US Consumer Borrowing Unexpectedly Drops by Most Since 2024

Total credit outstanding decreased by about $182 million in May after sizable increases in the prior two months, Federal Reserve data showed Wednesday. The median estimate in a Bloomberg survey of economists called for a $17.5 billion advance.

The pullback was driven by a $5.3 billion decline in credit-card and other revolving debt outstanding, which also marked the biggest decline since 2024. Non-revolving credit, such as loans for vehicle purchases and school tuition, rose $5.1 billion in May. The report doesn’t include mortgages. (…)

The data suggests Americans are beginning to pay down some debt after the biggest back-to-back increase in consumer borrowing in more than three years. (…)

China Weighs Limits on the AI Models American Companies Love Beijing talks to top companies about protecting their technology after it wins fans in Silicon Valley

Across Silicon Valley, models made by Chinese companies such as DeepSeek and Moonshot AI have become core to daily work at companies large and small, offering a less costly alternative and supplement to the products of OpenAI and Anthropic.

Chinese officials have recently held discussions with domestic AI labs that produce the country’s most powerful models about how to safeguard their valuable proprietary technology, according to people familiar with the matter. Beijing is concerned that sharing some of this technology could help adversaries or other malicious actors and ultimately get weaponized against China, the people said.

Beijing’s recent moves show it is revising its AI priorities in the race with the U.S.

Until recently, Beijing wanted to encourage the rapid spread of Chinese AI models globally, believing this represented a form of Chinese soft power. Many leading Chinese models are open-source, meaning anyone around the globe can download them without charge and generally use them with few restrictions.

The new thinking: Not all technology should be open to everyone.

Sharing of technology, including through open-source releases and research papers, may expose technical secrets. Some Chinese innovations, such as techniques that allow AI models to use computing power more efficiently, have been adopted by Western AI labs. (…)

Andy Fang, co-founder of DoorDash, said in an X post Tuesday that his company routed the most complex tasks to Anthropic’s cutting-edge model Fable—a public version of Mythos—while delegating lower-level workloads to Kimi K2.6 from China’s Moonshot AI. DoorDash separately said the approach delivered better performance at a lower cost than using two Anthropic models.

Staff at legal-services platform Harvey typically turn to top models from OpenAI, Anthropic and Google to complete high-stakes tasks, said Niko Grupen, head of applied research at Harvey. For simpler work, lower-cost alternatives from Chinese companies such as DeepSeek and Zhipu have become the go-to models.

Startup Vercel said DeepSeek’s share of AI usage rose to 23% in June from 1% in April on the Vercel platform, while DeepSeek’s share of AI spending stayed in the low single digits. (…)

Officials are considering requiring labs to defer public releases and restrict access by certain tiers of users, such as foreign entities, if a review determines their products contain sensitive technology, the people said.

China is also looking at tighter restrictions on exporting some AI technologies and on Chinese AI companies’ accepting foreign investment, some of the people said. They said officials have consulted recently with Chinese companies and researchers to identify areas where China has developed a technology edge.

Such discussions are still in the early stages, the people said. Reuters earlier reported some exchanges between Chinese officials and companies.

Any moves by China to tighten its grip on domestic models could risk alienating foreign users and slow global adoption, industry participants said.

Companies need to implement AI agents to be/stay competitive and Chinese models provide the lower cost tools to make the AI investments financially acceptable.

Without Chinese models to process the more mundane AI tasks, many companies will curb their AI enthusiasm, slow down innovations, at the risk of diminished competitiveness (more sophisticated or less costly products or processes) against others with larger budgets or access to lower cost tools.

This is the bigger risk to an otherwise exploding compute demand.

Without the ability to keep AI development costs reasonable, demand will decline and Western AI builders will need to accept lower margins and reduce their prices.

The US-China conflict risks jeopardizing AI growth.

Who has the best hand? The West with its better frontier models or China with its lower cost models?

According to this chart, US frontier models are now only 3% better than Chinese models:

But in the AI world, 3% is still verry meaningful.

Serious AI development must use the best frontier LLMs whose cost per million tokens (below) are multiples above Chinese costs (note the compressed scale). But without the lower costs tools, many will find AI investing too onerous.

In this cost/benefit race, the American/Chinese combo allowed for fast development. Without it, the whole AI pyramid is in danger.

And so is the US economy:

 

(KKR)

BCA Research illustrates how non AI-related demand is weak in the USA:

Meanwhile,

Fed Officials Flagged Risks That Would Warrant Higher Rates More officials pointed to the AI build-out as a source of persistent inflationary pressures

Federal Reserve officials broadly agreed at their meeting last month that they would need to raise interest rates if inflation stays elevated this year. They also agreed that they could stay on hold if price pressures fade soon, according to minutes of that meeting released Wednesday.

Which of those paths they take depends on something they haven’t resolved: whether the forces pushing up prices will last.

The minutes showed how they are increasingly focused on a source of inflation that barely figured in their debates a few months ago: the boom in artificial-intelligence investment. It was one of the forces, along with the war in the Middle East and tariffs, that could keep prices elevated and tip the Fed toward a rate increase, according to the written account. (…)

Still, investors read last month’s meeting, the first under Chairman Kevin Warsh, as a step toward higher rates because of interest-rate projections showing a larger contingent anticipating hikes. Nine of 18 meeting participants penciled in at least one rate increase by December; none had done so in March. Only one anticipated lower rates, down from 12 in March.

Even the most hawkish officials weren’t pushing to act. A few participants saw a case for raising rates at the June meeting, according to the minutes, but supported holding—a sign that the split visible in the projections was less about what to do now than about how the outlook might unfold.

Warsh didn’t submit a projection, but his repeated vow to deliver price stability—with no accompanying signal of patience—reinforced the impression of a committee leaning toward tightening. (…)

“Several participants commented that price pressures had become more broad based, with a large share of goods and services…experiencing substantial increases,” the minutes said. (…)

Several have noted that a year ago, the Fed could treat tariff-driven price increases as one-off and let them pass through without a policy response because the labor market was soft enough to justify the patience. Now, with hiring steadier and fresh cost pressures arriving from energy and AI at once, that same instinct to wait carries more risk that above-target inflation gets entrenched. (…)

The most important statement:

“Most participants pointed to scenarios in which, in the context of stable labor market conditions, inflation would remain elevated due to strong AI-related demand, the conflict in the Middle East, or the effects of tariffs. In such scenarios, almost all of these participants indicated that some policy firming would likely be warranted.”

Also meanwhile,

Chinese Officials Gauge US Truce, Elections in Low-Key Meetings

A pair of former Chinese diplomats visited the US last month to better understand the Trump administration’s stance on a fragile trade truce, according to people familiar with the matter, as China seeks to gauge how upcoming US elections might affect ties.

Former Ambassador to the US Cui Tiankai and ex-deputy UN representative Geng Shuang were among former and current Chinese officials who turned to American experts — including those previously with the Trump administration — in recent months to discern the White House’s intentions toward its adversary.

Their visit underscores how much uncertainty remains around the truce the two sides signed late last year and China’s concern that President Donald Trump may renege or refuse to deliver on some of its key components. (…)

In response to questions from Bloomberg News, the Chinese Foreign Ministry said more exchanges between China and the US are in line with the consensus of Xi and Trump.

“We hope that the US will work in the same direction and work with China to promote exchanges and cooperation between the two countries across all sectors and enrich the content of a constructive China-US relationship of strategic stability,” a spokesperson said. (…)

“Chinese officials are confident the truce can last, but confidence is not the same as certainty,” Da Wei, director of the Center for International Security and Strategy at Tsinghua University, said at a forum in Beijing. Da was recently in New York, where he said the mood among American scholars was “more pessimistic” than in China.

“Many people are asking how long this stability or truce can last,” he said. “My question is: Is this a truce or stability that is based on the goodwill of the two leaders or are there structural factors that shape this stability?” (…)

China’s Reflation Shows Signs of Peaking as War Shock Fades

Consumer inflation and the core gauge of prices both slowed more than expected from a year earlier, according to data released by the National Bureau of Statistics on Thursday. And while the effect of a low base from 2025 kept producer inflation on an upswing, factory prices declined 0.3% from May on a month-on-month basis, their first drop since July 2025.

Note: Apologies to subscribers to the free daily email, yesterday’s Daily Edge could not be sent out.

‘It’s a big wake-up call’: What market observers are recommending after Trump says Iran MOU ‘is over’

(…) “It’s a big wake-up call for the markets because the expectation was that following the MOU, we were likely ​to start to see the flow of oil coming back into the markets. And we saw inflation ‌expectations being dialed down.

“The way we’re looking at it now is what has changed materially is the (Iranian) oil waiver is gone. It’s removed a very key incentive for Iranian compliance.”

“Trump’s comments add that further layer of additional risk premium into the markets. But the reality is with Trump, you always have TACO (‘Trump always chickens out’) trade at play.”

“He was fast approaching the midterm election. The fact that he wanted to do this memorandum of understanding with Iran implied that he wanted to improve his ratings ahead ‌of the winter elections, and ​that is going to be a critical ‌factor for him to keep in mind.” (…)

“And do we take Trump literally ⁠or seriously? He says that the peace deal is over, but that U.S. negotiators can continue doing their work. We also know that ⁠Trump can turn on a dime. He could have an about-face today, tomorrow, next week, or perhaps later. I don’t see him waging war with Iran into the elections.”

US Consumers’ Inflation Expectations Rose in June, NY Fed Survey Says

Americans’ expectations for inflation over the near and medium term rose in June amid strong increases anticipated for medical care costs and rent, according to a Federal Reserve Bank of New York survey released Tuesday.

Consumers said they see inflation at 3.7% over the next year, up from 3.5% in May. Expectations for inflation in three years increased to 3.3%, the highest since June 2022, up from 3.1%. Estimates for inflation in five years remained steady at 3%. (…)

Importantly, Americans’ higher inflation expectations are in spite of consumers expecting gasoline prices to be just 1.5% higher a year from now.

The minutes of the FOMC April 28-29 meeting said: “While both market- and survey-based measures of inflation expectations indicated upside risks to the near-term inflation outlook, measures of medium- and longer-term inflation expectations remained well anchored.”

The anchor is slipping…

And Kevin Warsh said “this committee will deliver”.

Consumers’ inflation sentiment is no longer simply a reflection of gas prices. Same decoupling with the bond market as Apollo shows:

Trump’s $40 Billion Travel Slump

Since Donald Trump’s return to the White House last year, the US has fallen into a tourism trough. His on-again, off-again tariffs, toughened borders and visa policies, agents dispatched to American cities and deployments of troops abroad have made the country a far less inviting destination for global travelers.

As would-be overseas visitors went elsewhere or simply stayed home, the US tourism industry lost out on as much as $16.6 billion in 2025 and is headed for an additional $21 billion deficit this year — what it might have earned if it had maintained its pre-Trump market share — researcher Tourism Economics estimates.

“Travelers have choices,” says Aran Ryan, a director at the firm.. “Trump administration policies and pronouncements are the primary contributor in the US decline that we’re experiencing.”

The US should be experiencing a tourism tsunami: The World Cup is expected to draw more than 1.2 million foreign fans to 11 American cities this summer, and the appetite for travel remains voracious, with global tourism projected to grow 5.8% in 2026 despite simmering geopolitical tensions and sky-high airfares.

Although 71 million foreign visitors are expected to journey to the US this year — an increase of about 3% — a full recovery to the pre-pandemic record of 80 million, reached in 2018, isn’t expected until 2029, the year Trump leaves office, according to Tourism Economics’ projections.

Riding a post-pandemic wave of so-called revenge travel, worldwide tourism rebounded to its pre-Covid-19 level in 2024. The US has yet to reach that milestone, though international arrivals jumped a respectable 9% that year. But in 2025, as Trump returned to office, visits to the US dropped 5.5% even as the global market expanded 4.7% — a difference of more than 10 percentage points.

There was a similar disconnect in the president’s first term: Total foreign arrivals to the US edged up 1.3% annually on average from 2017 to 2019, versus an increase of 5.7% per year in the global market before the pandemic shut down travel in 2020. That stands in stark contrast to the Obama administration, when US growth averaged 3.6% per year, roughly in line with the rest of the world, and under Joe Biden, when US expansion was about 3 percentage points higher than the global level.

Since Trump’s return to the White House, there have been fewer visitors from western Europe and China — traditionally the highest-spending holiday makers. The UK, the No. 3 source market (after Canada and Mexico), is down 2.2% this year through May after rising just 0.5% in 2025, according to the US International Trade Administration.

German visits have plummeted 13% year-to-date, while Chinese arrivals dropped 4% in 2025 and an additional 1.5% this year. Those travelers are being supplanted by people from less wealthy countries. Ecuadorians are up 21%, Hungarians jumped 18%, and Colombians climbed 16% so far in 2026.

Foreign tourists spend an average of $4,000 per trip to the US, and as more of them stayed away, international visitor spending in the US fell 4.6% last year, according to the World Travel & Tourism Council. The US was one of just a handful of developed countries to post declines in international visitor spending in 2025, according to Tourism Economics, joining the likes of Haiti, Iran, Nigeria and Pakistan.

Proposed changes to the visa waiver system known as ESTA, available to citizens of more than three dozen developed countries, would require visitors to hand over extensive information regarding social media accounts, family contacts and email addresses from the past decade. The new rules would deter some 4.7 million people from the affected countries coming to the US — almost a quarter below the typical level — blowing a $15.7 billion hole in foreign-visitor spending, the World Travel & Tourism Council estimates.

 

For hoteliers from Miami to Atlanta, who spent millions of dollars upgrading properties in anticipation of a global stampede, the World Cup math isn’t adding up. Almost four-fifths of owners and operators say bookings in US host cities during the tournament are significantly below expectations, according to an April survey by the American Hotel & Lodging Association. And some said reservations were even trailing typical summer demand. (…)

The Global Business Travel Association, which represents corporate bookers and travel management companies, says more than half its members plan to decrease either near-term or long-term business travel to the US. Business visas issued to western Europeans fell 2% in the 16 months through April. And regions that posted growth last year — Asia, the Middle East and Oceania — have flipped to losses this year, according to the US International Trade Administration.

China Auto Sales Drop Even as Exports Help Cushion Slump

Retail sales of passenger vehicles fell 23.2% from the year before to 1.6 million units, China’s Passenger Car Association said on Wednesday. New-energy vehicle sales, including plug-in hybrids and pure-electric vehicles, slipped 9% year-over-year, though the segment still accounted for over 60% of the market for a third straight month.

Manufacturers leaned heavily on exports to offset the slowdown. Exports of new-energy vehicles surged over 150% to 499,000 units, signaling that foreign markets are absorbing excess production. Despite a one-fifth drop in domestic sales, BYD Co.’s total shipments rose 5.5%, driven by exports that almost doubled from a year ago.

Amid the broader domestic deceleration, Tesla Inc. posted one of the month’s strongest performances. The Austin, Texas-based automaker delivered 89,091 vehicles from its Shanghai factory in June, including 52,920 sold locally. That extended Tesla’s year-on-year growth streak in China to eight months and marked its best monthly total of the year.

The latest figures highlight a widening pressure on weaker brands. Zeekr crossed the halfway mark toward its annual sales target, while Xpeng Inc. fell significantly behind initial projections following a pullback in first-half deliveries.

The industry now enters the second half of the year under immense pressure, while the the domestic auto market is increasingly defined by intense competition in a mature, replacement-driven market, according to the PCA statement.

“Internal polarization across the sector continues to widen,” said Cui Dongshu, secretary general of the association, while warning that both costs and margins are under severe pressure, squeezed by skyrocketing automotive-grade chip prices and speculative bidding for lithium carbonate.

While the association lowered its 2026 retail sales growth forecast to a 14% contraction in mid-June, the group expects market sentiment to steadily turn a corner after July.

Anthropic 3Q26 Profit Over $1B: The Anthropic IPO Financials Sneak Peak

(…) we expect Anthropic to take advantage of their superior business model and margins to invest in further in new models that help extend their lead and monetization over closed and open source competitors. Anthropic has the ability to truly make OpenAI dance and we see Anthropic as the first $6T company as a base-case possibility if they continue to execute.

Pricing power, gross margins, business model, and profitability are all reasons for Anthropic to IPO first and put the impetus on OpenAI to open their financials and raise the necessary capital to compete and fund the massive AI buildout still to come. (…)

While other SemiAnalysis work focuses on the technical aspects of AI Labs, this piece will focus solely on the current and future financials, margin economics, and long-term outlook for Anthropic.

(…) we see over 100GW of compute demand between OpenAI and Anrthropic by the end of 2030. This will require over 90GW of net compute additions through 2030 vs 2.5GW added in 2025 and 5GW in 2026. Today, we estimate OpenAI and Anthropic have access to just over 6GW of compute. (…)

Subscriptions make up just 15% of total ARR [Annualized Recurring Revenue]. Compare this to OpenAI and you’ll see a stark contrast. During 1Q26, OpenAI was over 65% subscription based.

The business vs consumer split is even more stark. Just 5% of Anthropic ARR today comes from Consumer subscriptions vs our estimates of OpenAI around 40% at the end of 2Q26 (the CCO of OpenAI said in a blog that B2B was around 40% of total ARR in 1Q26). Claude Code’s success has levered Anthropic to the right tailwinds and business model. As we show in the margin section below, margins on the API business are vastly superior to consumer subscriptions. (…)

The AI Lab inference business model is simple. Revenue is dictated by token volume and blended token prices. Token volume has skyrocketed, but blended token pricing has declined over time. With higher token pricing on new models, how has blended pricing gone down? Cache rates, input/output ratios, and agentic vs chat workflow mix are all factors in determining blended token pricing. (…)

As more agentic work gets adopted, revenue growth is driven by rising token consumption. Agentic workflows consume significantly more tokens than the traditional chat function, and blended token prices have actually fallen drastically since 2024. The consumption curve, then, is doing all the work in driving revenue growth.

Broadly, the business model of a lab is to invest in building new models, buy or rent compute to train and serve those models, and then sell that compute to customers via tokens. Anthropic and OpenAI have very different ways of selling that inference compute historically.

Anthropic has wholly aligned their business model with high-margin, usage-based pricing while OpenAI is serving costly free and subscription-based users who can run up token usage that hog compute. While OpenAI continues to converge towards Anthropic with incremental ARR being much more B2B and API based, they still have the weight of consumer and subscription-based usage hitting their financials. (…)

Labs rent or buy compute from various sources and sell subscriptions or usage-based tokens on top of that. Labs have several levers to pull to get the most out of a given unit of compute they have for inference.

First, new models typically have higher prices. Second, different accelerator chips have different token throughputs and can be optimized over time to better serve inference. Tokens are the fundamental unit of measure for AI workloads that LLMs use to process data. Think of a token as a fragment of the query; it could be a word, part of a word, or punctuation mark in a text-based query.

These compute costs are largely fixed per unit of compute, so when you can either: 1) get more tokens through a given unit or 2) higher prices on the tokens you run through a given unit of compute, incremental margins approach 100%. This is why you see blended gross margins today in the mid 60% range vs –94% in 2024.

Labs have made great strides in inference efficiency in the last 12 months, which enabled ARR per MW of compute to drastically rise, and margins followed. Anthropic ARR per MW will reach the $60M range later this year. Just 9 months ago, this number was $16M. This is the powerful margin upside in lab financials at work: token throughput can be improved and better monetized while compute costs are largely fixed, leading to massive margin upside when revenue growth is strong. (…)

Anthropic, at $1B of GAAP EBIT in 3Q26 (6% margin) and growing NNARR faster than OpenAI, who is sitting at –100% EBIT margins, is a stark contrast. (…)

Anthropic is showing that AI Labs are wonderful, profitable businesses when run with focus and execution. We expect investors to reward them with ample access to capital to fund the next stage of the compute build-out when they go public. For OpenAI, it is paramount that they continue to flex early Codex and API momentum into better margins they can reinvest in model development and chart a path to better profitability for their build-out.

• Anthropic and OpenAI Face a New Threat from China

Silicon Valley has long defined victory in the global artificial intelligence race as being determined by capability. Whoever has the smartest and most powerful AI model leads the pack. But what if winning were instead defined by adoption by technologists? In that case, China is swiftly gaining ground.

Anecdotes and tentative data suggest that more American startups are reaching for cheaper Chinese models from companies like Hangzhou DeepSeek Artificial Intelligence Co. (better known as DeepSeek), Alibaba Group Holding Ltd. and Moonshot AI Pte Ltd. to cut costs, as coding with AI and so-called tokenmaxxing becomes prohibitively expensive. (…)

With top-of-the-line AI built into their workflows, some companies are now looking at ways to pivot to cheaper options, the logic being that you don’t need a chainsaw when scissors will do. Smaller models can handle most everyday tasks, like sorting and drafting routine emails, summarizing documents, answering common customer service questions or tagging, sorting and cleaning up data. Many of those affordable, lightweight offerings now come from China.

The latest flagship AI model from Hangzhou-based DeepSeek, for instance, costs about $0.87 per million output tokens, versus roughly $30 for OpenAI and $25 for Anthropic. Tokens are the units AI companies charge by, and can be thought of as billable chunks of text, roughly equivalent to a word each.

Since companies often switch between different AI providers it can be difficult to get a clear view of market share, but there are signs that startups and software developers are increasingly moving toward models from China. Vercel Inc., a cloud platform used for deploying websites and apps that also tracks which AI models are being used by programmers, found that DeepSeek’s share of AI traffic jumped to 17% from under 1% in May.

A similar platform, OpenRouter Inc., saw DeepSeek’s usage double between January and June 2026, making it the most popular model among its users. It said in a recent blog post that a group of Chinese open-source models including those from Xiaomi Corporation, MiniMax Group Inc. and Tencent Holdings Limited “all saw their share of tokens rise over the past six months,” and that the shift had come at the expense of Google and OpenAI.

OpenRouter only represents about 3% of global AI traffic, most of it coming from startups and independent developers — not the big corporate customers who account for the bulk of AI labs’ revenue. But it signals a direction of travel for coders and raises questions about whether larger companies might eventually make a similar change.

Lindy, an AI assistant startup based in San Francisco, recently said it had switched to DeepSeek from Anthropic’s Claude after AI costs for the 25-person company exceeded personnel costs. CEO Flo Crivello said the move saved his startup “millions,” though it would continue using Anthropic for coding “because of the absurd Max plan subsidy.”

Some larger companies like Airbnb Inc. and Anysphere Inc., the firm behind coding platform Cursor that is being acquired by SpaceX, have added Chinese models to their mix rather than swap out US ones entirely.

All of this is happening because the frontier innovation that gave US labs a moat against competitors has been seeping out to the world, with startups in Shenzhen and Shanghai distilling that tech (a technique some see as flagrant copying) to make cheaper and lighter versions.

None of this means China has won anything yet. The US still builds the most capable models, and the largest corporate customers aren’t shifting who they buy from any time soon.

But history has shown that adoption has its own kind of advantage. The more the world’s developers build on Chinese models, the more entrenched the habits, standards and dependencies of those models become for the world of programming and fast-growing startups. That’s perhaps a quieter contest than the race for the smartest tools, and it’s one the leading US labs may be losing.