
Nov 13, 2026 · 1h 13m
AI could create vast value without making model companies rich
#65 - The Tech Analyst - Benedict Evans explains Anthropic's economics
The episode separates AI’s potential economic impact from the harder question of which companies, infrastructure layers, and services will capture the returns.
- 1AI adoption can expand rapidly while hyperscalers and model providers earn commodity-like returns on enormous infrastructure spending.
- 2Model performance and prices may converge, shifting durable value toward applications, integrations, domain expertise, and professional services.
- 3Investors need evidence of sustainable usage, cost reductions, and new behaviors rather than speculative spending or inflated total-addressable-market forecasts.
Don't miss
Benedict Evans reframes the AI investment debate: enormous economic value may emerge while model companies capture only a fraction of it.
The brief
Benedict Evans and Steve use Anthropic’s IPO prospectus to examine a market where revenue growth, inference costs, training bills, pricing, and future demand support sharply different interpretations.
The infrastructure boom may resemble earlier telecom cycles: widespread adoption can create enormous utility without delivering attractive returns to every network, chip, or model provider.
GPU economics depend less on a fixed depreciation schedule than on workload, power availability, failure rates, chip efficiency, and the returns generated per watt.
As frontier models converge and prices fall, value may migrate toward applications and professional services that combine models with data, compliance, integrations, and domain expertise.
The standout tension is valuation: AI could transform software, search, cybersecurity, and drug discovery while conventional forecasts still miss where entirely new value pools emerge.
Evans’s investor test is practical rather than promotional: look for durable product-market fit, genuine usage beyond the technology ecosystem, falling costs, and new user behavior.
What was said on this episode
38 statements · 15 positive · 15 negative · 2 mixed · 6 neutral
Anthropic’s revenue grew at least tenfold in one year.
“the revenue has grown 10x in a year or more”
Listen at 4:34
Anthropic is approximately cash-flow break-even.
“the company is now more or less breaking even on a kind of cash flow basis”
Listen at 5:39
Anthropic’s inference costs stayed roughly flat while revenue increased.
“the cost of inference, which is the cost of revenue, has basically stayed flat as the revenue has increased NX”
Listen at 6:00
Inference providers can price according to customers’ return on investment.
“you can price at the customer's ROI”
Listen at 6:51
AI inference costs are falling roughly tenfold annually or faster.
“the cost is dropping sort of 10x a year or more”
Listen at 7:05
Future AI model pricing remains uncertain.
“you don't know what the pricing is going to look like”
Listen at 7:56
Anthropic’s future competitive differentiation remains uncertain.
“you don't know whether they're going to be able to differentiate and have some kind of competitive advantage over model companies”
Listen at 8:10
AI adoption could become widespread while prices collapse.
“it's entirely possible that you get an enormous amount of adoption for this stuff at quite low prices, that you get a pricing collapse”
Listen at 10:34
AI infrastructure may become low-margin commodity infrastructure without pricing power.
“the infrastructure itself is a low margin. is low margin commodity infrastructure with no pricing power”
Listen at 10:50
Persistent supply shortages could sustain high AI margins despite no differentiation.
“supply remains behind demand for quite a long time, and so you could have quite high margins with zero differentiation”
Listen at 11:48
AI value capture may shift toward application-layer software companies.
“the value goes further up stack”
Listen at 14:01
Future AI scientific advances remain unpredictable.
“we don't know what the science is going to produce next”
Listen at 15:05
AI progress and growth could flatten by next summer.
“there's like a worst case scenario, which is the whole thing kind of flattens out next summer”
Listen at 15:22
Hyperscalers’ capital expenditure has exceeded their free cash generation.
“the CapEx has outgrown the free cash throughout these companies”
Listen at 17:34
Replacing older chips can produce better overall returns per watt.
“it's a better overall return to throw those chips out and put in new ones that get better results per watt”
Listen at 18:52
Power constraints favor replacing chips with more efficient ones.
“if you're power constrained, then it makes sense to put in new chips that get a higher... return per watt”
Listen at 19:27
AI infrastructure could eventually be sold at marginal cost across the stack.
“everything at every layer in the stack was being sold at marginal cost”
Listen at 22:54
AI models currently lack a strong, general competitive moat.
“The simple answer is today, not really.”
Listen at 25:57
Anthropic derives 80% of revenue from models released within six months.
“80% of revenue comes from models built in the last six months”
Listen at 28:16
One-year-old AI models are commoditized and very inexpensive.
“the model from a year ago is such a commodity and it's super, super cheap”
Listen at 28:23
Frontier models may retain pricing power and differentiation for demanding use cases.
“for that, there's pricing power and maybe some differentiation”
Listen at 29:05
Application developers will switch models and clouds based on token prices.
“you're going to be looking at your price per token, which model running on which cloud. And you're going to be swapping them in and out.”
Listen at 29:28
AI usage is currently concentrated in software development but this concentration is temporary.
“this is very, very highly concentrated in software development. But I think that's, you know, that is transitory.”
Listen at 31:15
Natural-language business-question systems constitute a new software category.
“That's new software.”
Listen at 33:58
Anthropic cannot independently create all domain-specific enterprise applications.
“It's not Anthropic. Anthropic isn't going to be able to create all of those by themselves.”
Listen at 35:13
Useful enterprise AI requires purpose-built software, not merely access to a model.
“you're going to need to build an actual piece of software for this, not just give everybody fraud”
Listen at 38:40
Enterprise-wide AI deployment will not be achieved by Claude alone on a CFO’s laptop.
“It's not Claude running on the CFO's laptop. That's one thing you can say with certainty. That's not how any of that happens.”
Listen at 42:18
Conventional TAM exercises cannot reliably value transformative AI companies.
“you can't value things like that. You don't know what the new things are going to be.”
Listen at 44:36
Value created by AI applications may not accrue primarily to Anthropic or OpenAI.
“it's going to be really tough for the stuff that gets built with this all to accrue to Anthropic and OpenAI”
Listen at 47:31
Frontier model development may consolidate into one or two companies.
“it may be that that's what models look like. That it turns out that with each generation, they get harder and more expensive and then you only have one or two companies doing it.”
Listen at 49:50
AI currently has explosive product-market fit in software development.
“Clearly, we have explosive product market fit for software development right now.”
Listen at 53:44
Muse’s product-market fit outside Silicon Valley and at scale remains unproven.
“Tell me it's working outside the valley. Tell me it's working at scale and sustainably and really working.”
Listen at 55:16
OpenAI and Anthropic lack access to users’ broader behavioral data.
“OpenAI and Anthropic don't have any of that.”
Listen at 56:50
Google can provide search-like ChatGPT functionality, probably better.
“Google can do that too. And they can probably do it better.”
Listen at 58:19
Google, Meta, and Amazon have gained double-digit ad-revenue percentage points through AI optimization.
“they've all got like double digit percentage points gains in their ad revenue simply from optimizing”
Listen at 1:00:18
Apple is well positioned for AI devices through hardware, privacy, battery, camera, and silicon strengths.
“Who is good at making beautiful devices with a strong privacy story, great battery life, a great camera, and really strong custom silicon? Sounds like Apple to me.”
Listen at 1:01:33
Amazon will resist becoming a passive API provider for third-party agents.
“Amazon wants to own that customer relationship. They don't want to be somebody else's dumb API call.”
Listen at 1:03:23
Much AI-created value may become efficiency gains, lower prices, or consumer surplus.
“there'll be a bunch of value that doesn't accrue to anybody. It's just efficiency and maybe lower prices or consumer surplus.”
Listen at 1:08:36
Statements are attributed to the speaker as said on the episode and reflect their view at the time, not PodLume's. They are not advice.
Books & mentions
Some links are affiliate links — PodLume may earn a commission if you buy.
Weekly email
State of the Pod, every Monday
The week's most-discussed names and stories from podcasts, with the moments behind them. Free, five minutes, unsubscribe in one click.
