China’s Powerful New AI Surprises Investors, Fueling Tech Rout
A surprise breakthrough from Chinese AI startup Moonshot rippled through global markets Friday, sending AI and semiconductor stocks sharply lower as investors drew parallels with last year’s “DeepSeek moment” and questioned whether the industry’s enormous spending spree is becoming harder to justify.
The catalyst was Moonshot’s new Kimi K3 model, which the company said rivals the strongest offerings from OpenAI and Anthropic PBC. The launch was quickly dubbed a new “Kimi moment” by investors, echoing the market shock unleashed by Chinese startup DeepSeek’s breakthrough last year.
“People are worried that if US companies start using Chinese models more and Anthropic less, then Anthropic will invest less. That means those US firms will lower the capex and in the end chip demand will be affected,” said Vey-Sern Ling, managing director at Union Bancaire Privee. Such worries have abounded since April when DeepSeek launched its V4 model, while the release of K3 just deepened such concerns, he added. (…)
To be sure, benchmark scores don’t always translate into commercial success, and Moonshot’s claims have yet to be fully validated by the broader AI community. Markets also weathered last year’s “DeepSeek moment” as US hyperscalers pressed ahead with record AI spending, easing fears that demand for cutting-edge chips would falter. (…)
“It is a confirmation that China is firmly in the AI race with the US, particularly when cost efficiency is taken into account,” he said, adding that the tech industry is beset by a host of concerns, including AI capex returns, stretched valuations and signs of broader AI fatigue that could account for Friday’s selloff. (…)
Moonshot’s release shows that China’s AI race is evolving beyond a competition on cost alone. The startup priced its latest model at roughly the same amount as Anthropic Sonnet, a premium compared to other Chinese models based on its strong capabilities. Moonshot also claimed it surpassed Z.AI’s most advanced offering on coding tasks. That’s stoking concerns that competition in China’s AI sector is intensifying even as the country’s sector leaders race onto global markets. (…)
“The blended price for K3 is about half of the Claude Opus and GPT-5.5, which represents meaningful improvement of economics of the tokens.” (…)
Moonshot said its latest model has 2.8 trillion parameters and a 1 million token context window, gauges of its capability. Artificial Analysis ranked Kimi K3 ahead of Anthropic’s Opus 4.8 on some frontier benchmarks, making it the first Chinese open-weight model to achieve that milestone. (…)
Another way to look at the AI rankings. Note the 17-place jump from Kimi-k2.6:
LLM models will keep leapfrogging one another, continually improving their effectiveness and costs, boosting compute demand as AI agents gradually become ubiquitous.
Google’s Gemini 3.5 Pro, the company’s most powerful flagship AI model, is months behind schedule due to efforts to improve its capabilities, particularly in coding, Bloomberg News has learned.
Ten current and former employees say the delay has frustrated Google engineers and researchers, who fear the company is losing ground to rivals Anthropic, OpenAI, and Meta, which have released models that outperform Gemini in coding tasks.
A late-June attempt to improve Gemini’s performance by updating its training data yielded disappointing results, and Alphabet shares fell as much as 3.2% on the news.(Bloomberg)
Xi used his first appearance at the World AI Conference in Shanghai on Friday to urge the world to adopt an inclusive approach, encouraging collaboration without rivalries. “AI development should not be a solo performance by a single country, but a symphony of international cooperation,” he said, adding that safety risks must be contained.
His presence at the gathering, attended by scores of tech and government leaders, conveys a potent signal of China’s ambitions to dominate a technological sphere with the potential to revolutionize industry and economies — an effort that’s shot to the top of the nation’s agenda. Chinese models are winning over companies worldwide, with their share of US firms’ AI usage nearing a record 60% on the popular marketplace OpenRouter. (…)
“We should take seriously the various types of inherent and secondary risks that AI may trigger,” Xi told the gathering on Friday, calling for more regulations, monitoring and warning systems to ensure that AI is always under human control.
He also warned against “overstretching” the concept of national security to curb access to the technology. But Chinese officials have recently held discussions with companies including Alibaba Group Holding Ltd. — developer of the popular Qwen models – on how to mitigate the security risks posed by their increasingly powerful models, people familiar with the matter said.
The talks are early, with no enforcement planned, but restricting foreign access to top models was among the options raised, the people said. Reuters previously reported that Beijing was weighing curbs on overseas access. Alibaba and the Commerce Ministry didn’t respond to requests for comment.
Global AI governance has emerged as a new battleground for the world’s leading powers. With the cybersecurity threat of cutting-edge AI looming large, Washington has in recent weeks pressured prominent American labs such as Anthropic PBC to curtail foreign access to advanced models.
China is looking to build its own AI ecosystem that offers its citizens and global customers a cheaper alternative to US technology. It also wants to secure its own AI supply chain that can guarantee access for its companies and government agencies. As part of this effort, Beijing has earmarked 2 trillion yuan over the next five years to creating a network of inter-connected data centers across the country, Bloomberg News reported last month. (…)
On Friday, China’s top planning agency released an action plan for AI cooperation and development, outlining eight areas for countries to work more closely together, including data sharing, computing power, the open-source ecosystem and standards setting.
But the venture pales in comparison to the $725 billion that US leaders such as Meta Platforms Inc. and Microsoft Corp. are setting aside for AI this year alone. Chinese data centers in general cost less than in the US because of cheaper labor, component and construction costs, and local government incentives. (…)
Western companies’ challenge is to achieve respectable returns on higher cost investments competing with lower cost, lower price Chinese models. The race is on models but also importantly on applications and versatility. Closed models and closed ecosystems face significant challenges.
BTW, Xi also said: “We should seize this rare historic opportunity to encourage open-source, openness, and collaboration.”
Bloomberg’s Catherine Thorbecke is in Shanghai:
(…) It’s a rare moment indeed. The remarks take a clear side in the crusade for the future of AI: whether its evolution will progress in the hands of a few elite labs or the people. American firms have bet on the former, which makes good business sense. China is urging the rest of the world to join in the latter. (…)
In recent months, cheap Chinese AI has garnered attention for undermining the business models of US titans. But the same commoditization pressures domestic firms as well. With so many companies now giving away their technology, it makes it hard to differentiate or find real revenue streams.
There’s a real danger that these companies go the way of electric vehicles: endlessly undercutting and competing with each other. The potential is especially apparent here in Shanghai, where over 1,000 firms are exhibiting. As more of these startups go public — with a high-profile DeepSeek IPO reportedly in the works — it’s something more difficult for investors to overlook.
But it’s a much bigger warning shot to American model makers like Anthropic PBC and OpenAI. Just-as-good Chinese alternatives are punching holes in their moats and challenging astronomical valuations ahead of their mega-IPOs. (…)
Xi recognizes that the real metric of the AI race is diffusion. He’s playing the long game. Can the same be said for Washington?
If history is being written in Shanghai this weekend, and China’s vision for the AI future plays out, the consequences will be more profound than market share. The openness could just as easily be pulled back once it’s no longer needed to undermine US dominance. And by then, Chinese models will have already become the default.
US AI protectionism will continue in this increasingly digital world.
- China Can Still Win the AI Race With Inferior Technology It’s about economics, power as well as engineering.
(…) A standardized intelligence task on DeepSeek’s V4 Flash model costs about two cents. The same task on Anthropic’s premium model runs over two dollars, according to benchmarking site Artificial Analysis which compared them on different kinds of standardized tasks. The difference racks up fast for companies using these models at scale, and it’s increasingly hard to defend using pricier models. San Francisco’s Lindy AI said it slashed its AI spend by 90% after switching.
DeepSeek’s investor list reveals battery giants, state funds and e-commerce players who will likely help the startup access energy, chips and distribution. It’s a familiar Chinese playbook that has helped take the country’s steel, solar and electric vehicle products global. And it’s all backed up by China’s ambitious infrastructure buildout.
None of this is a guarantee that China will come to dominate the cutting edge of AI. But DeepSeek’s price advantage, plus international alarm about the US suspending foreign access to Anthropic’s advanced Mythos and Fable models, indicate that success isn’t just contingent on being the best.
Add the power issues, supply and costs:
The crunch time for America’s biggest power grid is intensifying as PJM Interconnection’s latest capacity auction fell more than 6 gigawatts short of what will be needed by June 2028, even as supply costs tied a record, driven by surging AI data center demand that is outpacing new generation construction.
PJM plans an emergency two-step process launching in September — bilateral contracting followed by a special auction — aimed at making AI hubs cover the costs of securing their own energy needs. (Bloomberg)
And watch robotics, a huge future compute demand vector:
Four of Japan’s leading industrial and manufacturing giants — Fujitsu, Fanuc, Kawasaki Heavy Industries, and Yaskawa Electric — are joining forces with Nvidia to accelerate AI-driven robotics development across factories, retail, logistics, and health care, in a direct bid to counter surging Chinese competition in humanoid and AI-enhanced robots.
The collaboration, born over a meal of tonkatsu, will see the four Japanese firms join Nvidia’s Cosmos Coalition, an alliance aimed at speeding adoption of Nvidia’s physical AI platforms.
The Model Is Not the Moat. The Flywheel Is: How to Read the LLM Race After Kimi K3
(…) what Kimi K3 really proves is not that model companies have no moat, but that the model itself is not the moat.
Raw model capability is becoming commoditized very quickly. A model can top the leaderboard today and be matched by competitors a few months later. Model capability still matters. It determines whether a company can sit at the table. But it is becoming harder for model capability alone to form a durable moat.
The real long-term value lies in the flywheel formed by the model, workflow, feedback data, customer relationships, and reinvested revenue.
There is also one detail that is especially easy to miss: an open-source model being widely used does not mean the model company captures the corresponding revenue. This is particularly important for Chinese AI model companies, because it determines whether model influence can actually become commercial value.
So the right question is not simply how strong Kimi K3 is. The more important questions are: how long can model capability leadership last? After open-source diffusion, who actually captures the revenue? Where do customer relationships and task feedback accumulate? And who can turn one model release into the starting point for the next iteration and the next stage of commercialization?
In many cases, the difference between the top model and the rest of the leading pack is no longer a generational gap. It is often a difference in benchmark design, task preference, and specific use case.
A single model lead is therefore more like an asset that depreciates quickly.
A model may look impressive at launch, but several months later, competitors can often catch up on a large part of its capability through architectural imitation, data improvements, post-training, distillation, and engineering replication. This is especially true in general Q&A, summarization, translation, and simple code generation, where the leading models are becoming increasingly similar.
This does not mean model capability is unimportant. Quite the opposite. Model capability is the ticket to the game. But a ticket is not a moat. What matters is whether a company can keep getting the next ticket, faster, cheaper, and more reliably than others.
US Consumers Singing, “Ain’t No Stoppin’ Us Now!
For a long time now, their doubters warned that a low savings rate, flatlining real disposable income, rising consumer debt, and mounting affordability challenges would force households to retrench. Instead, they continue to do what they do best, namely shop! A well-balanced labor market and the wealthiest retiring generation ever continue to power consumer spending.
Retail sales (including food services) rose 0.2% m/m in June after a 1.0% gain in May. The slowdown largely reflected a 5.3% drop in gasoline station sales as pump prices fell by roughly 50 cents per gallon.
Excluding gasoline but including food services, sales increased a solid 0.7%, with gains across the board.
Nonstore retail sales jumped 1.9%, the largest monthly increase in a year, likely boosted by Amazon’s Prime Day. Encouragingly, control group sales, a key input into GDP goods spending, rose by 0.5%. For Q2 as a whole, control-group sales advanced at a remarkable 9.2% annualized rate.
Several sales categories rose to new record highs in June, including discretionary areas such as motor vehicles, electronic shopping, and general merchandise. Food services, the only services category in the report, also climbed to a record high. (…)
Spending might cool in July as the World Cup ends.
The labor market remains in very good shape. Initial jobless claims fell to a 10-week low of 208,000 in the week ended July 10, while continuing claims eased to 1.81 million. Together, the data suggest layoffs remain very low. (…)
The Atlanta Fed’s GDPNow tracking model currently shows Q2 real GDP growth at 1.7% (saar). Much of the weakness reflects a surge in AI-related imports, with net exports expected to subtract 96 basis points from growth.
By contrast, consumer spending is projected to rise a solid 2.5% (saar), up from 0.5% in Q1, while final sales to private domestic purchasers is expected to increase 3.4%, up from 1.7%.
Ed is right celebrating strong spending. But let’s not totally dismiss the effects of inflation:
Also, let’s keep in mind that “a low savings rate, flatlining real disposable income, rising consumer debt, and mounting affordability challenges” are realities for most Americans.
- Gas Prices Will Stay Higher for Longer, Even if Oil Falls Refining capacity is strained, and fuel markets don’t have much of a buffer
Despite the resumption of U.S.-Iran hostilities, U.S. benchmark crude prices are hovering around $80 a barrel. That is about 18% above where they were before the Iran war began. By contrast, gasoline prices remain 32% higher at $3.94 a gallon, according to energy data firm OPIS.
So-called gasoline crack spreads, which measure the difference between gasoline and crude-oil prices, are averaging 90 cents a gallon so far this month. That is the highest level in four years, according to data from Novi Labs.
Why are fuel prices so much higher? One reason: fuel markets didn’t have the buffers that crude oil had. This means that even if there’s another ceasefire that allows oil to again flow, gasoline and diesel prices are likely to stay higher for longer.
The crude-oil market’s surprise shock absorber was China, which dramatically cut its imports. It imported just 5.7 million barrels a day in June, according to the International Energy Agency’s July oil market report, down from around 11 million barrels a day before the Iran war.
China hasn’t provided the same slack in the refined-products market. It is typically a net exporter of gasoline and diesel, but has reduced exports significantly since the war started, according to Rob Smith, global head of fuel retail at S&P Global Energy. Chinese refiners’ throughputs dropped to a six-year low in May, consistent with its decline in crude imports, according to the IEA.
Crude-oil supply was also boosted by hefty releases from strategic petroleum reserves around the world. Members of the International Energy Agency, a coalition of oil-consuming countries, released 2.4 million barrels a day in May and 1.5 million barrels a day in June. The vast majority of these releases were of crude oil rather than refined fuel, according to data from the IEA.
All of this crude oil needs to be processed to become gasoline and diesel. But refining capacity in two major fuel-export regions—the Middle East and Russia—has been severely curtailed by conflict. Global refineries processed 5.1 million fewer barrels a day in the second quarter compared with the same period in 2025, according to the IEA. (…)
Russia’s diesel export ban could have knock-on effects. If diesel prices keep rising globally, U.S. refiners may be motivated to boost diesel production and cut gasoline production unless the price of the latter rises, according to energy economist Philip Verleger.
Global gasoline inventories were 3% below the trailing-five-year average as of June and the gap is expected to widen to 4% this month, according to S&P Global. In the U.S., gasoline inventory is about 8% lower than the five-year average for this time of year.
Despite tight fuel inventories, global gasoline demand has remained resilient because of government policies to protect consumers from high prices, according to the IEA. (…)
Relief won’t be on the way for motorists for some time, even if the Strait of Hormuz opens up. One reason: Shut-in crude oil production can be resumed quickly, but turning refining capacity back on takes longer, according to S&P Global’s Smith. (…)
The problem there is that the U.S. doesn’t have a big strategic stockpile of gasoline or diesel. Unfortunately for drivers—and the U.S. economy—there is no easy way out of high gas prices.
The other problem is actual product shortages, not only for transportation but also manufacturing. Hormuz is still closed.
US Corporate Insiders Are Selling Stocks at a Near Record Pace
US executives are selling shares at the second-fastest pace in more than 20 years, a classic red flag to some investors because it suggests people with the most corporate knowledge are wary about markets.
Corporate insiders sold $77.6 billion of stock during the first half of 2026, a 20% increase from a year ago, according to EPFR Global Market Intelligence. The only time the selling spree was more intense was back in 2021, when markets were flush with pandemic-driven stimulus cash.
“Insider activity suggests executives are not especially eager to increase their exposure at current valuations,” wrote analysts including Winston Chua at EPFR.
In contrast, stock buying by corporate insiders has been subdued. They purchased just $6.9 billion worth of shares in the first half. That’s only modestly above the seven-year low of $6.7 billion recorded a year earlier. (…)
Trump to Cut US Stays for China Journalists, Risking Retaliation
US visas for foreign journalists will be trimmed to 240 days and mainland Chinese journalists will get just 90, according to new rules being published on Friday by the Department of Homeland Security. The 90-day limit would revive restrictions that President Donald Trump proposed at the end of his first term and Joe Biden later scrapped.
The US agency said the change would allow officials to more closely monitor visa holders and ensure their activities are consistent with the purposes of their entry. Extensions would be available, it said.
Chinese Foreign Ministry spokesman Lin Jian said his nation “firmly rejects the US’s discriminatory move targeting certain countries,” and urged the US to immediately halt the changes regarding journalist visas.
“China reserves the right to take reciprocal countermeasures,” he added at the regular press briefing in Beijing on Friday.
The US also put a new four-year limit on student visa holders, ending a decades-old policy that had allowed international students and exchange visitors to remain in the country for as long as their studies last. (…)
- Top Treasury Tax Official Ousted After Clashes With White House Over IRS Audits Kenneth Kies warned that political officials risked breaching a law that limits who can influence tax enforcement
The Treasury Department’s top tax policy official was forced out of his job after he warned that the White House was at risk of violating a federal law prohibiting senior officials’ involvement in IRS audits, according to people familiar with the matter. (…)
Kies at times clashed behind the scenes with White House officials, the people said. That included a recent meeting in which he contended that a potential White House request would violate Section 7217 of the Internal Revenue Code, one of the people said. That law prohibits the president, vice president, White House staff and certain agency heads from directly or indirectly requesting that the IRS conduct or terminate an audit or investigation of any particular taxpayer.
Violations are punishable with up to five years in prison and up to $5,000 in fines, and IRS officials have long seen the prohibition as an important shield against the kind of political interference that President Richard Nixon tried to impose on the tax agency.
IRS employees who receive requests they consider improper must report them to the agency’s inspector general or risk the same punishments. The 1998 law is untested in court, and enforcement in the near term would require action by the Trump administration.
It couldn’t be determined what White House requests Kies objected to and whether the administration plans to follow through with them after Kies departs. The taxpayers at issue could include individuals, corporations or nonprofit groups. Kies represented President Trump in private practice before joining the administration last year and has been recused from matters involving the president.
The president and some of his aides had become increasingly frustrated with Kies, some of the people said. Kies was at odds with White House officials over several issues during his tenure, including taxation of income from fantasy sports. He cited IRS rules about audits when asked questions about policies, even when officials weren’t asking about particular companies or audits, one of the people said.
Kies declined to comment. After The Wall Street Journal sought comment from the White House and the Treasury Department, several administration officials and Trump allies reached out to privately criticize Kies, arguing he was difficult to work with and didn’t do enough to advance the president’s agenda. (…)
Kies has been at the center of Republican tax-policy circles in Washington for nearly a half-century. (…)
In his Treasury job, Kies described himself as being pro-taxpayer, and he pushed for lighter regulations in many cases. (…)