Blue Wave politics, Anthropic compute strategy, and Moderna vaccine data discussed
Summary
Speakers analyze how Democratic congressional gains may struggle to reduce deficits due to presidential vetoes, while new governor elections could impact data center permitting. They highlight Anthropic's S1 filing showing training dominates compute usage, driving vertical integration to capture cloud margins. Additionally, they review Moderna's melanoma vaccine results and CVS Health's strategic spin-off of its retail business.
Episode sections
- 0:34 – 6:35
Oil Market Dynamics and Geopolitics
Speakers discuss oil traffic in the Strait of Hormus returning to pre-war levels despite high prices, attributing the $25/barrel price increase to rhetoric rather than supply/demand fundamentals. They predict a stalemate with Iran leading to a gradual price decline over 12-18 months as sanctions and embargoes persist without a nuclear inspection agreement.
- 6:35 – 9:14
US Political 'Blue Wave' and Fiscal Policy
The hosts analyze the Democratic gains in the House and Senate, noting a potential 40-50 seat House majority and a Senate near 60 seats. They argue that while Democrats will propose tax increases, President Trump's veto power will likely block them, creating a challenge for reducing the federal deficit despite the new legislative majorities.
- 9:14 – 10:30
Data Center Pipeline and Governor Elections
The election of Democratic governors in key states like Georgia, Iowa, Nevada, Ohio, and Texas raises questions about the permitting pipeline for data centers planned for 2028 and beyond. The speakers suggest this political shift might temper potential overbuilding of data center capacity due to regulatory coordination requirements.
- 10:30 – 11:07
AI Compute: Training vs. Inference and Anthropic's Strategy
Discussion centers on Anthropic's S1 filing showing 70% of compute is used for training, contradicting earlier assumptions that inference dominated. The hosts debate whether AI models will hit diminishing returns on training data or continue scaling, noting Anthropic's vertical integration into compute ownership to capture margins currently taken by cloud providers like AWS.
- 11:07 – 27:05
Healthcare: Medicare for All Speculation and CVS Split
The hosts joke about the possibility of a 'Trump Care' or Medicare for All plan emerging from the political landscape, though they note the difficulty of passing tax-funded expansions with a Republican president. They also discuss CVS Health's decision to spin off its pharmacy and retail business to focus on its insurance, PBM, and clinic operations, driven by state-level PBM regulations.
- 27:05 – 36:40
Moderna Cancer Vaccine Results and Oncology Updates
The speakers analyze leaked Phase 3 data for Moderna's melanoma vaccine, which showed a 42% reduction in metastasis/death risk, matching a prior study using Catruda. They explain that Moderna's stock dropped because the Catruda approach is cheaper and off-patent, though Moderna hopes for longer-lasting immunity. They also highlight positive news for lung cancer treatments cutting death risk by over 50%.
Assets discussed
| Asset | Kind | Stance | Context |
|---|---|---|---|
| Crude oil | commodity | neutral | The people that keep track of oil moving in or ships moving in and out of the Straits of Hormos say that the amount of oil traffic, and now increasingly products traffic to oil and jet fuel and stuff like that is getting to be the same or slightly exceed what it was before the war started in late February. |
| NVIDIANVDA | equity | neutral | And as as Mike was saying, difficulty of getting data centers through either local or state regulatory limits, is there a chance that between Nvidia and AMD and Samsung and Ix and Micron and Broadcom and Google with their own GPUs and Amazon with their own CPUs? |
| AlphabetGOOGL | equity | neutral | And as as Mike was saying, difficulty of getting data centers through either local or state regulatory limits, is there a chance that between Nvidia and AMD and Samsung and Ix and Micron and Broadcom and Google with their own GPUs and Amazon with their own CPUs? |
| AmazonAMZN | equity | neutral | And as as Mike was saying, difficulty of getting data centers through either local or state regulatory limits, is there a chance that between Nvidia and AMD and Samsung and Ix and Micron and Broadcom and Google with their own GPUs and Amazon with their own CPUs? |
| MicrosoftMSFT | equity | neutral | There's overlaps, obviously, in both, things like Microsoft Office. |
| UNITEDHEALTH GROUP INCUNH | equity | neutral | So the pharmacy was kind of dead weight for them, and they're cutting it off, and they're gonna turn into more of a united health group where they are the insurer, the PBM, and the point of care. |
“Stance” is how the speakers talked about the asset, not a recommendation.
Claims worth verifying
Oil traffic in the Strait of Hormus has returned to or slightly exceeded pre-war levels.
Mike Nicoletti ·0:34
How to verify: Check shipping data from sources like Kpler, Vortexa, or the US Energy Information Administration (EIA) for Strait of Hormus throughput data post-February 2026.
Crude supply is only off by 2 million barrels a day and demand is off by 1 million barrels a day compared to pre-war levels.
Mike Nicoletti ·1:54
How to verify: Review EIA Short-Term Energy Outlook or OPEC Monthly Oil Market Report for supply/demand balance sheets.
WTI crude was $60 at the start of the war and the 2027 strip price was $89.
Mike Nicoletti ·2:40
How to verify: Check historical WTI futures prices on Bloomberg or CME Group data for the relevant dates.
US commercial crude reserves are off by about 50 million barrels.
Mike Nicoletti ·2:32
How to verify: Check EIA Weekly Petroleum Status Report for US Commercial Crude Oil Inventories.
Democrats will have a 40-50 seat majority in the House.
Mike Nicoletti ·6:35
How to verify: Review official election results from the Clerk of the House or major news outlets' election projections.
Democrats will get close to 60 seats in the Senate, likely 57.
Mike Nicoletti ·7:29
How to verify: Review official election results from the Senate.gov or major news outlets' election projections.
Democrats won every statewide election in Texas, including Governor and Lieutenant Governor.
Mike Nicoletti ·7:51
How to verify: Review Texas Secretary of State election results for the 2026 election cycle.
Anthropic's S1 filing shows 70% of compute capacity is used for training versus inference.
Jason Wallace ·10:49
How to verify: Read Anthropic's S-1 registration statement filed with the SEC, specifically the 'Use of Proceeds' or 'Business' sections detailing compute allocation.
Moderna's Phase 3 melanoma vaccine showed a 42% reduction in risk of metastasis or death.
Mike Nicoletti ·27:47
How to verify: Review Moderna's press releases or clinical trial data presentations from the European oncology conference mentioned.
A separate study using Catruda showed a 42% reduction in risk of metastasis or death.
Mike Nicoletti ·29:27
How to verify: Review the published study or press release for the Catruda (likely Keytruda/pembrolizumab) timing adjustment study.
Do your own research
- Verify oil supply/demand data from EIA or IEA reports to assess geopolitical vs. fundamental price drivers.
- Review US Congressional budget proposals and presidential veto history to evaluate deficit reduction feasibility.
- Check state-level permitting regulations and governor election results in TX, OH, NV, GA, IA for data center impact.
- Analyze Anthropic's S1 filing for compute allocation breakdown and vertical integration strategy details.
- Compare Moderna's melanoma vaccine Phase 3 data with Catruda therapy efficacy and cost structures.
- Review CVS Health's spin-off announcement and state-level PBM ownership restrictions.
Full transcriptexpand
- 0:00
- Unidentified speaker: Welcome to Telltales.
- 0:01
- Unidentified speaker: I'm Mike Nicoletti.
- 0:02
- Unidentified speaker: Every week, along with my co-hosts Jason Wallace and Hunt Lawrence, we bring you the cash flow memo.
- 0:07
- Unidentified speaker: Get it at telltales.us.
- 0:10
- Unidentified speaker: For the next 30 minutes, we're going to dive deep into energy, technology, and healthcare.
- 0:15
- Unidentified speaker: So download the memo now and get started.
- 0:18
- Unidentified speaker: The following conversation is intended for informational purposes only.
- 0:22
- Unidentified speaker: You should always do your own work to determine if an investment is suitable for you.
- 0:34
- Unidentified speaker: The people that keep track of oil moving in or ships moving in and out of the Straits of Hormos say that the amount of oil traffic, and now increasingly products traffic to oil and jet fuel and stuff like that is getting to be the same or slightly exceed what it was before the war started in late February.
- 1:03
- Unidentified speaker: Price stays stubbornly high, but I think part of that was the rhetoric coming from Donald J.
- 1:13
- Unidentified speaker: Trump and also from the spokesman for the Revolutionary Guards who were in control in Iran.
- 1:22
- Unidentified speaker: I think with the Blue Wave, or whatever we want to call it, an election, I think there's a chance that the rhetoric on the US side goes down.
- 1:34
- Unidentified speaker: And I think the Iranians will probably reduce their rhetoric.
- 1:42
- Unidentified speaker: And the I think that the other people who keep track of production and demand are pretty much in line with these 26 numbers.
- 1:54
- Unidentified speaker: So the remarkable thing is, with all the publicity and up in prices, if you look at crude supply, it'll only be off about two million barrels a day from 103 to 101.
- 2:10
- Unidentified speaker: And oil demand or liquid fuel consumption is off by about a million barrels a day.
- 2:20
- Unidentified speaker: Inventory draw is around a million barrels a day.
- 2:24
- Unidentified speaker: Now it's a lot of oil.
- 2:25
- Unidentified speaker: A million barrels a day is 360 million barrels in a year.
- 2:30
- Unidentified speaker: And in fact, the U.S.
- 2:32
- Unidentified speaker: commercial reserves, which is the bottom here, are off about 50 million barrels.
- 2:40
- Unidentified speaker: So it's prices at the beginning of the war for WTI were $60.
- 2:51
- Unidentified speaker: And they traded, you know, the the 27.
- 2:55
- Unidentified speaker: Well, if we go to the bottom of Exhibit B, the last Thursday, it'll be down some today, but the 27 strip price was 89.
- 3:07
- Unidentified speaker: So there's from you know there's a full $25 a barrel, and all that happened was supply went down by 2 million barrels, and demand went down by a million barrels, and inventory was drawn a million barrels.
- 3:24
- Unidentified speaker: It seems like more should have happened to have the price for 27 be $25 higher than it was the price before the war.
- 3:37
- Unidentified speaker: But that's why I think this will calm down and pricing will be somewhat lower, maybe closer to 90 for the near month rather than 100.
- 3:51
- Unidentified speaker: And the reason I for confidence that the US won't do additional substantial bombing in Iran, and Iran won't respond by putting missiles and drones into gulf fire facilities, desalinization plants, crude loading facilities, refineries, and whatnot, is no one benefits from either one of those things happening.
- 4:18
- Unidentified speaker: I think that the possibility of Iran and the US having an agreement is hard because what the U.S.
- 4:30
- Unidentified speaker: wants is 24-hour day, seven-day-a-week inspection by IEAA, which I think is an extension of the UN, that does this on a worldwide basis, kind of checking up on people's how they're handling their nuclear fuels, that is something the Iranians are probably not going to agree to.
- 4:54
- Unidentified speaker: And I think until they agree to that, we'll have the sanctions and the embargo.
- 5:00
- Unidentified speaker: So it's a bit of a stalemate, but I think if the stalemate continues, I know I'm not predicting that things quiet down and the price of oil starts to trade at you know 80 or 75 or something.
- 5:15
- Unidentified speaker: But but look, if we look at the exhibit B, the the 28th price is 78.
- 5:22
- Unidentified speaker: Now the 27th price of 89 by today or whenever, you know, next tomorrow, they'll trade down to 85, 86, and the 78 will trade for 72 or 3.
- 5:35
- Unidentified speaker: So what people participants in the oil market are going to say, if you give it enough time, if you go to 12 or 18 months, chances are somehow Iran and the U.S.
- 5:45
- Unidentified speaker: reach some kind of accommodation.
- 5:47
- Unidentified speaker: That's what they're predicting.
- 5:49
- Unidentified speaker: And I think if you give it 12 or 18 months, that'll probably happen.
- 5:53
- Unidentified speaker: Iran's under a lot of pressure, they have a lot of inflation, but there are 90 million people there.
- 5:59
- Unidentified speaker: They're all being fed.
- 6:00
- Unidentified speaker: It isn't like North Korea where they have famines and stuff.
- 6:05
- Unidentified speaker: And to a certain extent, sanctions make them stronger because they have to, I mean, they have their own car industry, they have their own copper bonds.
- 6:13
- Unidentified speaker: I mean, it's quite remarkable what the Iranians have been able to do under sanctions.
- 6:18
- Unidentified speaker: I think the futures market has it about right.
- 6:20
- Unidentified speaker: If you give it 12 or 18 months, probably things will things will be some type of uh room for the Iranians to make an agreement, for us to make an agreement and for us to lift the embargo and lift sanctions.
- 6:35
- Unidentified speaker: But that to think that's gonna happen next month or the month after or six months from now is, I think, way optimistic.
- 6:42
- Unidentified speaker: As far as the exhibit aid goes, the cash flow statement for the government, the Democrats are gonna have a 40-50 seat majority in the House.
- 6:52
- Unidentified speaker: Once they finish counting all the House votes, that's extraordinary.
- 6:57
- Unidentified speaker: I don't think anyone's had that big a majority in a long time.
- 7:01
- Unidentified speaker: That'll allow Hakeem Jeffries as the Speaker of the House to be pretty active.
- 7:08
- Unidentified speaker: He won't have to worry about DSA's Democratic Socialists.
- 7:13
- Unidentified speaker: They will they're there he won't need them on most votes.
- 7:18
- Unidentified speaker: The Senate is different because any senator can stop any bill, but the the Democrats are gonna get pretty close to 60.
- 7:29
- Unidentified speaker: 60 you can close off debate.
- 7:31
- Unidentified speaker: And I think that it looks like the Democrats are gonna get to 57.
- 7:35
- Unidentified speaker: I mean, the scale of the losses or the gains by the Democrats went way beyond what the polls were predicting.
- 7:45
- Unidentified speaker: And for example, I mean, the people I interact with every day very often live in Texas.
- 7:51
- Unidentified speaker: The Democrats won every statewide election that happened yesterday.
- 7:59
- Unidentified speaker: Not only the governor, their lieutenant governor, it's one spot on the railroad commission.
- 8:04
- Unidentified speaker: I mean, it's just a complete wipeout of the Republicans.
- 8:08
- Unidentified speaker: What does this mean from the point of view of the cash flow statement on Xibiday?
- 8:14
- Unidentified speaker: The Democrats with healthy majorities in both the House and the Senate are going to have to be part of the solution.
- 8:22
- Unidentified speaker: And we simply cannot go back to the overspending that we had in 2021.
- 8:28
- Unidentified speaker: And we really need to curtail spending, they will propose increased taxes on corporations and on individuals.
- 8:39
- Unidentified speaker: But as long as Trump is, or some replacement, Republican replacement for Trump is president for the next two years, I don't think any bills increased taxes will escape the veto.
- 8:53
- Unidentified speaker: So how to get this cash flow statement so it doesn't help their shape is gonna be a real a real challenge.
- 9:01
- Unidentified speaker: And that's one it won't only be up to the Republicans now.
- 9:05
- Unidentified speaker: The Democrats control both houses, so it's something they're gonna have to topless as well.
- 9:10
- Unidentified speaker: Mike Jason, anything else to add on the blue wave?
- 9:14
- Unidentified speaker: Well, the let's talk about data centers because the blue wave wasn't just Congress, as you said.
- 9:21
- Unidentified speaker: It's also governors.
- 9:23
- Unidentified speaker: There were five races that I don't know if they're totally called yet, but I think they're all the margin's pretty high so far.
- 9:30
- Unidentified speaker: Georgia, Iowa, Nevada, Ohio, and Texas have all gone Democratic governors.
- 9:37
- Unidentified speaker: And that brings into question the pipeline for data center capacity that's expected for 28 and onward.
- 9:44
- Unidentified speaker: My understanding is pretty much everything planned for 2027 already has permits, but getting new permits will require some level of coordination from the governor's office as well as the ISO, the utilities in the area as well.
- 9:59
- Unidentified speaker: You know, that's the numbers I have for total pipeline are actually pretty high if you add those three or five states up.
- 10:08
- Unidentified speaker: I mean, Illinois at 33.
- 10:10
- Unidentified speaker: Yeah, there's quite a bit of data centers in planning.
- 10:14
- Unidentified speaker: So some of that gets thrown into question.
- 10:16
- Unidentified speaker: And maybe, you know, the a silver lining might be that we may be tempering a potential overbuild, because I think everyone's having a bit of trouble estimating at what point demand caps out.
- 10:30
- Unidentified speaker: And I think that's it's a reasonably difficult question to answer because as the models get better, new demand is created.
- 10:37
- Unidentified speaker: But as we saw from the Anthropic S1 and the updated version with Q3 data, about two-thirds, 70% of their compute capacity goes towards training as opposed to inference.
- 10:49
- Unidentified speaker: That kind of gives you an idea of how much is going into training these data centers versus just doing inference.
- 10:55
- Unidentified speaker: So all that is to say is that we're still building data centers the next couple of years, but 28 and 29 start getting thrown into question.
- 11:04
- Unidentified speaker: Jason, what about health care?
- 11:07
- Unidentified speaker: Mike and I were joking right before the call that Trump has the opportunity here to enact kind of the Medicare for All plan that we've been discussing a little bit.
- 11:19
- Unidentified speaker: And you know, the Democrats can put it forward, but they'd have to concede the name and it would have to be Trump care, and then maybe he signs on to it.
- 11:27
- Unidentified speaker: But broadly, healthcare is getting a bump today.
- 11:30
- Unidentified speaker: Absolutely.
- 11:31
- Unidentified speaker: Yeah.
- 11:32
- Unidentified speaker: Yeah.
- 11:33
- Unidentified speaker: Medicaid Medicaid spending should continue, um, maybe at the levels that were kind of you know COVID emergency level.
- 11:43
- Unidentified speaker: Yeah, I I believe the solution is sound like a broken record on this.
- 11:49
- Unidentified speaker: Bernie Sanders and Donald J.
- 11:51
- Unidentified speaker: Trump at the same podium announcing Medicare for everyone.
- 11:55
- Unidentified speaker: And by the way, since everyone's gonna be covered by Medicare, we can scrap the Medicaid program, we can scrap the the uh the Obamacare subsidies.
- 12:07
- Unidentified speaker: Uh just and then and then we'll have single-payer health care.
- 12:13
- Unidentified speaker: You you'd have to imagine that idea is gonna be put forward in the next two years.
- 12:17
- Unidentified speaker: Yeah, you would think.
- 12:19
- Unidentified speaker: And I I do think it would be less of a drain on the federal budget if we did that.
- 12:26
- Unidentified speaker: I think to make it palatable, you'd have to permit employer health plans, private, private insurance.
- 12:36
- Unidentified speaker: I mean, Medicare recipients can get you know extra insurance if they want to and they want to pay for it.
- 12:44
- Unidentified speaker: I think our system is so expensive and so uh unproductive and costs so much relative to anywhere else in the world.
- 12:55
- Unidentified speaker: I I just think it it's almost inevitable.
- 12:59
- Unidentified speaker: But we'll we'll see.
- 13:02
- Unidentified speaker: What let me let me pose a question.
- 13:06
- Unidentified speaker: If we look at the front end of the memo and you know, with the benefit of which we discussed last week of that of seeing the uh seeing the anthropic S1, is there a chance with electing all these Democrats?
- 13:21
- Unidentified speaker: And as as Mike was saying, difficulty of getting data centers through either local or state regulatory limits, is there a chance that between Nvidia and AMD and Samsung and Ix and Micron and Broadcom and Google with their own GPUs and Amazon with their own CPUs?
- 13:48
- Unidentified speaker: And is there a chance that we'll have way more chips than we have data center sites where we can put them?
- 13:56
- Unidentified speaker: Is that I think in the next year, but 28 or 29.
- 14:02
- Unidentified speaker: I mean, is that something that might happen?
- 14:04
- Unidentified speaker: I'm gonna I'm gonna use something you've said about the energy industry and that the the energy industry is very efficient in finding a way.
- 14:11
- Unidentified speaker: I think the semiconductor industry is similarly quite efficient.
- 14:15
- Unidentified speaker: So if there is demand somewhere, they will figure out how to get it to you.
- 14:19
- Unidentified speaker: And if that means data centers in space, it might be data centers in space.
- 14:26
- Unidentified speaker: If it means people buy inference machines for themselves, that's maybe we go that way.
- 14:34
- Unidentified speaker: But there's the we if we decide that we're not going to do data centers in the US at a level that's sufficient to meet demand, we will find other ways to get that inference capacity and training capacity on the ground.
- 14:51
- Unidentified speaker: Chase and I were talking earlier that one of the unfortunate maybe impacts of that is that it wouldn't be as accessible if you would, you know, instead had to purchase, as a consumer, have to purchase a machine to do this inference.
- 15:05
- Unidentified speaker: Or a company may set up their own data center to do this stuff.
- 15:10
- Unidentified speaker: It's a couple years out, a lot of things can change in a short period of time.
- 15:15
- Unidentified speaker: But I'm confident the semiconductor industry finds a way.
- 15:18
- Unidentified speaker: I got a question for Jason.
- 15:20
- Unidentified speaker: He's the expert between the three of us, especially expert versus May.
- 15:25
- Unidentified speaker: Jason, how likely is it that Anthropic then open AI and Gemini and Groc and Muse?
- 15:36
- Unidentified speaker: How likely is it that they're going to need the same amount of tokens that they're using now for training?
- 15:44
- Unidentified speaker: It's remarkable that Anthropic is used, you know, two-thirds of their available compute is used for training.
- 15:50
- Unidentified speaker: But will that slow down at some point?
- 15:53
- Unidentified speaker: Or what how how do we think about the need for compute to train as compared for inference, which is, you know, what we're you know, where in the end we've got to have inference and we've got to have people willing to pay for the tokens or the apps or whatnot.
- 16:11
- Unidentified speaker: But that two-thirds, one-third ratio is is that the way it's gonna look two years from now or three years from now, or is it uh does everyone get their models to a you know fixable level and stop spending so much money to train them?
- 16:25
- Unidentified speaker: Or what's your view?
- 16:27
- Unidentified speaker: Yeah, so I thought by the beginning of this year we had crossed a threshold where more compute was now being allocated to inference than training.
- 16:38
- Unidentified speaker: The Anthropic S1 clearly shows that my assumption was wrong.
- 16:43
- Unidentified speaker: And you look at their contracted compute for the future, and clearly they're betting that future frontier models continue to scale as they have in the past, which does mean exponentially more compute is required to generate and train that model.
- 17:03
- Unidentified speaker: So they're betting that that continues.
- 17:06
- Unidentified speaker: My belief is at some point, probably not far from where we are now, we hit diminishing returns for building ever better models.
- 17:18
- Unidentified speaker: My kind of theory is that a language model will be the expert in any given field.
- 17:27
- Unidentified speaker: And then once it's trained up, that it's the expert in every field, you know, it's the top, it's the top human-level knowledge in every field.
- 17:36
- Unidentified speaker: Uh we're kind of at the point where improving it becomes more difficult.
- 17:41
- Unidentified speaker: And really, one way we look at the these language models is that they're just the most efficient compression algorithm that we've ever invented.
- 17:49
- Unidentified speaker: It can, you know, it statistically picks words and it can regenerate ideas that it's been trained on from every part of the internet and every document it's gotten its hands on.
- 18:01
- Unidentified speaker: So we see Anthropic and OpenAI doing, you know, some kind of questionable things to get their hands on ever more data.
- 18:10
- Unidentified speaker: And at what point do they run out of data?
- 18:13
- Unidentified speaker: They've trained on it all.
- 18:15
- Unidentified speaker: They're synthetically generating data.
- 18:19
- Unidentified speaker: I think if when the majority of the data is generated by a language model already, then what happens to the quality of the models when you're just training on the output that you've already generated?
- 18:32
- Unidentified speaker: So I do think we hit a wall at some point.
- 18:36
- Unidentified speaker: Okay, but there's another angle to that.
- 18:38
- Unidentified speaker: If you look at the compute spend versus the inference spending in Anthropic over the last few years, the data that we have, the growth rate and the cost of training is growing faster than the growth rate and the cost of inference.
- 18:53
- Unidentified speaker: So that's one of the biggest negatives for a potential investor, you know, at least from one of the one of my reads through, is that this thing basically requires ever more capital to go into training in order to maintain the lead they have in order to be rewarded with the high margin inference dollars.
- 19:14
- Unidentified speaker: 100%.
- 19:15
- Unidentified speaker: Yeah.
- 19:15
- Unidentified speaker: And when there's a whole bunch of other people competing for that, that's bad.
- 19:20
- Unidentified speaker: That's why you might really like a situation where that market is more regulated because you could slow that pace.
- 19:28
- Unidentified speaker: So will they require more training in the future?
- 19:31
- Unidentified speaker: I'm not gonna make a statement on I'm not gonna make the call that Jason made.
- 19:36
- Unidentified speaker: I think you could make an equally different answer that says we can multiply the advancement of human knowledge much faster with these tools.
- 19:47
- Unidentified speaker: Pick whichever direction you want to go.
- 19:49
- Unidentified speaker: You can make the case that we're gonna peter out in this in how capable these modern models are, or you can make the case that this thing just keeps going.
- 19:56
- Unidentified speaker: I don't think I don't think the capabilities stop.
- 19:59
- Unidentified speaker: I think the model the requirement to train.
- 20:03
- Unidentified speaker: Yeah, the requirement to train and the language model hits diminishing returns, but everything built around it continues to make it more.
- 20:12
- Unidentified speaker: But as as training gets cheaper, we will train models for specific tasks because the inference costs will then be much lower.
- 20:20
- Unidentified speaker: And the the the example today is what did it cost?
- 20:24
- Unidentified speaker: Like $20 million to train the first GPT model?
- 20:27
- Unidentified speaker: And today you can do it for under $100.
- 20:30
- Unidentified speaker: Right.
- 20:30
- Unidentified speaker: So again, think three years out, what we have today will seem like a toy relative to what we have in three years.
- 20:41
- Unidentified speaker: Well, let me put it another way for Jason, and you might Claude is the most widely used, or it's the app that generates the most revenue.
- 20:50
- Unidentified speaker: Otherwise, how does Andropic have 17 billion of revenue on the third quarter?
- 20:56
- Unidentified speaker: So it's Claude work and Claude Code and the other Claude products.
- 21:01
- Unidentified speaker: For Claude to continue to have that revenue, does Andropic have to have the best or almost the best model?
- 21:12
- Unidentified speaker: Or could they pull back a little bit on their training and still be have lots of revenue from Claude apps?
- 21:21
- Unidentified speaker: No, they have to have one of the best models.
- 21:24
- Unidentified speaker: Because open source is always six, eight months behind them.
- 21:30
- Unidentified speaker: And then when you get to the point of if anthropic slows down and they don't keep that lead, eight months from now, the open source model is equally as capable, and you don't have to pay the margin to anthropic.
- 21:43
- Unidentified speaker: Your cost for generating tokens has been cut in half.
- 21:48
- Unidentified speaker: And you're gonna you're gonna use that one.
- 21:50
- Unidentified speaker: And you may use that one through a different harness.
- 21:54
- Unidentified speaker: And that's part of it.
- 21:55
- Unidentified speaker: Is it in software, two of the most important powers, if you're think of like what enables the company to earn excess profits over the long run are network effects and switching costs.
- 22:08
- Unidentified speaker: Switching costs predominantly being a function of enterprise software and network effects being a function of consumer software.
- 22:17
- Unidentified speaker: There's overlaps, obviously, in both, things like Microsoft Office.
- 22:20
- Unidentified speaker: So as much as they can, they should and will continue to try to build up these switching costs and network effects.
- 22:26
- Unidentified speaker: So half their business has that, but the other half, the tokens are generated through a cloud service provider.
- 22:32
- Unidentified speaker: And I think that that part is really at risk.
- 22:35
- Unidentified speaker: And also the margin that they receive there, I think, is a little bit of deceptive accounting on their part.
- 22:43
- Unidentified speaker: I think really they're not receiving an 85% margin.
- 22:48
- Unidentified speaker: It's more like 50%.
- 22:50
- Unidentified speaker: Well, here's a follow-up question, Jason and Mike, too.
- 22:54
- Unidentified speaker: Wouldn't that drop it be a better financial business if they weren't building their own compute, if they were just leasing compute?
- 23:03
- Unidentified speaker: Apparently they're the one of the reasons.
- 23:05
- Unidentified speaker: I mean, they have a hundred billion dollars on hand.
- 23:08
- Unidentified speaker: They want to do $100 billion IPLs.
- 23:11
- Unidentified speaker: What do you need $200 billion for if you're cash flow positive based on your, you know, on your current revenue or on your growing revenue?
- 23:20
- Unidentified speaker: And I think what basic need is they want to have their own data centers, their own compute as compared to dealing with Amazon, who was their early backer or Google or Microsoft or anyone else.
- 23:40
- Unidentified speaker: But I I Mike was explaining this to me on one of our morning calls.
- 23:46
- Unidentified speaker: And what Mike's answer was is that the data center or the cloud provider was taking half of the margin.
- 23:55
- Unidentified speaker: Is that a fair analysis, do you think?
- 23:58
- Unidentified speaker: I know the two of you have been working on that, but is that the logic of Andropic spending all this money to have their own capacity?
- 24:06
- Unidentified speaker: I think so.
- 24:07
- Unidentified speaker: Yeah.
- 24:07
- Unidentified speaker: So do you want me to walk through that?
- 24:10
- Unidentified speaker: Yeah, I think you should walk through the analysis.
- 24:12
- Unidentified speaker: What's happening is they sell through AWS or Azure or Google.
- 24:17
- Unidentified speaker: And if they sell for every hundred dollars of token service, that cloud is taking, I think it was 35 to 37% out for their marketplace fees because you're going through AWS Marketplace to book the compute to generate the tokens, you know, quote unquote from Anthropic, even though they're you're licensing the software in reality and it's running in Amazon.
- 24:42
- Unidentified speaker: So Anthropic books $100 in revenue when that happens, even though they receive $65.
- 24:49
- Unidentified speaker: And then from that, they at the end of the month, they get their bill from Amazon.
- 24:54
- Unidentified speaker: $15 of compute was used.
- 24:56
- Unidentified speaker: They pay that to Amazon.
- 24:57
- Unidentified speaker: So end of the day, Amazon has $50, Anthropic has $50, but Amazon still is, they have to, they're paying for their own compute out of that as well.
- 25:06
- Unidentified speaker: Right.
- 25:06
- Unidentified speaker: That $15 is an SGNA instead of cost of goods sold.
- 25:12
- Unidentified speaker: It's yeah, it's marketing spend is how they book it.
- 25:15
- Unidentified speaker: I think that's one of the reasons, but all but also Anthropic, the leadership there very clearly thinks they're building towards this artificial general intelligence.
- 25:27
- Unidentified speaker: And the risk they repeat over and over in their S1, and when they speak publicly, is they won't have enough compute.
- 25:33
- Unidentified speaker: So I think they fully believe in a future where they've taken over all of the white collar jobs.
- 25:39
- Unidentified speaker: They're doing all the knowledge work.
- 25:42
- Unidentified speaker: And I think that's what they're working towards.
- 25:44
- Unidentified speaker: So I think they they believe they need all the compute in the world because they will be doing all of the work in the world.
- 25:51
- Unidentified speaker: Do you do you agree with that?
- 25:52
- Unidentified speaker: That's what Dario says.
- 25:54
- Unidentified speaker: Yeah, if you take what they say literally, that's yeah.
- 25:57
- Unidentified speaker: Yeah.
- 25:59
- Unidentified speaker: No matter what they say in public, or no matter what they put on X, or no matter what they what stance they're taking to caution to tell people not to worry about them taking over all white collar work.
- 26:15
- Unidentified speaker: That's really what they think is where you see it, Jason.
- 26:19
- Unidentified speaker: Yeah, I mean that that's what they that's what they say.
- 26:23
- Unidentified speaker: Dario has said he believes he will be the only company in existence in the future, which is mind-blowing.
- 26:30
- Unidentified speaker: Who are their customers?
- 26:32
- Unidentified speaker: Now we've got to do some healthcare, Jason.
- 26:36
- Unidentified speaker: How do we I mean, does that mean you can run Lowly with 200 researchers and apps developed on a language model to take care of everything else in Lowly?
- 26:48
- Unidentified speaker: I think you need some robots doing the wet lab work as well.
- 26:51
- Unidentified speaker: But they're working on that too.
- 26:53
- Unidentified speaker: They're working on that too.
- 26:55
- Unidentified speaker: Yeah.
- 26:56
- Unidentified speaker: We did we talk about the no the the the leaked results out of the Moderna cancer vaccine.
- 27:03
- Unidentified speaker: So oh, and we have to cover that.
- 27:05
- Unidentified speaker: We have to cover that.
- 27:06
- Unidentified speaker: We'll we'll talk about whether you can run lowly with 200 200 Einstein type of people, we'll put that on till next week.
- 27:14
- Unidentified speaker: But let's consider the Moderna results because we yeah, we that we missed that last week.
- 27:21
- Unidentified speaker: Yeah, so as as part of the conference that happened a week or two ago in in Europe, the some data got leaked.
- 27:29
- Unidentified speaker: So when Moderna announced their success of their phase three cancer vaccine study for melanoma, it was they claimed success.
- 27:40
- Unidentified speaker: The market rewarded them for that.
- 27:42
- Unidentified speaker: And part of the press release said that they saw similar results as they did in phase two.
- 27:47
- Unidentified speaker: The phase two study showed that the vaccine had a 60% reduction in risk of metastasis or death from the cancer.
- 27:58
- Unidentified speaker: What the leaked results of the phase three study showed was that reduction in risk was really 42%.
- 28:05
- Unidentified speaker: So still effective.
- 28:07
- Unidentified speaker: The reason they sold off so hard and this is viewed as a failure is because another study that occurred last year produced similar results using a different method.
- 28:19
- Unidentified speaker: So what that study did was take Catruda, which is a engages your immune system to fight the cancer in a way essentially it's preventing the cancer from hiding from your immune system.
- 28:32
- Unidentified speaker: So the standard of care is that you'll have a tumor surgically removed and then start taking Catruda and you take a course.
- 28:40
- Unidentified speaker: So what the study did was look at if you bump that up, I can't remember if it's one week, two weeks, but you bump that up to where the course of Catruda starts before the tumor is removed, essentially your immune system is being primed to find it, recognize the cancer, and then build the memory of how to fight it.
- 29:01
- Unidentified speaker: So in both cases, whether it's your immune system learning from the actual tumor itself or the mRNA vaccine being built based on your cancer and then given to you to prime your immune system.
- 29:15
- Unidentified speaker: The end result was that your immune system knows what the cancer looks like.
- 29:19
- Unidentified speaker: So, in a fate of statistics, that study also showed exactly a 42% reduction in risk.
- 29:27
- Unidentified speaker: So reproduce the same exact result, two different methods.
- 29:32
- Unidentified speaker: The problem for Moderna now is that the Katruda version is just you bump the timing of the course of that medication up earlier versus Moderna was targeting a three or four hundred thousand dollar price tag for this cancer vaccine.
- 29:49
- Unidentified speaker: Katruda will be off patent.
- 29:54
- Unidentified speaker: Yeah.
- 29:55
- Unidentified speaker: There's follow-on versions, but yes.
- 29:56
- Unidentified speaker: So those things are only gonna get sold in the United States.
- 29:59
- Unidentified speaker: The cancer vaccine?
- 30:00
- Unidentified speaker: Yeah.
- 30:01
- Unidentified speaker: Yeah, yeah.
- 30:02
- Unidentified speaker: If at all, if at all.
- 30:04
- Unidentified speaker: Yeah.
- 30:04
- Unidentified speaker: And so what Moderna is really leaning on now is they hope that your immunity against, you know, your immune system's recognition of the cancer cells lasts far longer using their vaccine than than the Katruda, you know, bumping the therapy up.
- 30:22
- Unidentified speaker: And you can redose it.
- 30:24
- Unidentified speaker: So there's ways you to extend the efficacy, but they're studying that now, and that's going to take a lot more time.
- 30:33
- Unidentified speaker: And then just to repeat something you said last Wednesday or the Wednesday before is that the use of cancer vaccines, this technology on other cancers has been disappointed by and other participants.
- 30:48
- Unidentified speaker: And so, you know, it's not as it's not as it's not as notable or as optimistic for cancer patients and their families as would be ideal.
- 31:05
- Unidentified speaker: Yeah, it's proving to not be as effective as we had all hoped.
- 31:09
- Unidentified speaker: It definitely works.
- 31:10
- Unidentified speaker: It just not in every cancer and maybe not more effective than we can get using other methods.
- 31:18
- Unidentified speaker: Ryan.
- 31:19
- Unidentified speaker: But there was a lot of good news.
- 31:20
- Unidentified speaker: There's a tremendous amount of good news that came out of that oncology conference.
- 31:24
- Unidentified speaker: One of the one of the other big pieces of news was treatments for lung cancer.
- 31:29
- Unidentified speaker: There were two treatments announced.
- 31:31
- Unidentified speaker: They both cut the risk of death by lung cancer by greater than 50%.
- 31:36
- Unidentified speaker: One's for small cell lung cancer, the other's for non-small cell.
- 31:40
- Unidentified speaker: So it's great.
- 31:41
- Unidentified speaker: I believe both of those were ADCs, so antibody drug conjugates.
- 31:45
- Unidentified speaker: It's a you can think of it as a super-targeted chemotherapy.
- 31:50
- Unidentified speaker: So the big win there is that lung cancer produces the most deaths today, I believe.
- 31:57
- Unidentified speaker: It's not the most prevalent form of cancer, but it's harder to treat.
- 32:01
- Unidentified speaker: So it ends up having a higher mortality rate today.
- 32:05
- Unidentified speaker: So a lot of great news.
- 32:07
- Unidentified speaker: You know, it wasn't just the disappointment for Moderna, but there's a lot of great news in the world of oncology.
- 32:13
- Unidentified speaker: Uh the big news today is CVS announced they're going to split up the pharmacy and retail business from their the rest, you know, the the bigger part of the business, even though it's called CVS.
- 32:25
- Unidentified speaker: It's, you know, their insurer and the PBM and their line of clinics.
- 32:29
- Unidentified speaker: Right.
- 32:30
- Unidentified speaker: Yeah.
- 32:30
- Unidentified speaker: What do they do, Jason?
- 32:32
- Unidentified speaker: They spin it out to their stockholders or I don't know the details, and I don't know if they've even announced the details, but it's it doesn't look good for owning CVS the retailer.
- 32:42
- Unidentified speaker: You know, it's Walgreens hasn't had much success lately, you know, and CVS is probably in the same situation.
- 32:51
- Unidentified speaker: A lot of the state level legislation to try to disrupt the PBMs was focused on the PBM can't own the pharmacy as well.
- 33:01
- Unidentified speaker: So it's really targeted CVS.
- 33:02
- Unidentified speaker: So the pharmacy was kind of dead weight for them, and they're cutting it off, and they're gonna turn into more of a united health group where they are the insurer, the PBM, and the point of care.
- 33:15
- Unidentified speaker: What do you think the blue wave names just in a couple of minutes left, and then we can do more next week in terms of I mean, are they the Democrats probably on balance aren't very happy with with the Secretary of Health, but what what else you know will happen based on the Democrats running the Senate committees and the House committees and doing more and being active, you know, on on things that trouble them about the Trump administration's health care policy?
- 33:50
- Unidentified speaker: Yeah, that's a great question.
- 33:52
- Unidentified speaker: I think the easy answer is that there's gonna be more money allocated to healthcare, which is only gonna drive the costs up.
- 34:00
- Unidentified speaker: I haven't given a lot of thought to a lot of the policies that you know that Kennedy's been pushing.
- 34:09
- Unidentified speaker: They're, as we spoke about a couple weeks ago, it's they're gonna be hard-pressed to really roll back what's being done on the food side of things.
- 34:18
- Unidentified speaker: No one wants unhealthier food supply.
- 34:21
- Unidentified speaker: There's been a lot of finger pointing that there's been no improvement in chronic disease.
- 34:27
- Unidentified speaker: I mean, but chronic disease is built up over decades, so we're not gonna see that for a long time.
- 34:31
- Unidentified speaker: And there frankly hasn't been a lot done yet either.
- 34:35
- Unidentified speaker: I think the thing that I hope Democrats don't undo, and it wouldn't be a popular thing to undo, is some of the changes that occurred at the FDA around accelerated pathways for approval, priorities for rare diseases.
- 34:52
- Unidentified speaker: There's been a lot of things that kind of happen quietly behind the scenes that I think are really beneficial that are speeding up drug, you know, the clinical stage of drug discovery and bringing to market.
- 35:04
- Unidentified speaker: We will see more funding go back to early stage research.
- 35:08
- Unidentified speaker: I think that's a a guarantee.
- 35:12
- Unidentified speaker: We run over than we normally do, but then we're dealing with the day after the election.
- 35:17
- Unidentified speaker: We'll be here exactly seven days from now.
- 35:20
- Unidentified speaker: Everyone be well and stay healthy.
- 35:22
- Unidentified speaker: And thanks very much for listening.
- 35:24
- Unidentified speaker: Take care.
- 35:26
- Unidentified speaker: The views expressed on this podcast are the host alone and do not constitute an offer to sell or a recommendation to purchase or a solicitation of an offer to buy, any security nor a recommendation for any investment product or service.
- 35:39
- Unidentified speaker: While certain information contained herein has been obtained from sources believed to be reliable, neither the host nor any of their employers or their affiliates have independently verified this information, and its accuracy and completeness cannot be guaranteed.
- 35:52
- Unidentified speaker: Accordingly, no representation or warranty expressed or implied is made as to and no reliance should be placed on the fairness, accuracy, timeliness, or completeness of this information.
- 36:02
- Unidentified speaker: The host and all employers and their affiliated persons assume no liability for this information and no obligation to update the information or analysis contained herein in the future and may or may not hold positions in the securities mentioned.
- 36:15
- Unidentified speaker: The following conversation is intended for informational purposes only.
- 36:19
- Unidentified speaker: You should always do your own work to determine if an investment is suitable for you.
- 36:26
- Unidentified speaker: Thank you for joining us this week.
- 36:28
- Unidentified speaker: Please tune in again next week as we will be back on Wednesday.
- 36:31
- Unidentified speaker: As a reminder, nothing on this podcast should be considered investment advice.
- 36:35
- Unidentified speaker: You should always do your own work to determine if an investment is suitable for you.