Noise Cancelling with Neil Woodford

Neil Woodford explains why he won't buy Anthropic IPO due to valuation uncertainty and AI cost dynamics

40 min Listen to the original Episode page
AI InfrastructureAnthropic IPOOpen-Weight ModelsValuationUK EconomyCloud ComputingLabor Displacement

Summary

Neil Woodford discusses Anthropic's financials, noting $17 billion in sales but significant net losses and $895 billion in computing commitments. He declines to invest in the upcoming IPO because he cannot accurately judge intrinsic value given high uncertainty and opportunity costs. Woodford highlights that AI prices are falling tenfold annually, with open-weight models like those used by Revolut cutting costs by 90%. He argues that cloud providers are shifting to capital-intensive models while service sectors like UK banks may benefit from labor displacement.

Episode sections

  1. 0:00 – 5:20

    Intro and Anthropic Financials

    Neil Woodford and host John Adair discuss Anthropic's recent financial filings, noting $17 billion in sales but a net loss, and $895 billion in computing commitments. They question who profits from AI development given that model developers are currently losing money.

  2. 5:20 – 8:51

    Anthropic's Cost Structure and Model Lifecycle

    The speakers break down Anthropic's costs into inference, R&D, and future contracts. They discuss how models have short lifespans and how investment cycles overlap, making it difficult to isolate returns on specific model iterations.

  3. 8:51 – 13:07

    Open Weight Models and Price Competition

    Woodford notes that AI prices are falling tenfold annually, citing Revolut's shift to open-weight models to cut costs by 90%. He argues that while open models are cheap, frontier models offer exponential capability growth for high-value tasks.

  4. 13:07 – 22:24

    Cloud Providers and Capital Intensity

    The discussion shifts to cloud providers like Amazon, Google, and Microsoft, who are profitable but spending heavily on data centers. Woodford predicts a shift from high-margin, low-capital models to capital-intensive infrastructure businesses.

  5. 22:24 – 26:23

    Anthropic IPO and Valuation Uncertainty

    Woodford explains why he will not invest in the Anthropic IPO, citing an inability to accurately judge intrinsic value due to high uncertainty and opportunity cost compared to other investments.

  6. 26:23 – 33:58

    Future of AI Spending and Productivity

    Woodford argues that AI spending is at the start of a long industrial revolution, not near its end. He suggests that even if model developers don't capture all value, companies deploying AI will see massive productivity gains.

  7. 33:58 – 34:52

    Economic Returns in AI and Labor Displacement

    Neil Woodford argues that technology delivering transformation will not come for free, with economic returns accruing to the companies that own the technology. He notes that 'picks and shovel' suppliers will continue to do well due to rapid investment, profit leaks from AI companies may land in service-based industries like UK banks, insurance, and law firms that can shed highly paid labor.

  8. 34:52 – 38:04

    UK Economy and Anthropic IPO Valuation Framework

    Woodford highlights the UK's service-oriented economy as uniquely positioned to capture AI upside, citing Microsoft's reduction of 27% of its global workforce as evidence of labor transformation. Regarding the Anthropic IPO, he states he applies his standard valuation process: comparing market price to intrinsic value to determine if the gap compensates for risks.

  9. 38:04 – 40:23

    Monitoring Anthropic and Conclusion

    For Anthropic, Woodford prioritizes revenue and cash flow growth over quarter-on-quarter fluctuations, expecting detailed public information to help investors judge the valuation gap. The episode concludes with a summary of Anthropic's unit economics and a call to action for the Noise Cancelling app.

Claims worth verifying

Anthropic sold $17 billion of AI in three months and made a loss.

Unidentified speaker ·0:00

How to verify: Check Anthropic's IPO filing or financial statements for Q3 revenue and net income.

Anthropic has signed contracts to spend $895 billion on computing power.

Unidentified speaker ·0:07

How to verify: Review Anthropic's IPO prospectus or SEC filings for capital commitments.

Anthropic's sales were $16.96 billion in Q3, up 47% from the previous quarter.

Unidentified speaker ·2:32

How to verify: Check Anthropic's Q3 financial report or IPO filing.

Anthropic's operating loss was $567 million, attributed to $5 billion in short-term computing rentals.

Unidentified speaker ·2:50

How to verify: Review Anthropic's Q3 income statement and notes on operating expenses.

Anthropic's computing commitments increased from $518 billion to $895 billion.

Unidentified speaker ·3:09

How to verify: Compare previous data center episode notes with the new filing mentioned in the transcript.

Inference costs for Anthropic were $2.4 billion for $16.3 billion in sales in the first half of the year.

Unidentified speaker ·5:34

How to verify: Check Anthropic's H1 financial report for cost of revenue and sales figures.

Anthropic spent $9.7 billion on R&D in the same six-month period.

Unidentified speaker ·5:53

How to verify: Review Anthropic's H1 income statement for R&D expenses.

Epoch AI estimated OpenAI's GPT-5 was the top model for about four months with a 30% gross margin.

Unidentified speaker ·6:24

How to verify: Search for Epoch AI's research or reports on GPT-5 performance and margins.

Revolut moved coding work of 2,000 engineers to an open model on 56 NVIDIA chips, cutting costs by 90%.

Unidentified speaker ·9:14

How to verify: Check Revolut's public statements, blog posts, or news reports regarding their AI infrastructure shift.

Alibaba invested ~$50 billion in its Qwen model series.

Unidentified speaker ·11:18

How to verify: Review Alibaba's financial reports or public announcements regarding AI infrastructure investment.

Technology that delivers transformation will not come for free, and companies that own it will accrue appropriate economic returns.

Neil Woodford ·34:18

How to verify: Analyze long-term profitability and economic moats of dominant AI infrastructure and model developers (e.g., Microsoft, Google, Amazon) via financial filings (10-K) to see if they capture disproportionate returns relative to suppliers.

Microsoft offered redundancy to 27% of its employee base globally, totaling 60,000 workers.

Neil Woodford ·35:39

How to verify: Check Microsoft's official press releases or SEC filings (8-K) from October 2024 regarding workforce reductions.

The UK is the most service-oriented economy of any developed economy globally.

Neil Woodford ·36:07

How to verify: Compare the service sector's percentage contribution to GDP across developed nations using data from the World Bank or IMF.

Manufacturing is 8.5% of the UK economy.

Neil Woodford ·36:14

How to verify: Check UK Office for National Statistics (ONS) data on GDP composition by sector.

Anthropic is the first listing of a pure AI company.

Host ·36:56

How to verify: Review the prospectus and listing announcements for Anthropic's IPO to confirm its classification relative to other recent tech listings.

Anthropic keeps about 85 cents of every dollar of sales after running its models.

Host ·39:51

How to verify: Analyze Anthropic's S-1 filing or investor presentations for gross margin data and unit economics related to API usage and model inference costs.

Do your own research

  • Verify Anthropic's financial filings for sales figures, net loss, and $895 billion in computing commitments.
  • Check Epoch AI data for model ranking durations and lifecycle estimates.
  • Review Revolut's engineering cost reports regarding the shift to open-weight models and NVIDIA chip usage.
  • Analyze cloud providers' (Amazon, Microsoft, Google) cash flow statements to assess capital intensity shifts.
  • Examine UK economic data on service sector composition versus manufacturing (8.5%).
  • Research Microsoft's workforce reduction announcements and associated financial impacts.
  • Compare Anthropic's IPO pricing against intrinsic value estimates using standard valuation frameworks.
  • Investigate open-weight model investments by Chinese firms like Alibaba, Tencent, and DeepSeek.
Full transcriptexpand
0:00
Unidentified speaker: Anthropic, the company behind the AI model Claude, sold $17 billion of AI in three months.
0:07
Unidentified speaker: It still made a loss, and it has signed contracts to spend $895 billion on computing power.
0:13
Unidentified speaker: What it tells you is that the company is investing a substantial amount of money on its future technology iterations.
0:20
Unidentified speaker: Anthropic lists on the stock market this month, and the question for anyone looking at it is who ends up keeping the profit from AI?
0:27
Unidentified speaker: What is clear at the moment at the net level is that the model developers aren't, and neither are the openweight model developers either.
0:34
Unidentified speaker: That doesn't mean that they won't in the future, but right here, right now, the money's being made elsewhere.
0:39
Unidentified speaker: Neil's going to show you where the money goes and the $600 million Anthropic gave up in one quarter by cutting its prices.
0:46
Unidentified speaker: Would I invest in Anthropic?
0:48
Unidentified speaker: The answer is.
0:52
Unidentified speaker: Welcome to Noise Cancelling.
0:53
Unidentified speaker: I'm John Adair and I'm here with Neil Woodford.
0:55
Unidentified speaker: On today's episode, first, what Anthropic's accounts show about how an AI company makes and loses money.
1:02
Unidentified speaker: Second, where that money goes.
1:04
Unidentified speaker: Third, Neil's view and this week's argument over whether the AI spending can last.
1:08
Unidentified speaker: And finally, what it means for your money.
1:11
Unidentified speaker: Neil, $17 billion of sales in a quarter and making a loss.
1:16
Unidentified speaker: Have you seen a business like this before?
1:19
Unidentified speaker: There have not been many like that.
1:22
Unidentified speaker: Obviously, companies like OpenAI, SpaceX, you know, these have been huge companies with giant valuations and they've also been loss making, but it is it is very, very unusual.
1:38
Unidentified speaker: It's the fastest growing company in terms of revenue of all time, is that right?
1:42
Unidentified speaker: Yeah, it looks like it.
1:43
Unidentified speaker: Yeah.
1:44
Unidentified speaker: On the scale that we're measuring, yes, it's grown from zero to annualized revenues of forty-six odd billion dollars in the space of five and a half years.
1:55
Unidentified speaker: Neil writes about this kind of topic every week, and the easiest place to read him is in the Noise Cancelling app.
2:00
Unidentified speaker: It's free on iPhone and Android, and if you'd rather listen to the articles than read them, the app will read everything to you.
2:06
Unidentified speaker: There are two things to look out for later on in this episode.
2:08
Unidentified speaker: Anthropic's filing has a line that puts a number on how much cutting its prices cost it this quarter.
2:14
Unidentified speaker: And a fintech moved the coding work of about 2,000 engineers off Anthropic and OpenAI onto a model anyone can download, running on 56 of its own chips, and they cut their bill by about 90%.
2:27
Unidentified speaker: We'll get to both of those in the next part, but first of all, some numbers.
2:31
Unidentified speaker: Here's Anthropic's third quarter.
2:32
Unidentified speaker: Sales were $16.96 billion, up 47% on the quarter before.
2:38
Unidentified speaker: That's fast, but the quarter before that, sales grew about 145%.
2:43
Unidentified speaker: Computing power and infrastructure cost $9.9 billion, which is 58 cents of every dollar of sales.
2:50
Unidentified speaker: Anthropic made an operating loss of $567 million, and it put that down to $5 billion spent renting extra computing power at short notice.
3:00
Unidentified speaker: Then there's the bill it's already signed.
3:02
Unidentified speaker: When we did our data centers episode two weeks ago, Anthropic had committed to spend $518 billion on computing.
3:09
Unidentified speaker: The new filing says $895 billion.
3:13
Unidentified speaker: About $70 billion of that falls due in the next 12 months.
3:16
Unidentified speaker: For comparison, Anthropic's sales for the first nine months of this year were $33 billion.
3:22
Unidentified speaker: Neil, $895 billion of commitments against $33 billion of sales so far this year.
3:29
Unidentified speaker: Who do you think is carrying the risk on that?
3:31
Unidentified speaker: Is it Anthropic or the companies that it's signed with?
3:35
Unidentified speaker: The risks are sp are spread.
3:37
Unidentified speaker: Of course, you know, Anthropic shareholders are and the people who invest and then lend to the company will be on the hook to some extent for these commitments.
3:50
Unidentified speaker: Equally, the companies to whom Anthropic has has made these commitments are on the hook.
3:55
Unidentified speaker: If if Anthropic is unable to follow through on those commitments, then of course there will be uh contingent problems throughout that chain.
4:07
Unidentified speaker: And indeed the customers of those customers.
4:09
Unidentified speaker: So you know it the the the the risks have spread throughout the the the whole AI sort of industry.
4:18
Unidentified speaker: Um but but clearly the main the main issue would be if if it wasn't able to meet those commitments um then then anthropic shareholders and its bondholders would would would would be the most obvious risk takers in in in in that in that situation.
4:37
Unidentified speaker: But I mean you know I I I I think that there is a a little too rigid a view about these commitments, albeit that they are contracted, etc.
4:46
Unidentified speaker: The contracts can be renegotiated.
4:48
Unidentified speaker: If it if it serves the interests of all parties to to renegotiate a situation, then it will be renegotiated.
4:54
Unidentified speaker: So I don't see them as being written in stone, if you like.
4:58
Unidentified speaker: Um I I think that would be a slightly naive view of what these commitments really amount to.
5:04
Unidentified speaker: So that's what the financials say.
5:06
Unidentified speaker: But next we're going to look at why a company can sell $17 billion of something and still lose money.
5:12
Unidentified speaker: There are three separate bills, and only one of them is the cost of answering your questions.
5:16
Unidentified speaker: Then we'll get to Neil's view and what it all means for your money.
5:20
Unidentified speaker: To see how a company can sell $17 billion of something and still lose money, you need to separate out three bills.
5:27
Unidentified speaker: The first is the cost of answering your questions, the chips and the electricity used every time somebody uses one of these models.
5:34
Unidentified speaker: People in the industry call it inference.
5:36
Unidentified speaker: At Anthropic, it's cheap compared with the price.
5:39
Unidentified speaker: In the first half of the year, it cost $2.4 billion to serve 16.3 billion of sales, a gross margin of about 85%.
5:48
Unidentified speaker: A year earlier, the margin was just 12%.
5:51
Unidentified speaker: The second bill is the next model.
5:53
Unidentified speaker: In the same six months, Anthropic spent $9.7 billion on research and development, about four times what it costs to serve its customers.
6:01
Unidentified speaker: And most of that is computing power for training new models.
6:05
Unidentified speaker: The third bill is the contracts for computing in future years, the $895 billion.
6:11
Unidentified speaker: The second bill, though, is the one I want to focus on.
6:13
Unidentified speaker: Each model is effectively a product with a short life.
6:17
Unidentified speaker: It costs billions to train, it earns money for a few months as the best on the market, and then a newer model comes in and replaces it.
6:24
Unidentified speaker: The research group Epoch AI looked at OpenAI's GPT-5 and estimated it was OpenAI's top model for about four months, with a gross margin of about 30%.
6:35
Unidentified speaker: Neil, if each of these models has to pay back its training before the next one replaces it, how do you think about valuing a business like that?
6:43
Unidentified speaker: Well, if you could if you could take that snapshot of a company.
6:46
Unidentified speaker: So, like Anthropic, the the part of the problem with trying to understand whether anthropic's valuation is appropriate or not, is that it's really hard to isolate each sort of bucket of investment and to then make a judgment about whether that particular piece of investment made an appropriate return.
7:06
Unidentified speaker: Um, you can't really make that judgment because all of the investment cycles are overlapping.
7:11
Unidentified speaker: You know, anthropic is not just developing a model at a time and launching it and then seeing how well it goes and then starting to invest in a new model and launching that.
7:22
Unidentified speaker: This is not this is not a series process.
7:25
Unidentified speaker: It's the investment cycles are in parallel.
7:29
Unidentified speaker: So I'm absolutely sure, although I don't I don't know the detail here, but I'm absolutely confident that Anthropic is already building, investing in and fine-tuning um models that that that are several series ahead of its current market leading frontier model.
7:50
Unidentified speaker: And I think it would be wrong also to believe that that because a model is superseded by another one, everything that it spent on that previous model was wasted or has has to be written off.
8:01
Unidentified speaker: Clearly, what was learned and um what was embedded and learned from that previous model is forms part of the sorts of building block on which the new model and the one after that will be will will be developed.
8:19
Unidentified speaker: So um this isn't this isn't capital or this isn't investment that that you should think of as having been written off.
8:26
Unidentified speaker: They're not building something, they're not building, for example, a road to nowhere and and and that road is no longer used and therefore was a was a wasted investment.
8:36
Unidentified speaker: I I I think it's very much the case that the next new model builds on the the learnings and the sophistication and and all the technology that was embedded and invested in the previous one and the one before that.
8:51
Unidentified speaker: I think the second thing to understand is how quickly AI is getting cheaper.
8:55
Unidentified speaker: For the same level of capability, prices have been falling about 10 times a year.
9:00
Unidentified speaker: Anthropics filing puts a number on what that costs.
9:03
Unidentified speaker: Price cuts reduced its sales by about $600 million in the third quarter.
9:08
Unidentified speaker: In the last quarter of 2025, the figure was about 70 million.
9:12
Unidentified speaker: And customers are definitely moving.
9:14
Unidentified speaker: Revolut, the fintech, the largest private business in Europe, has put the coding work of about 2,000 engineers onto an open model, one that anyone can download, running on 56 of NVIDIA's newest chips in its own data center.
9:28
Unidentified speaker: It says the cost is about a tenth of what it was paying Anthropic and OpenAI together, and its engineers are getting work out more quickly.
9:35
Unidentified speaker: It still pays for OpenA and Anthropic, but for everything else, they're using a Chinese open weight model.
9:41
Unidentified speaker: Neil, if a company like Revolute can do most of the work for a tenth of the price, does the best model keep its premium or does it become a commodity?
9:51
Unidentified speaker: It's a really interesting question, and I'm not sure I've got a definitive answer to it.
9:56
Unidentified speaker: But um clearly there are there are now open weight models that are very cheap, that can do a lot of the things that many companies want their AI to do for them.
10:09
Unidentified speaker: But and there's a there's a big butt here, at the frontier, the capability of these models is is growing exponentially.
10:18
Unidentified speaker: And that capability is going to deliver all sorts of new applications, new customers with who are prepared to spend and invest in in that capability.
10:34
Unidentified speaker: So I think it would be wrong to assume that just because open weight models are are very cheap now in relation to Frontiers, uh Frontier models, that that everybody migrates to open weight models.
10:46
Unidentified speaker: Uh, as much as many of the Chinese developers, those models would hope that that would be what happens, I I I'm not sure that that at all, that that is the likely scenario.
10:56
Unidentified speaker: And let's not forget, of course, this is a long game.
11:01
Unidentified speaker: Uh, and the open weight model developers like Alibaba and Deep Seek and Tencent, etc., those Chinese model developers, and there are some other upstarts as well.
11:12
Unidentified speaker: These companies are not developing these models on thin air, they're investing billions.
11:16
Unidentified speaker: I mean, for example, I checked this morning.
11:18
Unidentified speaker: Alibaba, I think, has already invested some, the best part of $50 billion in developing its uh Quen model series.
11:26
Unidentified speaker: Tencent, another leading Chinese model developer, has invested less, but a substantial amount of money relative to the size of the business.
11:34
Unidentified speaker: And even DeepSeek, it's got a huge valuation now.
11:38
Unidentified speaker: Um, but DeepSeq raised $16 billion in an equity placing earlier this year.
11:44
Unidentified speaker: It is not developing open make models, as I say, on thin air.
11:47
Unidentified speaker: It is investing a huge amount of money in building these capabilities.
11:53
Unidentified speaker: And at some stage, these open weight model developers will need to make a return on that investment, just as all model developers will.
12:01
Unidentified speaker: So taking a snapshot now and concluding too much from the relative pricing architecture of this of this industry right now, I think would be way way too early, way premature.
12:13
Unidentified speaker: And of course, Revolute is keeping access to those frontier models for the hardest tasks.
12:18
Unidentified speaker: And that sort of tells you the answer.
12:21
Unidentified speaker: It's about choosing the right tools for the right job.
12:24
Unidentified speaker: Yeah, absolutely.
12:25
Unidentified speaker: So spending the spending the big bucks when you need to.
12:27
Unidentified speaker: But it's important to remember that the even the the price of the the open weight models is falling, but their capability is growing exponentially too.
12:39
Unidentified speaker: Um compute, the the cost of compute per per unit of electricity, um the the the efficiency, if you like, of compute relative to the consumption of electricity is improving dramatically as well.
12:55
Unidentified speaker: So uh we're we are on an exponential curve of improvement towards AGI and artificial superintelligence ultimately.
13:07
Unidentified speaker: If the AI companies aren't keeping the money, who is?
13:11
Unidentified speaker: Let's start with the companies renting them the computing power.
13:14
Unidentified speaker: Last quarter, Amazon's cloud business had sales of $46 billion up 39%, with an operating margin of about 39%.
13:22
Unidentified speaker: Google's cloud sales doubled to 30.7 billion with a margin of 36%, and Microsoft's cloud grew 48%.
13:30
Unidentified speaker: Those are profitable businesses, but they're spending more than they bring in to build data centers.
13:35
Unidentified speaker: Amazon invested $77 billion in the quarter against 42 billion of cash from operations.
13:41
Unidentified speaker: Microsoft has changed how it accounts for its data centers, and it now writes them off over 25 years instead of 15, which lifts its reported profits.
13:49
Unidentified speaker: Neil, the cloud companies are profitable on paper and spending more cash than they make.
13:55
Unidentified speaker: Which of those two facts matters more to you?
13:58
Unidentified speaker: Well, these are highly successful businesses.
14:00
Unidentified speaker: I wrote about this actually years ago, some years ago, at the start of this process.
14:07
Unidentified speaker: I one of the reasons that I was nervous about the valuations of the Mag 7, and I said this at the time, was that I felt that their business models would change with the AI Industrial Revolution.
14:17
Unidentified speaker: What I what I expected to see was these companies start to become, to move, if you like, shift from low capital intensity, very high margin, very cash generative business models to a different model, really.
14:32
Unidentified speaker: One that was far more capital intensive, um uh not so cash generative, and one that would would ultimately involve a huge amount of infrastructure investment.
14:43
Unidentified speaker: Now that's a different sort of type of business model, and one that ought not to be rated as highly as as those companies were rated before.
14:51
Unidentified speaker: And that was sort of roughly what I I said some while ago.
14:55
Unidentified speaker: Now I had no idea uh uh how quickly the the investment would scale.
15:00
Unidentified speaker: I had no idea that it would it would grow as rapidly as it has done, but but that sort of approach was one that I talked about some while ago.
15:10
Unidentified speaker: Anyway, the fact is that they are now spending more than than than they are receiving in cash flow.
15:18
Unidentified speaker: And you'd sort of expect that.
15:20
Unidentified speaker: What I think we've got to understand is is that none of these companies are going to voluntarily say, well, it's all too difficult at the moment.
15:29
Unidentified speaker: We're spending a lot of money, we're investing a huge amount, we're gonna stand back and sort of press pause button for a couple of years and then come back in.
15:36
Unidentified speaker: No company, no company that that that is in this race will be prepared to do that because they know that if they do, they are basically giving up.
15:48
Unidentified speaker: They are giving up in this competition.
15:52
Unidentified speaker: Um because the prize ultimately is who controls the customer, who owns the customer.
15:59
Unidentified speaker: When you do own that customer relationship, when you own the customer, then you can monetize the customer.
16:06
Unidentified speaker: Then you can start to um deploy all sorts of additional add-on services on top of the models.
16:14
Unidentified speaker: There are all sorts of additional services that that that a an AI company can then provide to monetize the relationships that they have with their customers.
16:27
Unidentified speaker: If you now jump off that off that race, you basically give up on that end game.
16:33
Unidentified speaker: You you you basically wave the white flag.
16:35
Unidentified speaker: And no company, certainly none of the leading companies in this race are gonna do that.
16:40
Unidentified speaker: So they will carry on spending, but there are gonna be external constraints on how much they can spend.
16:48
Unidentified speaker: Things that they don't control.
16:50
Unidentified speaker: For example, this week there's been a huge shift in in politics in the US.
16:56
Unidentified speaker: I think the five states um that have spent or invested the most in data centers now all have democratic governors.
17:07
Unidentified speaker: Those democratic governors have have a different political perspective.
17:11
Unidentified speaker: There may well be additional constraints on how quickly this capital can be deployed, how quickly AI data centers can evolve.
17:20
Unidentified speaker: We already know that there's a backlog, and that backlog may well grow.
17:23
Unidentified speaker: So it the idea that the growth in investment that we're that we're told to expect will come through seamlessly and without any hiccups, I think is wrong.
17:35
Unidentified speaker: I think there'll be constraints, as we've talked about this on previous podcasts.
17:38
Unidentified speaker: There'll be constraints in terms of the cost of capital, there'll be constraints in terms of the supply of of the sort of the ingredients that go into data center build.
17:47
Unidentified speaker: We know that the the supply of of memory chips or the um HBM chips, for example, uh is is constrained.
17:57
Unidentified speaker: We know other things are constraining the deployment of data centers, worries about their impact on the environment.
18:06
Unidentified speaker: Energy is the main, the most prominent limiting factor as well, access to energy, access to electricity.
18:13
Unidentified speaker: So, you know, the idea that all of those plans that are now in place will be spent uh on schedule, precisely at the moment they're they're scheduled, is I think it would be a mistake.
18:25
Unidentified speaker: There are gonna be constraints.
18:26
Unidentified speaker: I think the rate of growth of capital investment may well slow.
18:32
Unidentified speaker: But but no company is gonna volunteer to volunteer to sort of step away from this race.
18:38
Unidentified speaker: Um uh that would be really economic suicide, I think.
18:43
Unidentified speaker: And there's another pot of money that the AI companies are going after, which is the wage bill.
18:48
Unidentified speaker: Microsoft has offered voluntary redundancy to about 60,000 staff, 27% of its workforce, including in the UK, while planning to invest about $170 billion this financial year.
18:59
Unidentified speaker: The financial tracker layoffs.fyi counts about 94,000 tech job cuts this year that companies have put down to AI.
19:06
Unidentified speaker: And outside tech, Barclays expects to cut 5,000 jobs and save about £1.1 billion in 2027.
19:13
Unidentified speaker: Now the AI companies want a share of those savings, so they're setting up consulting arms.
19:17
Unidentified speaker: OpenAI has put more than $4 billion into one and bought an Edinburgh firm called Tomorrow.
19:22
Unidentified speaker: Anthropic set one up called Ode, and the FT reports is starting another in London.
19:28
Unidentified speaker: That puts them up against the big four.
19:30
Unidentified speaker: OpenAI is also selling adverts.
19:32
Unidentified speaker: It reached a $1 billion annual run rate by the end of August against a target of $2.4 billion.
19:37
Unidentified speaker: Neil, consulting and advertising.
19:40
Unidentified speaker: Are those the signs of a business finding its feet or one that just can't make enough from the product itself?
19:46
Unidentified speaker: I think very much the former, not the latter.
19:48
Unidentified speaker: I think the this is this speaks to what I what I I I've just hinted at earlier on this episode.
19:56
Unidentified speaker: Building an embedded relationship with your customer, and there are going to be millions of customers for these frontier models.
20:05
Unidentified speaker: Building that embedded relationship, creating exclusivity, if you like, in terms of the relationship, is what all of these hyper successful companies in technology have been very successful at doing.
20:20
Unidentified speaker: Meta, Amazon, Google, uh, Apple, etc.
20:24
Unidentified speaker: They they've captured the customer relationship and then grow the relationship by embedding that relationship, creating stickiness, if you like, in terms of that relationship.
20:35
Unidentified speaker: So I see what's happening now with these add-on services, for example, like consultancy, which I think is very smart, um create more stickiness between the model developers and their customers.
20:47
Unidentified speaker: So just imagine, for example, that that Anthropic recruits and and trains a load of uh consultants that will go into big businesses and embed Anthropic's relationship with that corporate customer.
21:00
Unidentified speaker: You you embed the technology in that customer, you create stickiness in terms of the relationship between the customer and the AI company, and and then you create all sorts of multiple opportunities to monetize that relationship.
21:13
Unidentified speaker: So I think this is this is how the the um the AI companies will evolve.
21:19
Unidentified speaker: And the idea that there will be one sort of killer application that will generate all of the return, I think is is in a way a mistake.
21:27
Unidentified speaker: I think there will be all sorts of different services that these AI companies will develop and evolve as they build crucial relationships, sort of very, very important interdependent relationships with with with their customers.
21:45
Unidentified speaker: This week, the argument over whether the spending can last got louder.
21:49
Unidentified speaker: Analysts and Omura, as reported in the press, expect the big tech company's free cash flow, which is the cash left after investment, to go from a surplus of about $289 billion last year to a deficit of about $144 billion in 2027 if their investment rises by 60% next year.
22:07
Unidentified speaker: A summary from a note from Bernstein says OpenAI has about $614 billion of payments due by 2030, and that there's a risk of too much capacity within two to three years.
22:18
Unidentified speaker: Other commentators add up to $3.3 trillion of commitments across the industry.
22:24
Unidentified speaker: On the other side, some investors are saying that the boom is nowhere near over.
22:28
Unidentified speaker: Neil, in our data centers episode, you said you wouldn't invest in the companies building the models.
22:34
Unidentified speaker: Anthropic lists on the stock market this month.
22:37
Unidentified speaker: Would you invest in it?
22:38
Unidentified speaker: Well, the answer to that is no.
22:41
Unidentified speaker: But the reason it's no, it's a very personal one for me because it reflects the way I look at investment.
22:49
Unidentified speaker: My investment process is focused very much on making valuation judgments.
22:57
Unidentified speaker: That doesn't mean that I only invest in value stocks, by the way.
23:00
Unidentified speaker: That that's not true.
23:02
Unidentified speaker: Valuation judgment depends on me being able to make an assessment about what the intrinsic value of a business is and comparing that intrinsic value with the market price.
23:14
Unidentified speaker: Now, with anthropic, the market price is the IPO price.
23:18
Unidentified speaker: What I've then got to establish is what is the intrinsic value of anthropic.
23:22
Unidentified speaker: Now, I just don't know enough about Anthropic's business model.
23:26
Unidentified speaker: I don't know enough about what's going to happen in the years ahead to make that judgment.
23:33
Unidentified speaker: And because I I can't make that judgment, I could guess at it.
23:37
Unidentified speaker: I can make all sorts of very, very sort of uneducated guesses about how the business evolves.
23:45
Unidentified speaker: But because there's no precision in that process and because there's no certainty, there's no certainty about the future anyway, but because there's so much uncertainty about that judgment, I can't really make the intrinsic value judgment at all accurately, and certainly not accurately enough to persuade me that that I can make uh an educated and informed guess about the current market value.
24:13
Unidentified speaker: And of course, it's also important to remember that in any portfolio there's an opportunity cost.
24:19
Unidentified speaker: If people want to invest in anthropic, and I'm sure there will be millions of people, millions of institutions that will want to invest in anthropic, but there's an opportunity cost involved in that decision.
24:30
Unidentified speaker: If you're investing in anthropic, you're not investing in other things where you might have much more certainty about the gap between the intrinsic value and the market value.
24:42
Unidentified speaker: That's that's that's why, and and and that what I just described is exactly the way I feel about anthropic relative to other opportunities.
24:53
Unidentified speaker: The opportunity cost of making that guess about anthropic is one I'm not prepared to take, given that I have much more certainty about the gap between intrinsic value and market value in a whole host of other opportunities.
25:07
Unidentified speaker: For someone like me who's, you know, looking at anthropic as a business that I have got a huge amount of respect for.
25:13
Unidentified speaker: I use their products every day.
25:15
Unidentified speaker: I couldn't do my job now without their product.
25:18
Unidentified speaker: Have you ever had a scenario like that where you you think that it's a fantastic company, you really like the company, what they stand for, but you just can't consider it as an investment because of the valuation?
25:31
Unidentified speaker: Yes.
25:32
Unidentified speaker: Yeah, all the time.
25:33
Unidentified speaker: And and there have been many companies that I haven't had respect for that I didn't like, and I didn't like the product or or didn't didn't rate the people running it, or that equally have been very attractive investments.
25:46
Unidentified speaker: It it the valuation is everything.
25:49
Unidentified speaker: Different companies uh deserve different valuations, and but when share prices stand at a significant discount to the appropriate valuation of that business as you judge it, then that has to be an attractive investment opportunity.
26:03
Unidentified speaker: Whether you really like the product, whether you really like the company or don't like it, in a way you've got to cleanse your sort of investment process from that prejudice.
26:12
Unidentified speaker: And you've just got to take a very sort of dispassionate clinical look at these companies and only really look at them through a valuation process lens.
26:23
Unidentified speaker: Well, of course, all of that is Neil's own view.
26:25
Unidentified speaker: None of it is a recommendation to buy or sell anything.
26:28
Unidentified speaker: So who's right about the spending?
26:30
Unidentified speaker: The analysts saying the cash is running out, or the people saying that the spending has barely started?
26:35
Unidentified speaker: I'm I'm I'm more inclined to sort of err towards the second.
26:39
Unidentified speaker: I I think we are at the start of this industrial revolution.
26:43
Unidentified speaker: We we as yet do not know where this is all going to end up.
26:51
Unidentified speaker: And I'm sure it doesn't end.
26:53
Unidentified speaker: It evolves and continues.
26:55
Unidentified speaker: There is no sort of end point.
26:57
Unidentified speaker: Even if we get to AGI within three years or two years or five years, you know, I don't, I don't know.
27:03
Unidentified speaker: There are all sorts of different opinions in the industry about what it is and when we get to that point, but that won't be the end.
27:10
Unidentified speaker: No, nobody's going to sort of down tools and stop developing this technology.
27:16
Unidentified speaker: Um, it will carry on developing, I think, at an exceptionally fast rate.
27:22
Unidentified speaker: So, so I think we are at the start of this industrial revolution.
27:26
Unidentified speaker: We're the the start of the adoption process.
27:29
Unidentified speaker: We we as yet do not know really the full capability of this technology, how it will transform productivity in the future, how companies will use it, how households will use it, how it will improve our lives, what impacts it will have.
27:44
Unidentified speaker: We don't yet know what they are.
27:46
Unidentified speaker: I'm much more persuaded we're at the start of this process.
27:50
Unidentified speaker: Nowhere near the end of it.
27:52
Unidentified speaker: Clearly, the investment, the scale of investment required is massive, but it doesn't yet dwarf other industrial revolutions in terms of percent of GDP, for example.
28:04
Unidentified speaker: I've I've said that I think the pace of the pace of CapEx deployment was slow because of the constraints that we've already talked about on this episode.
28:12
Unidentified speaker: But but the spending will continue, uh, and it will probably continue way beyond um the next two or three years.
28:20
Unidentified speaker: Okay, Neil, I want to put two hypothetical scenarios to you.
28:24
Unidentified speaker: First, picture two years' time, anthropic sales are still doubling year on year, and it stopped losing money even after paying for the next model.
28:31
Unidentified speaker: Does that make you wrong to not consider the IPOs?
28:34
Unidentified speaker: I think I think the answer to that question is yes, I would, I would be wrong, because by that time, if anthropic is transformed in the way you've just described, such that it's highly profitable, and this cycle of investment um continues, but but the returns on the the sort of launched models is so attractive that it's now making serious profits.
28:59
Unidentified speaker: Um that I think is a scenario in which the market will will be very bullish about Anthropic and will think, well, this company is going to be, first of all, a winner in this race.
29:10
Unidentified speaker: It will it will be it will be highly profitable, highly successful.
29:14
Unidentified speaker: It will be able to monetize multiple different sorts of earnings streams with its corporate and and and personal customers.
29:22
Unidentified speaker: Um yeah, and I think people will put a very high valuation on on that expectation.
29:29
Unidentified speaker: So in that scenario, probably I'll be wrong.
29:33
Unidentified speaker: So I know as an investor you always need to look at your decisions and make an assessment about whether you were right or wrong.
29:40
Unidentified speaker: What would you need to see in Anthropic's first set of results as a listed company to know that you'd been too cautious?
29:47
Unidentified speaker: I I I suppose it's it's it's same old things, really.
29:51
Unidentified speaker: Revenue growth, um, cash flow.
29:55
Unidentified speaker: I'd I'd like to have, I'd like to find a way to take that snapshot view that I was talking about earlier.
30:03
Unidentified speaker: I'd like to have the opportunity to sort of press the pause button, if you like, uh in a simulation, if you like, of of Anthropic, take a look at the business and look at, well, what did it invest in in that model or that series of models, and what were the returns, what were the returns on investment in in those models?
30:22
Unidentified speaker: That that will give you a sort of a perspective.
30:25
Unidentified speaker: And of course, what you might also want to do is is to look at uh the all of the additional things that Anthropic was thinking of in developing and launching as ways to monetize the relationships that it has built with what I imagine will be a very much larger number of customers in two years' time.
30:45
Unidentified speaker: Okay, here's the second hypothetical case.
30:47
Unidentified speaker: Let's imagine that token prices and the price of AI keeps falling 10 times a year and open models do most of the work.
30:55
Unidentified speaker: In that scenario, who is still making money in 2028?
30:59
Unidentified speaker: This is a really important question, and I think it's it's very hypothetical and highly unlikely.
31:08
Unidentified speaker: But in that scenario where the model developers, the frontier model developers are not are not able to monetize their technology, and the open weight models remain virtually costless, in that environment, who where is the return?
31:27
Unidentified speaker: Who who owns the return?
31:29
Unidentified speaker: Well, you have to look to the to the businesses that have deployed the technology because the technology, I'm utterly convinced, I'm utterly convinced that the companies that are deploying this technology will will become much more productive.
31:52
Unidentified speaker: And that that's not just going to influence a few companies.
31:57
Unidentified speaker: I think this will have whole economy effects that that people have yet to get their heads around.
32:03
Unidentified speaker: I think there's going to be a product revolution.
32:06
Unidentified speaker: I think there's going to be a productivity revolution um across many economies that where this technology is deployed.
32:14
Unidentified speaker: So in the scenario you've just just painted, all of the upside or much of the upside is accrued to the companies that apply this technology and see their productivity transformed.
32:27
Unidentified speaker: Basically, they're they're they're buying a service which enables them to massively increase their profitability for virtually nothing.
32:36
Unidentified speaker: Now, that doesn't make economic sense to me, but in that scenario, clearly, companies that deploy the technology make massively higher returns.
32:47
Unidentified speaker: Of course, competition will limit to some extent how much that return grows.
32:53
Unidentified speaker: And of course, then, you know, there are all sorts of geopolitical and societal shifts that will take place.
32:59
Unidentified speaker: Lots of people will be out of work, you know, taxation will have to change, companies will have to be taxed more highly, people will probably be taxed much at a much lower level.
33:11
Unidentified speaker: But what's very clear in that scenario is that the economy will create the wealth that will enable redeployment and redistribution to enable society still to function appropriately.
33:22
Unidentified speaker: So, so in in that scenario, there's the productivity revolution ripples across all developed economies and and generates massively higher returns, huge increases in wealth, which which are then redistributed in in an appropriate way, presumably, or hopefully.
33:43
Unidentified speaker: I do actually see that as the future scenario.
33:46
Unidentified speaker: This technology, I think, will deliver that sort of outcome, but but the first bit, which is that that the model developers don't capture the economics, I think that's unrealistic.
33:58
Unidentified speaker: At some stage, the pricing shifts.
34:01
Unidentified speaker: When when market dominance or market concentration allows, or when competition allows, it it just doesn't make economic sense to me for a technology that is able to deliver that sort of transformation, that technology does not come for free.
34:18
Unidentified speaker: It doesn't, it, you know, the people who if the companies that evolve, develop, and own that technology will accrue an appropriate economic return on that technology, in my view, in the future.
34:28
Unidentified speaker: I think in that scenario, the the the picks and shovel suppliers are still doing very well because I would imagine that the the scale of build, the scale of investment in this technology industrial revolution will will have continued at a rapid pace.
34:46
Unidentified speaker: So the picks and shovel suppliers will carry on doing well in that environment.
34:52
Unidentified speaker: Neil, if the profit leaks out of the AI companies, where do you think it lands?
34:56
Unidentified speaker: Which kinds of businesses does that point you towards?
35:00
Unidentified speaker: Well, I I suppose the best example would be in the in the UK.
35:04
Unidentified speaker: I think we got the early signs already of this technology rippling through the economy, but particularly in in service-based industries.
35:13
Unidentified speaker: So banks, insurance companies, lawyers, all of these sort of, all of these sorts, all of these businesses that have to employ lots of very highly paid people to do tasks that I think ultimately in some respects will be performed by AI, what what will happen is that those companies will will shed highly paid labor.
35:38
Unidentified speaker: We're seeing this.
35:39
Unidentified speaker: I mean, Microsoft's announcement of two weeks ago that it would that it offered redundancy to 27% of its employee base globally, 60,000 workers tells you all you need to know about how this technology is going to transform the corporate landscape for those companies that are able to deploy it and use it intelligently.
36:01
Unidentified speaker: For me, the UK is is is well placed, incredibly well placed.
36:07
Unidentified speaker: It is a very service-oriented economy, the most service-oriented economy of any developed economy globally.
36:14
Unidentified speaker: Manufacturing is 8.5% of the economy.
36:17
Unidentified speaker: The UK economy is in some ways uniquely well placed to capture the upside from deploying this technology.
36:25
Unidentified speaker: Now, this will come with all sorts of associated issues.
36:30
Unidentified speaker: There'll be lots of challenges, but the productivity miracle that I think will will flow from this will deliver a massive increase in economic growth, wealth.
36:42
Unidentified speaker: And that the question then becomes well, how is that wealth redistributed throughout the economy to enable it to continue to function appropriately?
36:52
Unidentified speaker: Well, a lot of people are going to get the chance to buy anthropic shares this month.
36:56
Unidentified speaker: It's the first listing of a pure AI company.
37:00
Unidentified speaker: We had SpaceX, that was a space exploration company.
37:03
Unidentified speaker: Without telling anyone, Neil, what to do?
37:07
Unidentified speaker: What would you want to have answered before you buy into any AI listing?
37:12
Unidentified speaker: It's the same, it's the same answer that I have to that I have to be satisfied with with every single investment I've ever made.
37:21
Unidentified speaker: What is the the market price?
37:23
Unidentified speaker: What is the price I'm being asked to pay for this investment?
37:27
Unidentified speaker: And then to compare that market price with an educated, informed, calculated guess at what is the intrinsic value of that business.
37:39
Unidentified speaker: If that gap is sufficiently large to compensate for the risks associated with the business and the risks associated with how certain the risks associated with its growth and development, uh, the competitive environment, etc.
37:54
Unidentified speaker: If that gap is sufficiently large, then of course it becomes an attractive investment.
37:58
Unidentified speaker: And that's the same process that I deploy when I invest in anything.
38:04
Unidentified speaker: What are you going to be watching over the next three months that will tell you which way this is going, if it was below its intrinsic value or not?
38:11
Unidentified speaker: Well, I want to see Anthropic continue to deliver the sort of revenue growth that we've with that we've seen already.
38:20
Unidentified speaker: Um I'm I'm less concerned about the quarter-on-quarter growth, but what you want to see over slightly longer time periods, what you want to see is is is that that Anthropic is able to deliver the sort of revenue growth and cash flow growth that that you would associate with a highly valued company.
38:42
Unidentified speaker: Now, if if there will be all sorts of signals that will will tell you whether the key metrics that you would focus on as an anthropic investor, there are all sorts of things that will tell you, well, that that that's a tick or that's a cross.
38:57
Unidentified speaker: You know, there will be a lot of information available to uh to the market on anthropic.
39:02
Unidentified speaker: It'll be a very closely followed, closely analyzed company.
39:06
Unidentified speaker: So everybody will have access to the information that they will need to make those judgments.
39:11
Unidentified speaker: Whether they'll make them in the same way as I would, of course, is is a move.
39:15
Unidentified speaker: Many won't, but but that though that information will be what I will use to make my judgments about whether that gap between market value and intrinsic value is appropriate.
39:29
Unidentified speaker: Neil, for people that made it this far into the episodes, where can they get more of your thinking on this?
39:34
Unidentified speaker: They can get uh more of my thinking on this on the noise cancelling website.
39:38
Unidentified speaker: So that's noise canceling.co um uh where they will find all sorts of research pieces that back up what we talk about in these podcast episodes.
39:49
Unidentified speaker: Well, AI is making money.
39:51
Unidentified speaker: Anthropic keeps about 85 cents of every dollar of sales after running its models.
39:56
Unidentified speaker: What Neil's question today is how much of that the AI companies get to keep once they've paid for the next model, the computing contracts and the price cuts.
40:04
Unidentified speaker: Watch Anthropic's listing and its first results and the cloud company's spending plans in the new year.
40:09
Unidentified speaker: Neil's full analysis is free in the Noise Cancelling app on iPhone and Android, and the app will read all of Neil's pieces to you, or you can start at noisecancelling.co.
40:18
Unidentified speaker: Neil and I will be back with more next week.
40:21
Unidentified speaker: Until then, take care.
This page was produced automatically: the audio was transcribed with nvidia/parakeet-tdt-0.6b-v3 and interpreted by Qwen3.6-35B-A3B-GGUF:UD-Q4_K_XL. Transcription and interpretation can contain errors; always confirm against the original episode. Nothing here is investment advice.

More from Noise Cancelling with Neil Woodford

All insights