How China Just Overtook America In AI Traffic

The Viral Orchid Launch Video [00:00:59]

• An AI startup called Orchid posted a viral launch video featuring an AI assistant helping a woman whose boyfriend forgot their anniversary.
• The AI books a reservation and gets flowers, leading the boyfriend to text, "Okay, okay, I can do this." The AI responds, "You can't, that's why I'm here."
• The video sparked vitriol, seen by some as an example of AI developers being out of touch with the public.

"Instead of trying to be a better significant other, you can just stay a piece of shit and have AI cuck your girlfriend."

Reframing the AI Discourse [00:02:23]

• The speaker argues against getting lost in the "cultural layer" of AI, such as marketing or individual user behavior.
• Instead, AI should be understood as an "arms race" between the US and China for intelligence.
• Intelligence is framed as the most critical battle for nations, underscoring the importance of understanding AI as a "weapons system."

"You need to understand it as a weapons system."

Human Intelligence vs. AI [00:03:38]

• Humans' dominant impact on the globe stems from their brains being organized for higher-level intelligence.
• This includes the ability for theory of mind, future planning, and mental manipulation to build objects.
• The focus should be on scaling intelligence, not just the "artificial" aspect, which could radically transform the world.

"Intelligence, flexible intelligence, is precisely the end goal if we want to move forward as a species."

The AI Financial Bubble [00:05:08]

• AI is described as being in a financial bubble, where asset prices rise to unsustainable levels detached from actual value.
• The speaker aims to argue from the "bull case" of AI, acknowledging the bubble concerns but presenting a different perspective.
• There's disagreement on whether the market is already "unhinged" or just beginning to tap into AI's potential value.

"A financial bubble is when the asset prices of a certain industry rise to unreasonably high fake levels that do not match the actual value of what's being built."

Hyperscalers and AI Infrastructure [00:06:42]

• Hyperscalers (Meta, Microsoft, Alphabet, Amazon) are the largest cloud computing and data center providers, building most AI infrastructure.
• Since 2021, these companies have invested heavily in AI development.
• Their ability to justify this spending stems from AI being a race to build the "next big thing" and their substantial free cash flow.

"Hyperscalers are the largest cloud computing and data center providers who have created most of the AI infrastructure."

Hyperscaler Cash Flow Plummets [00:07:29]

• Just two years prior, the four biggest hyperscalers generated roughly $210 billion in free cash flow annually.
• By 2026, their free cash flow had "plummeted to below zero."
• This wealth transfer is shifting from hyperscalers to semiconductor companies.

"But in 2026, their free cash flow has plummeted to below zero."

Optional Expenditures vs. Fundamental Flaws [00:08:13]

• Hyperscalers are "optionally very valuable" and "default alive" due to strong core businesses.
• AI spending is an "optional fashion" expenditure; they could halt it if it fails, returning to profitability.
• This is compared to Meta's bet on the Metaverse, which was later scaled back without destroying the core business.

"So understanding that what they're doing now, they're not investing in something that is destructive."

The Rise of Hyperscaler Debt [00:10:15]

• To fund continued AI spending, hyperscalers are increasingly taking on debt by issuing bonds.
• The annual volume of bonds issued by hyperscalers surged from roughly $20 billion in 2024-2025 to $150 billion in 2026.
• This trend is being likened to the dot-com bubble of the late 1990s and early 2000s.

"From 2024 to 2025, the annual volume of bonds issued from hyperscalers went from roughly $20 billion to $110 billion."

Debt-to-Value Ratio: A Saner Place? [00:11:45]

• While raw debt figures are extraordinary, the debt as a percentage of hyperscalers' value is around 4% for AI.
• This is significantly lower than the 30% seen at the height of the dot-com mania.
• The speaker finds this manageable level "shocking" and a sign of a saner situation compared to the dot-com bubble.

"Right now, they're at 4% in AI, so a very manageable number versus 30% at the height of the dot-com mania."

The Bull Case: AI as an Arms Race [00:13:04]

• AI is the single most important technology, and the US is in an arms race with China.
• The US government would likely backstop the industry to prevent collapse, ensuring infrastructure investments are utilized.
• This government support, combined with manageable debt-to-value ratios, suggests less risk than the dot-com bubble.

"The US government is going to step in and either prop that company up, or they're going to step in and help sell off the assets to the other companies."

Stock Market Valuations Sound Alarms [00:15:22]

• The CAPE ratio (Cyclically Adjusted Price-to-Earnings) is at 40, far exceeding the traditional 16-17X.
• A CAPE ratio of 40 was last seen just before the dot-com bubble burst.
• This high valuation suggests "crazy territory" for the stock market, even if the companies themselves are doing fine.

"The last time we were at 40 was right before the dot-com bubble."

The "Miracle" Revenue Projections [00:18:09]

• The top five hyperscalers project doubling revenue in three years while cutting $80 billion in operating expenses.
• This requires unprecedented cuts in sales, general, and administrative (SG&A) expenses to offset soaring depreciation.
• An accounting professor notes this scenario of SG&A offsetting depreciation increases has never happened before.

"I can't think of a scenario where that's ever happened before."

Jensen Huang's Data Center Pitch [00:19:19]

• Nvidia CEO Jensen Huang argues AI hardware doesn't depreciate as quickly as assumed (e.g., H-100s used for six years).
• He proposes viewing data centers as long-term rentals generating rent, not depreciating widgets.
• Huang backs this with a financing package, guaranteeing 25% of the value of certain chips against decline.

"The reality is that we have H-100s that came on six years ago that are still being used."

AI vs. Railroads: Revenue Now [00:21:23]

• Unlike railroads, which required infrastructure build-out before generating revenue, AI generates revenue immediately.
• Companies like Anthropic are experiencing unprecedented revenue growth.
• While older chips may decline in value, the immediate revenue stream and increasing compute costs present a different economic model.

"The revenue is growing at an incredible rate. The usage rate is growing at an incredible rate."

Potential Threats to AI Value [00:23:06]

• Increased competition from other companies entering the AI space could lead to an overproduction of compute.
• More efficient algorithms, potentially from China, could reduce the need for compute for the same output.
• These factors could cause chip values to decline, even if demand remains high.

"There are several scenarios that could cause the value of those chips to decline, not just a reduction in demand."

The Job Market and AI [00:24:50]

• The narrative that AI will eliminate jobs and require rehiring is dismissed as "bullshit."
• AI excels at certain tasks, while others are diminished; some jobs will be eliminated (e.g., data entry), others drastically reduced (e.g., driving).
• New jobs will be created, often requiring different skillsets, focusing on AI integration rather than resistance.

"The things that it does well grow by the day. The things that it does terribly get diminished by the day."

Companies Rehiring After Layoffs [00:26:32]

• Market research indicates 29-32% of organizations that cut jobs due to AI projections are restaffing those roles.
• This suggests companies are hiring individuals skilled in leveraging AI, not necessarily replacing laid-off workers with identical skillsets.
• The job market remains challenging, with "ghost jobs" and disappearing entry-level positions.

"Roughly 29% to 32% of organizations that cut jobs due to early AI projections have already started restaffing those exact roles."

China's AI Cost Efficiency [00:28:08]

• China is excelling at creating scalable, cost-efficient AI products with thinner margins to gain market share.
• They allocate tax dollars to industries and are adept at optimization.
• While not yet leading in "frontier models," China is rapidly catching up, posing a national defense concern for the US.

"China is extraordinarily good at, they'll allocate tax dollars to an industry to make sure that it's thriving."

Kimi K3: Open Source Challenge [00:29:40]

• Kimi K3, an open-weight model from China's Moonshot AI, can be downloaded and modified, unlike closed US models.
• This offers lower costs (up to 60% reduction in token cost) and comparable performance.
• While beneficial for users, it could impact US companies' revenue generation and bleeding-edge research.

"Unlike the closed models from OpenAI and Anthropic, Kimi K3 can be downloaded, modified and run on your own hardware without relying on Moonshot's servers."

Intelligence Per Dollar: The Real Race [00:30:56]

• Maximizing "intelligence per dollar" is a crucial metric, essentially meaning getting the most intelligence for the cheapest cost.
• This concept is vital for companies aiming to hire the "smartest people for the least amount of money."
• China's models are increasingly comprising AI traffic, driven by this efficiency.

"It's bang for buck just in AI words."

China's Dominance in Traffic [00:31:44]

• Chinese models' token traffic grew from 1.2% in 2024 to over half by summer 2026, surpassing American models.
• This shift could shrink revenue for US companies invested in OpenAI and Anthropic, impacting their spending on chips and infrastructure.
• A significant shift to Chinese models could accelerate the bursting of the AI bubble.

"According to Open Router, Chinese models went from 1.2% of token traffic in 2024 to more than half of it by the summer of 2026, overtaking American models in total volume."

Public Sentiment Turns Negative [00:34:16]

• While positive sentiment towards AI effects increased from 15% to 27% (2023-2025), negative sentiment rose from 40% to 47% in the same period.
• This negativity is attributed to anxiety about AI's impact on life, meaning, purpose, and economic prospects.
• Concerns extend beyond abstract fears to tangible issues like data centers impacting communities and the proliferation of "AI slop."

"However, in that same period, those who said it will have a very negative or somewhat negative effect went from 40% to 47%."

The Scandal of Destructive Scanning [00:36:38]

• AI companies, including Anthropic, have reportedly acquired and "destructively scanned" millions of physical books.
• This process involves removing book spines, scanning pages, and discarding the original copies.
• The rationale is that physical books are "verifiably not AI slop writing" and provide valuable, authentic data.

"The internal project, dubbed Project Panama, relied on a process called destructive scanning."

A Different Take on Book Destruction [00:37:55]

• The speaker argues that many books are not being read, and digitizing them for AI training preserves knowledge.
• Learning from AI patterns is seen as more valuable than simply reading the original book.
• This is framed as standing on the shoulders of giants, a historical method of knowledge progression.

"So somehow, someway, getting these books actually recorded to me is far more interesting."

The Emerging Market for Books [00:39:35]

• Companies like ISBN DB advertised services to help AI developers acquire large quantities of physical books for training.
• While ISBN DB claims the service was never launched, the proposal indicates a market for supplying books to AI labs.
• A Dutch bookseller received requests for thousands of academic titles, suggesting a targeted acquisition effort.

"The company advertised a service that would help AI developers acquire anywhere from 1,000 to 1 million printed books for large language model training."

Preserving Knowledge or Theft? [00:41:34]

• The speaker rejects the idea that uploading data into AI is inherently theft, comparing it to learning from human authors.
• Creating AI "TomBots" based on public figures' content is seen as a natural extension of learning and adaptation.
• The efficiency of ingesting human knowledge for broader access is viewed as a worthwhile trade-off.

"The fact that we've now made it far more efficient, that we've ingested all of human knowledge and given that access to everybody that has access to an internet connection is a trade-off that is well worth it."

The Scarcity of Authentic Data [00:43:47]

• AI companies are desperate for authentic human-made data, which is becoming increasingly scarce.
• This scarcity contrasts with investors betting on AI becoming intelligent enough to improve itself.
• The future of the AI spending boom depends on which of these narratives proves true.

"The industry's most valuable resource is still authentic human writing, and it's becoming increasingly scarce."

Where Are We in the Bubble? [00:45:04]

• The speaker's view is between pessimism and extreme optimism, acknowledging AI as a critical arms race.
• Government backstopping is likely, but debt accumulation relative to revenue growth is a key factor for investors.
• The current underestimation of AI hardware depreciation is a significant point, suggesting immediate revenue generation is possible.

"And so when you ask a question, not is this a bubble, but where are we in the bubble? Are we early innings? Are we late innings?"