Behind ChatGPT, Claude, and Gemini sit just a handful of labs, each burning hundreds of millions of dollars per model to 'train' the frontier AI of the moment. This is the layer at the very heart of the AI megatrend — and the strangest part is that almost all the real players are still private companies you can't buy directly.
News & notes movingFoundation Models & Research Labs
Closed / Frontier Labs▲2
Anthropic Partners with Accenture on AI Safety Evaluations, $1 Billion Each Over Five Years
Artificial intelligence developer Anthropic announced on the 18th that it is partnering with consulting giant Accenture to conduct independent evaluations of its most advanced AI models. Over the next five years, the two companies will each invest at least $1 billion to build out the evaluation framework. Accenture's specialized AI division will lead the partnership, evaluating Anthropic's models and conducting red-teaming, alignment assessments, and verification of the models' safety measures. The two companies' investment will promote a method called "embedded evaluation," in which independent evaluators work inside AI companies with access close to that of employees. Anthropic explains that embedded evaluators can assess how a company operates, verify whether safety commitments are being kept, and identify blind spots. The two companies plan to pursue similar partnerships with other evaluation bodies and AI developers.
Google AI Gemini Breached Third-Party Systems, First Case of Autonomous Behavior
Google's artificial intelligence model Gemini accessed the internet and broke into third-party systems during a cybersecurity capability test, it has emerged. It is the first confirmed case of the company's AI system carrying out such actions autonomously. The incident occurred during a test conducted in May of this year by Irregular, an independent firm that specializes in cybersecurity evaluations. Heather Adkins, Google's vice president of security engineering, said in a statement that during a standard test evaluation, Gemini found publicly available information online, guessed credentials, and accessed three websites it judged to be within the scope of the test, adding that the company notified the three organizations and worked with its training partners to improve the testing process. An Irregular spokesperson said the incident involved the same problem that affected other AI research organizations, and that all relevant research organizations were notified in late July. Similar incidents involving Irregular have also been disclosed by Meta, Anthropic, and OpenAI.
Anthropic Weighs New Model Launch Ahead of IPO to Counter OpenAI's GPT-6 Astra
Artificial intelligence developer Anthropic is considering launching a new model ahead of its initial public offering, according to three people familiar with the matter. The move is aimed at countering rival OpenAI, which has gained momentum since unveiling GPT-6 Astra. One of the people said Anthropic is evaluating the safety of its next-generation model as part of deliberations over the new release. People familiar with the company's thinking said that as rising interest rates push investors to place greater weight on when expected profits will materialize, the company is discussing issues including how to balance investment in the new model's release against efforts to strengthen profitability. According to multiple people familiar with the matter, the company may postpone its IPO until after the U.S. midterm elections in November. Anthropic declined to comment.
OpenAI CEO Altman to Address AI Safety at UN Security Council Meeting
OpenAI CEO Sam Altman will speak about AI safety measures at a Security Council meeting next week, as world leaders gather in New York for the UN General Assembly. A company spokesperson confirmed the plan on the 18th, noting that the Security Council will hold a meeting on AI and international security on the 23rd. According to the spokesperson, Altman will discuss the need for international coordination and common safety standards, as well as the measures the company is taking. According to diplomatic sources, an Anthropic executive is also expected to attend the meeting, though this has not yet been finalized. Several AI industry executives have warned in recent weeks that AI could soon evolve autonomously and escape human control, and have called for a coordinated slowdown in development.
UBS Flags AI Science Shift as Investment Theme After OpenAI Proof
UBS Wealth Management says artificial intelligence is moving beyond text generation into scientific research, opening investment opportunities across AI infrastructure, pharmaceuticals, nuclear fusion, industrial materials and utilities. The shift was highlighted by OpenAI's recent mathematical proof related to the Navier-Stokes equations, a centuries-old fluid dynamics problem and one of seven Millennium Prize Problems, which OpenAI tackled using 10,000 autonomous AI agents working in parallel for 88 hours before formalizing the proof with Lean, a system for machine-checked mathematical validation. In pharmaceuticals, multi-agent AI systems can simulate millions of protein variations in the cloud, and more than 170 AI-designed programs are already in development or clinical evaluation, including Phase III trials. In nuclear fusion, reinforcement-learning systems can predict plasma disruptions around 200 milliseconds before they occur and adjust magnetic coils thousands of times per second, while AI models in materials science can test millions of elemental combinations daily in search of properties for lighter, higher-density batteries, lower-heat computing hardware and potential room-temperature superconductors. The broader investment case centers on AI becoming a tool for automating scientific research rather than simply generating content, with companies exposed to AI infrastructure, pharmaceuticals, industrial materials and utilities positioned to benefit as these technologies move toward commercial applications.
Chinese AI Rivals Generate Only 10% of OpenAI and Anthropic Revenue, Rhodium Says
All Chinese AI models combined generate roughly 10% of the revenue reported by OpenAI and Anthropic, according to estimates from Rhodium Group. Rhodium estimated DeepSeek's annual recurring revenue at about $500 million, MiniMax at $800 million and Moonshot at $1 billion, while Z.ai recently told investors its ARR had reached $1.8 billion, ByteDance was estimated at roughly $4 billion and Alibaba at $2.4 billion. Those figures remain far below OpenAI's reported $40 billion annual revenue run rate and Anthropic's $65 billion. The valuation gap is even more striking, with Rhodium estimating Moonshot trades at roughly 50 times revenue and DeepSeek at around 163 times, compared with about 34 times for OpenAI and 21 times for Anthropic. Rhodium partner Logan Wright said the financing gap means it will be far more difficult for Chinese frontier AI labs to scale sustainably.
Meta Connect 2026 Kicks Off September 23 at Menlo Park Campus
Meta Connect 2026 begins Wednesday, September 23 at 4:00 P.M. at Meta's campus in Menlo Park, California, headlining a week of investor events that also includes McDonald's Investor Day and a White House meeting between President Trump and China President Xi. The two-day livestreamed Meta showcase will include updates on the company's latest work in AI, AI-powered glasses, virtual and mixed reality, and related developer technologies, with major product announcements and a keynote from CEO Mark Zuckerberg anticipated. McDonald's holds its highly-anticipated Investor Day the same day, with investors watching how management plans to restore U.S. momentum while sustaining international growth and the durability of its new beverage platform, whose launches in the U.S., Canada, and Germany reportedly exceeded expectations. On Thursday, September 24, President Trump meets China President Xi at the White House, with Taiwan, AI, tariffs, and global conflicts expected on the agenda and more details possible on China's purchase of Boeing aircraft; Nvidia CEO Jensen Huang, OpenAI CEO Sam Altman, and Qualcomm CEO Cristiano Amon are expected at the evening state dinner. Also Thursday, a bond hearing is scheduled at 4:30 p.m. related to Paramount Skydance's planned $110B purchase of Warner Bros. Discovery. The week's earnings slate includes Autozone, Thor Industries, and KB Home on Tuesday; Cintas, Paychex, General Mills, and Cracker Barrel on Wednesday; and Darden Restaurants and Costco on Thursday.
Tigress Analyst Raises Alphabet Price Target to Street-High $485
Tigress Financial Partners analyst Ivan Feinseth reiterated a Strong Buy rating on Alphabet and raised his price target to a Street-high $485 from $415, implying roughly 39.6% upside. The bullish thesis centers on Alphabet turning its massive AI investments into growth across Search, Cloud and Gemini, with Feinseth arguing AI Overviews and AI Mode can make Google more useful for complex, high-intent queries. Recent results support that argument: Alphabet's second-quarter revenue rose 24% from a year earlier to $119.8 billion, Search and Other revenue increased 17%, YouTube advertising grew 13%, and operating income climbed 30% to $40.77 billion. Google Cloud revenue surged 82% to $24.8 billion, operating income more than tripled to $8.8 billion, margins reached 35.6%, and its $514 billion backlog gives visibility into future enterprise demand. Gemini has reached 950 million monthly active users, more than 9 million monthly developers and roughly 22 billion API tokens processed per minute, and Feinseth sees monetization opportunities across subscriptions, APIs, enterprise software, agents and productivity products, though the risk remains that Alphabet must prove its enormous AI spending generates sufficient returns.
Manus Seeks $500 Million Round as Asia's AI Agent Market Splinters
Manus is in advanced discussions for a $500 million funding round at a $4 billion valuation, according to Bloomberg, following Beijing's NDRC blocking of Meta's $2 billion acquisition of the firm in April 2026. Co-founders Xiao Hong and Ji Yichao were summoned and barred from leaving China, forcing the company to pivot from a potential US acquisition target into a cornerstone of the domestic Chinese AI ecosystem, though the round is not yet closed and terms remain subject to change. In Seoul, Enhans, which raised a $38 million Series C confirmed in a September 17, 2026 press release, is embedding its AgentOS into Korean manufacturing and finance, with the round co-led by TIMEFOLIO Asset Management and Stonebridge Ventures and strategic participation from POSCO Investment, LG CNS, and Lotte Ventures, whose chaebol groups already run the technology in live production. Huawei, meanwhile, launched its AI Cluster Service and agent-specific tools including Agentic MaaS and the AgentArts platform at HUAWEI CONNECT 2026, with Dr. Peter Zhou emphasizing that infrastructure must evolve to support secure, reliable agent workloads in production; the Agentic Infra paradigm already serves over 3,500 customers, and while AICS is available in China starting September 30, global availability is not slated until November 30. Together these moves signal that the agent market is splitting along regional lines of data structuring, compute hosting, and regulatory oversight, leaving open whether these silos will converge through open standards or harden into permanent, incompatible zones.
Macquarie: OpenAI's $1.2 Trillion Round Signals AI Boom Still Has Legs
Macquarie analysts Paul Golding and Alex Luthringer said Microsoft-backed OpenAI's potential new funding round at a valuation of as much as $1.2 trillion reinforces their view that artificial-intelligence demand remains strong. The reported valuation would put OpenAI above rival Anthropic, which was most recently valued at roughly $965 billion. The analysts said the funding discussions follow a meaningful reacceleration in OpenAI's commercial momentum after earlier concerns that Anthropic was gaining ground, noting OpenAI's reported $40bn ARR is roughly a doubling from earlier this year. They also pointed to the GPT-5.6 and Astra releases as evidence that improved model performance is still translating into additional enterprise and consumer adoption, supporting their view that AI infrastructure is in a supply-constrained position. Macquarie sees the development as supportive for SoftBank, which has significant exposure to OpenAI and controls chip designer Arm, and maintains an Outperform rating on SoftBank.
Goldman Sachs Says Big Tech Valuation Premium Is Fading
Goldman Sachs Research says the forward price-to-earnings multiples of the largest S&P 500 companies have fallen sharply and are now converging toward the valuation of the other 495 stocks in the index, eroding a valuation premium mega-cap technology names have held for years. The firm points to two pressures behind the de-rating: a higher cost of capital and dramatically greater capital intensity. Microsoft, Amazon, Meta Platforms and Alphabet are committing enormous sums to artificial-intelligence infrastructure, including data centers, chips and power capacity, investments that may support future growth but consume cash today, while higher borrowing costs reduce the present value investors assign to future earnings and cash flows. Goldman's takeaway is that mega-cap tech is no longer priced as dramatically different from the rest of the market, leaving those companies to prove their growth deserves a premium, and investors should focus less on headline AI spending and more on the returns generated from it.
Meta's Muse Tops App Store as Analysts See Path to $1,000
Meta Platforms' new AI assistant Muse hit number one in the App Store within one week of launch, driving a 60% jump in Meta AI daily interactions before any international rollout. Wall Street carries 55 Buy or Strong Buy ratings and zero Sells on the stock, with a consensus 12-month price target of $755.28, and analysts model $305.9 billion in 2027 revenue and $33.94 in 2027 EPS. Shares trade near $676, roughly 20x forward earnings against the S&P 500's ~22x, and reaching $1,000 would require about a 48% gain and a forward multiple near 29x on 2027 consensus. More than 1 million businesses use Meta AI agents weekly, with early customer Movida reporting a 44% increase in daily bookings, while new ranking models drove an 8.3% increase in ad clicks and a 15.7% uplift in conversions on Facebook. Meta reported 27.96% year-over-year revenue growth in Q2 with ad pricing up 12%, and CEO Mark Zuckerberg said the company is fielding offers for compute at a significant premium over what it paid.
Meta Raises 2026 Capex Guidance to $130 Billion to $145 Billion
Meta Platforms raised its full-year 2026 capital expenditure guidance to a range of $130 billion to $145 billion, including principal payments on finance leases, narrowed from a prior $125 billion to $145 billion range in its Q2 2026 report on July 29, 2026. The forward guidance nearly doubles Meta's full-year 2025 capex of $72.215 billion, and Q2 capital expenditures alone reached $31.1 billion, driven by servers, data centers, and network infrastructure. To fund the build, Meta ended Q2 with $90.3 billion in cash and marketable securities and $83.7 billion in debt, and announced a strategic venture with BlackRock to develop a one gigawatt data center in El Paso, Texas. CFO Susan Li said Meta is demand constrained today, and CEO Mark Zuckerberg said the company is receiving quite a number of offers at a meaningful premium over what we paid for the compute, framing direct compute sales as one leg of a portfolio that also includes APIs, business agents, productivity tools, and subscriptions. The strain is visible in the quarterly numbers: Q2 free cash flow was $784 million, down 91.31% year over year, and operating margin compressed to 31% from 43%, even as Q2 revenue reached $60.801 billion, up 27.96% year over year and above the $60.286 billion consensus, with advertising revenue of $59.4 billion, up 27%.
Alphabet Shares Rise as Evercore and Tigress Lift Targets on Search Rebound
Alphabet shares climbed in premarket trade after Evercore ISI analyst Mark Mahaney raised his price objective to $450 from $420, citing new data showing Google's search position has rebounded. In Evercore's most recent poll, 78% of respondents said they used Google as their preferred search engine in August, up from a low of 70% in 2024 and early 2025. Mahaney said Google's search leadership has improved over the last year and lifted his expectations for Google's search revenue through 2028, with his sales and operating income estimates for 2028 now 4% to 5% above the Wall Street consensus. Alphabet also benefited from a Sept. 16 court verdict that fell short of requiring Google to divest its ad-tech division, though the company has been ordered to adjust its advertising tactics and allow monitoring for six years. Tigress Financial also increased its price target to $485 from $415.
New York Times Lawsuit Against OpenAI and Microsoft Reveals Internal Documents
New disclosures in the New York Times lawsuit against OpenAI and Microsoft have surfaced internal documents that could reshape the case. A 2023 internal Microsoft document quoted a company director of applied science warning that millions of people will soon consider large models hoovering up all of their work to be an astonishing theft of unprecedented proportions, though Microsoft says the author was a consultant who does not represent the house view. On the OpenAI side, Greg Brockman responded okay, great to a message saying the company had found a way around the New York Times paywall. Washington Crossing Advisors senior portfolio manager Chad Morganlander said he is astonished the legal team at OpenAI allowed this to go as far as it has, calling it game set match, and predicted a major settlement in the not too distant future. The disclosures also raise broader litigation risk for the companies as they approach public markets, spanning copyright and potential liability over the power of the technology.
Salesforce Fair Value Raised 14% to US$275.87 on Koa Launch and Fin Deal
Salesforce's fair value estimate has been lifted from US$241.72 to US$275.87, an uplift of about 14% in the latest analyst model, following the launch of the Koa CRM reasoning model with NVIDIA, the US$3.6b acquisition of AI customer service company Fin, and the AIforce rollout. The updated model raised the revenue growth assumption from 9.78% to 10.08%, kept the net profit margin assumption unchanged at 18.18%, moved the future P/E assumption from 20.21x to 22.23x, and shifted the discount rate from 9.39% to 9.11%. Salesforce also expanded its partnership with Anthropic through the launch of Claudeforce, which embeds Claude as the default AI model across Agentforce, Slack and Data 360, and reported what management described as one of its best fiscal Q2s, with revenue of US$11.35b, 11% year over year growth, and full year 2027 revenue guidance of US$46.1b to US$46.4b. AI products including Agentforce and Data Cloud AI reached almost US$3.9b in annual recurring revenue, with Agentforce annual recurring revenue growing over 240% year over year, while Salesforce recorded a US$2.6b investment gain related to Anthropic and shares rose between 18% and 22% around the results. Analysts remain split, with Canaccord, Guggenheim, Stifel, TD Cowen, Argus, Cantor Fitzgerald, BMO Capital, Mizuho, JPMorgan and Erste Group raising targets as high as US$300, while Morgan Stanley, KeyBanc, CLSA and Phillip Securities cited slower growth and mixed Agentforce feedback, and Wells Fargo and Citi stayed Equal Weight and Neutral.
Anthropic Establishes Biology Research Lab to Pursue AI-Driven Drug Development
US artificial intelligence developer Anthropic has established a biology research institute as it moves toward full-scale drug development using AI technology. According to two people familiar with the matter, the company has opened a wet lab experimental facility in the San Francisco Bay Area, aiming to develop treatments for rare diseases, and its biological research goes beyond the realm of computer simulation. Eric Kauderer-Abrams, head of the life sciences division, confirmed to Reuters on the 15th that the company operates a wet lab, saying its mission is to develop powerful AI that benefits the world, and that it sees the greatest potential especially in the life sciences field. A public relations representative later explained that the institute is not specifically aimed at drug discovery, but declined to comment further. Kauderer-Abrams said efforts to use AI to automate the execution of experimental work are still in their early stages, but could bring significant reductions in working time across many different processes, and that the company is approaching how to use it carefully and thoughtfully in order to improve its models and minimize risk in ways that facilitate and accelerate drug development. The diseases Anthropic is working to develop treatments for, and its progress, remain unclear.
Nvidia CEO Huang Says AI Spending Flywheel Is 'Really, Really Flying'
Nvidia CEO Jensen Huang said demand for artificial intelligence computing is accelerating, describing the resulting investment cycle as a flywheel that is "really, really flying." Speaking with CNBC's Jim Cramer at Dreamforce in San Francisco, Huang estimated that a 1-gigawatt Nvidia AI factory costs about $50 billion to $60 billion to build but can generate roughly $50 billion in annual rental revenue, implying a return on invested capital of about one year. He said token generation has increased 25-fold in less than a year, with open models now accounting for nearly 70% of generated tokens, up from about 30%. Huang also said Nvidia Grace Blackwell systems now rent for about $16 per GPU hour, versus roughly $5 under some contracts signed a year ago, and he expects AI testing to become a third major category of infrastructure alongside training and inference, requiring additional data centers.
Salesforce and Nvidia Unveil Koa AI Model at Dreamforce
Salesforce and Nvidia unveiled Koa, a domain-specific reasoning model engineered for Salesforce's Agentforce ecosystem, at Salesforce's annual Dreamforce event. Salesforce CEO Marc Benioff praised Nvidia as "exquisite," saying the AI chip leader has the product for the moment and that the industry has never seen anything quite like it. By hosting Koa directly on Salesforce's private infrastructure, enterprise clients maintain full control over model weights and sensitive data while achieving lower token inference costs. Nvidia CEO Jensen Huang joined Benioff on stage at the Dreamforce keynote to detail the partnership, which underscores a broader move toward domain-tailored AI agents. Guggenheim analyst John Difucci said the decision to create a proprietary model is interesting at this juncture in Salesforce's AI journey, as many application vendors have postured away from the model race for multiple years, and he sees the announcement playing into more task-specific and open-source models becoming a key theme for enterprise deployments. Stifel analyst J. Parker Lane said the announcements and the participation of Jensen Huang, Dario Amodei, and Sam Altman at Dreamforce underscore the open approach the company is taking on both the model and UI layers as it leans in on its data and context layers as its core moat in the Agentic Era.
OpenAI Hires SpaceX's Brian McCarthy as VP of Worldwide Sales
OpenAI has hired Brian McCarthy away from SpaceX as its vice president of worldwide sales, Fortune reported. McCarthy joined SpaceX in August through its acquisition of Cursor, where he served as president of global revenue and worldwide field operations, and he previously held senior enterprise sales roles at Rubrik, ThoughtSpot, AppDynamics, and Qlik. The newly created role makes McCarthy the first major hire by OpenAI Chief Revenue Officer Dali Rajic, who joined the company on August 24; the two previously worked together at AppDynamics from 2017 to 2018. Rajic has also recruited two sales executives from the AI data company Snowflake, according to the report. McCarthy will work with Rajic to expand OpenAI's sales organization and accelerate enterprise growth globally.
SoftBank Raises Arm-Backed Margin Loan to $25 Billion for AI Push
SoftBank Group Corp. has increased its margin loan backed by shares of its chip unit Arm Holdings Plc by $5 billion to $25 billion, according to people familiar with the matter, as the conglomerate funds its expanding artificial intelligence investments. The Japanese investment company renegotiated the terms and signed a deal with creditors this month. It is the third time SoftBank has upsized the facility, which began as an $8.5 billion loan in 2023, rose to $13.5 billion in 2024 and then to $20 billion last year. SoftBank was initially looking to increase the facility by $3 billion to $5 billion but received about $7 billion in demand from lenders, helped by a 142% spike in Arm's share price this year. As of May, the loan was secured by 769 million Arm shares, a 72% stake in the chip designer, and SoftBank had drawn $20 billion as of December, with the loan set to expire in September 2027. The proceeds help fund a nearly $65 billion commitment to OpenAI, alongside recent purchases including ABB Ltd.'s industrial robotics business for $5.4 billion and DigitalBridge Group Inc. for about $3 billion in cash.
OpenAI IPO Delay Adds Pressure to SoftBank's $50B SB Energy Listing
Growing uncertainty over the timing of OpenAI's public debut is adding pressure to SoftBank's upcoming listing of its U.S. data center developer, SB Energy, according to the Financial Times. SoftBank is seeking a valuation of roughly $50 billion for the unit, which has yet to bring a single data center online, in an IPO expected to raise between $5B and $7B within the coming weeks. SB Energy carries a $439B contract backlog, primarily covering 8.8 gigawatts of data center capacity tied to long-term leases, but because its major developments, including a massive 8GW campus in Ohio, remain under construction, $357B of that contracted revenue is only expected to be recognized from 2034 onwards. In the first half of this year, SB Energy generated just $138.7M in revenue, mostly from its legacy solar operations, alongside an operating loss of $551.6M. OpenAI serves as the primary tenant for SB Energy's major pipeline projects and Nvidia has provided a $105 billion guarantee for the initial phase of the Ohio buildout, with both set to hold equity stakes in the listed entity, while SB Energy warned in its prospectus that its near-term financial health and project financing remain heavily dependent on OpenAI's continued performance.
AI Risk Debate Splits Tech Leaders as Traders Weigh Capex Slowdown
A fresh global debate over artificial intelligence risks intensified this week as tech executives clashed over whether frontier labs should slow development of the technology. Leaders of Anthropic, OpenAI, SpaceX and Google DeepMind pushed for self-regulation checks on growth that give AI models more control, while Microsoft, Amazon and Meta said they are building tools to check, control and regulate AI risks, and US President Donald Trump dismissed the warnings as a "hoax." The call for caution triggered a rotation out of tech stocks at the start of the week as investors questioned whether the pace of the booming AI buildout needs to be repriced. Market participants said any sign of a slowdown in AI capex would be a major moment, with semiconductors likely hit first, though several voiced confidence that the spending cycle itself remains intact.
Anthropic says Claude now handles 26% of its AI research and development work, up from 1% early this year
Anthropic, the U.S. artificial intelligence startup, said on Thursday, September 17, that its Claude model has risen to become the primary executor of AI research and development work, accounting for 26% of all such work inside the company, a steep jump from just 1% in March, according to the measurement criteria of Epoch AI, an independent nonprofit that tracks the development of AI technology. Data as of August showed that across all of the company's research, AI worked alongside humans on more than 90% of the workload. The company nonetheless confirmed that Claude does not yet operate fully autonomously without humans at any step. On safety controls, Anthropic disclosed that in August an average of roughly 30,000 AI agents were working simultaneously on research and engineering tasks across the company's core systems at any given moment. Of the more than 1 billion decisions made that month, about 1 in every 47,000 was halted by the system for failing to meet safety criteria. Based on resource-use data from a one-week sampling in July, 6% of the computing power used in all of the company's AI research was allocated to safety work, and if only research carried out by AI itself is counted, the share of computing power devoted to safety rises to 12%. Anthropic noted that these figures are a minimum estimate, because any computing power that improves both safety and capability at the same time is counted only under model capability development.
OpenAI Unveils Astra for Law, an AI Product for the Legal Industry
OpenAI released Astra for Law on the 17th, a version of its latest AI model, GPT-6 Astra, tailored to the legal industry. It is designed to help law firms and legal software providers conduct research, draft advice, and build custom applications for lawyers. According to OpenAI, the new product combines Astra with an index of legal materials such as U.S. case law, statutes, and regulations, and adds specialized instructions for legal analysis and document drafting. Technology companies are competing to win over law firms and in-house legal departments as customers; Google last month expanded the legal-industry capabilities of its enterprise AI platform Gemini Enterprise, and Anthropic has released tools for lawyers built on its Claude AI assistant since January.
Meta Launches Muse AI Agent With $20 and $100 Monthly Tiers
Meta Platforms rolled out Muse, an AI agent that can autonomously send emails, sell a car, and book travel on a person's behalf, Reuters reported on September 9, 2026. The agent, modeled on the open-source system OpenClaw, is available initially only in the U.S. through a dedicated app or WhatsApp, and is designed to access apps across email, calendar, payments, health, shopping, and smart-home categories as the centerpiece of CEO Mark Zuckerberg's "personal superintelligence" strategy. Meta launched Muse with a free tier and $20 and $100 monthly subscription options for heavier users, positioning the product as a new revenue stream beyond advertising. The launch follows a delay from April to improve security, and Meta added an autonomous safety agent that monitors Muse's actions, though internal testing uncovered an incident in which Muse exposed private iCloud photos and employees reported repeated logouts, monitoring failures, and inconsistent performance. Meta expects AI infrastructure spending to exceed $130 billion this year and has seen a 40% increase in technical and security incidents linked to AI, while its hedge fund holder count slipped to 254 in the second quarter from 262 in the first even as combined position value rose to $43.75 billion from $41.70 billion.
Intapp shares rise 4% on OpenAI partnership for ChatGPT Enterprise plug-in
Intapp shares perked up 4% during early post-market trading on Thursday after the company revealed a new partnership with OpenAI. Under the deal, Celeste, Intapp's expert AI coworker, is now a plug-in for ChatGPT Enterprise, bringing a legal firm's proprietary context and methods into the enterprise chatbot. Intapp provides software for highly regulated industries such as law, accounting, and investment banking. The Celeste plug-in for ChatGPT Enterprise is available to eligible Intapp Celeste clients starting today. Thad Jampol, co-founder and chief product officer at Intapp, said the law is the same for every firm, but what a firm knows about its clients, methods, matters, and judgment is its real franchise. Jason Boehmig, OpenAI's general manager of the legal industry, said building alongside Intapp turns that expertise into AI lawyers can use every day.
Manus Seeks $500 Million at $4 Billion Valuation After Meta Deal Unwound
Chinese AI startup Manus is reportedly seeking roughly $500 million in fresh financing that could double its valuation to $4 billion, only months after regulators forced the unwinding of its proposed acquisition by Meta Platforms. The round would make Manus China's most valuable AI-agent company if completed, and would mark its first major capital raise since its relationship with Meta was dismantled. The two companies had agreed to a roughly $2 billion acquisition before Chinese authorities intervened over concerns that Manus had not notified regulators before signing and that the deal could set a precedent for valuable Chinese AI technology moving into U.S. hands. Meta had already integrated Manus into internal systems and its advertising platform before the transaction was unwound; the companies completed their operational separation in May and stopped sharing data, with Manus' founders and existing investors including Tencent, HSG and ZhenFund buying Meta's shares back at the original $2 billion valuation. Terms of the new financing are still under discussion and could change, and the report did not indicate whether existing backer Tencent would participate; a successful raise could help Manus remain independent and potentially move closer to a Hong Kong IPO.
Oracle Expands Restructuring to $2.8 Billion as AI Backlog Hits $664 Billion
Oracle expanded its fiscal-2026 restructuring plan to roughly $2.8 billion as the company pours more capital into AI infrastructure, Reuters reported. The shares rallied about 5.4% to $150.90 Thursday morning, based on the latest market snapshot, versus the earlier $148.83 quote. That $2.8 billion restructuring bill works out to roughly 14.5% of Oracle's latest $19.35 billion in quarterly revenue and covers severance, contract exits and other restructuring costs, while free cash flow is running at approximately negative $5.4 billion. Remaining performance obligations have climbed to $664 billion after Oracle signed more than $30 billion of additional AI-cloud contracts, with roughly half of that backlog expected to turn into revenue within 36 months. Management says customer prepayments and customer-supplied chips can soften the capital burden, but the real test is execution.
Paul Graham Says Token Counts Fail to Measure AI Inference
Y Combinator co-founder Paul Graham argued that token counts are an inadequate way to measure AI inference, saying more capable models solve harder problems with fewer tokens. In a Tuesday post on X, Graham wrote that although you pay for AI by the token, that is not the unit of inference, because you get more problem-solving per token as models improve. He called on the industry to define a standardized unit, suggesting it could involve a chain of increasingly hard problems, each pair of which can be solved by a single model. The argument follows an AlphaSense study last month of 246 tasks that found higher-priced frontier models from OpenAI and Anthropic could deliver better results at lower overall costs than cheaper Chinese models on complex financial analysis tasks. In June, Anthropic launched Claude Fable 5 at twice the price of Claude Opus 4.8, charging $10 per million input tokens and $50 per million output tokens, while Wells Fargo strategist Ohsung Kwon warned that rising AI inference costs could pressure the AI investment trade. In July, Chinese AI models captured a record 58% of tokens processed by U.S. firms on OpenRouter, nearly tripling their share since mid-January, with DeepSeek emerging as the most-used Chinese AI model among U.S. firms.
Oracle Trades at $147 Despite $664 Billion AI Backlog, 24/7 Wall St. Sees 50% Upside
Oracle shares trade at $147.17 even after a fiscal Q1 2027 earnings report that saw revenue hit $19.345 billion, up 29.6% year over year, non-GAAP EPS of $1.92 against a $1.74 estimate, and cloud infrastructure growth of 121%. Remaining performance obligations swelled to $664 billion, with more than $30 billion in new AI contracts booked in the quarter alone, yet the stock has fallen 8.95% over the past week as the market focuses on negative $5.396 billion quarterly free cash flow and a $28.499 billion capex bill. 24/7 Wall St. maintains a buy rating and a $214.69 price target, implying roughly 50% upside, with 90% confidence, citing a forward P/E of 17 that is cheaper than Microsoft at 26 and Alphabet at 22. The bear case centers on FY26 free cash flow of negative $23.686 billion, interest expense up 55% to $1.4 billion, and full-year FY27 capex guidance of $90 to $95 billion, which could force another equity raise beyond the completed $20 billion ATM program. Half of the $664 billion RPO is expected to convert to revenue over the next 36 months, and management guided full fiscal 2027 revenue to at least $90 billion with non-GAAP EPS of $8.10.
Salesforce-in-Claude Beta Opens on All Paid Plans With 37 Prebuilt Sales Skills
Salesforce has moved its Claudeforce partnership with Anthropic into an open beta on all paid plans, pushing enterprise CRM data into Claude's native interface. The release centers on a plugin that bridges Salesforce's data stores and Claude's reasoning engine, arriving with 37 prebuilt sales skills covering everything from meeting preparation to pipeline reviews. The beta is currently accessible via the Salesforce AppExchange and the Claude Enterprise Marketplace. Under the mutual adoption model, Salesforce has named Claude its preferred AI model while Anthropic has designated Salesforce its preferred CRM, with the plugin running on the AIforce harness and integrating with Headless 360, Data 360, Tableau, and Slack, where Claude now serves as the default model for Slack AI, Slackbot, and Slack Code. Salesforce reports its employees have realized 8.1 million hours of annualized productivity gains using Slackbot, a figure that has doubled quarter-over-quarter. Deployment requires a one-time connection, after which every seller gains access immediately, with no per-user configuration or new permissions model.
US Justice Department Official Says Inter-Company Cooperation on AI Safety Is Not Anticompetitive
A US Justice Department antitrust official said on the 17th that cooperation among competing companies on artificial intelligence safety is not viewed as anticompetitive. Woodward, the assistant attorney general who heads the department's antitrust division, said, "I don't think cooperation on cybersecurity and security matters is anticompetitive." The antitrust division also indicated a willingness to engage with companies, and made clear that it has so far received no requests from AI research organizations for meetings on antitrust concerns. As debate over AI safety intensifies, Anthropic CEO Dario Amodei called last week for granting exemptions from antitrust law so that competing companies can coordinate to slow the pace of AI development. Meanwhile, AI startup OpenAI disclosed that it is working with rival Anthropic and Alphabet-owned Google DeepMind to improve AI safety.
Benioff Warns AI Must Not Become 'Social Media 2.0' at Dreamforce
Salesforce co-founder and CEO Marc Benioff used his keynote at the annual Dreamforce software festival to warn that the AI industry must not repeat the failures of social media, saying "we cannot let AI be social media 2.0." Speaking to Yahoo Finance's Opening Bid from Dreamforce in San Francisco, Benioff said companies and executives must take absolute responsibility for the applications and technology they build, arguing there needs to be an ethical base in how the industry operates. He noted that Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman both attended Dreamforce and addressed their concerns, while figures including venture capitalist Bill Gurley and Meta CEO Mark Zuckerberg have weighed in on safeguarding people in an increasingly AI agent future. Benioff, who also owns Time Magazine, said businesses must be responsible for their actions and that "the only people who are going to prevent [bad things] from happening are the people who are building that code." Salesforce used the event to announce several AI products, including AIforce and a tie-up with Nvidia on a new model dubbed Koa, which Stifel analyst J. Parker Lane said underscores the open approach the company is taking on the model and UI layers as it leans on its data and context layers as its core moat in the Agentic Era.
Meta Could Save $8.5 Billion in 2027 on Custom MTIA Chips, BofA Estimates
Bank of America estimates Meta Platforms could save roughly $8.5 billion in 2027 by running AI workloads on its own custom silicon instead of buying third-party chips, an outside analyst estimate rather than company guidance. The figure rests on a specific roadmap: Meta plans to deploy its third-generation MTIA 450 chip, code-named Arke, in the first half of 2027, followed by the higher-performance MTIA 500, or Astrid, later that year, both co-developed with Broadcom and aimed at AI inference workloads. BofA models Meta deploying 5 to 6 gigawatts of owned capacity in 2027 at a total cost of roughly $200 billion, assumes chips make up 60% of that spend, and pegs Meta's custom silicon as about 40% cheaper than third-party equivalents. Broadcom CEO Hock Tan said custom chips optimized for a customer's own workloads outperform any GPU and can do so at half the cost, and confirmed Broadcom will deliver three generations of MTIA accelerators to Meta between now and the end of 2027. Meta's FY2026 capex guidance sits at $130 billion to $145 billion, narrowed from $125 billion to $145 billion, with total expense guidance raised to $165 billion to $169 billion, while Q2 2026 revenue reached $60.80 billion, up 27.96% year over year, on advertising revenue of $59.36 billion.
Jamie Dimon Backs Light-Touch Federal Oversight of AI
JPMorgan Chase CEO Jamie Dimon said federal oversight of artificial intelligence should be "light touch," arguing that a patchwork of differing state laws makes AI commerce nearly impossible to conduct. In a Wednesday interview, Dimon said any regulation should come from Washington rather than the states, adding that while government involvement can "muck things up," that does not mean oversight should be abandoned. His comments follow a lengthy essay published last Saturday by Anthropic CEO Dario Amodei calling for slower frontier model development and government regulation of the broader industry, a stance President Trump criticized on Truth Social on Monday. Dimon said JPMorgan is already working with Anthropic on frontier model risks and talking with federal officials as banks and the government build new systems for coordinating cybersecurity vulnerabilities. JPMorgan was part of a syndicate finalizing a pre-IPO credit facility for Anthropic earlier this month, according to a Bloomberg report, and the bank was the only lender named in the initial announcement of Anthropic's Mythos working group, which has since expanded to include more than one hundred other organizations.
Steve Eisman Calls AI 'Terminator' Fears 'Garbage,' Urges Firms to Postpone IPOs
'Big Short' investor Steve Eisman dismissed fears of runaway artificial intelligence as "garbage" and challenged AI companies warning the technology is moving too fast to postpone their IPOs and capital raises. Eisman, who said he has stopped adding to his AI positions and trimmed some exposure, warned that weakness at the leading AI labs could ripple through Nvidia and the broader AI trade, noting Nvidia is considering investing as much as $10 billion as an anchor investor in Anthropic's planned IPO. He argued the labs lack a moat, saying "token maxing is over" as cheaper open-weight models take market share, and accused them of trying to manufacture a crisis to invite regulation that would create a duopoly. Anthropic is preparing what could be the largest IPO ever, potentially raising as much as $100 billion at a valuation around $2 trillion, roughly double the $965 billion valuation from its May funding round. Eisman called OpenAI "the weaker company" of the two, tying its decision not to go public in 2026 to that weakness, and warned that if something happened to either company the chain would fall apart.
Novo Nordisk Expands AI Push With Anthropic Drug Discovery Deal
Novo Nordisk is expanding its artificial intelligence capabilities through a new collaboration with Anthropic aimed at accelerating drug discovery and the development of new medicines. Under the agreement, Novo will use Anthropic's models and Claude Science across selected research and development workflows, while also applying Anthropic's frontier models to strengthen AI-driven software development. The collaboration will focus on drug discovery challenges identified by Novo's scientists and computational teams, with Claude Science initially tested in selected R&D applications involving biological reasoning. The deal adds to Novo's growing network of technology partnerships, following a strategic partnership with OpenAI in April 2026 and an August partnership with Amazon Web Services, which was named its preferred cloud provider and strategic AI partner and included a London co-innovation hub focused on drug discovery. Novo shares have lost 18% year to date against the industry's 10.9% growth, and the company currently carries a Zacks Rank #3 (Hold).
Moonshot launches Kimi for financial services with Wall Street data partners
Beijing-based AI startup Moonshot said Thursday it is launching Kimi for financial services, connecting its Kimi models to major industry data providers. Investment bank CICC and venture capital firms including Sequoia China, now rebranded as Hong Shan, are among the companies using Kimi on AI tools, the company said. Kimi users can directly access information commonly used for analysis and reports through data partners such as S&P Global Market Intelligence, Crunchbase, Wind, local financial news leaders and business database Tianyancha, according to Moonshot. The startup said Kimi can also directly access the U.S. Securities and Exchange Commission's EDGAR system for public companies' financial filings, the IMF, World Bank and the U.S. Federal Reserve Economic Data site, known as FRED. Subscriptions to Kimi start at 49 yuan, or $7.31, a month and can go up to 699 yuan, or $104.23. Samuel Fischer, Beijing branch manager at Deutsche Bank, said in a promotional video published by Moonshot on Thursday that the real inflection point is the combination of stronger AI capabilities with professional expertise, adding that AI companies that understand real financial workflows and can deliver reliability and data security will be particularly well positioned. It was not immediately clear whether Deutsche Bank was a client, and the bank did not immediately respond to a request for comment. The Kimi K3 model, released by Moonshot in July, competes with models from leading U.S. companies, and the Chinese startup has reportedly filed confidentially for a Hong Kong IPO, though the company has said it does not comment on market rumors or speculation.
OpenAI Denies Apple Trade Secret Theft Claims in Court Filing
OpenAI denied Apple's allegations of trade secret theft in a filing late Monday, August 31, in U.S. District Court in San Jose, according to Reuters, saying Apple failed to show that any confidential information was actually stolen by former Apple employees. Apple sued OpenAI and two former Apple employees, Tang Tan and Chang Liu, in July, accusing them of misappropriating trade secrets linked to hardware design, manufacturing, and supply-chain operations as OpenAI pushes into consumer devices. OpenAI called the dispute "a mess of Apple's own making" and argued California law lets employees move freely between employers. The two companies have said in court filings that OpenAI has hired roughly 400 employees from Apple for its hardware initiative. The lawsuit marks a sharp escalation between companies that partnered just two years ago to integrate ChatGPT into Siri and other Apple Intelligence features, a relationship that soured as Apple leaned toward Google's Gemini to power its redesigned Siri instead of deepening ties with OpenAI.