Last updated on July 19th, 2026 at 01:34 am

Did you feel the Earth move? It depends on how close you are to the epicentre and the epicentre right now is not where it used to be. It’s not where it was even 10 days ago. A seismic shift is occurring that is going unnoticed and unreported on by most of Western media, largely now controlled by systems that are still using outdated framing to characterize where the most significant technology development and innovation is now happening. What you just felt was the anxiety level in the Valley, and in power centres across America, rise to heretofore unforeseen levels and a definitive shift in power dynamics. We are witnessing the end of an empire, sped along by its own worst impulses, in real time, and there are very good reasons to be afraid of the consequences.

This is the second such shift in 18 months. The first was the arrival of DeepSeek on an unsuspecting industry in January 2025. When DeepSeek released R1 that month, the reaction told you everything about what American markets actually believed. Within a week Nvidia shed roughly $600 billion in market value in a single trading day, the largest one-day loss any company had ever recorded, and the app climbed to the top of the US App Store. A lab operating under export controls, working with restricted chips and a training budget that rounded to pocket change by Valley standards, had matched frontier reasoning performance and then given the weights away under an MIT license. The panic was instructive. American AI valuations rested on the assumption that capability required scarcity, that the moat was capital and compute. R1 demonstrated that the moat was a story, and the market repriced the story in an afternoon.

DeepSeek was two models arriving in quick succession. V3, released in late December 2024, was the workhorse: a mixture-of-experts model with 671 billion total parameters, only 37 billion active per token, which is what made it cheap to train and cheap to run. It matched GPT-4o and Claude 3.5 Sonnet on most standard benchmarks. R1, the January release, was the reasoning model built on top of it, trained largely through reinforcement learning to produce long chains of thought before answering. It scored at or near OpenAI’s o1 on math and coding benchmarks, the tests the industry had agreed were the frontier. The comparison that mattered most was price. OpenAI charged sixty dollars per million output tokens for o1; DeepSeek charged about two, a gap of roughly 27 times for comparable performance. And where o1 hid its reasoning traces, R1 showed its full chain of thought, then published the training method in a paper detailed enough that labs everywhere immediately began replicating it. It was frontier reasoning, visible, documented, MIT licensed, at commodity prices.

Kimi K3 appeared on the leaderboards on July 16 and rearranged them. Independent evaluation places the new model from Beijing’s Moonshot AI fourth among 189 systems, behind only Claude Fable 5 and two configurations of GPT-5.6 Sol, ahead of Opus 4.8 and everything else American labs sell. It runs 2.8 trillion parameters and a million-token context window, it costs 15 dollars per million output tokens against Fable’s 50, and when its weights ship on July 27 it becomes the largest open model ever released. A lab most Western executives could not name a year ago now sits inside the top tier of the frontier, and it is giving the machine away.

The release landed the day before the World Artificial Intelligence Conference opened in Shanghai, where Xi Jinping delivered the first keynote a Chinese president has ever given at the event, a notable milestone in itself. He opened with Dartmouth in 1956, claiming the entire seventy-year arc of the field, and committed China to seizing what he called a rare historic opportunity through open source, openness, collaboration and sharing, with new capacity-building initiatives for the developing world. The timing between the model and the speech was choreography, and the message of the choreography was singular: the machine is the policy, the benefit is collective.

Minutes before Xi spoke, Donald Trump addressed his own nation in primetime. Predictably, he spent his time on elections and grievances. He claimed China had accessed 220 million voter files in the largest compromise of election data in history, released declassified assessments describing vulnerabilities election officials have spent years addressing, and pressed Congress on an elections bill his own party has declined to pass, with the midterms months away. His last several primetime addresses have covered a war, an economic blame assignment, and now a six-year-old grievance. One president used his stage to narrate the operating system of the next economy. The other used his to relitigate the last election and pre-litigate the coming one.

The Crash Beijing Chose

The structural difference underneath that split screen has been visible for five years, and it is China’s property crash. Beijing initiated the de-risking of its own real estate sector deliberately, accepting what officials called a stable contraction. Housing investment fell from 12.3 percent of GDP in 2020 to 6.1 percent in 2025. Developers defaulted at scale. Evergrande was liquidated. The speculators who inflated the bubble ate their losses, and the state declined to rescue them. A bubble the size of a civilization’s savings was absorbed slowly, deliberately, and without a Lehman moment.

The pain was real, and it is ongoing. Growth came in at 4.3 percent last quarter, the slowest in three years, deflation is in its tenth consecutive quarter, and youth unemployment sits near 17 percent. Chinese families, whose wealth concentrates in property carried a share of the burden. But the losses were distributed across the whole hierarchy, and they fell heaviest on capital. An entire system absorbed this bubble bursting, intentionally, with distributed impact.

Last month I wrote about the tribute system, the organizing logic of Chinese statecraft across centuries: hierarchical and reciprocal at once, the center conferring stability and benefit downward, the periphery contributing upward, the superior position carrying obligation. The property crash is that logic applied to a domestic emergency. The center absorbed the reckoning rather than pushing it down the chain. Compare this to 2008, when American losses concentrated catastrophically at the bottom, roughly ten million foreclosures moving through ordinary households, while the institutions that engineered the bubble were recapitalized within months and returned to record profits. Washington inverted the obligation. The periphery paid upward to rescue the center.

The redirection is the part the malaise coverage misses. While property investment halved as a share of the economy, exports grew 7.7 percent annually over five years as China climbed the value chain into autos, shipbuilding, and robotics, and state capital moved into R&D, energy, and AI. Moonshot raised two billion dollars in May at a valuation above twenty billion, in the middle of the deflation the West reads as decline. The capital that once chased apartment towers now funds frontier models released to the world for free.

What Xi Actually Said

Xi’s speech was an economic doctrine delivered as diplomacy. Open source as national commitment. AI as an international public good. Capacity-building for the Global South, backed by programs already running. The address named, in public and at head-of-state level, the strategy I have been tracking across this ecosystem all year: DeepSeek funded by a hedge fund and freed from monetization pressure, open weights propagating between labs, a distillation of one firm’s model improving another’s, Alibaba running Qwen as a substrate strategy for the developing world, gains compounding across the hierarchy instead of being hoarded as proprietary weapons. High-Flyer’s (the Chinese hedge fund) profits come from trading. For the lab’s first three years it funded DeepSeek off its own balance sheet: no venture investors, no public shareholders, no valuation to defend. When DeepSeek took its first outside money last month, the terms preserved the design: commercial investors were channeled into a limited partnership with no voting rights and a five-year lockup, and the only party granted direct equity was the state’s national AI fund.

American labs run the inverse structure, where the capital arrives as equity bought at record valuations and the investors’ return exists only if the models stay scarce and expensive, so the lab’s core asset is the one thing it can never give away.

In March I wrote that China’s 141-page Five-Year Plan, aligning ministries, provinces, industrial capital, and energy systems around 90 percent AI integration by 2030, was what strategic coherence looks like. This week the coherence went on stage in Shanghai, and it brought a 2.8 trillion parameter gift.

The American stack is already running on the gift. From Fortune, Cursor used Kimi models to help build its coding agent. DoorDash delegates work to K2.6. Thinking Machines used K2.5 to generate training data for its own model. Epoch AI puts the average lag of Chinese models behind the American frontier at an average of seven months, since the start of 2023. The dependency is accumulating order by order, workload by workload, inside the very economy whose leadership describes China purely as a threat vector.

Open Weights Meet an Overbought Market

K3 quickly stops being a benchmarks story and becomes a capital markets story. The American AI economy is priced for a world where frontier capability is scarce, proprietary, and expensive. OpenAI projects roughly fourteen billion dollars in losses this year, targets six hundred billion dollars in compute spending by 2030, and does not expect positive cash flow before the end of the decade. Every major American lab is losing money on purpose, betting on being the last one standing, and the bet is funded by the largest private valuations in market history plus a newly crowded public pipeline of AI infrastructure and aerospace listings, the two hottest categories of 2026 debuts.

That pipeline has produced its warning shot. SpaceX went public on June 12 in the largest IPO in stock market history, raising 86 billion dollars at 135 dollars a share, touched 225 within four days, and this week closed below its IPO price for the first time, roughly 1.2 trillion dollars off its peak valuation. Bloomberg reports the fall has dragged the weighted average return of this year’s US listings to 6 percent against the S&P’s 11, cooling exactly the sectors the AI boom needs to stay hot. The most hyped listing in history, spanning rockets, satellites, and AI, could not hold its price for five weeks.

Putting a free frontier model into that market is something few have grappled with. Every valuation in the American AI complex embeds an assumption about pricing power: that access to frontier capability will remain scarce enough to command a premium large enough to retire the burn. K3 attacks the assumption directly. Capability within seven months of the frontier, at a third of the price today and at zero license cost in ten days, is a solvent poured on scarcity premiums. It does not need to beat Fable to do damage. It needs to be good enough, cheap enough, and open enough that enterprise buyers renegotiate, and it’s all three.

The United States has never had to contend with this dynamic. Japan competed on manufacturing cost. The Soviet Union never competed commercially at all. No strategic rival has ever attacked American capital markets by giving away the crown jewels, deflating the scarcity that underwrites trillions in valuation as an act of foreign policy. There is no tariff against a free download, no export control on a gift moving in the other direction, and no precedent in the American playbook for a competitor whose winning move is generosity. The axis shifted this week. The market that noticed first was not the leaderboard. It was the tape.

Critics may say it’s just one model, but it isn’t. Critics will say but China, but repression and to that thinking people will say: have you seen what’s been happening in the United States lately? Are you aware of the level of entitlement and corruption? And there’s good reason for concern. Dying empires are messy and far from benevolent. The current US administration has demonstrated loud and clear it will do everything necessary to silence its critics and hang onto power, and nothing in the infrastructure so far has indicated anyone is willing to stop it.

The US economy hangs in the balance of the AI arms race, and its dominance just visibly ended. If anything should hit you right now it’s the pointlessness of the Mythos psychodrama, the precariousness of Elon Musk’s multiple positions and the vulnerability of the Valley overall. People have long said that if an AI bubble pops, there will be another “nuclear winter”, but that’s not what’s happening. The epicentre has shifted. Competition is no longer the anchor. The sovereignty consequences are enormous, as is the commercial impact.

Everyone else is playing catch-up now against a technology sector that is fed by domestic policy and domestic use insulating it from what every other frontier model has to confront: the cost of compute, the cost of competition, the cost of sovereignty. We will feel the consequences play out over the next few months.