❯ Anthropic Reportedly Plans October Listing; Investors See $2 Trillion Valuation, an IPO Record

[QUIET-PERIOD PRICE] The Financial Times, citing six investors in Anthropic, reports that the company expects to list in October, with valuation expectations starting at $2 trillion — surpassing SpaceX, which listed in June at $1.77 trillion, and becoming the largest IPO in history. The company confidentially filed its IPO documents with the U.S. Securities and Exchange Commission in June and is currently in the quiet period, having raised nearly $100 billion year-to-date.

[REVENUE SLOPE] Underpinning that figure is the revenue curve: annualized revenue has climbed from $1 billion at the end of 2024 to $47 billion in May, and investors expect it to reach $100 billion to $120 billion by December — more than 10x growth for the full year. After the latest funding round in May, the valuation had already reached $965 billion, at one point surpassing OpenAI.

[CAVEAT] Note the distinction: the $2 trillion figure is an investor expectation, not a company quote. According to the report, Anthropic executives have given no valuation target even in private discussions; the final price will be hammered out between underwriters and institutional buyers during the roadshow.

[PRICING ANCHOR] Should it materialize, this would be the first time the public market has put a price tag of this size on a pure-play model company, and all AI companies’ primary and secondary market valuations would take it as a benchmark from then on. Fund managers holding stakes in OpenAI and xAI will have one more tradable anchor at hand when they next revalue their positions to fair value.

▪ SIGNAL The $2 trillion is a psychological price level investors backed out from the revenue slope; the October pricing window will test whether that slope still holds in the public market.

❯ Bloomberg: OpenAI Annualized Revenue Tops $40 Billion, Doubling From End of 2025

[PACE] Bloomberg, citing people familiar with the matter, reports OpenAI’s annualized revenue has surpassed $40 billion, roughly doubling from the end of 2025. President Greg Brockman said in an internal note that in July alone, annualized revenue expanded more than 20% month over month. Bloomberg places this acceleration in the same context: the company is building momentum for an anticipated IPO.

[REVENUE] Growth isn’t driven by ChatGPT subscriptions alone: per Bloomberg, the advertising business has barely started yet is already contributing revenue, with the coding agent Codex and enterprise-facing ChatGPT Work named as the two specialist engines. Previously, OpenAI’s revenue was almost entirely staked on consumer subscriptions as a single line — this is the first time it is simultaneously holding advertising and enterprise software as two cash-flow streams, and the roughly $20 billion annualized base at end-2025 was precisely the starting point for this doubling.

[MATCHUP] Read alongside the previous item: OpenAI at $40 billion annualized, Anthropic expected by investors to reach the $100-billion scale within the year — the gap narrative between the two is reversing, and the first comparable financial statement Wall Street receives will come from whichever files its prospectus first. Microsoft, Amazon, and Google Cloud, which supply computing power to both, will also find their bargaining leverage in the next round of compute contracts hinged on these two revenue curves.

▪ SIGNAL Both leading labs are pushing revenue figures into the media ahead of their listings; the main battleground of the fundraising narrative has shifted from model capability to the income statement.

❯ Google Launches Gemini 3.7 Flash, Cuts Intro Price in Half, Targets Coding and Agents

[LEAD] Google launched Gemini 3.7 Flash on August 13, just 3 weeks after 3.6 Flash debuted. Intro pricing is $0.75 per million input tokens and $3.75 per million output tokens — half the prior generation’s launch price. The rate holds through the end of the year, then reverts to $1.50 and $7.50 in 2027. Product lead Tulsee Doshi called it Google’s “most intelligent flagship model,” aimed at coding and agent use cases.

[IMPACT] The gains are concentrated in coding and automation. Per official results: DeepSWE v1.1 rose from the prior generation’s 49.0% to 65.3%, FrontierCode 1.1 improved from 34.4% to 43.6%, and AutomationBench nearly doubled, from 17.0% to 30.4%. Third-party benchmark firm Artificial Analysis gave it an intelligence index of 56 — 4 points higher than the model released three weeks ago — and placed it on the “intelligence vs. latency” Pareto frontier.

[CONTEXT] Over the past 3 months, Google had already shipped two Flash models; this is the third. Developer Simon Willison flagged the awkward pricing design: the intro price is set to double on December 31, yet at a three-week iteration cadence, who will still be using this generation five months from now? The answer, in all likelihood, is no one. Google only wants today’s adoption.

[DETAIL] What gets squeezed is the mid-tier price band: Claude Sonnet, GPT-5.6 Terra, and a host of open-source models in the same class all need to justify why they cost more. Teams building agent applications will have to recalculate their per-task cost comparison tables this week.

▪ SIGNAL A release every three weeks, a price halved with each release — Google is wielding iteration speed as a weapon, turning mid-tier model margins into a war of attrition.

❯ DeepSeek Releases V4-Pro at One-Seventh the Price of Kimi K3

[EXTREME PRICING] DeepSeek released its strongest model, V4-Pro, on August 13, with API pricing at $0.435 per million input tokens and $0.87 per million output tokens — versus $3 and $15 for Kimi K3, putting its input price at only about one-seventh of the rival’s. The new version is built around agentic capability upgrades: reasoning effort can be toggled across Low, High, and Max tiers, and it natively supports OpenAI’s Responses API, claiming one-click integration with Codex.

[BENCHMARKS] The capability gap remains: per Artificial Analysis, V4-Pro scores 53 on the intelligence index versus 60 for Kimi K3, with the latter leading on several coding benchmarks. But on the firm’s per-task cost metric, V4-Pro is about 93% cheaper than K3 — K3 averages $0.84 per task, while V4-Pro costs just a few cents. Every previous DeepSeek generation courted developers with “good-enough performance, disruptive pricing,” but this generation’s price gap is the widest yet.

[ANOTHER HAND] Another hand is at work behind the low price: per VentureBeat, V4-Pro’s launch came with API price adjustments, with cache-hit pricing rising more than tenfold during peak hours. For agentic applications that reuse context heavily, the actual bill may not be lower. Sticker prices down, cache prices up — DeepSeek is shifting its revenue structure toward heavy users.

[RECALCULATION] The low-price, volume-grabbing route hasn’t changed, but this time developers no longer calculate the sticker price — they have to calculate the real bill under their own call patterns. For agent teams with a high cache-hit ratio, the migration decision has only gotten more complicated: first re-run the past month’s call logs through the cache ratios to recalculate costs, then decide whether to stay or go.

▪ SIGNAL List price cut to a tenth, cache prices up tenfold — DeepSeek has turned a price war into a price-structure war and shifted the full burden of price comparison onto developers.

❯ DeepSeek open-sources agent framework Harness with an “everything is a plugin” philosophy

[STAR SURGE] On the same day it released its model, DeepSeek open-sourced the agent framework DeepSeek Harness v0.1 under the MIT license, opening it to developers worldwide for testing. GitHub stars crossed 33,000 within hours of the repo going live, and it can be launched with a single npx command.

[ARCHITECTURE] The core thesis is “everything is a plugin”: models, tools, skills, sessions, sandboxes, file systems, loops, orchestration, even the UI — all are implemented as replaceable, recombinable plugins. Underneath sits a meta-framework called Cordis, with no privileged kernel that plugin authors are forced to go through. The team was assembled in May, invited open-source developers to beta test in early August, and officials state plainly that the current version will include breaking changes and is not yet a production-grade platform.

[ECOSYSTEM] VentureBeat immediately cast it as the open-source rival to Claude Code. In light of the cache-pricing changes in the previous item, the intent is easy to read: price models down to the floor, then make the money back from the heavy agent usage that grows up around Harness. Startups building coding-agent toolchains need to re-measure their moat against a free MIT-licensed foundation. The next variable to watch is the growth rate of the plugin ecosystem — how many of those 33,000 stars convert into active plugin authors will determine whether this move pays off.

▪ SIGNAL With models free and frameworks open-sourced, the truly scarce assets left in the agent race are usage entry points and dominance of the plugin ecosystem.

❯ OpenAI Previews Ultrafast Service Layer, GPT-5.6 Sol Up to 14x Faster

[SHIFT] OpenAI is previewing a new API service layer, Ultrafast: powered by Cerebras wafer-scale chips, it runs GPT-5.6 Sol at output speeds up to 750 tokens/second, up to 14x faster than standard processing, with intelligence on par with the standard version. It is currently in limited release to a select group of customers, with gradual rollout as capacity expands.

[HARDWARE] The speed comes from Cerebras’s wafer-scale engine architecture: each wafer-sized chip carries 44GB of on-chip SRAM, keeping model weights resident on-chip and bypassing the memory-bandwidth bottleneck of conventional inference hardware — previously, speeds like this appeared only in small open-source models; this is a first for a frontier flagship. According to benchmarks officially released by the two companies: the full 2,500-question “Humanity’s Last Exam” took just 11 hours to complete, while the same exam took Claude Fable 5 more than three days; on the knowledge-work benchmark GDP-Val, end-to-end speedup is 5.6x.

[SUPPLY] In agent scenarios where a single task strings together dozens of model calls, latency itself is the product — this tier is built for them. The bigger signal is in the supply chain: OpenAI has moved production inference for the frontier model onto non-Nvidia chips, and what Cerebras gets is a flagship-model endorsement, worth more than any benchmark score.

▪ SIGNAL For the first time, a frontier model runs on non-Nvidia chips as an official service layer — opening a gap in the procurement landscape for inference hardware.

❯ Anthropic Reportedly in Talks to Acquire Israeli Startup Decart for $6 Billion

[PRE-IPO RESTOCK] Bloomberg, citing people familiar with the matter, reported that Anthropic is in talks to acquire Israeli startup Decart for roughly $6 billion — which, if completed, would be its fifth acquisition this year and its largest ever. The transaction is not finalized and could still fall through.

[TARGET PROFILE] Decart was founded in 2023 by the Leitersdorf brothers and a third co-founder. It has two product lines: a world model that generates interactive video in real time, and optimization software that lifts GPU utilization and lowers training costs. In May, it closed a $300 million round led by Radical Ventures, with NVIDIA and Adobe Ventures participating, at a valuation near $4 billion — making the $6 billion offer a roughly 50% premium.

[WHAT IT BUYS] Per the report, if the deal closes, Decart’s team would be folded into Anthropic’s inference and performance division. What is being bought is not the world-model narrative but the engineering capability to squeeze more output from existing compute. Set against a $100 billion revenue expectation and an October IPO timeline, the money buys the “gross-margin improvement path” page of the prospectus.

▪ SIGNAL A $6 billion spend on a GPU-optimization team shows that, for leading labs, saved compute now prices better than new compute.

❯ CXMT Overtakes Tencent to Top China Market Cap 17 Days After Listing

[FIRST-EVER] Memory chipmaker ChangXin Memory Technologies (CXMT) reached a market cap of $524 billion, surpassing Tencent’s $511 billion and claiming the No. 1 spot in China just 17 days after listing—the first time in 35 years of mainland China’s stock market that a semiconductor company has taken the top seat.

[IPO FRENZY] The company listed in Shanghai on July 27, raising $8.6 billion, with retail subscriptions oversubscribed more than 212 times and a first-day surge of 466%. CXMT is the world’s fourth-largest DRAM maker, behind SK Hynix, Samsung, and Micron. It plans to build a sixth major fab and aims for 30% of global DRAM share by 2030.

[SHIFT] The other half of the overtaking story is Tencent: its Q2 report showed capital spending on AI infrastructure surged 176%, sending the stock lower. For the same AI spend, the market is pricing chip sellers and chip buyers in opposite directions—in the memory up-cycle, profits are moving upstream, and A-share investors have voted with real money.

[VALUATION] The hard question is ahead: a $524 billion market cap still corresponds to a catch-up player’s capacity and technology gap. Once the memory cycle peaks, CXMT’s share price will have to be supported by progress toward that 30% share goal by 2030, and the valuation ceiling for A-share semiconductor names will rise and fall with it.

▪ SIGNAL China’s No. 1 market cap has shifted from a platform company to a memory maker, as capital’s AI pricing focus moves from the application layer to physical capacity.

❯ SK Hynix Bets $720 Billion on World’s Largest Memory Chip Cluster

[MEGA BLUEPRINT] CNBC’s first on-site visit to SK Hynix’s Yongin cluster: total investment of about $720 billion — 600 trillion won poured into the Yongin semiconductor cluster, plus another 100 trillion won for the Cheongju expansion — the company says this will be the world’s largest memory fab network, with the first fab entering production in February 2027, three months ahead of the original May schedule.

[RECENT MOVES] Turning to the near-term actions: a week earlier, the board had just approved $38.1 billion to build two new fabs — the Yongin Y2 plant with 35.2 trillion won in investment, its first cleanroom operational in June 2029, focused on HBM and advanced DRAM; the Cheongju M17 plant breaks ground in February 2027, with its first cleanroom opening in December 2028. AI compute demand pushing memory prices steadily higher is the direct confidence behind this full-throttle expansion.

[COLLISION] Read against the previous item: CXMT is calling for 30% share by 2030, and SK Hynix answers with $720 billion in capacity. Both sides’ production windows are pinned to 2027 to 2029 — when the HBM supply-demand balance flips, the answer lies in these fabs’ ramp-up curves, and memory customers’ long-term contract negotiations must start picking sides now.

▪ SIGNAL Chinese and Korean memory fabs are simultaneously maxing out capex; the risk of HBM tipping from shortage to oversupply after 2027 is already written into both sides’ construction timelines.

❯ Alibaba Adds Commercial Licensing Threshold to Open-Source Flagship Qwen3.8-Max

[LICENSING] Alibaba’s flagship open-source model Qwen3.8-Max is no longer releasing its weights under Apache 2.0, switching to a custom license: companies operating “Model-as-a-Service” or “AI work assistant” businesses — and their affiliates — with revenue exceeding $50 million in any consecutive 12-month period must obtain a separate commercial license from the Tongyi Qianwen team before using the model or its derivatives.

[TERMS] According to the South China Morning Post, the fine print has two more layers: products with more than 100 million monthly active users or more than $20 million in monthly revenue must display model attribution; companies with annual revenue below $50 million can download and use the model commercially for free, and purely internal use is likewise exempt as long as the model’s capabilities are not exposed to third parties. In contrast to the previous Qwen series, which was uniformly released under the permissive Apache 2.0, this is a clear strategic U-turn.

[MONETIZATION] Weights still ship; big customers pay — the line Alibaba has drawn is aimed squarely at inference cloud providers and wrapper services that build their business on redistributing open-source models. Model-hosting platforms that owe their start to Qwen now carry a new line item in their cost model: licensing-fee risk. And for downstream enterprises doing model selection, the word “open source” will from now on require reading the license before drawing conclusions.

▪ SIGNAL China’s era of free open-source models is now tiered by revenue — open weights and commercial licensing are henceforth two separate conversations.

❯ Apple Proposes Up to 15% Commission on External Link Purchases; Epic Says It Should Be Zero

[FORCED FILING] In the Epic v. Apple case, Apple on August 13 submitted its U.S. external-link purchase commission proposal to the court per Judge Gonzalez Rogers’ instruction: 15% for standard apps, 10% for partner programs covering video, news, and subscription renewals, and 5% for Small Business Program apps — corresponding to the original 30% in-app purchase commission.

[COMPARISON DEFENSE] The backdrop: the judge previously issued an injunction requiring Apple to let developers steer users to external purchases; Apple’s appeal to the Supreme Court for a stay was denied, and on August 11 its request to delay submitting the rate proposal was rejected as well. Apple’s core defense is peer comparison: Google Play charges 20%, 15%, and 10% in three tiers for external links, “and Epic accepted those rates” — implying its own 15% is already the low end of the market.

[ZERO-FEE CLAIM] Epic immediately pushed back: under the Ninth Circuit’s “necessary costs” definition, Apple should charge 0% for external link purchases. Next, the judge will examine whether Apple’s cost evidence can support these three rate tiers, and Apple must file its brief with the Supreme Court on September 14. Developers’ external link pricing strategies can only truly begin once both proceedings conclude.

▪ SIGNAL The dispute has narrowed from “whether external links are allowed” to “how many percentage points the cost evidence can justify,” and Apple’s services business profit margin rides on this evidentiary question.

❯ JD.com Q2 Revenue Dips 2.9%, Non-GAAP Net Profit Beats with 20.8% Growth

[REVENUE DIP, PROFIT RISE] JD.com posted Q2 revenue of RMB 346.4 billion (about USD 51.1 billion), down 2.9% year-on-year, a decline the company attributed to a high comparison base; non-GAAP net profit was RMB 8.9 billion, up 20.8% year-on-year, beating analyst expectations.

[PROFIT DRIVERS] According to the filing, the two sources of profit improvement are specific: food-delivery losses were halved and JD Retail margins expanded; meanwhile, R&D spending rose 40% year-on-year, with funds shifted from the food-delivery subsidy war of the past year-plus toward technology investment. CEO Sandy Xu called it a “clear inflection point” in the company’s profit trajectory.

[MARKET REACTION] Shares fell about 4% after the report — the profit repair was acknowledged, but concerns on the revenue front remain. Food delivery has pivoted from burning cash for market share to narrowing losses; in the second half, the subsidy pace of rivals Meituan and Alibaba is the swing factor.

▪ SIGNAL JD’s revenue-down, profit-up report signals the delivery war is cooling; the relay question now is what will restore growth to the core retail base.

❯ Didi Q2 core orders surpass 5 billion, international transaction value jumps 61%

[PROFIT] Didi’s core platform orders in Q2 reached 5.052 billion, up 13.2% year over year, with total transaction value of RMB 133.9 billion, up 22.2%; Bloomberg reports revenue rose 11% to about $9.3 billion, with net profit of approximately $128 million, ending two straight quarterly losses.

[TWO CURVES] According to the company’s earnings report, China mobility transaction value reached RMB 90.4 billion, up 9.5% year over year, marking 14 consecutive quarters of growth; international business orders totaled 1.403 billion, up 29%, with transaction value surging 61%, covering 14 countries and regions and serving over 100 million users, after sustained profitability in the first half.

[TRACK SHIFT] With domestic growth settling into a single-digit steady state, international operations have become Didi’s valuation elasticity—a 61% transaction growth rate set against the penetration headroom in Brazil and Mexico. Four years after delisting, Didi once again has a growth story to tell the capital markets.

▪ SIGNAL For the first time, the structure of stable domestic operations and aggressive overseas expansion has simultaneously landed in the earnings report. Didi’s next open question is when it will return to a mainstream exchange.