Back to home
This issue or one story

What would you like to share?

Share the full issue or choose a story worth reading.

AI Daily Insights16 stories in this issue

Jensen Huang offers to return as Nvidia CEO on X. The board chair rules it out about six hours later

01 TALENT

❯ Jensen Huang offers to return as Nvidia CEO on X. The board chair rules it out about six hours later

Two statements, five days after the resignationOn the morning of November 17 Pacific time, Nvidia co-founder Jensen Huang issued his first public statement since resigning. Writing on X, he said he was willing to serve again if the board agreed. About six hours later, Nvidia chair Harvey Jones opened the company’s fiscal third-quarter call by saying the board had unanimously accepted Huang’s resignation and did not intend to reinstate him.

Huang apologizes and offers to return in the same statementHuang called Nvidia his life’s work, acknowledged that his personal conduct had fallen short of company standards and committed to cooperating with the board’s independent investigation. He said employees, investors, partners and customers had directly urged him to return. He attached a condition to his offer: if the board concluded that his return served shareholders, employees, Nvidia and the United States, he would be honored to answer those calls.

Calls for his return had already gone publicHuang led Nvidia for 33 years. WIRED reported that some employees publicly called for his return and that Elon Musk, Replit CEO Amjad Masad and Y Combinator CEO Garry Tan backed a petition. Huang’s post put the founder on record as willing to return after others had publicly made the case for him.

The chair restates the circumstances of the departureOn the earnings call, Jones went back to November 12. Nvidia said that day that Huang had voluntarily resigned as president, chief executive and director, effective immediately, after disclosure of personal conduct inconsistent with its code, standards and culture. Jones reiterated that directors accepted the resignation unanimously and explicitly addressed the possibility of reinstatement.

Kress runs the company while the search continuesUnder the transition plan, Colette Kress is serving as interim CEO while retaining the CFO role. Jones said the board would consider both internal and external candidates, seek a timely appointment and conduct a thorough search. He highlighted the national-security issues the next chief will have to navigate in the chip industry. The call then moved to operations, with Kress fielding a question about retaining employees through the transition.

▮ SIGNALBy publicly ruling out its founder’s return, the board put the transition in the hands of Kress, who must keep the business and its people on track, and a permanent successor who will have to manage both the chip business and its external constraints.

02 MARKET

❯ Nvidia revenue doubles to $115.8 billion and its fourth-quarter forecast reaches $130 billion

Sales beat estimates by $6.7 billionNvidia released fiscal third-quarter results for the period ended October 25 on November 17. Revenue of $115.8 billion was up 103% from a year earlier and 20% from the prior quarter, beating the $109.1 billion consensus by $6.7 billion. GAAP net income reached $65.3 billion and gross margin was 74.8%.

The new racks are running at several cloud providersData center revenue reached $107.6 billion, up 110% year over year and 21% sequentially, accounting for about 93% of total sales. Nvidia ramped production and began shipping Vera Rubin NVL72 racks, which are already running at CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud and Nebius. The segment includes several generations of products, and initial Vera Rubin shipments mark a new delivery milestone.

Cloud services connect systems to paying usersGoogle Cloud has opened its Vera Rubin-based A5X instances, and Oracle Cloud has made NVL72 systems and Vera CPUs generally available. Model developers can obtain access through cloud providers. Anthropic, for example, has expanded its access to Grace Blackwell and Vera Rubin capacity through those providers. The rollout spans Nvidia’s rack deliveries, deployment by cloud operators and access for end customers.

The next quarter has an extra week and a lower margin forecastInterim CEO and CFO Colette Kress delivered her first results in the new role. Nvidia expects $130 billion in fourth-quarter sales, plus or minus 3%. The midpoint exceeds consensus and is about 12% above the third quarter. The next period has an extra week, and Nvidia expects gross margin between 71% and 72%, down from 74.8% in the latest quarter.

▮ SIGNALVera Rubin has moved from shipments to cloud access, while the higher sales forecast and lower margin outlook show why revenue and profit should be tracked separately through the rollout.

03 LEGAL

❯ The Wall Street Journal says the SEC is examining whether Nvidia’s disclosures misled investors

A civil inquiry into investor disclosuresOn November 17, The Wall Street Journal cited people familiar with a civil SEC investigation that includes whether Nvidia’s disclosures misled investors. The report came five days after Jensen Huang resigned. No findings have been made public.

Singapore billings once accounted for about one-fifth of quarterly salesNvidia previously broke out Singapore revenue by customer billing address. In the quarter ended July 2025, that measure came to roughly $10.2 billion, or about 22% of total sales. Nvidia has said some customers are billed through Singapore while products typically ship elsewhere. The company subsequently stopped reporting Singapore separately.

How to read a geographic revenue figureBilling addresses show where customers arrange purchases and payment. If investors use the same figure to infer where hardware is delivered or used, the distinction between billing and delivery changes the answer. Removing the line item also takes away a number that investors could compare directly with earlier quarters.

The independent board inquiry continues as wellIn the risk section of its earnings release the same day, Nvidia separately cited an ongoing investigation by an independent board committee and potential effects from government and regulatory inquiries. The board oversees its own review, while the securities regulator is examining disclosures to investors. Neither inquiry has published findings.

▮ SIGNALWhen one geographic line once accounted for more than one-fifth of revenue, billing criteria, delivery destinations and subsequent reporting all affect what investors can see about that business.

04 POLICY

❯ Nvidia creates a standing board committee with a direct reporting line for compliance teams

Standing oversight moves to board levelNvidia on November 17 established a National Security and Export Control Oversight Committee. It will provide continuing oversight of company policy, compliance and global trade risk. Colette Kress was already serving as interim CEO, and the board had started its search for a permanent chief.

Two existing teams gain a direct reporting lineThe national security team and global trade compliance team were already operating within Nvidia. Under the new arrangement, both report directly to the board committee and continue reporting to Kress on day-to-day operations. The people doing the work gain direct access to directors, while executives continue to handle daily decisions and execution.

Three directors will meet at least quarterlySuzanne Nora Johnson will chair the committee, with Stephen C. Neal and John O. Dabiri joining her. It will meet at least quarterly, both with management and in executive session. It will also oversee development of the two teams and work with independent experts in law, security and trade compliance.

External engagement is part of the mandateChair Harvey Jones said advanced silicon and AI raise new security questions and that Nvidia intends to work with government and industry partners. The committee gives directors a standing way to review how internal teams handle issues and stay in contact with authorities whose requirements affect the business.

▮ SIGNALNvidia’s new reporting line gives directors direct visibility into when a risk was identified, how it was escalated and who is responsible for dealing with it.

05 CAPITAL

❯ Barron’s says the Huangs moved more than $10.5 billion in Nvidia shares into a family trust in March

A March transfer draws attention in NovemberBarron’s reported in November, after reviewing US securities filings, that Jensen and Lori Huang transferred nearly 60 million Nvidia shares to an irrevocable family trust on March 18, 2026. The shares were worth more than $10.5 billion at the time. The transfer occurred eight months before the report, which followed Huang’s resignation.

The transfer stems from two trusts set up in 2016The couple initially placed roughly $51 million of Nvidia stock in the trusts a decade ago. The two annuity trusts matured on schedule in March, and the remaining shares passed to the family’s irrevocable remainder trust. The $10.5 billion figure measures the stock at the time it moved and reflects Nvidia’s appreciation over those years.

Annuities go back to the grantor and the remainder goes to beneficiariesA grantor-retained annuity trust, or GRAT, pays an annuity to the grantor over a fixed term. Assets left over pass to beneficiaries under the trust’s terms. Estate lawyers interviewed by Barron’s explained that, when the applicable requirements are met, qualifying appreciation can pass to heirs free of federal estate tax.

The shares appreciated while the trusts were activeBarron’s said the trusts were worth more than $1 billion by 2021 and doubled in value again two years later. Dow Jones Market Data put Nvidia’s share-price gain over the life of the arrangement at more than 220-fold. The $10.5 billion measures the shares when the trusts transferred them in March.

▮ SIGNALLong-term share appreciation and a scheduled trust maturity can combine to produce a large ownership change, making the dates central to how the filing is read.

06 MODEL

❯ OpenAI releases GPT‑6.2 Astra and Sol and cuts flagship API token prices by 20%

Two models arrive in work products and the APIOpenAI released GPT‑6.2 Astra and Sol on November 17. Astra targets multistep research, professional documents and engineering tasks, while Sol is designed for everyday work. Both reached ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and developers can access them through the API.

Complex deliverables require both source work and usable outputOpenAI gave the models a fictional company’s accounts and operating assumptions and asked for an editable Excel valuation through 2031 with checks that its three financial statements reconciled. Astra put segment forecasts, valuation and sensitivity analysis together so the assumptions were easier to trace. In OpenAI’s reported Agents’ Last Exam results for multistep work, it scored 5.2 points above GPT‑6 Astra at less than one-third of the cost per task.

Pricing and usage both affect the billAstra costs $8 per million input tokens, $0.40 for cached input and $40 for output. Standard input and output prices are 20% below GPT‑6 Astra. Sol stays at $2 per million input tokens and $10 for output. On long tasks that revisit the same files, cache reuse and the amount of intermediate output also affect the bill.

Faster processing reaches more users, with blockers still worth checkingPro 500 users can run both models in Ultrafast, while Pro 200 subscribers start with Sol. Eligible organizations need administrator enablement. OpenAI found that both models were more likely in some tests to omit blockers such as a missing file or unavailable search tool and added Codex instructions to reduce those omissions. On complex work, users still need to know which materials the deliverable covers.

▮ SIGNALAstra makes a premium model affordable for more complex assignments, but teams still have to measure the cost of a usable deliverable, including the bill, review and rework.

07 CAPITAL

❯ Anthropic may price above its $78 IPO range as advisers discuss a valuation of up to about $1.6 trillion

The roadshow is followed by discussion of a higher priceThe Wall Street Journal cited people familiar with the matter as saying Anthropic may price its IPO above $80. The marketed range for the Claude developer was $70 to $78, implying a valuation of around $1.33 trillion. Advisers have discussed roughly $1.5 trillion to $1.6 trillion. The offering has yet to set a final price.

The base deal includes sales by Anthropic and existing shareholdersThe November 10 preliminary prospectus lists 825 million shares in the base offering, around 801.4 million from Anthropic and 23.6 million from existing holders. At $78, the base sale would total about $64.4 billion, including roughly $62.5 billion in proceeds from new shares. Proceeds on existing shares go to the selling holders. These figures are before offering expenses.

The $26 billion placement follows the same priceAlongside the public deal, AMD, Broadcom, Google, Micron, Microsoft, Nvidia and a TSMC investment entity agreed to buy $26 billion of another share class at the final IPO price. Anthropic plans to list on Nasdaq as ANT. The prospectus also gives underwriters an option for up to 123.75 million additional shares.

A higher price would change the two legs differentlyPricing the public offering higher would raise more money on the same number of shares. The private placement is defined in dollars, so a higher price would buy fewer shares for the same commitments. A price above $80 remains under discussion after the roadshow, with listing and closing still ahead.

▮ SIGNALClaude’s market valuation will be anchored by public demand and strategic buyers paying the same price per share, with different implications for proceeds and ownership.

08 MODEL

❯ Microsoft launches MAI Code 2 at Ignite and brings its own coding models to GitHub Copilot and Foundry

Microsoft puts its own coding models in front of developersAt Ignite on November 17, Microsoft introduced MAI Code 2 and Code 2 Flash alongside models specialized for security, imaging and speech. Both coding models reached general availability that day in GitHub Copilot and are also available through the Foundry enterprise AI platform.

The two variants target different workloadsMicrosoft positions Code 2 for multifile implementation, changes requiring repository context and complex debugging. Flash targets explanations, narrow fixes and frequent small edits. Teams can select either model in Copilot or access them through Copilot Fusion. On Microsoft’s chosen coding benchmarks, it says Code 2 delivered about 98% of GPT‑6.1 Sol’s performance at around 40% of the cost.

In-house and third-party models share a platformFoundry is adding third-party models too, including GPT‑6.2. Microsoft also introduced MAI Licensed Weights for qualifying customers, giving some enterprises an option to deploy model weights in environments they control more closely. The models in Copilot and Foundry are already usable, while weight licensing is aimed at selected customers.

The security models plug into an existing productMicrosoft also described how its Cyber 2 family integrates with the Perception enterprise security product. The coding models go into Copilot and the security models connect to Perception. In both cases, Microsoft is matching specialized models to a defined workflow in an existing product.

▮ SIGNALMicrosoft’s leverage is in distribution and task routing: Copilot and Foundry can use its own models for frequent, cost-sensitive work while keeping outside flagships available.

09 CAPITAL

❯ Isomorphic Labs raises $5 billion in Series C funding to move its first AI-designed drug candidates toward trials

Two rounds in six months, with funding now reaching clinical developmentAI drug developer Isomorphic Labs announced a $5 billion Series C on November 17, co-led by Greenoaks and Sequoia. New investors include a16z. Alphabet invested again alongside existing backers including Thrive Capital, MGX, Temasek and the UK Sovereign AI Fund. The company raised $2.1 billion in Series B funding in May.

The platform supports both owned and partnered programsDemis Hassabis founded the company in 2021. Its internal IsoDDE platform supports its own discovery teams and joint programs. The company advances some programs itself and brings its design capabilities into partners’ development plans for others. Both routes require staged validation before human trials.

The company reports positive preclinical resultsIn its financing announcement, Isomorphic Labs cited positive preclinical results in internal oncology and immunology programs and in partnered programs. This is the work done before studying candidates in people. The company wants to advance its first programs into trials, where safety and potential benefit can be assessed.

The $5 billion funds both platform and development capabilitiesThe company will keep investing in its design platform while building infrastructure to take candidates toward clinical development. At this stage, it needs the organizational capacity to manage further development and studies. The release lists models, drug programs and clinical infrastructure as uses of the proceeds.

▮ SIGNALOne use of the funding is to bring candidates to the stage where human data can test their value, allowing Isomorphic Labs to judge its platform in part by how its programs progress.

10 PRODUCT

❯ Google DeepMind adds a robotics plugin to Antigravity that takes tasks from simulation to real robot arms

The new release is a robotics workbenchGoogle DeepMind launched the Antigravity robotics plugin on November 17. Gemini 4 Hydrogen had arrived the previous week for tasks sensitive to speed and cost. The plugin applies it to hardware setup, simulation and physical control in a single workspace.

The flow begins with hardware setupIn Google’s example, a developer connected an I2RT YAM arm and external depth camera, then had the plugin install software, find the camera, calibrate the arm and test a motion. Google says setup took about ten minutes in that example and demonstrated the flow with both a budget arm and an industrial dual-arm system. Real deployments require supported hardware and a suitable camera and workspace.

Conveyor sorting runs in simulation firstGoogle demonstrated the workflow on snack sorting. Given a whiteboard sketch and images, the model recreated a conveyor and arm in MuJoCo, tested grasping over multiple simulation runs and then transferred the behavior to physical hardware. The developer used conversation to adjust grip force and sorting rules on the real system. Google says the full demonstration took about an hour.

The plugin and model have different access routesGoogle AI Pro and Ultra subscribers can install the plugin in Antigravity. Developers who only need the model can use Hydrogen through the Gemini API. The two routes serve on-site development and software integration respectively, with the complete hardware workflow demonstrated through the Antigravity plugin.

▮ SIGNALLowering the cost of moving between simulation and hardware lets teams test more candidate motions and reserve time on the physical arm for real-world validation.

11 LEGAL

❯ California bars Waymo driverless taxis from freeways, and Waymo suspends freeway rides nationwide

The state order leads to a wider pauseCalifornia’s Department of Motor Vehicles barred Waymo’s driverless taxis from state freeways on November 17, citing an imminent safety hazard. Waymo then paused freeway rides in its other markets to evaluate performance. It operated commercially in 15 US cities at the time, and surface-street service continues.

Three cars entered an active construction closureThe DMV told TechCrunch that three Waymo vehicles entered an active closure on US 101 near Burlingame. A separate event on Interstate 280 in San Francisco involved highway patrol vehicles. Waymo reported no injuries or collisions on US 101. Regulators pointed to repeated incursions into work zones where people, vehicles and equipment were present.

Similar incidents recur after a software updateWaymo began freeway service in late 2025, suspended it following similar incidents in May, and resumed in July after updating its driving software. The DMV noted that the latest work-zone incursions came after earlier restrictions and corrective measures. Waymo is working with officials in California and elsewhere and has announced no restart date.

Airport and cross-city trips lose their faster routesFreeways connect San Francisco with operating areas along the Peninsula and toward Silicon Valley, shortening airport and intercity trips. Waymo says it still serves more than half a million rides weekly on surface streets. It can keep carrying passengers, but trips that relied on freeways will need different routes.

▮ SIGNALRestoring freeway service hinges on handling temporary detours and construction closures reliably, especially after similar incidents recurred following the earlier update.

12 PRODUCT

❯ macOS 27.3 beta lets users restrict AI assistants to selected apps or folders

The beta lets users choose the scope of accessOn November 16, Apple described permission changes in the second macOS 27.3 beta. When software seeks to control other apps or read files broadly, macOS will explain what full access allows and guide users toward specific apps or folders. The change also covers non-AI software that requests those permissions.

Device permissions are split in twoDevice Control and Data Access is being divided into System & App Protection Override and Accessibility. Users can limit which apps the first permission can control, and can narrow filesystem access to selected folders. An assistant that only needs documents in one project folder can be given access scoped to that task.

Users can require a fresh confirmationApps can be configured to ask every time they need access. The Mac menu bar shows an indicator when protection-override permissions are being used to control another app. Apple previously highlighted that broad access for AI agents can expose email, messages, browsing history and other files.

Existing grants carry over, and new installs bring developer workExisting permissions remain in effect, with the changes appearing on new requests. Backup and assistive technology developers can request entitlements to preserve relevant workflows. Other developers need to adapt onboarding and handle cases where users grant access only to some apps or directories.

▮ SIGNALApple is making permissions task-specific, so the usefulness of desktop agents increasingly depends on matching each workflow to the access it actually requires.

13 TALENT

❯ The Information says xAI cut its Grok Imagine video team and shifted focus toward coding and agents

About 60 roles are cut and the tool remains onlineThe Information previously reported that SpaceX’s AI division eliminated about 60 Grok Imagine roles, leaving a small team to maintain the service. On November 17 the outlet again described the change as a shift in resources from video generation toward coding. Imagine remains accessible.

Video had looked like an expansion opportunity earlier in the yearImagine turns text or still images into videos. In March, after OpenAI pulled back from Sora, Musk said publicly that he would double down on Imagine. The subsequent move to maintain the product with a smaller staff marks a change from the expansion signaled in the spring.

Coding and Grok Bot get more attentionThe reporting identified coding tools and Grok Bot as priorities. SpaceX acquired xAI in February and bought coding tool Cursor in August. The reshuffle comes after those businesses came under one group, which now has both code-editing products and agents that can handle a range of digital tasks.

Maintaining the service and developing it carry different costsGenerating coherent video across many frames takes ongoing inference capacity. Training new versions, improving quality and adding features require more engineers and compute. The reported staff changes point to a pullback in development, and the scope of future updates has not been announced.

▮ SIGNALKeeping Imagine online preserves access for current users, while shifting development capacity toward coding and agents will change what the company can build and ship next.

14 MARKET

❯ JD.com revenue rises 3.5% to RMB309.6 billion and net income attributable to ordinary shareholders grows about 22%

Revenue grows and the group returns to operating profitJD.com reported third-quarter results on November 17. Net revenue rose 3.5% to RMB309.6 billion. Net income attributable to ordinary shareholders grew about 22% to RMB6.5 billion. Group operating profit was RMB5.2 billion after a RMB1.1 billion loss a year earlier.

Services grow faster than product salesGroup product revenue rose 3.1%, while services grew 4.7%. Core JD Retail posted RMB15.1 billion of operating profit versus RMB14.8 billion a year earlier, with margin edging from 5.9% to 5.8%. Group marketing expenses fell 26.2%, while fulfillment costs, spanning warehousing, customer service and stores, rose 7.4%.

Retail adds an AI shopping assistant while delivery investment narrowsOn September 9, JD launched the Dongdong shopping assistant in version 16.0 of its app to match complex requests with products and services. Management said new-business losses narrowed year over year and sequentially, and that food-delivery investment fell as operations improved and revenue sources broadened. The release does not quantify Dongdong’s contribution to quarterly growth.

The European business joins the accounts after Q3On November 12, JD completed its €1.335 billion tender offer for about 59.8% of shares and votes in CECONOMY, parent of European electronics retailer MediaMarktSaturn, and began consolidating it that day. JD wants to connect its technology and logistics to the local retail network. The acquired business starts appearing in consolidated results in Q4.

▮ SIGNALJD’s operating improvement combines a broadly steady retail margin with narrowing new-business losses. The European deal adds another source of revenue, making segment results important for tracking efficiency.

15 CHIP

❯ GigaDevice reports customer progress in custom memory for edge AI and expects faster revenue growth in 2027

Management outlines when the projects should contribute revenueAt its November 17 third-quarter briefing, GigaDevice said its custom memory business had made progress with key customers and projects in several areas, including edge AI. It expects the business to contribute revenue in 2026 and grow briskly in 2027. Management did not give a revenue target or growth rate.

Edge AI brings some memory demand into the deviceEdge devices run some models or applications locally and need to hold models, software and working data. Manufacturers have to balance space, read and write speeds, cost and power. Custom projects arise when memory needs to fit a particular device, and the opportunity extends to markets beyond edge AI.

Customer projects still need to turn into productsProject work can involve adapting specifications, validating samples and moving into production. Revenue at scale normally appears as parts ship with the finished device, so customer adoption and shipment can be separated by time. GigaDevice gave an aggregate update without naming customers, specifications, validation stages or production schedules.

The 2027 outlook covers the overall custom-memory businessManagement’s revenue outlook applies to the custom-memory business as a whole. Edge AI is one area where customer projects have progressed, and the company did not break out its share of current revenue or next year’s growth. Subsequent revenue or shipment disclosures could show how much of the project work is converting into sales.

▮ SIGNALCustom memory becomes a volume business when customer projects pass device validation and enter production, with finished-device shipments ultimately determining their value.

16 CAPITAL

❯ LangChain raises $300 million at a $3 billion valuation and plans a November 20 LangSmith update

Valuation rises from $1.25 billion to $3 billion in about a yearLangChain announced a $300 million Series C on November 13 at a $3 billion post-money valuation, up from $1.25 billion in October 2025. Existing investor CapitalG led. GIC and DFJ Growth joined, with IVP and Sapphire Ventures participating again. The company separately plans to announce LangSmith capabilities on November 20.

The framework helps build agents and LangSmith handles what happens in productionLangChain started with an open-source framework for connecting models, data and external tools. Once an agent is live, a failure could come from a model decision, a tool response or one step in a longer workflow. LangSmith records calls and results so teams can trace what happened in production, including for agents built with other frameworks.

Monitoring data is also used to verify fixesLangSmith groups its work around running, improving and governing agents. Teams use observability and evaluation to spot failures, while LangSmith Engine can suggest code or prompt changes and watch production traffic for regressions. The company says Clay monitors hundreds of millions of runs monthly and more than 7,000 organizations use LangSmith, including Toyota, Nvidia and Salesforce.

The November 20 event is about the productThe upcoming capabilities build on the existing platform. Once companies maintain agents on a recurring basis, platform adoption can turn on the ability to reproduce an incident, evaluate a new version and maintain records for review. LangChain plans to detail the additions at its November 20 event.

▮ SIGNALLangChain is extending from initial development into ongoing operations, where customers may return to the same platform for releases, troubleshooting and governance.

Pass along the stories worth reading.
Get each issue in your reader: Feedly Inoreader RSS feed