This article is written based on public information and is intended for informational exchange only, and does not constitute any investment advice.
In the last week of September this year, Peter Thiel, who has invested in OpenAI and Anthropic, received the Axel Springer Award in Berlin and held a conversation with Mathias Döpfner.
Applying his monopoly theory from the internet era outlined in "Zero to One" to examine today's model companies, he made a highly contrarian judgment:
The Google of the model era, a company capable of capturing monopoly value, may never emerge in the AI era.
1. Value Capture Dynamics in the Internet Era: Network Effects and Winner-Take-All
If we compare the model era with the internet era:
When the internet first took off, many assumed anyone could start a company. As it turned out, the network effect economy was extremely strong, giving rise to a group of super-large platforms and forming a winner-take-all landscape.
Google, Amazon, Meta, Tencent, ByteDance, Alibaba, and countless others.
The model era might be exactly the opposite:
The market is huge, but the endpoint looks more like intense competition rather than a few platforms taking it all. A large portion of the value will fall into the hands of users in the form of consumer surplus (Konsumentenrente), with a more even distribution.
You pay a USD 20 subscription fee and receive USD 2,000 worth of medical or knowledge services; most of the value accrues to you.
Note that Thiel is also expressing an "I wonder" sentiment.
This does not contradict his 2014 book "Zero to One," where he stated that "competition is for losers" and "good companies must build monopolies":
For a company to be valuable, it must do two things simultaneously: create value X, and then capture Y% of it.
Airlines create a massive X, but their Y is close to zero.
Google creates a relatively smaller X, but its Y is extremely high.
Looking at the current landscape, the X at the model layer could be enormous, but the Y might be very low.
Value is being created, yet companies like OpenAI, Anthropic, Zhipu, Moonshot AI, and DeepSeek cannot retain much of it.
An old example he used in a Stanford class in 2012 remains highly relevant:
The average one-way ticket price for U.S. airlines was USD 178, earning only 37 cents per passenger, while in the same year, Google's revenue reached USD 50 billion with a profit margin of about 21%, and its market capitalization exceeded the combined market cap of all U.S. airlines.
If you fail to form a monopoly, you might end up burning through all of the investors' money.
Warren Buffett once said that if he had a time machine, his first move would be to travel back to the Wright brothers and shoot down their airplane with a rocket.
Thiel is not saying that AI (Artificial Intelligence) is unimportant; rather, he is questioning whether today's red-hot model layer might (maybe) become the new airline industry:
Generating immense social utility, yet company profits are competed away.
2. Why the Internet Easily Breeds Monopolies
Internet platforms have numerous moats:
Network effects: Each additional user makes the platform more valuable to other users. This applies to social media, payments, marketplaces, and search.
Data flywheel: The more people use it, the better the product becomes; the better the product, the more people use it.
Extremely low marginal costs: Serving one additional user costs almost nothing; scale itself is a barrier.
Attention gateways: Controlling search, news feeds, and operating systems allows for rent extraction.
Winner-take-all matching markets: Once a platform becomes the default choice, it is extremely difficult for the runner-up to survive.
This is why monopolistic companies like Google, Meta, Amazon, Alibaba, and Tencent emerged:
They created massive consumer surplus X (many products are free or extremely cheap for users—think of Honor of Kings), while capturing a large chunk of value Y through advertising, commissions, and cloud services.
Users gained value, and the platforms became immensely valuable, often reaching trillion-dollar valuations.
The reason the internet did not turn into the airline industry is that platforms stacked network effects, data flywheels, near-zero marginal costs, and gateway control together.
Search remains the clearest example to this day:
Google's global search share still hovers around 90%, and Google combined with YouTube remain the two most visited sites globally.
Advertising revenue is equally concentrated:
In 2025, Google, Meta, and Amazon collectively captured 56% of U.S. advertising revenue, and research institutions project this proportion will rise slightly in 2026.
Cloud and traffic gateways have not been decentralized either; AWS, Azure, and Google Cloud still collectively hold about 60% of the cloud infrastructure market, and the three handle nearly 8% of global internet traffic.
Users often use these platforms for "free," but attention, merchants, and developers are paying rent, meaning the internet simultaneously generates massive consumer surplus and massive platform market capitalization.
Tencent Holdings is the most valuable Chinese internet company listed today, and every single point above applies to it.
3. Why Thiel Believes the Model Economy Might Fail to Capture Value
In Thiel's view, the model layer does not look like a monopoly business from any angle.
First, models are more like substitutable intellectual commodities rather than unique gateways.
For instance, search has only one default box, and Tencent's social relationships are extremely difficult to migrate.
However, the cost for users to switch models is very low: the same task can be asked of DeepSeek, Claude, Workbuddy, or Zhipu.
Product differentiation is often just a window of a few months. The capabilities among models are also rapidly catching up.
Closed-source leads, open-source catches up, and the next player overtakes again—this cycle has repeated several times.
Epoch AI found that since 2023, the cost of achieving the same capability level has dropped by approximately 47% per quarter, equating to becoming about 13 times cheaper annually, a faster historical decline rate than DNA sequencing, computing hardware, lithium batteries, and electricity.
Once capabilities converge, a price war becomes almost inevitable.
Airlines similarly create immense value, yet their profits are competed away: this is precisely the example Thiel himself loves to use.
Second, users want results, not specific model brands.
Whether writing code, editing contracts, translating, or generating images, users care about "good enough, fast enough, and cheap enough." Brand loyalty at the tool layer is typically very weak. Once the tool layer becomes commoditized, the value surplus flows to the buyers.
Stanford HAI reported that the inference cost for the GPT-3.5 tier dropped from about USD 20 per million tokens in November 2022 to USD 0.07 in October 2024, becoming roughly 280 times cheaper. By September 2026, the token price index for frontier models had fallen by about 80% compared to the March 2023 baseline.
In fact, as long as there are sufficiently good open-weight models, closed-source models struggle to charge an "intellectual monopoly tax."
A recent financial breakdown disclosed by Hock Tan illustrates this perfectly: global inference infrastructure costs are approximately USD 200 billion annually, while model-related revenue is about CNY 150 billion; token volume is roughly split half and half between closed-source and open-weight, yet revenue is about CNY 120 billion for closed-source and CNY 30 billion for open-source. Open-source has anchored the prices down.
Third, model switching happens within APIs and workflows, not within social relationships. Switching models is not like switching WeChat Moments, nor is it like switching e-commerce platforms.
Enterprises can connect the same RAG, the same agents, and the same permission systems to another model. If network effects exist, they are more likely to grow at the workflow, data, and permission layers, rather than in the foundation models themselves.
Following this logic, the model era is more like electricity, container shipping, or aviation: society becomes wealthier, but power plants and airlines do not necessarily become Google.
Perhaps, in Thiel's eyes, the Google of the model era will never emerge.
4. This Does Not Mean "There Are No Monopolies in AI"
It is crucial to point out that Thiel is referring to fewer winner-take-all platforms, not the absence of windfall profits across the entire value chain.
For instance, Thiel has always believed that the chip layer is the piece of the cake ignored by U.S. investors:
NVIDIA's total revenue for fiscal year 2026 was USD 215.9 billion, of which USD 193.7 billion came from data centers; by the second quarter of fiscal year 2027, data centers accounted for 93% of the company's revenue.
In terms of computing power, Epoch AI estimates that in the fourth quarter of 2025, NVIDIA still accounted for about 68% of the AI chip supply, Google's in-house chips accounted for about 19%, and the remainder was split among Amazon, AMD, and Huawei.
Second, looking at the computing power center layer, electricity, land, HBM, and advanced packaging also resemble rent monopolies:
This is also why Volcengine, Alibaba Cloud, and Baidu Cloud remain highly valued today, and a key reason why DeepSeek is beginning to extend into computing power centers: everyone can see the value.
Elon Musk's SpaceX AI space computing power is similarly an extension of this logic.
Third, the value of distribution gateways still has the potential to capture value:
Models can be commoditized, but default search boxes, office suites, mobile assistants, and enterprise identity systems do not have to be. Microsoft embedding models into Office and Google embedding models into search are following this exact path.
Tencent, Alibaba, and ByteDance are all heavily betting on Workbuddy, Doubao Office, Qwen Office, and even Qwenbook and Doubao phones, reflecting the same underlying logic: establish gateways, and the more front-loaded, the better.
If models cannot form monopolies, gateways still hold value: it might not exactly be a repeat of the 360 wars, but it is close enough.
Finally, at the workflow context layer, companies that Thiel favors could still emerge:
Once models are embedded into operating systems with permissions, auditing, industry data, and extremely high migration costs, what customers switch is no longer just the API, but the entire organizational process.
Copilot may not be as user-friendly as Claude, but the primary consideration for international enterprises is not efficiency, but security, auditing, and context;
Following this principle, the context advantages of Kingsoft Office WPS, DingTalk, and Feishu still exist. The idea that SaaS will be eliminated remains a mere illusion for AI investors.
Palantir, co-founded by Thiel, operates exactly at this layer. Therefore, while he says the model layer might be highly competitive, he simultaneously emphasizes that the urgency of U.S. enterprises to transform with AI is accelerating the company's enterprise sales; these two statements do not conflict.
Three things can hold true simultaneously:
Competition at the model layer is fierce, with thin margins;
Chips, electricity, and scarce capacity remain highly profitable;
A few companies that embed models into irreplaceable workflows may still achieve monopolies in Thiel's sense.
5. What Exactly Does "Value Goes to Consumers" Mean?
The value captured by consumers, as referred to by Thiel, is called consumer surplus in economics:
If you are willing to pay USD 20 for a good answer but actually only pay 20 cents, the difference is the value you capture.
The manifestations will be highly specific:
For individuals: The collapse in costs for writing, programming, translating, learning, and drafting legal documents.
For enterprises: The same number of people can do more, or the same output can be achieved with fewer employees.
For pricing: Shifting from high-priced subscriptions to usage-based pricing, and then being driven down further by open-source and price wars.
For distribution: The internet era was characterized by "astronomical market capitalizations for a few platforms + free usage for users"; the model era is more likely to be "large revenues but average profits for many companies + universally cheaper costs for users and work scenarios."
"More evenly distributed value" does not mean the disappearance of the wealth gap; rather, it means value no longer primarily accumulates as net assets for two or three platforms, but is scattered across the cost reductions of countless buyers.
Before everyone delves into the politics of employment, we might as well emphasize one more point:
An increase in consumer surplus does not equal an increase in worker surplus. Users are cheaper as buyers; the same group of people, as white-collar supply, may face replacement or wage suppression.
6. The Contrarian's Contrarian: Where Might Thiel Be Wrong?
First is the recently popular concept of regulatory capture:
Within the closed-source model camp, Dario's calculation is that if only licensed closed-source models can legally serve U.S. or European users, the model layer will revert to a protected oligopoly, and the open-source threat will cease to exist.
This is precisely the core of the open-source camp's criticism of Altman and Amodei: using safety narratives to legislate, keeping latecomers out, and recapturing a portion of the consumer surplus.
If this path succeeds, Thiel's premise of "intense competition" falls apart.
Second, distribution gateways may still be winner-take-all.
Models can be commoditized, but default gateways do not have to be. Whoever controls the operating system, office suites, mobile assistants, browsers, and enterprise identity systems can still extract rent. Microsoft embedding models into Office and Google embedding them into search are following this exact path. Monopolies may shift from "model companies" to "companies that already own distribution."
This brings us back to Tencent's internet muscle memory.
Once continuous learning, private data, and closed loops emerge, the moats will return.
Today, everyone is competing with the same batch of public data and similar architectures. If a system can continuously learn within a customer's own business, become more insightful the more it is used, and become impossible to migrate away from, the model will no longer be a replaceable API, but will turn into a system with network effects.
This is true for Claude, and it is true for Zhipu as well.
The capital barriers for training and inference remain extreme.
There may only ever be a few players capable of competing at the frontier. Oligopolistic competition does not equal perfect competition. Aviation is also an oligopoly, yet profits can be terrible; but in an oligopoly like chips, profits can be excellent. The key is not "how many," but whether differentiation can be sustained.
From this perspective, Kuaishou's Kling and Seedance can achieve the same.
Of course, Thiel's own investment portfolio has not been built on the premise that "models have no profits."
Founders Fund / the Thiel ecosystem has touched upon OpenAI and Anthropic, while also placing greater emphasis on chips and infrastructure. In fact, he treats "the model layer might be highly competitive" as a probability (I Wonder).
But if he is right, the investment logic in the AI era will be the exact opposite of the 2010s:
Do not default to the assumption that "building foundation models is building the next Google." That is pursuing Y in a layer that may become commoditized; historically, this is exactly the position of airlines.
For today's investors, the more worthwhile questions to ask are:
If the company you invested in removes a specific model from its product, will the customers still be there?
Are the barriers in data permissions, workflows, industry licenses, and on-site implementation, or just in prompts and packaging?
Is the company you invested in selling replaceable tokens, or an irreplaceable system?
Where does the shortage truly occur—at the layer of computing power, electricity, talent, compliance, or distribution?
7. For Capital, An Era of Uncertainty
If the scenario Thiel describes comes true, it also means a more uncertain era for investment is arriving:
For users, this world is more friendly: the same level of intelligence is becoming increasingly cheaper, choices are more abundant, and the degree of being locked into a single platform is declining.
For companies and capital, the future may be even more brutal. An increasing number of people in the AI era will create a massive X, yet still capture only a tiny Y. Their market capitalization narratives will look grim, even if the social utility is enormous.