One AI, two price tags: America bets on rent, China bets on irrigation
The same week China's leader offered AI to humanity, OpenAI was preparing a $1 trillion IPO. Neither side is telling the whole truth, and the numbers say both are being perfectly rational.
On 17 July, Xi Jinping stood up at the World AI Conference in Shanghai, his first keynote there, and described AI as an asset encapsulating humanity's collective wisdom, one that should benefit all of humanity. The same day, Moonshot unveiled Kimi K3, soon published as the largest open-weights model ever released. A few weeks earlier, OpenAI had confidentially filed with the SEC toward a listing at up to $1 trillion, Goldman Sachs and Morgan Stanley at the wheel.
Picture it: the same technology, at the same moment. One civilization is floating it on the stock market; the other is presenting it as a common good of humankind. Neither is telling you the whole truth, we will get to that, but the unsettling part is that both are being perfectly rational. To understand what is happening, don't follow the technology. Follow the money that pays for it.
Show me who pays the bills and I will show you the price of intelligence. The American machine has to sell it dear, because hundreds of billions of dollars of private capital are waiting for a return. The Chinese machine works to make it nearly free, and recovers the value everywhere else. Here are the verified numbers behind each bet, and the collision they are set on.
Follow the money: the verified numbers behind the two machines
1. America's capex wall: $725 billion this year, a trillion next
Amazon, Microsoft, Alphabet and Meta have guided to roughly $725 billion of combined 2026 capital expenditure, mostly AI infrastructure, up from about $410 billion in 2025, and analysts already see 2027 topping $1 trillion. Goldman Sachs projects $7.6 trillion of cumulative AI build-out between 2026 and 2031. Every layer of that stack, cloud, data center, model, eventually has to return the capital.
2. OpenAI's arithmetic: $2.20 out for every $1 in
OpenAI booked $13.1 billion of revenue in 2025 against a roughly $21 billion operating loss, and reported an adjusted operating margin of −122% in Q1 2026, $1.22 of operating loss per dollar of revenue. Its own projections don't reach profitability before about 2030. In March it closed a $122 billion round at an $852 billion post-money valuation (Amazon $50B, Nvidia and SoftBank $30B each, Microsoft participating). That capital is not philanthropy; it expects a return, which means access must be charged.
3. The price ladder, and the $1 trillion question
Since 9 February, US free-tier ChatGPT carries advertising; above it sit the $8 Go tier (also with ads), $20 Plus and $200 Pro. With 900 million weekly users (announced in February), 2025 revenue works out to under $15 per user per year, about $28 on the current ~$25 billion run rate. An IPO near $1 trillion would price OpenAI at roughly 40 times that run rate, for a company that loses money on every dollar it collects. Internal targets say $280 billion of revenue by 2030.
4. China's biggest AI spender is heavily outspent
ByteDance, China's largest AI investor, is weighing 400–500 billion yuan (about $56–70 billion) of 2026 capex, under 40% of what a single average US hyperscaler spends. The bill is already real: ByteDance's 2025 net profit fell more than 70% on AI spending (the widely repeated "$50 billion profit" was a projection, not the outcome). Even adjusting for lower Chinese costs, the financing gap is enormous.
5. Yet the under-financed side is the one cutting prices
On 24 April DeepSeek released V4 open-weights under MIT license, free for commercial use, with API pricing roughly 10–30x below US frontier models. Kimi K3, unveiled at WAIC on 17 July, full weights published 27 July, is the largest open-weights model ever released. Zhipu's GLM-5.2 went MIT on 13 June. Export controls forced efficiency: Elon Musk told The Economist that Kimi K3 approaches the strongest models on relatively little compute, and that with abundant compute Chinese labs "will be the leaders."
6. Two market fevers, but they are not buying the same thing
Hong Kong made OpenAI's 40x look prudent: Zhipu rose about 1,615% in five months to a $112 billion peak on 724 million yuan (~$107M) of 2025 revenue, over 1,000 times sales, then fell about 45% while MiniMax lost more than half its value; over HK$400 billion evaporated in weeks. America is pricing the model itself as an asset, a toll booth on intelligence. China is pricing the ecosystem the model irrigates.
7. Where the Chinese money actually lands: not in the chatbot
Doubao counted 345 million monthly users, more than Qwen and DeepSeek combined, with near-zero direct revenue, only adding paid tiers on 24 June. Meanwhile Tencent's Q1: advertising up 20% (credited to its AI recommendation engine), cloud-led services up 20%, $8.6 billion of quarterly net profit, while its chatbot stays free with ~$140 million of promotion behind it. Free is not the absence of a business model: the model attracts the usage; infrastructure, ads and cloud collect the money. Zhipu, for its part, raised API prices 83% into 400% demand growth (company-reported figures).
8. Both empires are quietly rationing their best product
In mid-June the US Commerce Department ordered Anthropic to suspend foreign access to its most advanced models, Fable 5 and Mythos 5, on national-security grounds; Fable returned on 1 July behind a new jailbreak-blocking classifier, while Mythos remains limited to about a hundred vetted US organizations. Two days after the order, Zhipu open-sourced GLM-5.2 worldwide and its stock closed up 33%. But on 7 July, Reuters revealed Beijing had convened Alibaba, ByteDance and Zhipu to discuss restricting foreign access to advanced Chinese models too, proposals for now, not policy. Today's openness installs the ecosystem and the dependence; tomorrow, Beijing can decide where openness stops.
9. The scoreboard that may decide it all: diffusion, not invention
Stanford's AI Index puts the best-US-versus-best-Chinese model gap at 2.7%, down from more than 30 points in 2023, against a ~23:1 US advantage in private AI investment (Chinese state funds excluded). Ipsos finds 85% of Chinese respondents see AI's benefits outweighing its drawbacks, versus 38% of Americans; Microsoft ranked the US 24th worldwide in AI adoption in January (21st by May). On 18 July, 142 protests against data centers ran across 42 US states. Jeffrey Ding's research argues general technologies reward the power that diffuses them, not the one that invents them, and China's AI-Plus doctrine targets 70% penetration of key sectors by 2027, 90% by 2030.
Push both logics to the end and they collide. The American machine needs intelligence to stay expensive, not out of greed but arithmetic: hundreds of billions of private capital must be serviced. The Chinese machine works, with patient public backing, to make intelligence nearly free and collect the value in hardware, cloud, advertising and robotics. And against that offensive, the classic trade weapons fail: you can stop a Chinese electric car at customs; you cannot tariff an open-weights file that copies to a thousand servers in seconds. Every open release drags down the ceiling US labs can charge.
Neither dream is safe. Goldman's $7.6 trillion of projected capex will one day need revenues to match, and ad clicks won't cover it. On the other side, free is getting expensive: Doubao started charging, ByteDance's capex now exceeds even its projected profits, and Zhipu raised prices 83%. Which leaves the variable neither Wall Street nor Beijing puts in its spreadsheet: whether a society actually accepts the technology. The winner will not simply be whoever builds the best intelligence, it will be whoever makes its price, financial, energetic and social, acceptable at the scale of a civilization.
Questions people ask
- Is OpenAI really going public at a $1 trillion valuation?
- OpenAI confidentially filed IPO paperwork with the SEC (confirmed in June 2026), with Goldman Sachs and Morgan Stanley leading and reports pointing to a listing targeting an $850 billion to $1 trillion valuation. At roughly 40 times its ~$25 billion annualized revenue, while still loss-making, the pricing is a bet that the model layer becomes the toll booth of intelligence.
- Why are Chinese AI models free or open-weights?
- Two rational reasons. US chip controls forced Chinese labs to squeeze more out of less compute, and their ultimate backer, the state, asks models to diffuse through the economy rather than return capital directly. Companies then monetize the usage the free model attracts: cloud, hardware, advertising and API scale. Analysts at MERICS compare it to China's solar and EV playbook: subsidize, diffuse, win on volume, build the ecosystem.
- Are Chinese AI models as good as American ones?
- Nearly. Stanford's AI Index measured the gap between the best US and best Chinese models at 2.7%, down from more than 30 points in 2023, and rankings place Moonshot's Kimi K3 within about three index points of the strongest US frontier models. The remaining US edge rests mostly on compute and chips, the material layer, not on an unbridgeable software lead.
- What happened with Anthropic's Fable and Mythos models in June?
- Following a US Commerce Department export-control directive citing national security, Anthropic suspended access to Fable 5 and Mythos 5 in mid-June; Fable 5 returned on 1 July with a new jailbreak-blocking safeguard, while Mythos 5 remains available only to a small set of vetted US organizations and agencies. It is the clearest sign yet that Washington treats frontier model weights as a strategic asset.
Written 16 August 2026 from an owner-supplied French video essay; every figure was re-verified against primary or first-tier sources on 16 August 2026 (SEC/IPO reporting by CNBC and Bloomberg; OpenAI's funding announcement; hyperscaler capex guidance; Goldman Sachs Research; Xinhua's full text of Xi's WAIC speech; Caixin, SCMP and Tencent's Q1 filing for the China figures; Stanford AI Index 2026; Ipsos AI Monitor 2026; Microsoft's Global AI Diffusion report; Reuters for the MOFCOM meeting and the 18 July protests; Anthropic's own statements on the Fable/Mythos suspension). Corrections to the source material: ByteDance's 2025 net profit was not "$50 billion", that was a projection, and actual profit fell more than 70%; Zhipu's June drawdown was about 45% (the steeper fall was MiniMax's); Musk's Kimi K3 remarks were made to The Economist, not "to economists" (a translation slip); Kimi K3's full weights shipped 27 July, ten days after its WAIC unveiling; GLM-5.2 went open-source on 13 June, not July; OpenAI's −122% Q1 margin is an adjusted, non-GAAP figure (GAAP 2025 was worse); "under $15 per user per year" is derived from 2025 revenue, the current run rate implies about $28. The video's attribution of Xi's "wisdom of humanity" formula to DeepSeek could not be confirmed and is omitted, as is an unverified Arab League cooperation center. Beijing's discussed export restrictions are proposals under consideration, not adopted policy. Disclosure: TaskNorth's knowledge base recommends Claude models, and Anthropic, whose export-control episode this article covers, is the maker of models used in building this site.
Trying to work out which AI tools fit your task? Describe the outcome and get a Blueprint: the tools, the prompt, and the steps, with pricing we verified ourselves.