Original content by Dongdong / Produced by

Author | Yang Yang

Earlier this year, US stocks were battered by an "AI panic trade."

In the week after Anthropic released its AI plugin, the market value of US software and services stocks evaporated by nearly $1 trillion. Traders' subtext was blunt: SaaS's core business can be done by AI on the side.

But pinning the panic solely on SaaS underestimates the bigger picture. AI's impact on internet business models became concrete after Muse launched.

Personal agents like Muse have already bared their fangs at replacing super traffic gateways. Once the market saw this substitution possibility, gateway-type platforms broadly declined, with Booking and Airbnb both dropping nearly 20% within less than a month.

Muse bypasses product interfaces—comparing products, ordering food, hailing rides for users, handling everything from start to finish. Once gateways are bypassed, the underlying traffic distribution power is stripped away, and the share of core monetization methods like advertising and transaction commissions may naturally be carved out.

However, the same impact lands with different severity on Chinese and American companies.

The US internet is open and fragmented. Traffic platforms like Meta and Google only handle referrals, with users ultimately completing transactions on independent sites. With scattered scenarios, giants can't easily lock things down, and agents can skip the giants and place orders directly on public independent sites.

China is a different story. China's internet has long been closed-loop across various scenarios, with platforms completing the entire process from traffic gateway to transaction completion themselves. This closure can directly block personal agents.

AI's difficulty in disrupting Chinese internet companies far exceeds that of American companies. But high difficulty doesn't mean no risk—many domestic companies have hidden concerns about "supplierization."

The more users rely on AI for decisions, the looser traffic gateways become, leaving platforms with only heavy-asset segments like logistics, warehousing, instant delivery, payment clearing, and after-sales service. "Supplierization" weakens platforms' bargaining power, pushing profit structures downward.

Looking around, the one least likely to lose is still Apple.

/ 01 / The US Internet Business Model Has Been Shaken

On the eve of Anthropic's IPO, the market's highest valuation expectation for it was $2 trillion. For comparison, Meta's market cap is $1.58 trillion. A company that only recently commercialized is valued above the global social media giant.

The valuation multiple gap is even more striking: annualizing this year's first-half revenue, Anthropic trades at 61x PS, Meta at 8x, and Google at under 10x.

What capital is betting on is clearly not just intrinsic value. It's buying a possibility of replacement: new technology tearing down old gateways, large models taking over search, matching, content production, and bit by bit eating into internet giants' most core profit pools.

Especially after personal agents emerged, the odds of this bet look higher.

On September 8, Meta's personal agent product Muse launched, open only to the US and Canadian markets. Even so, it quickly topped the App Store free charts, surpassing 10 million downloads in 40 days—faster than ChatGPT did back then.

What Muse does is take over steps users originally had to do themselves: comparing prices, placing orders, making payments, booking flights, hotels, restaurants, filling out forms, sending emails, scheduling, and dealing with customer service. Zuckerberg designed its revenue from two sources: transaction commissions and subscriptions.

Once Muse launched, Meta's identity changed: from a repeatedly thwarted "American Baidu" to a leader in personal agents. Within a week, its market cap increased by $290 billion.

The decline on the other side came quickly. After the market understood what Muse pointed to, gateway-type platforms broadly fell. Google and Amazon dropped 3.80% and 2.24% in a single day, on the mild end; Booking and Airbnb both fell nearly 20% in September.

Why can one product launch cause a batch of platforms to fall simultaneously? The reason isn't Muse itself, but what it points to.

The internet business is fundamentally built on the attention economy. DAU and user time are the tax base; take rate is the tax rate. All products desperately acquiring users and fighting for time are, in the end, first enclosing land, pulling people into their interfaces, then converting traffic into money through advertising and e-commerce.

But what Muse does is cut off the traffic gateway at the tax source. When personal agents bypass product interfaces—comparing products, ordering food, booking tickets, hailing rides for users, handling everything from start to finish—the internet's most fundamental traffic distribution power is taken away.

On this matter, a16z was prescient. In its 2025 annual report, it wrote: "The KPI of screen time is ending." Perhaps in the future, DAU should be replaced by another metric: Task Completion Rate, or TCR.

Once the yardstick changes, the business model changes with a generation. Only this time, the generational shift lands on Chinese and US markets at different paces.

/ 02 / What Domestic Internet Faces Is Deflationary Concern

Muse isn't new. Doubao could say: what you have, I've long had.

As early as December 2025, the Doubao phone had nearly replicated all of Muse's functions. But it didn't make a splash, and domestic internet asset stock prices showed no reaction.

The reason isn't complicated: the moment it entered super apps, it hit "abnormal login environment" and accounts were forcibly logged out. The same product triggering completely different stock reactions in China and the US comes down to ecosystems.

US internet services are open and fragmented. Meta and Google play referral roles, with users ultimately completing transactions on independent sites. Even Amazon, strong as it is, has less than 40% market share in US e-commerce. More often, to buy something or book a room, people open different official websites, register accounts, fill in addresses, enter card numbers, and complete payments.

With no closed-loop scenarios and each platform handling only one segment, no one can shut the door for the entire web. Agents thus have room to operate in the open ecosystem, bypassing traffic platforms and costing them advertising revenue; even transaction sites like Expedia could be downgraded from gateway to supplier, with commissions and bidding ads suffering accordingly.

China has a different ecosystem. China's internet has long been closed-loop across various scenarios, with platforms completing the entire process from gateway to transaction themselves. This closure can directly block personal agents at the door. In other words, disrupting Chinese internet companies is far harder than American ones.

But high barriers don't mean no hidden concerns. AI is giving internet platforms deflationary worries.

Advertising is the internet's core monetization model. This business can be simplified into a multiplication: number of ad slots × price per ad slot. AI's risk is precisely that both ends move simultaneously.

First, quantity. The physical basis of ad slots is dwell time—news feeds, splash screens, detail page recommendations are all sold by impressions and duration. Once users get used to delegating decisions, impressions decline and inventory compresses.

Then, unit price. Ad prices are built on choice overload and information asymmetry—users can't choose, merchants can't explain clearly, and that's where the premium comes from. What AI does is exactly eliminate information gaps, pushing unit prices toward convergence.

Deeper still is the change in identity.

Internet giants' past high profitability was built on super traffic gateways and the traffic distribution power derived from them; once user decisions increasingly rely on AI, this gateway loosens, leaving platforms with only heavy-asset, high-cost segments like logistics, warehousing, instant delivery, payment clearing, and after-sales service. Supplierization weakens platforms' bargaining power, and profit structures may shift downward.

As the internet teeters, terminal devices represented by Apple have instead become a safe haven.

/ 03 / Only Apple Is Hard to Beat

This year, no matter how US internet assets have fluctuated, Apple has sat firmly on the fishing platform.

Year to date, Apple is up 26%, with market cap standing at $5 trillion—the elephant still dances. Over the same period, Google is up 1%, Meta down 4%, Amazon up 7%.

Apple can be a safe haven because it's the one least impacted by AI.

The phone is the irreplaceable personal computing terminal today, and a gateway more fundamental than AI and the internet. No matter how well models run, they still end up installed in this device. So no matter who ultimately wins in AI, Apple won't lose.

This position also lets it dodge a capital consumption war.

Internet giants don't have this choice. To avoid being killed by AI, they can only desperately do AI—if they succeed, they remain core assets of the next era; if someone else succeeds, they get revolutionized.

The harder AI works, the faster money burns. Many internet giants have already exhausted their free cash flow. According to Wall Street forecasts, the combined free cash flow of Amazon, Alphabet, Microsoft, and Meta in Q3 is expected to drop to about $4 billion, compared to an average of $45 billion per quarter since 2022.

Tech giants with insufficient cash flow to burn have also become heavily indebted. This year, Google issued $31.5 billion in bonds; Amazon issued $37 billion in March and raised about $25 billion more in July.

By comparison, Apple, which doesn't need to scramble in AI, is living exceptionally comfortably. While internet companies frantically snap up GPUs, Apple's capital expenditure is only one-tenth of internet giants', and its free cash flow remains as high as $100 billion.

No one knows who will win the AI war, but that Apple most likely won't lose is a consensus among many.