❯ Anthropic Q2 Revenue Tops $11.5 Billion, Adjusted Operating Profit Turns Positive for First Time

[LEAD] Anthropic disclosed to investors that Q2 revenue exceeded $11.5 billion — up more than 14x from $787 million in the same period of 2025 and more than double the $4.73 billion in Q1 — while adjusted operating profit turned positive, according to Bloomberg. It marks the company’s first profitable quarter at the operating level, and the figures land just as investment banks price its IPO.

[BEAT] The scorecard beat the company’s own internal numbers. The internal forecast leaked in May had projected Q2 revenue of $10.9 billion, adjusted operating profit of $559 million, and a margin of about 5.1%; actual revenue came in roughly $600 million above that projection. Separately, as previously reported, the company’s annualized revenue run rate crossed $47 billion in May, with enterprise customer expansion the primary driver and Claude’s coding product line contributing the bulk of incremental growth.

[PRICING] Anthropic filed its IPO application with the SEC in June and reportedly plans to list in October, with investors discussing valuation levels as high as $2 trillion — which, if realized, would surpass SpaceX to become the largest IPO in history. On annualized Q2 revenue of roughly $46 billion, $2 trillion implies a price-to-sales ratio above 40x; Fortune has already questioned whether fundamentals can support that number. The bull-bear divide will be settled by book-building results in October.

▪ SIGNAL The earlier-than-expected breakeven — more than the revenue doubling — is what holds up the $2 trillion pricing table.

❯ SpaceX Closes $60B Acquisition of Cursor, Team Merges into SpaceXAI

[CLOSING] Securities filings show SpaceX has completed its $60 billion all-stock acquisition of Anysphere, the parent of AI coding tool Cursor. Anysphere shareholders received approximately 389.3 million SpaceX shares in exchange — just two months after the deal was officially announced in June. Both sides call it the largest startup acquisition by value in history.

[INTEGRATION] The Cursor team is being folded wholesale into SpaceXAI, deployed across four product lines — Grok Build, Grok Bot, Grok API, and Cursor itself — and gaining compute access to the Colossus supercomputer. As previously reported, Morgan Stanley estimates the deal could add up to $13 billion in revenue for SpaceX by 2027. The closing also doubled the personal wealth of Cursor’s two co-founders, tracking the rise in SpaceX shares.

[LANDSCAPE] The big three AI coding tools now all have a giant backer: Claude Code is backed by Anthropic, Codex by OpenAI, and Cursor has secured SpaceX’s compute and distribution. For coding-tool companies still raising independent funding, the story has been flattened — the first question investors will now ask is what you can compete on against these three.

▪ SIGNAL The decisive factor in coding tools has shifted from product experience to the compute and parent-company distribution behind them.

❯ Zhipu Unveils GLM-5.3: Cybersecurity Benchmark Tops Mythos 5, Weights Delayed Two Weeks

[KEY POINTS] Zhipu released GLM-5.3 on August 14: the base reuses GLM-5.2’s 743B model completely unchanged, with all gains coming from scaled-up post-training; the company self-reports 84.5% on the CyberGym cybersecurity benchmark, slightly above Anthropic Mythos 5’s 83.8% (not independently verified), with open-source weights to follow in two more weeks.

[RATIONALE] The delay is not a capacity issue. Zhipu says the model’s exploit-chaining capability exceeded expectations during training, so safety assessment and hardening must be completed first; the most sensitive cybersecurity features are only open to users verified through the “Trusted Access Program”. The company also disclosed that during model testing it found a “potentially serious vulnerability” in the coding tool Cursor — a type of finding that has typically come from professional security teams, not model vendors.

[REACTIONS] Independent research firm SemiAnalysis said it “significantly surpasses all U.S. open-source models”; Allen Institute researcher Nathan Lambert, meanwhile, cautioned that the scores carry the usual “benchmark optimization” controversy. Enterprise security teams have one more thing on their plate starting today: the attack-surface baseline for evaluating open-source models must be redrawn around GLM-5.3; this will also affect procurement decisions for security products — the defensive toolbox can put it to use immediately.

▪ SIGNAL Post-training squeezes frontier capability out of the same base model; pre-training is no longer the only ticket to catch up.

❯ Anthropic risk report reveals stronger internal Model 2, says no plans to release it publicly

[DISCLOSURE] Anthropic’s latest risk report discloses that an internal model codenamed Model 2 shows “clear improvements” over flagship Mythos 5 on most internal tasks, but the company states it has “no current plans for external release”; the same report also upgrades the risk rating for model inaccuracy in high-stakes scenarios from “very low” to “low,” citing recent cybersecurity incidents.

[INTERNAL] According to the report, Mythos 5 and Model 2 are already being “heavily” used internally for coding, agent tasks, and data generation, though this performance jump is a smaller step than the one from Opus 4.6 to Mythos earlier this year; Axios followed up and confirmed the report’s contents. The company also says it has observed the model’s automated R&D capabilities accelerating — which can speed up technical progress, and equally so if it falls into the wrong hands.

[ASYMMETRY] When the strongest model is kept in-house and never released, the outside world’s capability coordinate system begins to distort: competitors calibrate their catch-up targets against the released version, customers evaluate purchasing decisions against the released version, while the true frontier sits one notch further inside the lab. The leaderboards drawn up by evaluation bodies are really measuring “the part each company is willing to put out,” so every estimate of the frontier gap now carries an extra haircut.

▪ SIGNAL Frontier labs’ capability disclosure is turning into selective disclosure — leaderboards can’t measure the true frontier.

❯ Alibaba Open-Sources Qwen3.8 Weights: 27B Multimodal Model Runs on Consumer GPUs

[OPEN SOURCE] Alibaba’s Tongyi team released the full Qwen3.8 weight suite under the Apache 2.0 license: the headline Qwen3.8-27B is a natively multimodal dense model that officially beats the larger Qwen3.7-Plus overall, especially in real-world coding and office workflows; the flagship Qwen3.8-2.4T-A95B (2.4 trillion total parameters, 95 billion active) opens in tandem.

[LOCAL] The 27B carries a 262K-token native context, expandable to 1M, and handles images, documents, and long video; target hardware is a high-end consumer card with 24GB VRAM, and the quantized build runs in 17GB of RAM. The Qwen open-source family already holds the No. 1 share in the local inference market; on release day, more than ten inference platforms completed integration, and developer Simon Willison, after hands-on testing, called its generation quality the best tier among the local models he has seen.

[RACE] Landing the same day Zhipu AI’s GLM-5.3 swept the feeds, the two Chinese labs each staked a claim within 24 hours — one on “strongest open-source coding,” the other on “the most capable local small model.” Enterprises buying APIs must redraw the build-versus-buy line, and developers assembling local rigs have to redo their price-comparison tables this week — the inference-cost line is being pushed down by both in tandem.

▪ SIGNAL The 27B beats its own larger predecessor — capability density is improving faster than parameter stacking.

❯ OpenAI Annualized Revenue Breaks $40 Billion, Doubling From End of Last Year

[KEY POINTS] Bloomberg reports that OpenAI’s annualized revenue run rate has broken through $40 billion, roughly doubling since the end of 2025; President Greg Brockman said in an internal memo that in July alone the run rate rose more than 20% month over month. The growth engines are subscriptions, early-stage advertising, the Codex coding agent, and the ChatGPT Work enterprise product.

[COMPARISON] Rival Anthropic previously disclosed a run rate of $47 billion in May, but the two private companies use different statistical calibers, so the figures cannot be directly compared. While revenue sprints, management is bleeding: according to The Information, Chief Revenue Officer Denise Dresser is leaving after just 8 months in the role — the second executive to depart this week — with the company at a critical window in its race to go public.

[IPO RACE] These figures hit the press on the same day as Anthropic’s $11.5 billion quarterly report, with both companies sprinting toward an IPO. On one side, revenue is doubling; on the other, executives are departing in quick succession. When institutional investors price OpenAI, they must pick one of these two columns as the anchor — executive stability, for the first time, has become an input on the valuation model alongside the revenue curve.

▪ SIGNAL Revenue doubling and executive departures flash on the same screen — OpenAI’s IPO story is harder to price than the numbers themselves.

❯ Nvidia Filing Discloses ~$21B Stake in SpaceX, $30B in Intel Shares

[HOLDINGS] Nvidia’s latest 13F filing shows that, as of the end of June, it held 122.8 million shares of SpaceX Class A stock worth about $21 billion — its second-largest position; the largest is approximately $30 billion in Intel shares. The filing also shows the company committed up to $2 billion to Musk’s xAI and $5 billion to Intel in 2025.

[ORIGIN] The SpaceX stake was not bought directly: Nvidia originally took the equity stake to lock in xAI’s chip purchases, and in February this year SpaceX absorbed xAI in a merger at a $1.25 trillion valuation, converting Nvidia’s position into SpaceX stock. SpaceX’s earlier listing gave such holdings a public market price for the first time, and Alphabet and AMD — which filed their 13Fs the same day — also surfaced their respective SpaceX holdings.

[CIRCULARITY] The circular investment — chipmakers taking equity in leading customers, customers spending the proceeds on chips — is now written straight into the 13F. Nvidia’s balance sheet is tied to the capex of the top AI buyers on the same rope; if any downstream player stalls, it hits both revenue and investment income line items at once. From this quarter on, analysts modeling Nvidia will need to pull one more holdings table into the model.

▪ SIGNAL How much of Nvidia’s downstream demand comes from customers it itself bankrolls — the 13F gives the first verifiable accounting.

❯ Apple Teams Up with Alibaba to Train China-Dedicated LLM, May Become First Foreign Firm Cleared in China

[EXCLUSIVE] Reuters, citing three people familiar with the matter, reports that Apple — with Alibaba’s backing — has been training its own large language model for the Chinese market. If approved, it would make Apple the first foreign company authorized to offer its proprietary AI model in China — a course reversal from its earlier strategy of borrowing external models in the country.

[COMPLIANCE] The partnership was publicly confirmed by Alibaba Chairman Joe Tsai in February 2025. Last month, China’s cyberspace regulator approved the integration of Qwen into China-market Apple Intelligence, covering iPhone, iPad, Mac, and Vision Pro, with Baidu’s technology also part of Apple’s AI offering in China. The feature’s rollout had been repeatedly postponed for compliance overhauls, while Huawei steadily gains ground in the premium handset segment, keeping Apple’s China share under sustained pressure.

[TEMPLATE] When China-market iPhones will be fully equipped with AI features directly shapes assessments of Apple’s upgrade cycle in China. For other foreign firms, once this “self-trained model + local partner” approval path is navigated successfully, it becomes a ready-made template to copy — Google and Meta’s China teams should be studying this approval precedent point by point this week before deciding whether to follow.

▪ SIGNAL The compliance pathway for foreign companies building generative AI in China now has its first working example — and the template’s value outweighs Apple’s own sales figures.

❯ Saudi Sovereign Fund Builds SpaceX Position in Q2, $26.3B Becomes Its Largest US Holding

[POSITION] Saudi Arabia’s Public Investment Fund (PIF) 13F filing shows it initiated a new position in SpaceX in Q2, with an end-of-quarter market value of approximately $26.34 billion — one of the largest holdings on the disclosed list; it also held around $5.09 billion in Electronic Arts, $5.26 billion in Uber, $1.18 billion in Lucid, and about $43.7 million in Claritev.

[JUMP] The sovereign fund, managing over $900 billion, had been shrinking its US equity portfolio in recent years and shifting capital back to domestic projects; its entire US stock portfolio previously stood at only about $12 billion — a single SpaceX position more than doubled the portfolio.

[ENTRY] SpaceX’s public listing gave the sovereign fund a one-shot public-market gateway to round out its “AI + space” exposure, and Saudi Arabia bought it straight into the top holding. Whether other sovereign funds follow into the same target will directly affect SpaceX’s valuation support — next quarter’s 13F will have the answer.

▪ SIGNAL Sovereign funds no longer route AI exposure through funds and startups — they buy the largest target directly in the public market.

❯ Pony.ai and Uber Expand Partnership, Will Deploy Over 2,000 Robotaxis in Europe

[EXPANSION] Pony.ai and Uber announced an expanded partnership on August 14, planning to deploy more than 2,000 robotaxis across Europe: extending from the existing commercial service in Zagreb to four additional European cities, then into the Middle East. Specific cities and timelines were not disclosed; the companies said details will be announced in phases.

[FOUNDATION] The Zagreb service launched in late March this year, with the fleet owned and operated by Croatian mobility company Verne. The companies call it Europe’s first commercial robotaxi service. The division of labor is three-tiered: Pony.ai supplies L4 autonomous driving technology, Uber supplies the ride-hailing platform, and the local partner handles day-to-day fleet operations — heavy assets stay local, letting the two companies expand asset-light.

[GAP] Waymo has yet to enter Europe, Tesla’s European approvals remain pending, and Chinese players are seizing the regulatory vacuum to get vehicles on the streets first. European city regulators — which companies they approve next, and how quickly — will decide whether these 2,000 vehicles become a first-mover advantage or a one-off.

▪ SIGNAL On the European robotaxi table, the first to sit down is a Chinese technology provider paired with an American platform.

❯ DeepSeek open-sources agent framework Harness: every component is pluggable

[RELEASE] DeepSeek has open-sourced its agent framework DeepSeek Harness (CLI name: dsh) v0.1 developer preview under the MIT license; a single npm command gets it running. The design philosophy is “everything is a plugin” — models, tools, skills, sandboxes, orchestration loops, and even the UI are all replaceable.

[POSITIONING] The framework is built on its previously released Cordis meta-framework. The repository explicitly warns of breaking changes ahead, and it is not yet a production-grade platform. According to VentureBeat, it is positioned as an open-source counterpart to Claude Code’s underlying infrastructure. Around the same time, DeepSeek also listed a higher-priced V4-Pro on its API, pushing forward both the open-source framework and the paid model tracks in parallel.

[FRONTLINE] The competition in coding agents is moving down from models to the harness layer. Teams building their own agents now have, for the first time, a fully pluggable reference implementation, adding a free option to their shortlist. For commercial agent platforms that charge for closed runtimes, pricing pressure will emerge from this layer first, leaving only model capability gaps as the remaining moat.

▪ SIGNAL Beyond models, the agent runtime is becoming the next layer to be conquered by open source.

❯ Alibaba to Sell Lingxi Games to Trustar Capital at Over $1.5 Billion Valuation

[DEAL] Bloomberg, citing people familiar with the matter, reports that Trustar Capital, the private equity arm of CITIC Capital, has become the preferred buyer for Alibaba’s gaming business Lingxi Interactive Entertainment, in a deal that could be valued at over $1.5 billion. Negotiations are ongoing and no agreement has been finalized.

[ASSET] Lingxi’s pillar is Three Kingdoms Tactics, its highest-revenue flagship. The asset was publicly put up for sale starting in June, with an initial ask of around $1.03 billion. Trustar beat out multiple bidders, including strategic buyers from the gaming industry, lifting the valuation nearly 50% in two months.

[DIVESTITURE] This is the latest in Alibaba’s continued divestiture of non-core assets under Wu Yongming, with the cash and manpower being recovered now concentrated into AI and cloud capex. The same week, Alibaba released the full Qwen3.8 weight family — a sale, a release: the same arithmetic.

▪ SIGNAL Swapping gaming assets for AI ammunition — Alibaba’s portfolio contraction says more than any slogan about where its real business lies.