AI News

TMTPost AGI Morning Brief: Anthropic Launches IPO Roadshow, OpenAI Quarterly Revenue Tops $10 Billion, Google Releases…

TMTPost · AGIUpdated Nov 12 09:50

AI summary

TMTPost AGI Morning Brief summarizes multiple reports around November 12: Anthropic launched its IPO roadshow on November 10, planning to issue 825 million Class A shares at $70 to $78 per share, raising up to $64.35 billion, with a fully diluted valuation cap of $1.405 trillion, and intends to list on Nasdaq on November 19 under the ticker ANT. OpenAI's revenue in the third quarter of 2026 exceeded $10 billion for the first time, up nearly 50% quarter-over-quarter, with enterprise business doubling quarter-over-quarter and a gross margin of 33%. Google opened access to the Gemini 4 Carbon reasoning model the same day and released the Gemini Robotics 2.5 vision-language-action model.

Body · AI translation

Silicon Valley Frontier:

1. Anthropic launches a $64.35 billion IPO roadshow, with a $1.4 trillion valuation betting on AI's coming of age

1. Anthropic launched its IPO on November 10, 2026, planning to issue 825 million Class A shares at $70 to $78 per share, raising up to $64.35 billion; the fully diluted valuation ceiling is $1.405 trillion, nearly half again as much as the $965 billion valuation at its May funding this year. It plans to price on November 18 and list on Nasdaq on November 19 under the ticker "ANT".

2. Financially, it shows high growth and high losses: for the nine months ended September 30, 2026, revenue was $33.24 billion, up about 1,200% year on year, with a run-rate of about $81 billion at the end of October, but the net loss over the same period was $104 billion (versus $25.1 billion in the same period of 2025, much of it from fair value changes in convertible bonds and derivatives); on an annualized revenue basis the implied price-to-sales ratio is about 17 times, and valuation expectations have already been cut from a peak view of $2 trillion to $1.4 trillion.

3. The IPO binds the entire AI industry chain through capital instruments: agreements were signed with seven companies—AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC—under which strategic investors buy $26 billion of Class N shares in a private placement at the offering price without lock-up restrictions, with about 20% allocated to retail; the article argues that the listing marks the AI industry's shift from "burning money and telling stories" to "letting the financials do the talking", and that those who buy in at high prices in the secondary market face the risk of a valuation re-rating.

2. Anthropic storms the office heartland: Claude Sheets takes on Excel, going head-to-head with Microsoft and Google on the eve of its IPO

1. Product and market reaction: On November 10, Anthropic launched the Claude Sheets, Whiteboards and Data trio plus the soon-to-launch Claude Meetings, striking directly at the core office scenarios of Microsoft Office 365 and Google Workspace; after the news emerged, Microsoft's stock fell 3% intraday, and dropped another 2% after Bloomberg's report came out.

2. IPO and performance: Anthropic filed its S-1 on October 26 and plans to list on Nasdaq on November 19 (ticker ANT), issuing 825 million shares at $70 to $78 per share to raise up to $64.4 billion, with a valuation ceiling of about $1.4 trillion; it is also raising another $26 billion through a private placement, for a total of up to $90.4 billion; as of October 31 its annualized revenue reached $81 billion ($71 billion at the end of September and $65 billion in July).

3. Coopetition and cash-out: Microsoft and Google are both investors and challenged rivals; Microsoft has told UK Microsoft 365 Copilot tenants that Claude will be enabled by default from December 9; meanwhile, seven co-founders will sell a combined $1.75 billion of stock in the IPO, each about 1.5% of their total holdings, cashing out about $250 million each at the midpoint of the offering range.

1. On November 8, Nvidia's board announced the launch of an independent investigation, prompted by reports that founder and CEO Jensen Huang had become a target of a foreign intelligence operation; on November 9, a shareholder class action was filed in a California federal court, alleging that during the 13-month class period from October 9, 2025 to November 8, 2026, the company inadequately disclosed its relationship with Singapore client Megaspeed, in an alleged violation of U.S. securities law.

2. Market reaction: Nvidia fell 4.27% to $178.67 on November 9, then fell another 5.47% to $168.89 on November 10, with a market cap of about $4.07 trillion, approaching its 52-week low of $164.27 (52-week high of $243.37); its market cap has evaporated about 26% in the past month, and the article says this is not an atypical AI-cycle pullback but a re-rating driven by governance and national security risks.

3. Core logic: The lawsuit focuses on "inadequate disclosure" rather than unlawful conduct, and the key to victory lies in the "materiality" of the information; the board's stance of putting U.S. national security above shareholder interests is said to raise the cost of capital (higher discount rates, constrained growth options), and the national security risk discount may spread from Nvidia to AMD, Broadcom, TSMC and other AI semiconductor chain players.

1. On November 11, Baird downgraded Nvidia from "Outperform" to "Neutral", cutting its target price by 60% from $500 (set in May 2026) to $200; the reasons were limited visibility of data center demand in 2027, supply constraints, U.S. chip export restrictions to China and a rising probability of a CEO transition. Notably, before the downgrade the stock had already fallen about 29.4% from its October 6 closing high of $239.24, closing at $168.89 on November 10; this re-rating looks more like a rise in risk premium than a demand-side blow-up.

2. This is more a "visibility shock" ahead of a "demand shock": Nvidia's fiscal 2026 second-quarter revenue was $96.22 billion, up 106% year on year, with data center revenue of $89 billion, up 117% year on year, and it guided third-quarter revenue to $108 billion (plus or minus 2%); Baird, however, expects hyperscalers' AI capex growth to slow to about 40% or slightly less in 2027, and flags the timing risk that 10% to 30% of advance orders could be cut.

3. The indicators that need to be verified in focus going forward are: utilization after customer deployment, inventory levels, and return on capital expenditure (that is, whether chips are continuously used rather than piling up), as well as risks at the policy and governance level—export controls may prompt customers to "design out" American chips, the company's guidance for next quarter assumes no data center revenue from China, and the investigation surrounding Jensen Huang has also brought executive succession and strategic continuity into valuation considerations.

1. The AI infrastructure bottleneck is shifting from "insufficient compute" to "mismatch in power architecture": Nvidia's data center revenue in the second quarter of fiscal 2027 (ended July 26, 2026) was $89 billion, up 117% year on year. Compute is concentrating into large clusters at extremely high speed, while the construction cycles for power generation, transmission, substations, grid connection and other links are longer, becoming a key constraint on whether AI projects can go live on schedule.

2. Electricity demand shows structurally high growth but is highly concentrated spatially: the International Energy Agency expects global data center electricity consumption to rise from about 415 TWh in 2024 (about 1.5% of the global total) to about 945 TWh in 2030 (about 3%), with an average annual growth rate of about 15%; the United States may add about 240 TWh by 2030 compared with 2024 (an increase of about 130%), and the United States and China together will contribute nearly 80% of the global increment before 2030. A report by the U.S. Lawrence Berkeley National Laboratory shows that U.S. data center load has grown about threefold over the past decade and is expected to possibly double or triple by 2028.

3. The center of gravity of industrial value and the logic of competition are shifting toward "power architecture": scarcity has changed from "how much power there is" to "where power is available, when it is available, and whether it can be stably connected and expanded." The benefit weighting of transformers, switchgear, power distribution, liquid cooling, energy storage, energy management software and power engineering services is rising; flexibility arrangements such as workload shifting, energy storage and on-site generation, as well as long-term power purchase agreements and interruptible loads, are changing from optimization items into the "entry ticket" for data centers to obtain grid connection, control financing risk and maintain asset value.

VI. OpenAI's quarterly revenue exceeds $10 billion for the first time, enterprise business doubles

1. Revenue and structural transformation: In the third quarter of 2026, OpenAI's revenue exceeded $10 billion for the first time, up nearly 50% quarter on quarter and about three times the same period last year, with an annualized scale of about $50 billion; the enterprise business doubled quarter on quarter, pushing the ratio of consumer to enterprise revenue to 50/50, and consumer revenue alone in a single quarter exceeded that of all of 2025.

2. Cost and gross margin bottleneck: gross margin is only 33%, far below the 75% to 85% of typical enterprise SaaS and also below Anthropic's nearly 85%; inference costs rose from $8.4 billion in 2025 to $14.1 billion in 2026, and cash burn is expected to be about $27 billion in 2026 and $63 billion in 2027, showing the marginal cost expansion characteristic of "the more it sells, the more it burns."

3. IPO reconciliation and the rules of the new track: OpenAI submitted a confidential IPO registration statement on May 22, 2026, and Anthropic has also submitted one, with targets to list between autumn 2026 and early 2027; as GAAP standards are uniformly adopted, the core of investors' attention shifts from "run rate" to gross margin and the share of enterprise revenue, that is, whether gross margin can rise from 33% and whether enterprise revenue can hold its position at 50%.

VII. OpenAI puts Google's twenty-year business into ChatGPT

1. Launch and mechanism: On November 10, 2026, OpenAI launched "Businesses in ChatGPT" (beta), through Business Hub allowing businesses to claim and update their profiles, upload product feeds containing prices and inventory, and view three types of aggregate data: number of mentions, website clicks and "discovery topics"; at the same time, it opened "Business Agent" to eligible U.S. e-commerce and online service businesses, covering its 1.2 billion weekly users, essentially turning "how businesses appear in answers" from a black box into a measurable, purchasable commodity.

2. Monetization path: measurement is the bait, advertising is the hook—OpenAI has tested ChatGPT ads in the United States since February 9, 2026, and by August 31 the annualized advertising revenue run rate reached $1 billion (in less than seven months), and by October 21 more than 100,000 advertisers had run ads in 89 countries; the free Business Hub is intended to funnel large numbers of non-advertising businesses into the paid advertising ecosystem, reusing Google's path of "free tools first, then pay for traffic."

3. Competition and impact: OpenAI is directly benchmarking Google; in the quarter ended September 2026, Alphabet's search business revenue was $63.4 billion, accounting for 49.7% of total revenue, falling below half for the first time; at the same time, third-party GEO startups (OtterlyAI, Peec AI, Profound), because they rely on simulation and estimation while the platform holds real usage data, are seeing their information gap narrow, and their role has been downgraded from "category definers" to "platform ecosystem supplements," with their survival space shifting to areas not yet covered by platforms, such as cross-platform monitoring and prompt attribution.

VIII. Google's Gemini 4 Carbon debuts: tackling the hard problems, the longer it thinks the more accurate it gets

1. On November 10, Google officially opened Gemini 4 Carbon, positioned as "the strongest reasoning model to date," specializing in complex software engineering, multi-step agentic tasks, scientific research, and cybersecurity; it forms a "one hard, one soft" division of labor with the efficiency model Argon—use Argon for everyday coding and looking up documentation, and only bring Carbon in for hard tasks such as refactoring legacy systems and auditing codebases. The model was first introduced by DeepMind's Koray Kavukcuoglu on October 28, and in its initial phase was opened only to selected cybersecurity defenders through the Fairwind Program, before later being rolled out to developers, enterprises, and consumers.

2. Security and scientific research are the scenarios where Carbon establishes its authority: under guardrail-free evaluation, it achieved 88.2% on SEC-bench Pro vulnerability discovery, 54.8% on ExploitGym vulnerability exploitation, and 71.0% on CWE-bench patching; its success rate in internal secure code generation is more than double that of Argon, and the Chrome team said its performance in generating high-quality patches is more than ten times that of frontier models from other labs; the Wiz team used it to scan out a zero-day vulnerability in under two hours, compressing vulnerability research that originally took weeks into a single afternoon. On the scientific research side, in multi-agent mode, Carbon mines distant homologous sequences, increasing the number of retrieved homologous sequences by an average of 40 times on 7 low-homology protein families in the CASP17 blind test set.

3. The cost of deep thinking is equally clear: the third-party organization Artificial Analysis recorded that Carbon outputs about 200 million tokens in its intelligence index evaluation, about three times the median of 64 million for similar models, with an average cost of about $7.76 per task at standard pricing. Pricing shares the launch discount with Argon, at $2 per million input tokens and $10 for output, with cached input priced 95% lower than input, rising to $4 and $20 after the discount period. In addition, Google has adopted a phased rollout—defenders first, then trusted partners, and finally the general public—to address the double-edged-sword risk that the model could be used to find vulnerabilities and write attack payloads; its security benchmark scores have not yet been fully reproduced by an independent third party.

IX. Google releases its strongest robotics model Gemini Robotics 2.5: it has not only learned to tie a bow, but has also successfully taken up a post on the Mercedes-Benz production line

1. DeepMind released the Gemini Robotics 2.5 vision-language-action (VLA) model on November 10, 2026; across evaluations on more than 200 tasks, task success rates improved by 30% to 100% and completion speed by 20% to 30% compared with the previous generation, and it achieved fine manipulation such as tying a bow, picking up food with tongs, and pipetting, marking embodied intelligence's shift from demonstration to deployment.

2. On the industrial deployment side, on two real production lines at Mercedes-Benz's Kecskemét plant in Hungary and Sindelfingen plant in Germany, Apptronik's Apollo 2 humanoid robot completed automotive kit sorting with a success rate of 96% (8 errors out of 200 attempts); DeepMind director Kanishka Rao said closing the last 4% is a "steep road," and it remains at the PoC level, with a gap to 99.9% industrial-grade reliability, so large-scale unmanned operation will still take time.

3. The performance leap relies heavily on the Gemini 4 Argon simulation engine released on October 13, 2026, which greatly compresses training cycles (it reaches 87% on pure reasoning in the ERQA benchmark); the current model still runs in the cloud, with latency and network-disconnection risks, and the demo video was sped up 2x. Google is taking the model-layer "Android" route and has not yet opened a public API, weights, or training code.

X. CoreWeave rebounds 17%: can AI orders pass through the financing black hole?

1. CoreWeave's third-quarter revenue was $3.661 billion (up 168% year over year, beating expectations of $3.55 billion), with an adjusted EBITDA margin of 62% and a backlog of orders to be delivered of about $132.5 billion (continuing to expand from $104 billion at the end of June); the beat eased market concerns about a slowdown in AI infrastructure investment, and the stock price rebounded by about 17% at one point after the earnings report.

2. The other side of improving growth is financial pressure: third-quarter net interest expense was $846 million (versus $310 million in the same period last year), and the net loss widened to $709 million; the 62% adjusted EBITDA margin cannot cover depreciation, capital expenditure, and interest, reflecting the nature of its "asset-heavy, strongly financing-dependent" infrastructure operations rather than a purely high-margin cloud service.

3. The core valuation metric is cash generation capability after financing: the company's 2026 capital expenditure guidance is $37 billion to $39 billion, and it expects to receive about $9 billion in customer prepayments and deposits over the next three months; investors need to track the conversion speed of the order backlog, the conversion of adjusted profit into cash, whether financing costs can decline with scale, and the volatility risk brought by customer concentration.

1. On January 13, 2026, Microsoft released the "Community-First AI Infrastructure" initiative, pledging that data centers will bear the cost of their own electricity use and related grid infrastructure, not seek local property tax abatements, replenish more water than they withdraw, and train and hire local residents; in essence, it is an attempt to write "whoever benefits, pays" into the rate design for very large users in exchange for local social license.

2. Data from the International Energy Agency shows that global data center-related electricity generation will rise from 460 TWh in 2024 to more than 1,000 TWh in 2030 and 1,300 TWh in 2035, with renewable energy meeting nearly half of new demand over the next five years; because new load is concentrated in a few regions with land, network and grid-connection conditions, the impact on local grids, permitting and residential electricity prices is far higher than the global average, and corporate purchases of green power certificates cannot replace local real-time supply capacity.

3. The biggest commercial impact of the initiative is to move public infrastructure costs back onto Microsoft's balance sheet: upfront capital expenditure and long-term fixed costs rise, the cash recovery cycle lengthens, and costs may be passed through to Azure's utilization and compute pricing; if the industry widely adopts "whoever adds load bears the cost," some marginal projects will be delayed, and the AI compute growth curve may shift from capital-driven to demand-driven, with success depending on contract details such as minimum payments, long-term capacity reservations and delay liability rather than on principled commitments.

XII. Amazon Japan's algorithmic price limits ruled illegal

1. Japan's Fair Trade Commission (JFTC) formally approved Amazon Japan's remediation plan on November 11: it may no longer use algorithms to exclude sellers whose prices are higher than the lowest price for the same product on other major platforms from the "Buy Box"; when no Buy Box offer meets the price conditions, it must display on the product page a purchasable offer including shipping at the lowest price and an entry point, and this mechanism will last five years; it is behavioral remediation rather than a fine.

2. The investigation began on November 26, 2024, and is the third time Amazon Japan has been targeted by the JFTC (2016 involved pricing clauses, 2018 involved demanding "cooperation fees" from suppliers); according to the Japan External Trade Organization (JETRO), Amazon Japan's share reached 28.2% (fiscal 2021), and the regulatory difficulty lies in platforms hiding coercive intent inside algorithms and replacing contractual terms with "automatic system determination."

3. Global platform antitrust is shifting from fines and breakup threats to behavioral remedies: the U.S. FTC sued Amazon in 2023, the EU Digital Markets Act imposes ongoing obligations on "gatekeepers," and China's State Administration for Market Regulation issued fines over "choose one of two," all targeting single instances of conduct; the JFTC constrains algorithms through a five-year verifiable and auditable long-term mechanism obligation, which will continue to affect Amazon's traffic allocation, seller structure and logistics revenue.

XIII. Musk: SpaceX has ordered HJT equipment with an annual capacity of 100GW from Singulus to build an "intelligent sun"

1. Order event and scale: On November 10, 2026, SpaceX signed a turnkey supply contract with German equipment maker Singulus for a complete heterojunction (HJT) solar cell production line, and Musk stated explicitly on X that it is for an annual capacity of 100GW, with the first production line to be delivered in the first half of 2027 and located at the Bastrop factory in Texas. According to the IEA, global new PV installations exceeded 600GW in 2025, with China at nearly 370GW; this 100GW capacity ceiling is equivalent to about one-sixth of global annual additions and nearly one-third of China's.

2. The Starmind project and HJT selection: Starmind is SpaceX's space AI compute satellite network, and it has applied to the FCC to deploy up to 1 million satellites in sun-synchronous orbit; the AI1 satellite has a wingspan of 75 meters, average power consumption of 175kW and peak of 250kW. HJT was chosen because of its high conversion efficiency, radiation resistance and thin-wafer capability, matching space's stringent requirements for efficiency and weight. SpaceX provided funding support to Singulus in March 2026 and added a subordinated guaranteed loan in October, showing a vertical integration path of "invest first, order later."

3. Industry impact: Singulus had revenue of 48.3 million euros in 2025, and the value of the first production line alone exceeds 60 million euros; its announcement positions itself as "the leading Western supplier of next-generation solar cell technology," reflecting SpaceX's supply chain security considerations against the backdrop of a global PV supply chain concentrated in China. Brokerage research reports believe that p-type HJT is best suited to space scenarios due to its radiation resistance, thin-wafer capability and low silver consumption, and that space data centers may become the second growth curve for the HJT technology route.

Domestic progress:

XIV. Hong Kong-Zhuhai-Macao Bridge port cross-border e-commerce exports exceed 100 billion yuan

1. Scale and growth: In the first 10 months of 2026, exports through the customs cross-border e-commerce management platform at the Zhuhai Highway Port of the Hong Kong-Zhuhai-Macao Bridge reached 112 billion yuan, up 33.1% year on year, exceeding the full-year 2025 total two months early (in 2025, cross-border e-commerce exports at the port were 110.98 billion yuan, up 92.5%); average daily export parcels rose from about 1.6 million in 2025 to more than 3 million, with parcel volume growing faster than cargo value, per-item value declining, and orders becoming fragmented and high-frequency.

2. Mechanism and geography: Growth relies mainly on the 9610 "list-based release, consolidated statistics" regulatory model, which has reduced enterprises' customs clearance costs by an average of more than 30% and improved timeliness by about 50%; in the six years since the Zhuhai cross-border e-commerce comprehensive pilot zone was approved, its scale has grown more than 166-fold, and logistics in the Pearl River Delta has shifted from "eastern goods exiting east" to "western goods exiting west," with Cainiao, Shein, JD.com and Pinduoduo having already set up warehousing and logistics facilities in Zhuhai and Hengqin.

3. Pain points and outlook: Starting April 1, 2026, customs nationwide will promote cross-customs-area returns for cross-border e-commerce retail exports, which, together with "TIR + cross-border e-commerce" transport and "pre-positioned regulatory warehouses," addresses the three major pain points of difficult returns, slow transport and long customs clearance; however, the channel is highly dependent on the aviation capacity and route policies of Hong Kong and Macao airports, and faces regional competition from Shenzhen and Guangzhou Baiyun Airport, among others; full-year 2026 bridge cross-border e-commerce exports are expected to reach about 130 billion yuan.

Open source trends:

XV. Nvidia donates OpenShell to open source, and its aim is not agent security

1. Event and strategy: On November 10, 2026, at the KubeCon North America conference in Salt Lake City, Nvidia launched three sets of moves at once—donating the agent security runtime OpenShell to the CNCF Sandbox, contributing code to Google-led Agent Substrate and integrating it with Dynamo inference scheduling and KAI Scheduler, and adding three open-source components, NVKE, NVKR and Snapshot, to the DSX AI Factory Platform—aiming to move the main competitive battlefield from GPU chips up to the orchestration layer of AI factories.

2. Layered implementation: At the governance layer, OpenShell provides deterministic control of agent behavior, and officials say it has been deployed at hundreds of organizations; at the scale layer, Agent Substrate runs millions of agents on Kubernetes at about 10x density and forms a scheduling loop with Dynamo and KAI Scheduler; at the foundation layer, NVKE and NVKR are in public preview, and Snapshot compresses inference replica recovery to 3.6 seconds to 42.7 seconds, 8 to 36 times faster than cold start, supporting 1TB checkpoints across 8 GPUs.

3. Impact and landscape: Nvidia trades open source for ecosystem, shifting its moat from "closed code" to "ecosystem dependence"; the cloud-native ecosystem and developers benefit from expanded choice, while vendors that levy a "closed tax" on the agent runtime and scheduling layers through proprietary protocols come under pressure; under Google-led Agent Substrate, the two sides share an interest in making Kubernetes the AI foundation, and the direction of competition and cooperation depends on whether profits ultimately flow to the chip layer or the orchestration layer.

Source

TMTPost · AGI