and stop
5:08 being as valuable. So in 2025 before
5:11 accounting for the cost of all of this
5:12 capex the graphics cards which probably
5:14 makes the cost of inference
5:16 significantly higher than the published
5:21 numbers. OpenAI spent $260 for every $1
5:25 in revenue. In 2024, using the same data
5:27 that was shared with shareholders,
5:31 OpenAI apparently only spent $2.34 per
5:33 $1 of revenue. So, the gap between what
5:36 OpenAI spends and how much money it
5:38 makes is actually currently growing
5:40 instead of getting smaller. And
5:42 according to inputs from other
5:44 companies, it is the same story with
5:46 others as well. Economies of scale are
5:49 not applying here because there is a
5:52 physical cost to generating the extra
5:54 amount of compute required to process
5:57 queries in newer models like GBT5. And a
6:00 big reason here is that the models are
6:02 getting progressively better because the
6:04 algorithm just uses more and more
6:06 compute. It chucks more and more compute
6:09 at every query instead of the efficiency
6:12 of that compute increasing. So each time
6:15 you submit a query to Open AI, it then
6:16 generates a whole bunch of internal
6:19 subquery. So the model has a chance to
6:20 structure its answer. You know, when it
6:22 says thinking, is doing all of those sub
6:24 queries. It has a chance to
6:25 self-correct, to doublech checkck
6:27 things, to make sure that the answer is
6:30 exactly what the user is using, asking
6:32 for, etc. before it spits the answer
6:35 back to the user. And this approach
6:36 costs a lot of money. And every time a
6:38 new model comes out, it does more and
6:40 more of this other stuff, which costs
6:41 even more money. And each time that the
6:45 model gets better, the cost therefore of
6:47 running that model jumps as well. But as
6:50 everyone was busy looking at Sora 2,
6:53 OpenAI quietly closed another funding
6:56 round 2 days ago. OpenAI sold 6.6
6:58 billion of shares to select investors
7:00 and employees valuing the company at 500
7:02 billion. They apparently authorized up
7:05 to $10 billion worth, but only sold
7:08 about 2/3 of that. If OpenAI was a
7:09 public company, this valuation would
7:11 place them just above Exon Mobile and
7:14 Netflix as the 19th largest company by
7:17 market capitalization in the world. Just
7:18 for comparison, in the first half of
7:21 this year, Netflix made 21.5 billion
7:24 worth of revenue, exactly five times as
7:26 much as OpenAI. And Netflix has a net
7:29 margin of 28%, which means that their
7:33 net profit is well positive and growing
7:37 instead of minus 130%. and going lower.
7:39 Open AAI says that they have 17.5
7:41 billion dollars worth of cash as of the
7:44 end of the first half of the year. And
7:46 at the rate of capex and general cash
7:48 burn, they will probably need to do more
7:50 funding rounds in the next 12 months, 24
7:52 months, and so on. And this is precisely
7:56 where this story about AI hype gets
7:58 seriously weird and interesting. But
8:01 just before I dive into this bit, trust
8:03 me, it's crazy. I wanted to briefly
8:05 mention that my investor playbook course
8:08 is on sale right now, today and tomorrow
8:10 only. If you're watching this later than
8:13 that, I'm sorry. You'll have to wait for
8:14 another opportunity. The course is all
8:17 about analytical investing. It teaches
8:19 everything from technical investing
8:21 knowledge, all the jargon, a full
8:23 breakdown of how to read financial
8:25 statements in a lot of detail. There's
8:28 literally many hours, multiple hours of
8:31 content just on that alone. And then the
8:32 course covers how to build valuation
8:34 models, good practices, investing
8:37 tactics, and a lot more. You only pay a
8:39 one-off fee to join. There's no monthly
8:41 subscription, but then you get access to
8:43 all the new content as it comes for
8:45 free. I am expanding the course. I have
8:46 some big plans for the next couple
8:48 months. And you get to join live
8:51 sessions with me where you can ask me
8:53 questions. You can listen to a deep dive
8:55 on a different topic every week. And if
8:57 you can't join the live session, you can
8:58 always watch those later as they get
9:01 recorded. If you want to sign up, make
9:03 sure you use the code flash10 at the
9:05 checkout to get your 10% discount.
9:07 That's only available today and tomorrow
9:09 only. And thank you so much in advance.
9:11 I really look forward to seeing you in
9:13 the live sessions, including next week.
9:16 So, here is the bit that I found really
9:18 interesting about OpenAI. Just a few
9:20 days ago, Nvidia announced that they
9:25 will invest $100 billion into OpenAI. On
9:28 September 22nd, 2025, Nvidia announced
9:30 this $100 billion investment in OpenAI
9:34 to build 10 gawatt of AI data centers, a
9:37 scale that would require 4 to 5 million
9:40 GPUs, nearly double of Nvidia's annual
9:43 output. So, here is the way this works.
9:46 Early investors go and give a bunch of
9:49 money to OpenAI in funding. Open AAI
9:51 then goes and hands over that money to
9:53 Nvidia to buy a bunch of graphics cards
9:55 because that's their business model.
9:56 They use those graphics cards. Those
9:58 graphics cards are currently selling at
10:00 four to five times the price that they
10:03 were before the AI boom started. But it
10:05 still costs Nvidia the same amount of
10:08 money to produce them. So Nvidia sits
10:10 there printing cash. They make a giant
10:13 boatload of money. Nvidia's balance
10:15 sheet explodes. Their current assets go
10:19 from $16 billion to $12 billion in just
10:23 four years. But wait, even after that
10:26 meteoric rise, Nvidia still doesn't have
10:29 a hundred billion to invest in OpenAI.
10:31 Their total amount of cash and cash
10:32 equivalents and short-term investments
10:35 is $57 billion in their most recent
10:37 filing. So, what's happened here? Well,
10:40 it's a circular game. Imagine you have
10:43 this imaginary, completely non-real
10:47 company called company O. Company O
10:49 wants to raise their valuation so that
10:51 the founders can make a ton of money by
10:53 cashing in and selling their shares.
10:56 Their valuation is based on hype and a
10:58 essentially a multiple of the amount of
11:01 capex investment that they make. The
11:02 bigger the data centers, the more
11:04 graphics cards are in there, the more
11:06 their shares are worth. And you can see
11:09 how huge these data centers are getting.
11:11 OpenAI is already using more energy than
11:13 entire countries like Slovenia. And the
11:16 energy usage is only going up in a
11:18 vertical line. But we're of course not
11:20 talking about Open AI. We're talking
11:22 about this completely imaginary company
11:26 called Company O. So Company O loses a
11:28 lot of money. The amount of money they
11:32 lose goes up over time per user, per
11:34 dollar that they make. and doesn't the
11:36 company doesn't have enough cash to keep
11:38 buying millions of graphics cards to
11:40 stick into all those data centers that
11:41 they're building. So, the boss of
11:44 company O goes and has a little chat
11:47 with the boss of company N that makes
11:49 the graphics cards and they say, "Look,
11:52 we don't have enough cash to keep buying
11:54 your cards, and we're not expecting to
11:56 have the cash in the next few years, if
11:58 ever. So, how about we do a bit of
12:01 collusion and we give you shares in our
12:04 company in exchange for an investment of
12:06 $100 billion. We know that you don't
12:09 have $100 billion, but check this. You
12:12 give us just $1 billion to get us
12:15 started. We take it and then we give you
12:17 the $1 billion right back to buy the
12:19 graphics cards. You then take the $1
12:22 billion that we just gave you and you
12:24 give it back to us again, the second
12:26 billion. And we just rinse and repeat
12:28 this. a hundred times over and over
12:30 until the total investment is a h 100red
12:34 billion. In the end, shares of company O
12:36 are going to go absolutely through the
12:37 roof because we're going to have the
12:39 biggest data centers stuffed with
12:41 millions of company N's graphics cards.
12:45 And then company N can sell those shares
12:48 of company O that would make way more
12:50 money for them than by just selling the
12:52 graphics cards in the first place.
12:54 That's basically roughly what's
12:56 happening here. Nvidia is now not only
12:58 selling the picss to the people going up
13:01 the mountain in search for gold, but
13:03 they figured out a way of also making a
13:05 load of money from the Klondike hype
13:08 itself as well. Yeah, good for them. And
13:10 if you thought that it's really
13:12 beginning to take the piss, just listen
13:15 to the most recent Jeff Bezos interview
13:16 from 2 days ago.
13:17 One of the things that's going to happen
13:19 in the ne it's hard to know exactly when
13:21 it's 10 plus years, but I bet it's not
13:24 more than 20 years. We're going to start
13:26 building these giant gigawatt data
13:29 centers in space. So these giant
13:31 training clusters,
13:33 those will be better built in space
13:36 because we have solar power there 24/7.
13:39 Uh and it's and and the solar power
13:43 there is there no clouds and no rain, no
13:46 weather. So you can build we will be it
13:50 will we will be able to beat the cost of
13:52 terrestrial data centers in space in the
13:54 next couple of decades.
13:56 So according to Jeff Bezos in the next
13:58 few years companies will be building out
14:02 entire data centers like huge vast
14:04 buildings with millions of graphics
14:07 cards in them in space because you get
14:10 solar power 247 there. This of course
14:12 makes complete sense and is in no way
14:14 completely [ __ ] delusional. The fact
14:17 that you can use energy storage, nuclear
14:19 power, our other forms of power such as
14:21 wind or tidal for a fraction of the
14:24 price that it would cost to lift all of
14:26 the stuff you need into orbit and then
14:28 keep it in orbit. That fact is just
14:31 irrelevant. If you ever wonder if there
14:34 is a bubble happening at any point in
14:38 time, if you hear leaders in the space
14:40 not only continuously generating hype
14:41 because that's generally what they do,
14:44 but also making these types of
14:46 statements that are just clearly
14:48 insanely stupid. And when you hear at
14:51 the same time large volumes of fanboys
14:54 amplifying this and clapping wildly when
14:56 they hear this [ __ ] that's a pretty
14:58 shorefire way of knowing that you're
15:00 entering some kind of peak hype
15:01 territory. The problem with stating the
15:04 obvious, as I am right here, is that you
15:06 inevitably attract the tech bro crowd.
15:08 Just go and read the comments who will
15:10 loudly complain that you don't know what
15:11 you're talking about. You know, you
15:13 don't understand the future. It's the
15:16 technology, man. In fact, I am a really
15:19 big fan of AI technology. I use it every
15:21 day, just like I was a big fan and an
15:23 active user of the internet during the
15:25 com crash. And at the same time as being
15:28 a fan, I could also appreciate that the
15:30 company valuations and the hype around
15:32 those valuations was complete nonsense.
15:35 The two can happen at the same time. As
15:37 more people are posting clips of short
15:39 videos generated by Sora, you know that
15:41 one good clip that they generated after
15:44 generating hundreds of unusable clips,
15:46 the cost of generating these clips is
15:49 only going up. And there is a weird take
15:52 from everyone that soon there is going
15:56 to be entire movies created by AI. I
15:58 think we're nowhere close, but soon is
16:00 happening. It's happening soon. And we
16:02 will all want to watch these movies
16:04 apparently because you'll just be able
16:05 to watch exactly what you want. And we
16:07 will want to go and look at art
16:09 generated by AI. And we will want to
16:12 listen to music generated by AI. But
16:15 here is an interesting thoughtprovoking
16:19 example for you. Chess AI is complex but
16:22 way simpler than things like generating
16:24 a video. So chess AI has basically been
16:26 solved. It's been here a lot longer. It
16:28 happened a lot earlier. It became better
16:30 than humans of playing chess somewhere
16:32 around 20 years ago and today it's not
16:34 even remotely close. On every online
16:36 chess platform, there are chess bots
16:39 that you can play against. There are
16:41 chess bots for different skill levels,
16:43 for different types of games, chess bots
16:45 that imitate the style of famous chess
16:47 players or even famous non-chess
16:49 personalities, just celebrities, and you
16:52 can play against them whenever you want.
16:54 You don't have to wait for them to move.
16:56 You don't wait to have to be matched to
16:58 another player. You can take as long as
17:00 you need. They won't drop out of games.
17:01 They won't act like a dick to you
17:03 because they are a bot. They will always
17:06 say nice things. And you know what?
17:08 Absolutely nobody plays bots. There are
17:10 millions of people who play chess
17:12 online. None of them want to play bots.
17:14 All of the humans on these platforms
17:17 prefer playing other imperfect humans
17:21 because humans like and enjoy and seek
17:23 out the raw imperfection and the random
17:26 nature of how the human mind works. They
17:29 enjoy interacting with the real human at
17:31 the other end. Even if there is almost
17:34 no visual difference in terms of how the
17:37 chess game online looks on your screen,
17:39 I think about this a lot. These new
17:41 generations of AI are going to be
17:44 absolutely revolutionary in increasing
17:46 the productivity of humans. A good
17:48 developer can produce much higher
17:50 volumes of code at a higher quality than
17:53 before using AI. A CGI artist will
17:55 probably be able to do a lot more good
17:57 work in a shorter space of time by
18:00 incorporating AI in their workflow. It's
18:01 like the personal computer and the
18:03 smartphone revolution all over again.
18:06 But it can also be true at the same
18:08 exact time as acknowledging this that
18:11 the hype around the technology far
18:13 surpasses even these very lofty benefits
18:16 that are definitely coming. And the
18:18 valuations of companies in this space
18:20 can also be in a massive bubble even
18:23 though long-term AI is here to stay. It
18:25 will be here for decades and it will
18:28 benefit all of us long term. The S&P 500
18:30 is once again at all-time highs thanks
18:33 to this ongoing growing AI hype wave.
18:36 The Shellipe ratio has just broken 40
18:39 for the only the second time in history.
18:41 The only other time, the only previous
18:44 time being just before the dot crash
18:46 happened. Investors, the majority of
18:48 whom have only ever seen bull markets
18:51 and have never seen an actual stock
18:54 market crash. The two months in 2020
18:55 doesn't count. The last one happened in
18:58 2008, are tripping over each other.
18:59 These investors are just wanting to
19:01 invest in lossmaking companies at
19:04 increasingly absurd valuations. It will
19:06 be very interesting to see how all of
19:08 this unravels. Stay tuned. I will be
19:11 here to share my thoughts unfiltered on
19:12 everything to do with the economy and
19:14 with the stock market. I know I've been
19:17 busy over the last few months. Sorry. I
19:19 will make more content. Remember, my
19:21 course is on sale until midnight
19:24 tomorrow, Sunday, midnight New York
19:27 time. If you want to get get live weekly
19:29 sessions where you can ask me any
19:31 questions that you want or suggest
19:33 topics for deeper dives. If you want to
19:35 understand more about how to do
19:37 analysis, how to build valuation models,
19:39 how to look at companies as an analyst,
19:41 investing strategy in the good times and
19:44 the bad, you can get 10% off my investor
19:46 playbook course. Use the code flash 10
19:49 to get 10% discount. Only available
19:51 today and tomorrow. And thank you to
19:53 those of you signing up. It means a lot.
19:54 I really do appreciate it.
19:57 I will see you in the next live session next
Chrome Extension