Category: 3. Business

Second, solar energy’s marginal fuel costs are zero, meaning that it costs nothing to produce every additional unit of electricity beyond the original cost of installing the panels and the ongoing cost of maintaining them.

Third, solar panels are modular: They come in small sizes available at constant fixed prices, making it easier to create a decentralized grid. In contrast, thermal or nuclear power stations are usually large with high fixed costs.

Bottlenecks in solar panel supply are unlikely, in part because there is plenty of excess capacity in the sector, with China’s production potential alone covering 200% of global demand in 2024. “Any slowing in solar growth is likely to come from reduced policy support and from power supply volatility rather than from solar panel supply bottlenecks,” says Struyven.

This material is provided in conjunction with the associated video/audio content for convenience.  The content of this material may differ from the associated video/audio and Goldman Sachs is not responsible for any errors in the material. The views expressed in this material are not necessarily those of Goldman Sachs or its affiliates. This material should not be copied, distributed, published, or reproduced, in whole or in part, or disclosed by any recipient to any other person. The information contained in this material does not constitute a recommendation from any Goldman Sachs entity to the recipient, and Goldman Sachs is not providing any financial, economic, legal, investment, accounting, or tax advice through this program or to its recipient. Neither Goldman Sachs nor any of its affiliates makes any representation or warranty, express or implied, as to the accuracy or completeness of the statements or any information contained in this material and any liability therefore (including in respect of direct, indirect, or consequential loss or damage) is expressly disclaimed.

© 2025 Goldman Sachs. All rights reserved.

Donald Trump has called on Intel’s chief executive to resign, alleging Lip-Bu Tan had ties to the Chinese Communist party, sending the stock of the US chipmaker falling.

“The CEO of Intel is highly CONFLICTED and must resign, immediately,” Trump posted on Truth Social about Tan. “There is no other solution to this problem. Thank you for your attention to this problem!”

Shares in Intel dropped more than 3% in early trading. The company did not immediately respond to a request for comment on Trump’s post.

Trump’s comments came a day after he threatened a 100% tariff on imported semiconductors and chips, which could favor Intel as a US-based semiconductor firm.

Trump did exclude Taiwan Semiconductor and Apple, companies that have said they plan to increase their investments in US manufacturing.

Apple chief Tim Cook announced from the White House that the company would invest $100bn in US chip fabrication.

Trump’s criticism of Intel, which has lagged chipmakers such as Nvidia in producing graphics-processing chips suitable for AI applications, comes after Arkansas senator Tom Cotton wrote a letter to company chairman Frank Yeary expressing concern over Tan’s investments and ties to semiconductor firms that are reportedly linked to the CCP and the People’s Liberation Army, the party’s military arm.

Cotton, a Republican, asked Intel’s board if Tan had divested his interests and questioned if Tan’s previous leadership of Cadence Design Systems, a company that last month said it had sold products to China’s National University of Defense Technology, a violation of US export controls.

Cotton said Tan controlled dozens of Chinese companies, at least eight of which had ties to the People’s Liberation Army.

In a statement, the company said: “Intel and Mr Tan are deeply committed to the national security of the United States and the integrity of our role in the US defense ecosystem. We appreciate Senator Cotton’s focus on these shared priorities. We look forward to addressing these matters with the senator.”

Tan, 65, is regarded as an industry veteran in technology and venture capital. He was tapped to lead the once-dominant personal computing and laptop chipmaker in March as part of a turnaround effort. Intel’s market valuation is about $89bn, compared with $4.4tn for rival Nvidia.

Tan, 65, has said he wants to sell off Intel assets that are not core to the company’s revitalization, cut jobs and to delay or cancel projects to reduce operating expenses. Intel has received about $8bn from the Chips and Science Act for US investments, the Biden-era legislation framed as a national security imperative to reduce US dependence on foreign chip production.

A number of foreign-based chipmakers have recently announced they are boosting US chip production , including Taiwan’s TSMC and South Korea’s Samsung.

Beneficiaries of the Chips Act besides Intel include TSMC, Micron Technology, Samsung, GlobalFoundries and Texas Instruments. But the $1bn grants are dwarfed by more than $400bn in pledged private-sector investments.

The Trump administration wants Congress to scrap the Chips Act, arguing that tariffs are a more effective incentive for companies to build fabrication plants on US soil, pointing to TSMC’s decision to expand its chipmaking capacity from three to six plants in Arizona.

In the newest edition of a report that the CDC uses to provide estimates on the percentage of ultra-processed food that Americans consume each year, the results were enlightening for both children and adults in the country. About 62% of children’s and teens’ daily calories are coming from ultra-processed foods, closely followed by adults, who have an intake that equals about 53% of their daily calories, according to the report.^1,2

The report is part of HHS Secretary Robert F. Kennedy Jr’s continued initiative to reduce the amount of ultra-processed food in the United States. The secretary of HHS has previously talked about his goals of reducing the number as much as possible, including asking for food companies to stop using certain food dyes in their products.³ Kennedy has also pointed to ultra-processed foods as the cause of the “chronic disease epidemic” in America, believing that a reduction in the consumption of such foods would lead to a healthier country.

“Ultra-processed foods” has become a catch-all phrase for foods that require involved methods to produce them or foods that are synthesized with other compounds, which can help them last longer on the grocery store shelf. Ultra-processed foods can include foods like frozen pizza and instant noodles but also store-bought bread.⁴ Previous reviews have found that there is a link between ultra-processed foods and mortality from any cause, including heart disease and mental health conditions.

The CDC report aimed to assess how many people in America were consuming these ultra-processed foods, as it could enlighten the government and the people of the US alike on how much work needed to be done to reduce the intake of these foods. The report was based on the National Health and Nutrition Examination Survey that was conducted between August 2021 and August 2023.¹ The definition of ultra-processed food was based on the NOVA classification developed in Brazil, which classifies ultra-processed foods as those that are industrial creations that are made with little whole food.

Burgers, hot dogs, and peanut butter and jelly sandwiches were among the top sources of ultra-processed foods in children and adults, according to the report. Salty snacks, sugary drinks, and baked goods were also among the most frequently cited sources of ultra-processed food consumption. Those with higher incomes also tended to eat fewer ultra-processed foods compared with those with lower incomes.

“Although youth and adults consumed the majority of their calories from ultra-processed foods in the past decade, a decrease was seen in ultra-processed food consumption among youth and adults between 2017–2018 and August 2021–August 2023, and a decrease was seen among adults from 2013–2014 to August 2021–August 2023,” the CDC reported.¹

Marketing, said Marion Nestle, professor emerita of nutrition, food studies, and public health at New York University, is a foundational part of decreasing use of ultra-processed foods, as most foods are marketed toward kids. “They’re seen as cool and are iconic and you’re lucky to eat them, because that’s how they’re marketed,” she said in a statement.²

The new report emphasizes the need to reduce ultra-processed foods in the diets of Americans but necessitates that experts pick out the actual harmful ultra-processed foods rather than all such foods, as some ultra-processed foods are beneficial. However, as the majority of ultra-processed foods can be harmful, it’s important to define which foods carry health risks as the administration moves toward lowering the number of ultra-processed foods in the diets of all Americans.

References

1. Williams AM, Couch CA, Emmerich SD, Ogburn DF. Ultra-processed Food Consumption in Youth and Adults: United States, August 2021–August 2023. NCHS Data Brief. 2025;536:1-11. https://www.cdc.gov/nchs/products/databriefs/db536.htm

2. Lovelace B Jr. Ultra-processed foods make up the majority of kids’ diet, CDC report finds. NBC News. August 7, 2025. Accessed August 7, 2025. https://www.nbcnews.com/health/health-news/ultra-processed-foods-make-majority-kids-diet-cdc-report-finds-rcna223481

3. Nowell C. Inside RFK Jr’s conflicted attempt to rid America of junk food. The Guardian. July 8, 2025. Accessed August 7, 2025. https://www.theguardian.com/environment/2025/jul/08/rfk-jr-junk-food

4. MacMillan C. Ultraprocessed foods: are they bad for you? Yale Medicine. July 10, 2024. Accessed August 7, 2025. https://www.yalemedicine.org/news/ultraprocessed-foods-bad-for-you

Born June 18, 1940, in Grundy, Virginia, where the Mountain Mission School is located, McGlothlin received his undergraduate and law degrees from the College of William & Mary. McGlothlin founded the United Company, based in Blountville, Tennessee, in 1970 as a coal production company. It later diversified its four-state operations into oil and gas exploration services, investment management services and real estate development. He sold the company in 1997 but later repurchased it seven years later.

The World Health Organization (WHO) has officially designated Health Canada, the Ministry of Health, Labour and Welfare/Pharmaceuticals and Medical Devices Agency (MHLW/PMDA) of Japan, and the Medicines and Healthcare products Regulatory Agency (MHRA) of the United Kingdom as WHO-Listed Authorities (WLAs), a status granted to national authorities that meet the highest international regulatory standards for medical products.

With these latest designations, WHO expands the growing list of WLAs, now involving 39 agencies across the world, supporting faster and broader access to quality-assured medical products, particularly in low- and middle-income countries (LMICs).

In addition, the Republic of Korea’s Ministry of Food and Drug Safety (MFDS) – one of the first regulatory authorities to complete the WLA assessment for both medicines and vaccines in October 2023 – has had its listing scope successfully expanded, now covering all regulatory functions.

“This recognition reflects the deep commitment of these authorities to regulatory excellence,” said Dr Tedros Adhanom Ghebreyesus, WHO Director-General. “Their designation as WHO-Listed Authorities is not only a testament to their robust regulatory systems but also a critical contribution to global public health. Strong and trusted regulators help ensure that people everywhere have access to safe, effective, and high-quality medical products.”

Around 70% of countries worldwide still face significant challenges due to weak or inadequate regulatory systems for evaluating and authorizing medical products. The WLA framework promotes regulatory convergence, harmonization and international collaboration, allowing WHO Prequalification Programme and regulatory authorities, especially those in LMICs, to rely on the trusted work and decisions of designated agencies. This collaboration supports efficient use of limited resources, enabling better and faster access to quality-assured life-saving medical products to millions more people.

“The principle of reliance is central to WHO’s approach to regulatory systems strengthening and a cornerstone for effective, efficient and smart regulatory oversight of medical products,” said Dr Yukiko Nakatani, WHO Assistant Director-General for Health Systems, Access and Data. “WHO-Listed Authorities are key enablers in promoting trust, transparency, and faster access to quality-assured medical products, especially in low- and middle-income countries.”

In a world where health threats, including substandard and falsified medical products, know no borders, WLAs also serve as critical pillars of preparedness and equity, making life-saving products available more broadly, rapidly and efficiently.

The designations follow a rigorous performance evaluation process carried out by WHO using its globally recognized benchmarking and assessment tools. These evaluations were reviewed by the Technical Advisory Group on WLAs (TAG-WLA), which convened in June 2025.

Canada, Japan and the UK’s regulatory authorities were previously recognized as Stringent Regulatory Authorities (SRAs). Their designation under the WLA framework marks an important step in moving beyond the old SRA system, while ensuring continuity and stability in global procurement processes of quality-assured medical products.

Launched in 2022 to replace the previous SRA model, the WLA initiative provides a transparent and evidence-based pathway for global recognition of regulatory authorities to facilitate regulatory convergence and reliance. It builds on decades of WHO leadership to help countries work together more closely on regulating medical products, speeding up access to safe, effective and quality-assured medical products for people around the world. 

Study sample

Data for this study were obtained from a cohort of couples preparing for pregnancy at Guangzhou Baiyun District Maternal and Child Health Hospital. The objective was to assess the impact of pre-pregnancy BMI on TTP. The study was registered with the China Clinical Trials Registry (ChiCTR) under registration number ChiCTR2300068809, with the initial registration on January 3, 2023. Further methodological details are available in previous publications [17].

Couples who participated in the National Free Pre-pregnancy Checkup Program (NFPCP) between January and June 2022 were included. Telephone follow-ups were conducted 13 to 15 months post-examination to inquire about their pregnancy preparation status and subsequent pregnancies.

The inclusion criteria were as follows: (1) Female partner aged between 20 and 49 years, and male partner aged 22 years or older; (2) Couples who were not pregnant at the time of the initial assessment; (3) Both partners self-reported an intention to conceive and were not using contraception during the examination.

The exclusion criteria included: (1) Female partners who tested positive for cytomegalovirus or Toxoplasma gondii IgM antibodies, or if either partner had syphilis, HIV, or any other medical condition requiring treatment that would delay conception; (2) Couples who declined participation or were unwilling to cooperate with the follow-up survey; (3) Couples who were already pregnant during the examination month; (4) Couples who were planning to undergo or had previously used assisted reproductive technology (ART).

After applying the exclusion criteria, a total of 1,684 couples were included in the study. A more detailed analysis was conducted on a subsample of 1,506 couples in which the female partner reported a regular sleep pattern (Fig. 1).

Ethical approval

The study was approved by the Medical Ethics Committee at Guangzhou Baiyun District Maternal and Child Health Hospital. Every participant provided written informed consent before enrolling in the study, and verbal consent was obtained again during follow-up for participation in the telephone interview.

Exposures and outcome

In this study, the exposure factors refer to various sleep behaviors, including irregular sleep patterns, sleep onset time, sleep duration, insomnia, and perceived insufficient sleep. Sleep behaviors during the preconception period were obtained through telephone follow-up by trained medical professionals. Definitions and measurements were as follows:

Irregular sleep patterns were defined as inconsistent or highly variable sleep and wake times in daily life during the preconception period. Specifically, participants were asked: “During your preconception period, did you have irregular sleep patterns (inconsistent bedtimes and wake times varying by >1 hour on ≥3 days per week)?” (Yes/No).

Sleep onset time was defined as the time women subjectively reported falling asleep during the preconception period, and was recorded in a decimal format in which 60 minutes equaled one unit (e.g., 23:30 was recorded as 23.5). If sleep occurs past midnight, each additional hour is incremented by 1 (e.g., a bedtime of 1:00 AM is recorded as 25). Participants were asked: “During your preconception period, what time did you usually fall asleep?”

Sleep duration refers to the average total sleep time per day during the preconception period. Participants were asked: “During your preconception period, how many hours of sleep did you get in a typical 24-hour period?”

Insomnia was defined as persistent difficulty initiating or maintaining sleep, or experiencing early morning awakenings, among women attempting to conceive. Participants were asked: “Did you experience persistent difficulty falling asleep, staying asleep, or waking too early ≥3 times/week?” (Yes/No). Perceived insufficient sleep was defined as a frequent self-reported experience of unrefreshing sleep or persistent fatigue upon awakening during the preconception period, irrespective of sleep duration. Participants were asked: “During your preconception period, did you wake up feeling unrefreshed or experience persistent fatigue despite adequate time in bed on ≥3 days per week?” (Yes/No).

The primary outcome was TTP.

1. (1)

   TTP for pregnant couples = (date of last menstrual before pregnancy – date of last menstrual at examination)/30 + 1;

2. (2)

   TTP for unpregnant couples = (date of last menstrual at follow-up – date of last menstrual at examination)/30.

In accordance with previous studies on TTP calculation methods [18, 19], if conception is confirmed and occurs within the following cycle, one cycle is added to the TTP to account for the actual time of conception. Periods during which couples paused pregnancy attempts were subtracted from the total. All reported pregnancies were confirmed via clinical tests at the hospital.

Covariates

Covariate selection was based on the Health Behavior Theory (HBT) [20], which focuses on the influence of individual behaviors on health outcomes. HBT is grounded in the understanding that lifestyle factors, such as diet, physical activity, sleep patterns and smoking, play a crucial role in health outcomes, including reproductive health. Guided by previous research [8, 21], we identified potential variables that may influence the study outcome. We then applied the Least Absolute Shrinkage and Selection Operator (LASSO) regression method for variable selection to reduce multicollinearity, thereby enhancing the robustness and predictive performance of the model. The final variables included in the model were BMI, age, occupation, education, tobacco exposure (no, yes), frequent consumption of takeaway food (no, yes), duration of electronic device use, regular menstruation (no, yes), and exercise frequency. The ages of the couples were recorded either at the time of examination or when they began preparing for pregnancy following the examination, and were categorized into four groups:”≤25 years,””26-29 years,””30-34 years,”and”≥35 years.”This classification was based on both research regarding fertility and age, and actual reproductive policies and trends. BMI was calculated as weight (kg) divided by the square of height (m). Height and weight measurements were taken by trained medical professionals using a smart, interconnected height and weight measurement device. Participants were required to stand barefoot in a neutral posture, without shoes or outerwear, and weight data was recorded once it stabilized. According to the guidelines of the Chinese Obesity Working Group (WGOC), the BMI thresholds were defined as follows: underweight < 18.5 kg/m², normal weight 18.5–23.9 kg/m², overweight 24–27.9 kg/m², and obesity ≥28 kg/m². For analysis, BMI was categorized into three groups: underweight, normal weight, and overweight/obese. Occupation was classified into three categories:”Manual”,”Office”, or”Others”. Education referred to the highest level of education attained, including high school or vocational school and below, college, bachelor’s degree, and postgraduate education. Tobacco exposure was defined as active smoking or passive exposure for an average of five minutes or more per day. Frequent takeaway consumption was defined as eating takeaway at least once a day. The duration of electronic device use was treated as a continuous variable, representing the average daily hours spent using mobile phones, tablets, computers, or watching television during the pre-pregnancy period. Regular menstruation status was determined based on the doctor’s inquiry and judgment during the examination. Exercise frequency was categorized by the number of moderate-intensity physical activity sessions (lasting over 30 minutes) per week, with categories: “<1 time per week,” “1–3 times per week,” and “>3 times per week.

Statistical analysis

All statistical analyses were performed using R software (version 4.0.0). Group differences were assessed using the χ² test for categorical variables, the Wilcoxon rank-sum test for continuous variables with a non-normal distribution, and one-way analysis of variance (ANOVA) for continuous variables with a normal distribution. Normally distributed continuous variables were presented as mean ± standard deviation (SD), non-normally distributed continuous variables were described using the median and interquartile range (IQR), while categorical variables were described as frequencies and percentages.

The Cox proportional hazards regression model was employed to examine the association between various sleep behaviors and TTP, estimating fecundability ratios (FRs) and corresponding 95% confidence intervals (95% CIs). An FR > 1 indicated a shorter TTP and higher fertility, whereas an FR < 1 suggested prolonged TTP and reduced fertility.

During covariate selection, we first identified potential factors influencing TTP based on previous studies and theoretical frameworks. These variables included age, BMI, occupation, education level, tobacco exposure, and alcohol consumption for both partners. For women specifically, these included primiparity, frequent consumption of takeaway food, duration of electronic device use, coffee consumption, menstrual regularity, and exercise frequency. LASSO-Cox regression was then applied, incorporating an L1 penalty into the Cox proportional hazards model to shrink certain regression coefficients to zero, thereby automatically selecting the most important variables.

For each sleep behavior subgroup, the LASSO model was fitted using the following steps: First, a survival object was constructed, with TTP as the survival time and pregnancy success as the survival outcome. Categorical variables were then processed, and a predictor matrix for the Cox regression model was built. We used the glmnet() function to fit the LASSO-Cox regression model across a range of regularization parameters (λ), and the coefficient path was visualized. The optimal λ value was determined using 10-fold cross-validation, and the error curve was plotted. The final set of covariates included in the Cox regression model was composed of those with nonzero coefficients at the optimal λ value (lambda.min). The final covariates retained across most sleep behavior subgroups included BMI, age, occupation, and education level for both partners, as well as tobacco exposure, frequent consumption of takeaway food, duration of electronic device use, regular menstruation, and exercise frequency for women. Collinearity diagnostics showed that all VIF values for sleep variables and covariates were less than 5.

Regarding the missing data, the regular sleep group had 0.27% missing data for sleep onset time. These missing values were handled using multiple imputation by chained equations (MICE), with 50 imputations. The imputation model included the following covariates: BMI, age, occupation, and education level for both partners, as well as sleep onset time, sleep duration, insomnia, perceived insufficient sleep, tobacco exposure, frequent consumption of takeaway food, duration of electronic device use, regular menstruation, exercise frequency, and TTP for women.

The proportional hazards assumption of the Cox regression model was tested using Schoenfeld residuals. In the analysis, the variable”irregular sleep patterns”violated the proportional hazards assumption in the Cox regression model. Therefore, we applied an extended Cox proportional hazards model incorporating a time-dependent covariate to explore the time-varying effect of irregular sleep patterns on TTP. In this study, irregular sleep patterns were initially assessed as a binary categorical variable, with participants classified as having either a regular (coded as 0) or irregular (coded as 1) sleep-wake pattern based on self-reported data. To examine the potential time-varying effect of irregular routines on TTP, this variable was converted into numeric form and included as a time-dependent covariate in the Cox proportional hazards model. Specifically, to account for violation of the proportional hazards assumption, irregular sleep patterns were modeled using both a baseline main effect and a time-interaction term constructed via the tt() function in R, defined as a product of the variable and the natural logarithm of follow-up time (i.e., log(TTP + 1)). This allowed us to assess potential deviations from the proportional hazards assumption by capturing time-varying effects of irregular sleep patterns on time-to-pregnancy. After model fitting, the coefficients for both the main and interaction terms were extracted along with their variance-covariance matrix. These values were used to compute time-specific log hazard ratios and their standard errors across a 12-month interval. The log HRs were exponentiated to obtain corresponding hazard ratios and 95% confidence intervals, and the trend was visualized using the ggplot2 package in R.

In the regular sleep patterns group, we investigated the effects of sleep onset time, sleep duration, insomnia, and perceived insufficient sleep on TTP, while in the irregular sleep patterns group, we did not further analyze specific sleep behaviors due to the limited sample size (n = 178). In the regular sleep patterns group, three models were established: Model 1 without any adjustments, Model 2 adjusting for demographic characteristics of both partners, and Model 3 adjusting for all covariates. In the Cox regression model for the regular sleep group, Holm-Bonferroni correction was applied to adjust the p-values for sleep onset time, sleep duration, insomnia, and perceived insufficient sleep.

Restricted cubic spline (RCS) analysis was conducted to evaluate the nonlinear effects of different ranges of sleep duration and sleep onset time on TTP. The number of knots in the RCS model was selected based on Akaike Information Criterion (AIC). In our study, we compared the age and occupational characteristics between couples lost to follow-up and those included in the analysis, and found no statistically significant differences. Finally, sensitivity analyses were conducted to explore whether live birth impacted the findings. Statistical significance was defined as p < 0.05.

OpenAI on Thursday released the fifth generation of the artificial intelligence technology that powers ChatGPT, a product update that’s being closely watched as a measure of whether generative AI is advancing rapidly or hitting a plateau.

GPT-5 arrives more than two years after the March 2023 release of GPT-4, bookending a period of intense commercial investment, hype and worry over AI’s capabilities.

In anticipation, rival Anthropic released the latest version of its own chatbot, Claude, earlier in the week, part of a race with Google and other competitors in the U.S. and China to leapfrog each other on AI benchmarks. Meanwhile, longtime OpenAI partner Microsoft said it will incorporate GPT-5 into its own AI assistant, Copilot.

Expectations are high for the newest version of OpenAI’s flagship model because the San Francisco company has long positioned its technical advancements as a path toward artificial general intelligence, or AGI, a technology that is supposed to surpass humans at economically valuable work.

It is also trying to raise huge amounts of money to get there, in part to pay for the costly computer chips and data centers needed to build and run the technology.

OpenAI CEO Sam Altman described the new model as a “significant step along our path to AGI” but mostly focused on its usability to the 700 million people he says use ChatGPT each week.

“It’s like talking to an expert — a legitimate PhD-level expert in anything, any area you need, on demand,” Altman said at a launch event livestreamed Thursday.

It may take some time to see how people use the new model — now available, with usage limits, to anyone with a free ChatGPT account. The Thursday event focused heavily on ChatGPT’s use in coding, an area where Anthropic is seen as a leader, and featured a guest appearance by the CEO of coding software maker Cursor, an important Anthropic customer.

OpenAI’s presenters also spent time talking about safety improvements to make the chatbot “less deceptive” and stop it from producing harmful responses to “cleverly worded” prompts that could bypass its guardrails. The Associated Press reported Wednesday on a study that showed ChatGPT was providing dangerous information about drugs and self-harm to researchers posing as teenagers.

At a technical level, GPT-5 shows “modest but significant improvements” on the latest benchmarks, but when compared to GPT-4, it also looks very different and resets OpenAI’s flagship technology in a way that could set the stage for future innovations, said John Thickstun, an assistant professor of computer science at Cornell University.

“I’m not a believer that it’s the end of work and that AI is just going to solve all humanity’s problems for it, but I do think there’s still a lot of headroom for them, and other people in this space, to continue to improve the technology,” he said. “Not just capitalizing on the gains that have already been made.”

OpenAI started in 2015 as a nonprofit research laboratory to safely build AGI and has since incorporated a for-profit company with a valuation that has grown to $300 billion. The company has tried to change its structure since the nonprofit board ousted Altman in November 2023. He was reinstated days later.

It has not yet reported making a profit but has run into hurdles escaping its nonprofit roots, including scrutiny from the attorneys general in California and Delaware, who have oversight of nonprofits, and a lawsuit by Elon Musk, an early donor to and founder of OpenAI who now runs his own AI company.

Most recently, OpenAI has said it will turn its for-profit company into a public benefit corporation, which must balance the interests of shareholders and its mission.

OpenAI is the world’s third most valuable private company and a bellwether for the AI industry, with an “increasingly fragile moat” at the frontier of AI, according to banking giant JPMorgan Chase, which recently made a rare decision to cover the company despite it not being publicly traded.

The inability of a single AI developer to have a “sustained competitive edge” could increasingly force companies to compete on lowering the prices of their AI products, the bank said in a report last month.

——

The Associated Press and OpenAI have a licensing and technology agreement that allows OpenAI access to part of AP’s text archives.