❯ Huang admits affair and cooperates with independent inquiry as Nvidia addresses potential US national-security concerns
From reporting to admissionThe episode began with reporting about Huang’s private relationship with another business executive. Bloomberg says he had previously characterized the contacts as a normal business relationship. On November 8 local time, he wrote to Nvidia employees admitting infidelity and accepting full responsibility. Nvidia’s board announced an independent inquiry the same day, with Huang pledging support and cooperation. The company subsequently made the employee letter public.
Family and professional responsibilitiesHuang acknowledged that his conduct hurt his wife and children, betrayed their longstanding support and damaged the life they had built together. He also told employees, directors, partners and shareholders that leading Nvidia was the privilege of his life and that he had fallen short of the responsibility accompanying their trust. He pledged to work to regain that trust, accepted personal responsibility and asked the public to respect his family’s privacy.
Potential national-security concernsThe board disclosed that the reporting also raised potential US national-security concerns. Huang addressed those concerns in his letter, supporting the review and pledging to protect Nvidia and support US national security. The board identified US national and economic security as central to the interests of Nvidia and its shareholders. It said it currently had no reason to believe national security or company secrets had been compromised and that the inquiry would examine all available information.
Independent review and continued leadershipA committee of independent directors will oversee the investigation with assistance from outside independent counsel, placing the review outside day-to-day executive management. The board expressed continued confidence in Huang and his leadership team and said it would provide updates as appropriate. Huang has pledged full cooperation. Bloomberg reports that he will remain Nvidia’s CEO pending the inquiry’s findings.
The founder's significanceNvidia is a major supplier of AI computing infrastructure, and Huang remains directly involved in technology strategy, major customer relationships and management. Its flat organization acts quickly around his decisions, closely connecting his judgment with the company’s execution. Bloomberg reports that Nvidia has no public succession plan or obvious successor. As the company prepares to report earnings later this month, its board must maintain operational continuity while using an independent investigation to address governance questions facing customers and shareholders.
▮ SIGNALHuang’s reputation is closely tied to Nvidia’s technology decisions and customer relationships, making management trust an important issue for customers committing to long-term compute investment.
❯ DeepSeek Harness reaches 1.54 million daily users and about 92 trillion daily tokens as plugin marketplace takes shape
Less than three monthsThe DeepSeek Harness team said on November 5 that daily active users accessing its official API reached 1.54 million as of November 4, consuming about 92 trillion tokens a day. Version 0.1 launched on August 13 and the desktop application followed on September 29. Both usage figures cover users accessing the official API.
Plugins extend the workHarness provides a runtime for models to execute tasks, with plugins connecting tools and services. Third-party plugins numbered about 35,000, and roughly 60% of active users had created or installed at least one. The extension system lets the same agent framework serve different workflows while giving developers a route to offer their products and specialist capabilities to existing users.
Building distributionThe team plans an official marketplace to simplify plugin uploads and discovery, alongside improvements to API stability, compatibility and security. A creator-support program is also being prepared, with details to evolve through community feedback. It plans to support authors through token incentives, technical exchanges and recommendations.
▮ SIGNALThe 35,000 plugins give DeepSeek a source of product expansion beyond its own models, with distribution and security shaping whether that developer base becomes a durable ecosystem.
❯ Apple eyes proactive Siri as early as next year and revises iOS 28 development process
Following goals into actionBloomberg’s Mark Gurman reports that Apple wants Siri to operate more like an agent that can keep pursuing tasks, potentially as early as 2027. Such an assistant would track goals, determine the next action and continue working in the background. The implementation and release timing remain under development.
Pressure to move fasterGurman says CEO John Ternus recently met privately with software chief Craig Federighi and urged faster progress on the software roadmap and AI capabilities beyond Siri improvements. Personalized Siri, delivered with iOS 27, improved understanding and the use of personal information. The next ambition is to reduce how often users must issue a separate command for each step after setting an overall objective.
More room to exploreApple is also changing the first development milestone for iOS 28 and other systems. The usual four weeks of development and four weeks of integration become six weeks of coding, exploration and feature work followed by two weeks of integration. That gives engineers more room to experiment, while Apple’s existing frameworks, including App Intents, offer a potential foundation for connecting application functions and cross-app actions.
▮ SIGNALApple already has frameworks connecting application functions, making its next Siri challenge the orchestration of complete tasks users are willing to delegate.
❯ Anthropic takes over $3.4 billion of generator commitments for 1.7 GW of data-center power
Taking over existing ordersThe Financial Times reported on November 8, as relayed by Bloomberg, that Anthropic agreed in late September to pay about $3.4 billion for commitments covering 1.7 GW of heavy-duty gas-fired generators from Texas supplier OnePWR Solutions. Anthropic initially intended to buy power from OnePWR, then opted to take over its Caterpillar equipment orders.
Power becomes a core inputThe generators are intended to supply data centers, with 1.7 GW describing generating capacity. Anthropic must align the delivery of power, buildings and computing equipment to create usable capacity for training and operating models. Direct control of equipment orders gives it a greater role in infrastructure scheduling, along with additional deployment and project-coordination responsibilities.
Growing commitmentsAnthropic previously disclosed more than $895 billion in spending commitments over future years, largely tied to computing and infrastructure. Taking over generator orders extends its effort to secure greater control over supply. For AI services that continually consume computing resources, the timing of reliable power determines when servers can enter service and influences the cost of new capacity.
▮ SIGNALTaking over generator orders gives Anthropic more control over power delivery while making its expansion more directly dependent on executing energy projects.
❯ ByteDance considers Brazil data-center expansion larger than its existing $10 billion project
Seeking a phase-two developerBloomberg reported on November 6 that ByteDance is considering an expansion of its data-center project at Pecem port in northeastern Brazil’s Ceara state and is seeking a developer for the second phase. People familiar with the plans say the expansion could exceed the $10 billion, 300 MW project under construction. Preliminary talks also cover a similarly sized facility outside the port serving Brazilian customers.
Export and local marketsThe Pecem facilities qualify for tax benefits linked to data exports, while the separate project would target domestic demand. ByteDance previously announced plans to invest about 200 billion reais in the port complex in phases, with first-phase construction starting in early 2026. Both the expansion and the separate facility remain at the developer-discussion stage, and the company had not responded to Bloomberg’s request for comment.
Power and cable geographyThe Pecem region has substantial wind and solar resources, while nearby Fortaleza is a major South American submarine-cable hub. Power availability and international connectivity jointly support computing services for overseas customers. With Chile and Argentina competing for similar investments, Brazil’s competitiveness also depends on the tax, energy and communications conditions available to each project.
▮ SIGNALPecem combines renewable resources, submarine-cable connectivity and export-related tax benefits, illustrating how compute location decisions depend on all three together.
❯ SpaceX proposes Starbase power plant and supercomputer with about $24 billion in initial investment
A project in public recordsA City of Starbase hearing notice says SpaceX proposes to build, own and operate a gas-fired power plant and super-compute facility, with estimated initial capital investment of about $24 billion. The proposed site covers roughly 444 acres east of Brownsville between Boca Chica Boulevard and the Port of Brownsville. Construction may proceed in phases and include future expansion.
Incentives await reviewThe city commission plans a December 9 public meeting to consider a property-tax abatement agreement and a related economic-development agreement. SpaceX is the applicant, the initial investment figure is an estimate and the incentives await consideration. The commission will hear public views on the proposed tax-abatement and economic-development agreements.
Coordinating power and computeThe Information previously reported that public records, recruitment listings and construction activity pointed to a Starbase supercomputer. The municipal document adds an explicit plan to develop generation and computing together, while GPU counts, customers and commissioning dates remain undisclosed. On-site generation can help align power and data-center construction, while making fuel supply, equipment delivery and local permits important determinants of timing.
▮ SIGNALBy combining generation and computing in one project, Starbase extends SpaceX’s compute planning across land, power and data-center construction.
❯ Apollo’s two financing arrangements target about $75 billion of AI chips leased to SpaceX and Anthropic
Two arrangements combinedAccording to The Information, as summarized by Dealroom, two Apollo arrangements involve about $75 billion of AI chips. One would buy about $40 billion of Nvidia Vera Rubin chips through a subsidiary and lease them to SpaceX. The other concerns approximately $35 billion of fundraising involving Broadcom and Blackstone to buy Google TPUs for lease to Anthropic.
Lease payments service financingThe transactions use dedicated entities to raise money and own chips, with customer lease payments servicing the debt. Customers obtain equipment access through long-term leases, while the dedicated entities hold the corresponding assets during those leases. The comparison covered more than 100 neocloud chip-financing deals since 2023, with some convertible bonds, credit facilities and other equipment financing excluded.
Asset managers own the equipmentGPU-focused cloud providers commonly combine fundraising, equipment ownership and computing delivery. Apollo’s arrangements bring an asset manager directly into large-scale chip ownership. The resulting exposure depends on customers’ payment capacity, contract duration and equipment residual values. Apollo can also earn fees for arranging transactions and managing funds.
▮ SIGNALApollo can participate in compute expansion through financing and fund management, with tenant credit and chip residual values determining the durability of investor demand.
❯ Goldman tallies $296 billion of hyperscaler debt as AI issuance lifts financing premiums
Supply exceeds the old forecastMarketWatch reported on November 6, citing Goldman’s credit-strategy team, that year-to-date hyperscaler bond supply reached $296 billion, above its earlier $250 billion full-year forecast. Meta, Microsoft, Oracle and others borrowed another roughly $67 billion since late October. Goldman expects about $440 billion of such supply in 2027, with financing access varying by issuer and market conditions.
Extra yield for buyersIssuance pricing provides a concrete measure of funding pressure. A report carried by Tiger Brokers summarizing Goldman’s work says AI-related deals averaged about 18 basis points of new-issue concession in November so far, versus less than 7 basis points for other deals. This concession is the additional yield offered relative to comparable existing bonds. Concentrated supply can require issuers to give investors more price compensation.
Borrowing costs shape constructionThe same account says Goldman sees credit-spread risk skewed toward further widening through early December, implying potentially higher compensation above benchmark rates. If construction spending keeps running ahead of operating cash generation, additional bond supply and higher financing costs can jointly affect project returns. Goldman expects the usual December supply slowdown to offer some relief, while near-term spread risk remains skewed toward widening.
▮ SIGNALHyperscalers are competing for the same pool of bond-investor capital, making credit-market absorption an increasingly important influence on the pace of AI expansion.
❯ Nous releases Arete-27B, using about 52.8 million tokens to adapt Qwen3.8-27B’s response preferences
Adapting open weightsNous Research released Arete-27B, a post-trained version of Alibaba’s Qwen3.8-27B intended to adapt its value preferences and response behavior. The team also released model weights, training data and evaluation tools for reproduction and scrutiny. With roughly 27 billion parameters, the project demonstrates further customization of an existing foundation model.
Two supervised passesThe team trained for two epochs on about 52.8 million tokens using supervised fine-tuning with example answers, without reinforcement learning or architectural changes. Language-model weights were updated, the vision encoder was frozen and the multi-token prediction head was disabled. Teacher answers came from the open model Inkling, with training prompts checked for overlap against the evaluation set.
Maintaining task performanceNous reports 65.60% on its internal Hermes Bench, compared with 65.77% for base Qwen. Under the same local configuration, Terminal-Bench 2.1 scores were 70% and 67%, respectively, with the team interpreting the results as preservation of underlying capabilities and noting configuration differences as a possible explanation for local gains. The release also provides tools for examining changes in response preferences. Developers can obtain weights, data and the training description.
▮ SIGNALReleasing weights, training data and evaluations together lets users examine both the behavioral customization and the preservation of underlying task capabilities.
❯ Sand.ai announces $200 million Series A+ to advance video models and VidMuse
Three hundred million dollars in ten monthsVideo-generation company Sand.ai announced a $200 million Series A+ on November 5 at a valuation above $1 billion, bringing funding over the previous ten months to about $300 million. CoStone Capital and Wangsu Science & Technology co-led, with existing investors participating. This is the company’s third financing round this year, bringing together strategic investors, yuan funds and US-dollar investors.
Models meet paid creationFounded by Cao Yue in early 2024, Sand.ai developed autoregressive video model MAGI-1 and joint audio-video model GAGA-1, followed by the MAGI-2 Preview research release. Its paid application VidMuse supports music videos, product advertising and anime dramas, connecting generation tools with practical creation workflows. The company also plans to release the full MAGI-2 version in the near term.
Revenue and compute expand togetherDealStreetAsia reports that VidMuse launched in January and surpassed $10 million in annual recurring revenue within two months. The metric annualizes recurring revenue at the prevailing pace and describes the subscription business at that point. The company plans to deploy the proceeds into foundation-model research and training, inference infrastructure, product development and global expansion.
▮ SIGNALVidMuse gives Sand.ai a paid distribution channel for its models, making this round a bet on coordinated progress in capability, generation cost and the creation experience.
❯ GenRobot AI discloses $100 million Series A+ closed in Q3 for robot-training data systems
Closed in the third quarterGenRobot AI disclosed on November 2 that it completed a $100 million Series A+ in the third quarter. Participants included SAIC Motor, Baidu, Shoucheng Holdings, Galaxy General Robotics and Malaysia’s Eastworlds, alongside financial investors. The disclosed investors also include GL Ventures, CITIC Goldstone, CCB International and Huatai Zijin.
Recording human operationGenRobot supplies equipment and software for robot-development data, recording people performing household, manufacturing and logistics tasks. Its systems capture visual information together with hand and body movements, giving developers demonstrations that connect the environment with the sequence of actions. Those records can support both robot training and evaluation on similar tasks.
Making recordings usableCustomer needs span capture and subsequent processing, with product value linked to how accurately data represents actions, environments and task outcomes. Its Matrix platform synchronizes, cleans and labels the collected records, helping development teams produce usable training and evaluation material.
▮ SIGNALBy connecting human demonstrations to training workflows, GenRobot addresses recurring task-data demand that supports a market for capture hardware and processing software.
❯ SpaceX targets November 13 for Starship Flight 15 and first mission to catch both stages
Aiming for both stagesSpaceX is targeting Starship Flight 15 no earlier than November 13 at 4:45 p.m. US Central Time, or November 14 at 6:45 a.m. Beijing time, subject to approval and launch readiness. The mission plans to return both Starship and Super Heavy to the launch site for tower-arm catches on the same flight.
Returning after orbital workThe mission plans to deploy 26 Starlink V3 satellites and fly roughly 14 orbits at about 275 kilometers altitude over more than 22 hours. The booster returns shortly after launch, while the ship reenters and approaches the tower after its orbital mission. This would also be the first return-and-catch attempt for the V3 Super Heavy booster.
Health checks govern the returnControllers must confirm that systems needed for reentry, the landing burn and tower approach are healthy before proceeding with a ship return to Starbase. Three satellites will carry cameras to inspect the ship’s heatshield and inform that decision. Earlier flights caught previous-generation boosters, while this mission tests the additional chain from orbital work to ship recovery. The schedule remains subject to readiness changes.
▮ SIGNALReturning the ship safely after orbital work is central to advancing whole-system reuse and reducing the cost of each transport mission.