Blog

Chinese chain Luckin Coffee opened its first two U.S. locations this week, betting that mobile-only ordering and creative flavors can lure customers away from Starbucks.

Both new Luckin stores are based in Manhattan, and at the midtown location on Wednesday, Sam Liu took a sip of her jasmine cold brew.

“I’ve never tried anything like it,” she said.

Liu said she’d hoped for more seating — the small shop has only three tables — and was initially confused by Luckin’s in-app ordering system, which means customers can’t order directly from a barista.

“I thought I just order at the counter, but I realized everyone was standing around looking at their phone,” Liu said.

Luckin is China’s largest coffee chain, with more than twice as many locations as Starbucks there. Its two New York City stores are its first foray outside Asia, where it has over 24,000 locations across the region. By comparison, there are over 17,000 Starbucks in the United States.

Its CEO, Guo Jinyi, called the U.S. “a strategically important market” for the company’s expansion in a press release heralding the two new locations Wednesday. “We are excited to introduce a diverse and unique coffee experience to American consumers.”

The company, which didn’t respond to a request for comment, has touted its ambitions to expand globally but hasn’t publicly detailed its next moves in the U.S. or other markets.

The chain has gained success overseas through creative drinks like alcohol-infused coffees and fruit lattes, along with its smartphone-centric ordering model. The app-based approach makes it easier to track inventory, send personalized appeals to consumers and serve drinks quickly, said John Zolidis, an analyst who tracks Luckin and Starbucks at the brokerage firm he founded, Quo Vadis Capital.

“Luckin was able to develop an incredible muscle with regard to product innovation, and they have been very creative in China,” he said.

Zolidis said how Luckin fares on Starbucks’ home turf will depend on its ability to differentiate its menu from other major U.S. coffee chains and smaller, independent cafes. Its American lineup already includes distinctive drinks like blood orange cold brew and coconut lattes.

“These orange drinks, or one of their most successful, a coconut cloud latte — that’s how you get trial [customers] from the U.S.,” Zolidis said.

Luckin faced financial troubles during the pandemic. It was delisted from Nasdaq in 2020 after its stock plunged following an internal investigation that found an executive had falsified revenue reports. The company filed for bankruptcy in the U.S. the following year but emerged from proceedings in 2022 and its sales have soared since, reaching $4.7 billion worldwide in fiscal year 2024, a 38.4% increase from 2023.

Starbucks, by contrast, is struggling in both the U.S. and China. Its same-store sales in the U.S. declined 2% and its sales in China 8% in fiscal year 2024, and it reported in April that its quarterly profit was half of what it pulled in for the same period last year. The Seattle-based chain is reportedly looking to partially sell its business in China while revamping its U.S. strategy to focus on customer experience and human connection, in contrast with Luckin’s model.

“We veered away from, I think, owning the idea of the ‘third place,’ the coffeehouse experience, making sure that the customer was front and center,” Starbucks CEO Brian Niccol told NBC News in June.

A Starbucks spokesperson declined to comment.

Zolidis said that whereas Starbucks aims in both the U.S. and China to appeal to customers looking for higher-end coffee served in an inviting setting, Luckin has successfully positioned itself as the “everyman’s coffee” in China, with low prices and small, grab-and-go storefronts.

After taking the train in from Hoboken, New Jersey, to check out the new one in midtown, Samantha Coy said the trip was worth it. She had enjoyed Luckin in China previously and was eager to order one of its fruit drinks.

“I’m surprised Starbucks hasn’t tried to bring that over to the U.S.,” Coy said. “I hope they stay open.”

Zolidis said he thinks Luckin is well-positioned to gain a foothold in America.

“They’ve been able to operate and grow incredibly quickly in the Chinese market, much faster than I would have thought possible, and they’ve been able to sustain it and develop a strong financial model so they can fund their expansion in the U.S.,” Zolidis said. “They wouldn’t be coming here to try it if they didn’t think they had a shot of owning part of the market.”

By Gordon Gottsegen

Retail investors’ dip-buying remains as strong as ever despite tariffs, Middle East tensions and economic uncertainty

Tariffs, market volatility and war in the Middle East couldn’t slow down the buying spree by individual investors, as they traded a record amount of stocks in the first half of the year.

Retail investors cumulatively bought around $3.4 trillion worth of equities over the first half of 2025, according to data from Nasdaq. At the same time, they sold about $3.2 trillion worth – bringing the total traded to over $6.6 trillion.

This represents a strong bias toward buying into the market versus taking money out, despite high levels of market volatility in the first half of the year. Tariff announcements from President Donald Trump spooked markets, while investors weighed the possibility of global trade wars leading to an economic slowdown and higher inflation. The Dow Jones Industrial Average DJIA and S&P 500 SPX entered a correction, while the Nasdaq Composite COMP fell into bear-market territory. Some investors said it was “the toughest investment climate” they ever experienced.

Yet retail investors’ behavior showed that they were bullish in the face of all of this. Cumulative net retail inflows hit $137.6 billion into U.S. single stocks and exchange-traded funds the first half of the 2025, according to Nasdaq.

Data from capital-markets research firm Vanda Research differed slightly from the Nasdaq data, but it showed the same general trend. Vanda found that investors net purchased $155.3 billion worth of single stocks and ETFs in the first half of 2025. This was the largest-ever net inflow of retail investor cash since Vanda started keeping track in 2014. Inflows surpassed the previous high of $152.8 billion reached in the first half of 2021, when the meme-stock mania and pandemic stimulus checks drove hordes of everyday investors into the stock market.

In 2025, buying was driven by two things, Vanda said: the “American exceptionalism” trade and a record amount of dip-buying in response to Trump’s “liberation day” tariffs. Buzzy U.S. stocks – like Nvidia (NVDA), Tesla (TSLA) and Palantir (PLTR) – topped the charts of the most actively traded tickers throughout the first half, but retail investors also poured significant amounts of capital into index-tracking ETFs like SPDR S&P 500 ETF Trust SPY and Invesco QQQ Trust Series I QQQ.

Read more: Fourth of July holiday highlights 4 reasons ‘American exceptionalism’ isn’t going anywhere

The average daily inflow was roughly $1.3 billion, according to Vanda, which represents a 21.6% jump from the average in 2024.

This level of stock-buying hasn’t exactly hurt investors’ performance either. Vanda estimated that the average retail portfolio was up 6.2% so far in 2025, which was closely in line with the 6.1% that the S&P 500 gained in the first half of this year.

“Retail investors remain a major force in the market. Participation is at record highs, the dip-buying bias is fully intact, and engagement with single names – particularly high-beta and leveraged plays – continues to rise. Performance is holding up, and risk appetite is anything but shy. Nothing seems to stop this retail train,” Marco Iachini, Vanda Research’s senior vice president of research, wrote in a note.

-Gordon Gottsegen

This content was created by MarketWatch, which is operated by Dow Jones & Co. MarketWatch is published independently from Dow Jones Newswires and The Wall Street Journal.

  (END) Dow Jones Newswires

  07-05-25 0800ET
Copyright (c) 2025 Dow Jones & Company, Inc.

The six-month, full-time paid program offers aspiring AI researchers from adjacent fields like physics or neuroscience a pathway into the AI industry, rather than recruiting individuals already deeply invested in AI research and work. According to Jackie Hehir, OpenAI’s research residency program manager, residents aren’t those seeking Ph.Ds in machine learning or AI, nor are they employees of other AI labs. Instead, she said in a program info session, “they’re really passionate about the space.”

So what’s in it for OpenAI? Hot talent at cut-rate prices. While the six-figure salary puts OpenAI residents in the top 5% of American workers, it’s a bargain in the rarefied world of AI, where the bidding war for talent has some companies tossing around nine-figure bonuses.

By offering a foothold into the AI world, OpenAI appears to be cultivating talent deeply embedded in the company’s mission. This strategy, spearheaded by CEO Sam Altman, has long been part of the company’s approach to retaining employees and driving innovation. One former OpenAI staffer described the employee culture to Business Insider as “obsessed with the actual mission of creating AGI,” or artificial general intelligence.

Mission driven or not, OpenAI’s residents are also compensated handsomely, earning an annualized salary of $210,000, which translates to around $105,000 for the six-month program. The company also pays residents to relocate to San Francisco. Unlike internships, the program treats participants as full-fledged employees, complete with a full suite of benefits. Nearly every resident who performs well receives a full-time offer, and, according to Hehir, every resident offered a full-time contract so far has accepted. Each year, the company welcomes around 30 residents. 

The qualifications for residents at OpenAI are somewhat unconventional. In fact, the company claims there are no formal education or work requirements. Instead, they hold an “extremely high technical bar” at parity to what they look for in full-time employees as it pertains to math and programming.

“While you don’t need to have a degree in advanced mathematics, you do need to be really comfortable with advanced math concepts,” Hehir said.

As OpenAI attempts to build talent from the ground up, its rivals, namely Meta, are pulling out all the stops to poach top AI talent with reports alleging that Meta CEO Mark Zuckerberg personally identified top OpenAI staff on what insiders dubbed “The List” and attempted to recruit them with offers exceeding $100 million in signing bonuses.

Meta’s compensation packages for AI talent can reportedly reach over $300 million across four years for elite researchers. The flood of cash has ignited what some insiders call a “summer of comp FOMO,” as AI specialists weigh whether to stay loyal to their current employers or leave for record-breaking paydays.

Zuckerberg’s methods have had some success, poaching a number of OpenAI employees for Meta’s new superintelligence team. In response to news of the employees’ departure, OpenAI’s chief research officer, Mark Chen, told staff that it felt like “someone has broken into our home and stolen something.”

Meanwhile, OpenAI CEO Sam Altman called Meta’s recruitment tactics “crazy,” warning that money alone won’t secure the best people. “What Meta is doing will, in my opinion, lead to very deep cultural problems,” Altman told employees in a leaked internal memo this week. 

Ultimately, cultivating new talent, rather than trying to outbid the likes of Meta, may prove a more sustainable path for OpenAI in its quest to stay highly mission-oriented while supporting an industry grappling with a scarcity of top-tier talent. Estimates suggest there are only about 2,000 people worldwide capable of pushing the boundaries of large language models and advanced AI research. Whether the talent cultivated by Altman and OpenAI will remain loyal to the firm remains unknown. But Altman says that AI “missionaries will beat mercenaries.”

Participants

Twenty-four healthy, community-dwelling older adults, consisting of 15 men and 9 women with a mean age of 72.1 ± 4.1 years, participated (Table 1). The eligibility criteria of this study were as follows: age ≥ 65 years; ability to walk independently without assistive devices; and absence of conditions that could significantly influence gait, such as neurological disorders (e.g., Alzheimer’s disease, Parkinson’s disease, or stroke), severe cardiovascular or respiratory impairments with symptoms during daily activities (e.g., heart failure, chronic obstructive pulmonary disease), or musculoskeletal problem that disable independent gait (e.g., joint replacement, spinal surgery, or advanced arthritis).

Sample size calculation

The required sample size was determined based on the population within-subject standard deviation (PWSD). The number of subjects was determined to estimate PWSD within 10% of the population value ((frac{1.96}{sqrt{2nleft(m-1right)}}leq0.1,;m=the;number;of;observations;per;subject)) using the variance of PSWD ((frac{{sigma }_{w}}{sqrt{2nleft(m-1right)}}, {sigma }_{w}=PWSD)) [21]. A sample size of 24 was required for all of the distances with nine or more observations per subject (m ≥ 9) except for the two longest distances (4.9 and 5-m).

Muscle mass and strength assessments

Participants underwent bioimpedance analysis using an InBody S10 device (InBody Co., Ltd., Seoul, South Korea) to determine height-adjusted appendicular skeletal muscle mass. Muscle strength was assessed by measuring handgrip and isometric knee extension strength. Handgrip strength was measured using a Takei 5401 Digital Dynamometer (Takei Scientific Instruments Co., Ltd., Niigata, Japan) in a standing position with the elbow fully extended. Isometric knee extension strength was evaluated using a TKK-5710e tension meter (Takei Scientific Instruments Co., Ltd., Niigata, Japan); during measurement, participants were seated on a chair with a dynamometer anchored to it, maintaining knee flexion at 90°. Both measurements were conducted bilaterally, with each side assessed twice and a 1-min rest period between attempts. Participants were instructed to exert maximum effort for each measurement, and the highest reading was used in the analysis. All procedures were conducted by a single trained assessor following the recommendations of Asian Working Group for Sarcopenia and the European Working Group on Sarcopenia in Older People [9, 10].

Physical performance assessments

Physical performance was evaluated using the Short Physical Performance Battery (SPPB) [22], the 30-s chair stand test [23], the five-times sit-to-stand test [24], and the timed up-and-go test [25]. All assessments were conducted by a single trained assessor in a spacious setting under consistent environmental conditions, following the protocols of the Asian Working Group for Sarcopenia and the European Working Group on Sarcopenia in Older People [9, 10].

10-m gait speed test and data acquisition

Participants walked along a 10-m walkway, which included a 2-m acceleration zone for a dynamic start and a 2-m deceleration zone at the end. They were instructed to walk at their usual pace on a hard surface while wearing comfortable footwear. The 10-m walk was repeated three times, with a minimum rest period of 2 min between trials. Recordings were captured using an Apple iPad Pro 11 2nd Generation (Apple, Inc., Cupertino, CA, USA) equipped with RGB cameras arranged perpendicularly to the walking path at a distance of 3.8 m and a height of 0.8 m. Videos were recorded in the sagittal plane (resolution: 800 × 600 pixels; 30 fps; Fig. 1).

Gait analysis using 2D pose estimation

A customized pose estimation model (ViFive, Inc., Boulder, CO, USA) was used, which tracked 14 key body points using an architecture adapted from a standard stacked hourglass model [26]. We introduced multiple objectives to enhance the context, accuracy, speed, and stability of the model, which are vital for musculoskeletal assessment. The classification model included a random forest classifier with optimized features to increase accuracy and speed while reducing the model size. Pixel-per-meter estimation used markers at 2 and 8 m (Fig. 1). The CoM of each subject was determined using the weighted sums of the body segment centers of mass (Fig. 2a).

Illustrative case. a The movement pattern of the center of mass over time as estimated via pose estimation. b Gait speed of each segment according to the measurement distance (1.0–5.0 m). The x-axis represents the percentile of total walking distance (%), and the y-axis represents gait speed (m/s). c Distribution of gait speed according to the measurement distance

Gait speed estimation

Gait speed was measured using two independent methods for validation: manually with a stopwatch and using pose estimation algorithms. Manually assessed speed was determined by an evaluator using a stopwatch to record the time taken for the subjects to pass by the markers set at 2 and 8 m. Pose estimation gait speed was calculated by dividing the distance covered between frames by the elapsed time using either the CoM or the leading foot as reference points. CoM-referenced measurements simulate those obtained via conventional motion capture system, whereas leading foot-referenced measurements simulate those made using walkway or pressure sensors such as GAITRite® (CIR Systems Inc., Franklin, NJ, USA).

Gait speed measurement validation

Gait speeds measured using a stopwatch and pose estimation were compared using a linear mixed-effects model, with speed over 6 (manual) or 5 m (pose estimation) per trial as the dependent variable and with subject random effect to account for multiple tests from each subject. The intraclass correlation coefficient (ICC) was used to evaluate absolute agreement between gait speed measurements obtained via the two methods for the same walking trials.

Change of uncertainty with measured distance

A 5-m walk video of a skeleton with 14 key points was extracted from each recording using our pose estimation algorithm. This was further edited by cropping at 0.1-m intervals to generate 4.9- to 1.0-m segments. One 5.0-m walk video generated two 4.9-m segments, three 4.8-m segments, and so forth, up to 41 segments for a 1.0-m walk, culminating in 861 segments of varying distance (Fig. 2b,c).

The variability of gait speed across the measured distances was defined as the within-subject standard deviation (WSD) for each measured distance, calculated as the square root of the mean-square error in a one-way analysis of variance, where groups combined subjects with distance intervals. Three gait speed data from three measurements were collected for each group to avoid underestimating within-subject variation due to overlapping distances when distance intervals are not considered. For example, for a 4.7-m walk, four gait speed measurements were obtained at 4.7-m distances (0–4.7, 0.1–4.8, 0.2–4.9, and 0.3–5.0 m), and within-subject variation at a 4.7-m distance was estimated by considering different distance intervals.

Determination of minimum required distance

To determine the minimum required distance, we utilized WSD at each measured distance. Given that confidence intervals (CI) quantify variability, we computed the half-width of the CI using WSD and the critical value corresponding to the chosen confidence level. Specifically, the 95% CI was calculated as 1.96 × WSD, and the 90% CI as 1.64 × WSD. For the measurement to be clinically meaningful, the half-width of the CI, reflecting gait speed variability, had to remain below the MCID of 0.1 m/s [27, 28]. Thus, the minimum required distance was defined as the shortest distance at which this criterion was met, ensuring that gait speed measurements remained within an acceptable range of variability.

Factors affecting gait speed variability

CoM trajectory was plotted as distance against time for each test. Linear regression analysis provided a trend line and the mean squared error (MSE) for each subject. As MSE quantifies deviations from the trend line, lower MSE values indicate less variability in gait speed, leading to a shorter minimum required distance. We investigated whether epidemiological, anthropometric, or clinical variables were associated with MSE using linear regression following Pearson’s correlation for continuous variables and point-biserial correlation for dichotomous variables to identify subject characteristics influencing the minimum required distance. All processing and statistical analyses were conducted using MATLAB R2023b (MathWorks, Natick, MA, USA) and SAS 9.4 (SAS Institute, Cary, NC, USA), with statistical significance set at p < 0.05.