❯ Nvidia shares close 4.3% lower after independent inquiry begins, as Jensen Huang remains CEO and skips two industry events
The company moves from responding to reports to an independent inquiryOn November 8, Nvidia’s board announced that a committee of independent directors, assisted by independent outside counsel, would investigate matters involving founder and CEO Jensen Huang raised in recent reports. Nvidia had previously described the contacts as routine. Huang then acknowledged an extramarital relationship and apologized in a letter to employees. The board said Huang would cooperate, that it had no reason to believe company secrets had been compromised, and that it retained confidence in him and the leadership team. Huang remains CEO, and the board has given no timetable for the inquiry.
The market close and changes to his scheduleNvidia shares closed about 4.3% lower on November 9. That day, voice AI company ElevenLabs confirmed to TechCrunch that Huang would no longer appear at its November 11 summit in New York. He was also dropped from the schedule of Anthropic’s CEO Forum. Nvidia has invested in ElevenLabs, and these appearances were among the ways its CEO engaged with AI startups.
Shareholders bring a disclosure caseAfter the market closed on November 9, investors filed a proposed securities class action against Nvidia and Huang in federal court in Northern California. The plaintiffs challenge what the company previously told investors, alleging that its statements about its relationship with customer Megaspeed were misleading. The complaint has just been filed, and the court has yet to rule on the proposed class or the claims. The suit and the board’s internal inquiry are separate proceedings.
Why succession has come into focusAccording to The New York Times, Huang owns about 3.56% of Nvidia, sits on its 10-member board, and has more than 60 direct reports. He often brings employees from different levels into the same project meeting to tackle questions spanning technology, products, and customers. That flat structure gets information to the CEO quickly and puts many decisions involving several teams in his hands. Nvidia has announced no leadership change.
▮ SIGNALThe board’s practical test is whether it can keep chip road maps, systems deliveries, and major customer relationships moving during an independent review when those decisions so often converge on Huang.
❯ Culper’s May short report resurfaces, reviving its questions about Nvidia’s relationship with major customer Megaspeed
Why an old report is being read againAfter Nvidia’s board announced its independent inquiry, TechCrunch revisited a May 13 report on Nvidia by short seller Culper Research on November 9. Culper disclosed a short position in Nvidia when it published the report, and one of the businesses it examined was Singapore data center company Megaspeed. It described the relationship between Nvidia CEO Jensen Huang and former Megaspeed executive Alice Huang as unusually close. The report discussed those contacts in its examination of Nvidia’s relationship with a major customer.
Culper started with the customer’s rapid expansionIn its account of the research, Culper drew on corporate registrations, financial information, and supply-chain records to ask how a data center customer had scaled up purchasing so quickly. TechCrunch’s account of the records cited by Culper put Megaspeed’s combined liabilities and equity at about $33 million at the end of 2023 and about $3 billion a year later. A substantial share took the form of refundable deposits. That was a balance-sheet measure, and Culper used it to raise questions about the source of the funds and the customer relationship.
Where commercial and personal ties meetIn discussing Nvidia’s treatment of Megaspeed, Culper cited occasions when the executives appeared together or met in business settings, and argued that Nvidia should explain the relationship. Those were observations and judgments by a short seller. For a company that buys hardware, builds data center capacity, and rents it to customers, the business questions include whether its funding can sustain operations, whether its purchase demand is durable, and whether the supplier applies consistent rules to its customers.
The short report and the board inquiry are at different stagesCulper’s disclosed position in May meant it could profit from a decline in Nvidia shares. This week Jensen Huang acknowledged an extramarital relationship in a letter to employees without naming the other person. The independent board committee has yet to report findings. There are no public conclusions on whether a personal relationship affected either company’s business approvals.
▮ SIGNALWhen a fast-growing chip customer is also a relationship the CEO handles personally, independently checking its funding, end demand, and internal approvals gives the supplier a firmer basis for commercial decisions.
❯ Anthropic launches Fable 5.2 and Mythos 5.2 at 20% lower standard API rates, with Fable generally available and Mythos restricted to verified organizations
One underlying model, two routes to customersOn November 9, Anthropic released Claude Fable 5.2 and Mythos 5.2 for coding, reasoning, and research that cannot easily be broken into fixed steps in advance. They share an underlying model and use different safeguards. Fable 5.2 is open to Claude API customers and is available through major cloud platforms. Mythos 5.2 is restricted to organizations Anthropic verifies. The company still recommends starting with Opus 5.5 for everyday coding and knowledge work.
The unit cost comes downThe two models’ standard API rates are $8 per million input tokens and $40 per million output tokens, 20% below the 5.1 series. Anthropic says the new models generate output up to 30% faster, and both accept up to one million tokens of context. Developers can give a model a large codebase and its documentation in the same task, and can raise the reasoning effort when a problem warrants more work.
The total bill still depends on reasoning effortIn Artificial Analysis’s testing, Fable 5.2 scored 65 at maximum effort, up from Fable 5.1’s 49. It cost about $26.50 per task at that setting, roughly twice what Fable 5.1 cost in the same comparison. Higher effort can lead to more generated tokens, so the lower list price does not directly translate into a lower bill for a piece of work.
What a long task can look likeAnthropic demonstrated the scale of work it has in mind with a prototype operating system. Fable 5.2 started with an empty folder and coordinated more than 1,000 subagents over eight days to write, test, and review the code. Research teams can use that kind of capability to explore ideas and build prototypes, with the finished work still subject to the project’s reliability checks. Eligible paid Claude users receive additional Fable usage during the launch promotion, while API customers pay for usage.
▮ SIGNALThe value of a flagship model increasingly depends on how far it can carry difficult work toward acceptance, so buyers have to weigh the total bill against the rework and human handoffs it can save.
❯ Google unveils TPU 8tx and 8i Max to tackle hardware bottlenecks in confidential inference and long-context AI
Four systems in one TPU generationOn November 9, Google Cloud announced two new TPU systems. Google designed these accelerators for AI computing, initially for its own workloads and now also for cloud customers. In April it announced the 8t for large-scale training and the 8i for real-time inference. The 8tx and 8i Max address sensitive data and long inputs, respectively. Google infrastructure chief Amin Vahdat told Bloomberg that both systems will be available in 2027.
8tx extends protection to data in useSensitive information can be encrypted in storage and in transit, but a model still needs to access it to do its work. The 8tx combines 8t silicon with fifth-generation AMD EPYC host processors and isolates data and model weights in hardware while they are in use. Customers with strict compliance requirements can run a proprietary model and fine-tune it on a small scale with internal data on the same system, reducing the need to move that data between clusters.
8i Max packages two chips togetherWhen models read an entire codebase, hours of video, or records from multiple agents, feeding that information to the processors can become the bottleneck. Google says putting two 8i chips in a single package doubles compute, on-chip static memory, and high-bandwidth memory throughput relative to the 8i. Google also says the system can fit more computing hardware into the same rack footprint, useful for services that routinely process long inputs.
Customer workloads now shape the hardwareGoogle says TPUs have nearly 100 customers, and it has started letting customers deploy them in their own data centers. The 8tx and 8i Max respond to separate requests for stronger data isolation and larger context windows. Reusing the underlying silicon and changing the package, network, or host helps Google deliver a specialized system sooner than designing a new chip for each workload.
▮ SIGNALGoogle is increasingly competing with whole TPU systems, where customers buy the combination of security, memory bandwidth, and rack efficiency that their workloads need.
❯ Apple starts selling a $49.99 AI usage pack for Creator Studio subscribers, starting in Logic Pro
An on-demand purchase inside a professional subscriptionOn November 9, 9to5Mac found that the latest Logic Pro update let Apple Creator Studio subscribers buy extra AI usage for a one-time $49.99 charge. Logic Pro is the first app with the purchase option. On a Mac, subscribers can open Intelligence Features in the app menu, select Show Usage, and see the option to buy more.
Why music software comes firstApple announced the Creator Studio updates in October. Logic Pro’s beta assistant uses a mix of Apple and third-party models to do things such as setting up a vocal track from one prompt or helping with an active project. The update also includes on-device audio analysis that automatically classifies and tags sounds so producers can find suitable instruments and loops.
The monthly allowance comes first and unused extras roll overUnder Apple’s terms, the Logic Pro assistant is powered by Google Gemini, and more complex conversations can consume more of the allowance. Eligible features draw from the monthly allowance before using the purchased balance. Leftover purchased usage rolls over and expires 12 months after purchase. The pack is tied to the Apple Account used to buy it and cannot be shared with family members through Family Sharing. The first supported markets are the U.S., U.K., Canada, Brazil, France, Germany, India, and Mexico.
Other apps follow their own update schedulesFinal Cut Pro, Pixelmator Pro, and iWork are also getting Creator Studio updates, spanning tasks such as natural-language image edits and assistants that work inside documents. The purchase option currently verified is in Logic Pro. Apple has yet to announce a single pack covering all Apple Intelligence features. For creators whose work comes in bursts, the setup lets them add capacity in a busy month and carry the rest into later projects.
▮ SIGNALApple is keeping the core tools inside the subscription and charging heavier users separately for extra calls to selected AI features, with one purchase potentially covering several projects.
❯ ByteDance releases Seedream 5.5 in Doubao and Dreamina for repeated image edits and transparent assets
From getting an image to revising it repeatedlyOn November 9, ByteDance released Seedream 5.5. It is in Doubao, Doubao Work, and Dreamina, with a trial also available in Volcano Engine’s Ark platform. The preceding Seedream 5.0 Pro already edited images and handled complex layouts with text. Version 5.5 focuses on repeat edits, because passing each output into the next round can gradually distort parts of an image that were meant to stay fixed.
What ByteDance changed in a 100-round demonstrationThe Seed team demonstrated 100 consecutive edits to an image of the same person, changing clothing, accessories, and objects on a desk while still being able to adjust a ring or eyeglass frame near the end. The new model prioritizes preserving pixels outside the requested area and handles occlusion, perspective, and shadows more accurately. It also generates transparent or semitransparent assets directly, so fine edges such as feathers can be layered into other designs.
Billing counts input and output imagesThe Beijing News reported that 5.5 generates images faster than 5.0 Pro and that the forthcoming Ark API will continue charging separately for input and output images, with no charge for text tokens. Its example put one 2K square image edit at about RMB 0.28, less than half the previous version’s cost. Repeated edits incur multiple charges, so the total bill depends on the number of rounds and what each produces.
The first apps are live and a Pro version is plannedCreators can select 5.5 in the first set of apps. The team also improved realistic lighting, materials, and small text rendering, and plans a Pro version for more complex inputs and instructions. For a product photo, poster, or series of assets, the gains can save work such as repairing text, cutting objects from their backgrounds, and realigning elements after generation.
▮ SIGNALFor an image model to become part of a production workflow, a design that survives several rounds of revisions matters more than an attractive first result.
❯ Early Qwen4 models reach a game development benchmark, with Flash at 49.9 and 27B’s audio score corrected
Two early models enter an independent testOn November 8, GameFeel Bench added Qwen-4-Flash-Early. The downloadable leaderboard shows 49.9 for Flash at Max effort and 47.2 for Claude Opus 5.5 at XHigh. An audio scoring correction moved Qwen-4-27B-Early from 47 to 44.6 at XHigh, close to Claude Sonnet 5.5’s 45 at Max.
The task is to improve the feel of an existing gameThe benchmark uses 12 unreleased games and asks models to add content, fix issues, or rebalance gameplay. It considers ideas, game feel, visuals, and audio alongside code. The author says that humans rate some categories and know which model produced each result. Flash’s showing in gameplay and ideas helped lift its overall score, while the results for any one task depend on the project and the scoring criteria.
The cost figures come from the evaluator’s own billThe author said he paid about $0.40 per task for Flash and $0.16 for 27B on an early-access API. The benchmark’s $2.60 for Opus 5.5 uses subscription billing. Those numbers reflect the test settings and accounts used. Public API rates, request volumes in a different project, and the ability to run a model locally will depend on what is released.
Public release details are still openAlibaba Cloud said in September that Qwen4 was in training. The benchmark shows that an evaluator could use early versions, while Alibaba has yet to publicly confirm the full lineup, a release date, or production pricing. The tests give game teams a look at what a smaller model can do on specific creative tasks. Repeating the test on the public versions will require access to those versions.
▮ SIGNALA lower per-task cost can give a game team room to try more mechanics in parallel, provided the model can produce something playable in the team’s own codebase.
❯ SpaceX proposes a gas plant and supercomputer in Starbase, Texas, with $24 billion in estimated initial investment
A city notice puts a scope and cost on the proposalA public hearing notice from Starbase, Texas, describes a SpaceX proposal to build, own, and operate a gas-fired power plant and a supercomputing facility. The initial phases are estimated to require about $24 billion in capital investment. The notice allows for a phased build and future expansion and gives no separate budgets for the plant, computing hardware, or other work.
An industrial site and a local tax applicationThe proposed site spans about 444 acres east of Brownsville, between Boca Chica Boulevard and the Port of Brownsville, in the Cameron County Enterprise Zone. SpaceX has applied for a local property tax abatement. The Starbase City Commission is scheduled to hold a December 9 hearing and consider both the tax relief and a related economic development agreement. Both are still pending.
Power goes into the same construction plan as computingReuters reported that SpaceX already rents computing capacity to customers including Google and Anthropic. The Starbase proposal puts the power plant and supercomputer in one project, which could reduce the wait for a connection to the main Texas grid. On a campus built in phases, server deliveries, electricity, and cooling still have to line up. The public notice names no customer for the new site.
The hearing deals with local agreementsThe notice gives no size or duration for the proposed tax relief. It also leaves out the plant’s generating capacity, the date for the first phase to go into operation, and the planned server count. The December hearing covers local tax and economic development arrangements, while construction schedules for the plant and computing facility will determine when each phase can deliver capacity.
▮ SIGNALBringing generation into the same project gives SpaceX more control over when expensive hardware can go live and makes a power plant delay a problem it has to manage directly.
❯ ByteDance seeks a developer for a second Pecém data center phase that could exceed the first phase now under construction
The company is still looking for a developerOn November 6, Bloomberg reported, citing people familiar with the matter, that ByteDance is considering expanding its data center at Brazil’s Port of Pecém in Ceará and is looking for a developer. Those people said the expansion could exceed the first phase now under construction, which Bloomberg puts at about $10 billion and 300 megawatts. The budget, power requirements, and construction date for the second phase remain open.
The first phase is under way with a wind power dealByteDance announced a phased development at the port last year and construction began early this year. In May, Reuters reported that developer Omnia signed a 20-year power supply deal worth about $2 billion with renewable generator Casa dos Ventos, drawing on wind farms in northeastern Brazil. Omnia described the first phase as 200 MW of IT capacity with electricity demand of about 300 MW. It expects operations to start in the third quarter of 2027 and expand through 2029.
Why build at a portNearby Fortaleza is a landing point and hub for subsea cables that can carry data to other markets, and Ceará and neighboring states have substantial wind resources. The port’s tax treatment also depends on exporting data. Which customers the site serves and how it sends their data therefore matter to its eligibility.
An off-port site could serve Brazilian customersBloomberg also reported early talks about a similarly sized data center outside the port for customers in Brazil. That proposal is separate from the second phase at Pecém and remains at an early stage. If both advance, ByteDance could use different facilities for cross-border and domestic services, with each needing its own development and power arrangements.
▮ SIGNALPecém brings electricity, international connectivity, and tax benefits for data exports together. The best site for serving Brazilian customers may be somewhere else.
❯ Anthropic could start IPO marketing on November 10 as some investors discuss a $1.1 trillion to $1.3 trillion valuation
From filing to formal marketingBloomberg reported on November 9 that Anthropic could start marketing its IPO as early as November 10 Eastern time. It filed its prospectus publicly on October 26, updated it with third-quarter results on November 2, and plans to trade on Nasdaq as ANT. On the reported schedule, pricing and a debut could come in the week of November 16. The company has yet to announce a price range, share count, or listing date.
Market conditions have shifted the valuation discussionInvesting.com, citing Bloomberg, reported that after last week’s sell-off in AI shares some prospective investors discussed $1.1 trillion to $1.3 trillion, down from $1.5 trillion to $2 trillion discussed earlier this month. Those figures value the whole company. Its proceeds will depend on the offer price and how many new shares it sells. Indications of demand during the roadshow will help set both.
Revenue has grown, while operations remain loss-makingAnthropic’s amended prospectus shows that revenue in the quarter ended September rose to about $17 billion from about $1.4 billion a year earlier, while the company had an operating loss of $567 million. Anthropic earns money from Claude subscriptions and enterprise use of its models, and says revenue is primarily driven by usage. More calls can raise revenue, but the company needs to secure computing capacity in advance.
Contractual payments and customer usage run on different schedulesThe filing shows about $895.4 billion in purchase and lease commitments, almost all for computing and infrastructure. About $69.8 billion is due in the 12 months after the end of September 2026. About $213.8 billion of the total relates to arrangements that can be terminated or depend on conditions being met. Some contracts have minimum purchases that cost money even if usage falls short. Buyers have to judge how fast demand can fill the capacity Anthropic has already signed up for.
▮ SIGNALPublic investors will price a business whose revenue moves with customer usage and whose supply is contracted well in advance. The gap between those schedules has a direct bearing on valuation.
❯ Hardware testing software maker Nominal raises $240 million in Series C, tripling its valuation to $3 billion in eight months
A second raise in eight monthsOn November 9, Nominal announced a $240 million Series C at a $3 billion valuation, led by Lightspeed Venture Partners and Insight Partners. In March, Nominal announced an $80 million Series B-2 extension led by Founders Fund at a $1 billion valuation. The new round backs a push to take software originally built for hardware test teams into design, manufacturing, and operations.
From retrieving a memory card to monitoring a flight in real timeTests of aircraft, robots, and other hardware generate sensor readings, video, component versions, and test parameters. Nominal puts them on a common timeline so engineers can see what changed before and after an anomaly. In Nominal’s Odys Aviation case study, the flight test team previously retrieved data from storage cards after landing. With Nominal, the test site and headquarters could view the same sensor readings in flight and look up ground and flight tests in one catalog.
An AI assistant can refer back to the evidenceNominal says it serves more than 80 commercial and government organizations and has introduced Nominal Analyst. Engineers can ask it to identify unusual results, compare them with earlier runs, and assemble analysis they can check against the original records. Engineers still make key decisions and approve changes. The software can run in the cloud, on premises, or in isolated environments for teams that cannot send data to a public cloud.
Test evidence can follow a product into manufacturingOnce a machine goes into production, configuration and test records from its prototype can help diagnose similar problems on a production line. Nominal plans to offer more engineering applications on top of the same data platform, and customers can use its interfaces to build tools of their own. For Nominal, expansion into manufacturing and operations can put more of an existing customer’s work on the platform.
▮ SIGNALNominal already sits where engineering teams collect test data, which lets it carry traceable AI analysis into manufacturing and maintenance without asking customers to assemble a new dataset first.
❯ Alibaba chip unit T-Head reportedly plans a 2027 Hong Kong IPO to raise $2 billion to $3 billion
A reported fundraising rangeInternational Financing Review, as reported by Zhitong Finance, says Alibaba chip design unit T-Head is planning to list in Hong Kong in 2027 and aims to raise about $2 billion to $3 billion. Alibaba has yet to confirm a timetable or announce offering terms. CEO Eddie Wu raised the possibility of an independent listing in March. The new report adds a venue, timing, and a fundraising range.
What the shipments and revenue figures measureOn Alibaba’s March earnings call, Wu said T-Head had started volume production of its own AI chips, had shipped about 470,000 by February, and was running at roughly RMB 10 billion in annualized revenue. Shipments were cumulative, while the revenue figure described the scale of the business at the time. For Alibaba Cloud, the chips also provide a supply of computing power for its own models and services sold to customers.
A new chip is due for commercial rollout in early 2027Alibaba Cloud unveiled the Zhenwu V900 in September for training and inference, with volume production and commercial release planned for the first quarter of 2027. It says the V900 has 216 GB of memory and delivers three times the performance of the M890. Alibaba also announced an upgraded server design that combines computing, networking, and storage hardware. Cloud customers need working clusters to run large models, and networking and storage affect the compute they can actually use.
A listing would put the chip business in view on its ownAlibaba Cloud says the Zhenwu family has served more than 650 customers across industries including automotive, finance, energy, and manufacturing. T-Head supports Alibaba’s own cloud and model operations while finding uses in those markets. If a listing goes ahead, investors will need more detail on revenue by product, the mix of internal and external customers, and delivery schedules to assess the business on its own.
▮ SIGNALAlibaba Cloud gives T-Head a place to deploy new chips at scale. An independent listing would put more attention on how broadly the hardware and related services can sell beyond the group.
❯ Coinbase’s U.S. national trust charter takes effect, bringing institutional custody under federal supervision
From a New York charter to a national charterOn November 9, Coinbase announced that the Office of the Comptroller of the Currency had given final approval to charter Coinbase National Trust Company. The OCC’s filing record shows the application was received in October 2025, gained preliminary conditional approval in April 2026, and took effect on November 6. Coinbase has provided custody through a company that received a New York limited-purpose trust charter in 2018.
What institutional customers entrust to a custodianDigital asset custody includes safeguarding the keys that control the assets, tracking holdings, and processing transfers authorized by clients. The OCC’s April decision also describes related services available to custody clients and plans to migrate the existing business to the new entity. The national charter gives banks and asset managers a common federal framework against which they can assess their own vendor requirements.
Stablecoin payments have a separate pilotAlso on November 9, Coinbase and PNC expanded their partnership. The pilot lets eligible corporate and institutional clients receive USDC, with Coinbase supplying payment APIs and PNC supplying banking and treasury services. It builds on an earlier collaboration that gave eligible private bank clients access to digital asset trading and extends the relationship to corporate collections.
Stablecoin issuance remains a reported possibilityBloomberg, citing people familiar with the matter, reported that the federal trust structure could also help Coinbase issue its own stablecoin or provide issuance and management services to other companies. Coinbase has announced no such product or timetable. The charter in effect covers institutional custody, and the PNC pilot accepts the existing USDC.
▮ SIGNALCoinbase can pursue more institutional custody mandates with a federal charter while putting an existing stablecoin to work in corporate payments through a bank, building two ways to serve financial institutions beyond trading.
❯ Take-Two’s quarterly net bookings fall 18% to $1.6 billion as NBA 2K27 misses expectations and full-year guidance holds
The quarter falls short of both Wall Street and company expectationsOn November 9, Take-Two released fiscal 2027 second-quarter results for the period ended September 30. Net bookings fell 18% year on year to $1.60 billion from $1.96 billion, below the company’s previous $1.62 billion to $1.67 billion range and the $1.65 billion market estimate cited in Bloomberg’s reporting. Take-Two uses net bookings to track digital and physical sales and other commercial activity. Net revenue recognized under accounting rules was $1.50 billion.
NBA 2K27 sells more than 4 million copies and still falls shortOn the earnings call, Take-Two said NBA 2K27 had sold more than 4 million units since its September 4 release and delivered the series’ second-highest launch bookings, below its internal forecast. Bloomberg reported that management cited the attention going to GTA VI ahead of its release as one factor. The mobile business performed as expected, while the Grand Theft Auto franchise exceeded the internal forecast for the quarter.
Recurring spending declined and more content is on the waySpending on virtual currency, add-ons, and other in-game purchases accounted for 81% of total bookings and fell 9% year on year. Take-Two expects NBA 2K27 to use themed seasons, college basketball content, and MyTeam updates to increase engagement and spending in the second half of the fiscal year. For an annual sports title, what returning players spend after launch matters alongside early unit sales.
The annual outlook still depends heavily on the holiday quarterTake-Two reiterated full-year net bookings guidance of $8.0 billion to $8.2 billion for the year ending March 2027 and projected $3.35 billion to $3.50 billion for the quarter ending December. Rockstar’s GTA VI remains scheduled for November 19 on PS5 and Xbox Series X|S, with management saying preorders exceeded its initial expectations. The third-quarter forecast represents about two-fifths of the full-year target. Holiday sales of the new game and continued spending in existing ones will help determine whether the company meets that target.
▮ SIGNALTwo major titles from the same publisher can compete for players’ time and budgets. Take-Two is counting on GTA VI’s launch and NBA 2K27’s follow-on spending to contribute at different points in the year.