❯ Anthropic Could Publicly File as Soon as Late August, With Raise Said to Rival SpaceX’s Record

[FILING SPRINT] Bloomberg reports Anthropic is expected to publicly file its listing documents as soon as late August, targeting a raise that matches or exceeds the record set by SpaceX — which raised $75 billion in its initial offering, rising to $86.2 billion after the over-allotment option was exercised, the largest IPO in history. The company confidentially filed with the SEC as early as June, the report says, with an underwriting syndicate comprising Morgan Stanley, Goldman Sachs and JPMorgan.

[REVENUE CURVE] Underpinning that scale is a steep revenue curve. Bloomberg, citing people familiar with the matter, says Anthropic’s annualized revenue reached $65 billion in July, up from $47 billion at the start of the year and only about $10 billion for all of 2025. A May funding round valued the company at $965 billion, already past OpenAI’s $852 billion. The same day, The Information separately reported that the company is preparing super-voting shares for co-founders including Amodei, to hedge against control erosion after multiple rounds of dilution.

[IPO PRICING] What’s really being put on the table is the public pricing of a frontier lab. Until now, every valuation emerged from closed-door bargaining in the private market. After this filing step, the cash-burn curve, gross-margin structure and customer concentration all have to be laid open for the secondary market to see. For fundraisers racing on the same track, Anthropic’s opening performance will directly rewrite the floor for their next round of negotiations.

▪ SIGNAL The two months between the confidential filing and the public filing mark the first time an AI lab is required to explain, under the same accounting standards, exactly how it makes money.

❯ Nvidia Locks Up Poolside License for $6 Billion, 109 Employees Receive Job Offers at Once

[LICENSE & POACH] Newcomer reports Nvidia struck a $6 billion non-exclusive technology license with code-model company Poolside, invested $1 billion at a $12 billion pre-money valuation, and extended job offers to 109 Poolside employees. Poolside’s two founders stressed in an internal memo that this is not an acquisition, nor an acqui-hire, and the company remains independently run by them.

[STRUCTURE REPLAY] The structure is nearly identical to the playbook Nvidia used to secure Groq’s technology license: buy the license, poach the team, keep the shell company, sidestepping the antitrust review a formal acquisition would invite. Poolside, founded in 2023 by former GitHub executives Warner and Kant, hit a setback this April — cloud-compute provider CoreWeave backed out of a 15-year anchor lease and $2 billion in financing for the 2 GW “Horizon” data-center project in West Texas, briefly stalling the project.

[WHAT'S BOUGHT] Spending $6 billion on a non-exclusive license alone shows Nvidia wants model capability that can go into production immediately, and the people who wrote those models — not the balance sheet. With a non-exclusive license, Poolside can in theory sell the same technology to another buyer, but the core authors have already taken their seats at Nvidia. For other model companies still raising capital, the valuation anchor has been reset: a technology license can now be priced on its own at half of a company’s valuation, while founders keep the company to pitch the next round.

▪ SIGNAL A company valued at $12 billion sold its most valuable piece for $6 billion on its own — and what’s left can still raise at $12 billion.

❯ Stripe Acquires Model Routing Platform OpenRouter in Deal Reported at Over $7 Billion

[CLOSED] Payments company Stripe has closed its acquisition of OpenRouter, the model aggregation platform, with Bloomberg and The Wall Street Journal reporting a price of over $7 billion — Stripe’s largest acquisition to date. OpenRouter provides a unified interface that lets developers switch freely among more than 400 models from OpenAI, Anthropic, Google, Meta, and others.

[PREMIUM] That price marks a 5.4x jump in three months from OpenRouter’s $1.3 billion Series B valuation in May of this year. The Wall Street Journal first disclosed the talks in July, when a price near $10 billion was being floated. OpenRouter CEO Atala describes the company as “the Stripe of AI” — customers aren’t locked into any single provider.

[SETTLEMENT] Stripe is betting that mixed multi-model usage becomes the norm, with tokens turning into a settlement unit in enterprise procurement. What’s really being rewritten is model vendors’ pricing power: when switching costs fall to a one-line config change, control over developer budget allocation shifts from model vendors to the routing layer. The same day, reports said OpenAI is aggressively discounting on OpenRouter to win developers — that path is already being used.

▪ SIGNAL A payments company buying a model router is betting that token settlement, like card payments, will need a neutral clearinghouse.

❯ Broadcom in Talks for $60B+ in Debt to Fund Custom-Chip Infrastructure for Anthropic

[FUNDING] Broadcom is in talks with private-credit lenders for debt financing of more than $60 billion to procure custom AI chips and supporting infrastructure, with Anthropic as the primary customer, according to Bloomberg. The structure includes a $60–70 billion senior secured tranche and roughly a $30 billion junior tranche, with Broadcom backstopping part of the senior debt. The overall package could reach $100 billion. Blackstone and Apollo are in talks to participate.

[PRECEDENT] This is not the first time the three have worked together. In June, they closed $35 billion in financing through a vehicle called AI XPV, in which investors bought the custom chips and leased them directly to Anthropic. The new round is nearly three times that size and lands right around Anthropic’s IPO filing. Broadcom has repeatedly used this structure this year to take custom-chip orders from clients, while private-credit lenders get a rare AI exposure backed by physical assets. These figures remain per Bloomberg’s reporting and have not been confirmed by any party’s announcement.

[RISK BEARERS] The funding strain of chip purchases has been moved off Anthropic’s balance sheet and onto private-credit investors. Repayment on these deals depends entirely on the rent Anthropic pays going forward, which in turn depends on whether Claude’s revenue growth holds up. Once the company is public, if growth downshifts, the spreads on this debt will move before the share price does — and the holders are the pension and insurance funds behind Blackstone and Apollo.

▪ SIGNAL AI compute bills are showing up less on buyers’ financial statements and more in someone else’s credit portfolio.

❯ Report: Nvidia to Ship Small Batches of LPU to China Before Year-End; Company Denies Same Day

[CONFLICT] The Information reporter Qianer Liu, citing two Nvidia employees, said the company plans small-batch shipments before year-end of an LPU processor aimed at Chinese customers, and that the product complies with U.S. export-control rules. Nvidia publicly denied it the same day; a spokesperson said, “We have no LPU sales in the Chinese market today, and there is no China-only LPU product on the roadmap.”

[CONTEXT] The chip described in the report comes from technology Nvidia licensed from Groq, working with GPUs to specifically accelerate conversational inference. Nvidia reportedly rewrote the software so the LPU can work with processors available in the Chinese market — because the next-generation Vera Rubin system is export-restricted and cannot enter China. Earlier, Jensen Huang had just publicly said the company has “basically ceded” the Chinese market to Huawei.

[CONTRADICTION] The puzzling part is the wording of the denial: it says “not on the roadmap,” while the report describes a shipment plan at the employee level. This makes it harder for Chinese customers to decide on their next round of inference-compute procurement — keep waiting for Nvidia’s compliant products, or shift budgets entirely toward domestic alternatives. The cost gap between the two paths will surface within a few quarters.

▪ SIGNAL On one hand, it says it has ceded the Chinese market; on the other, it’s reported to be building a China-only chip. This split in messaging is itself a product of export controls.

❯ Alibaba quarterly net profit down 75% YoY, AI applications segment posts 13.861 billion yuan quarterly operating loss

[EARNINGS] Alibaba’s latest quarterly revenue rose 9% YoY to about $39.64 billion, in line with market expectations, while net profit fell 75% YoY to about 10.54 billion yuan. The profit collapse stems from two sources: heavy investment in AI infrastructure and weak domestic retail consumption. Cloud business revenue grew 45%, driven by AI demand.

[COST BREAKDOWN] This earnings report, for the first time, discloses results under four new segments, laying out the AI bill separately: the AI Labs & Applications segment posted a quarterly operating loss of 13.861 billion yuan, primarily from inference costs for the Qwen application; capital expenditures reached 67.678 billion yuan, versus 38.676 billion yuan in the same period last year, with the incremental spending going to AI infrastructure, procurement cycle fluctuations, and rising chip prices. The company’s free cash flow has been negative for two consecutive quarters.

[WHO PAYS] China’s major tech companies, for the first time, have written how much AI actually burns into an independent line item on the financial statements. For secondary-market investors, the basis for judgment shifts from the vague “cloud revenue growth rate” to whether inference costs can be amortized at scale; for Alibaba management, the pressure falls on the specific point in time when AI applications will stop losing more than ten billion yuan in a single quarter. Until then, the cloud business’s impressive growth will continue to be offset by this loss line.

▪ SIGNAL The 45% growth in cloud business and the 13.8 billion yuan loss in AI applications are two sides of the same coin; the only divergence is which side materializes first.

❯ Waymo Develops In-House 1000 TOPS Automotive Chip, Reducing Dependence on Nvidia and AMD

[IN-HOUSE] Google-owned Waymo has unveiled an in-house automotive AI chip with compute power exceeding 1000 TOPS, manufactured by TSMC on a 5-nanometer process, and already mass-produced and installed in the next generation of robotaxis. It is an ASIC dedicated to autonomous driving, with performance in the same range as Nvidia’s latest autonomous driving platform.

[WHY] The official blog cites latency as the rationale: the chip must ingest data from Waymo’s self-developed sensors and run models that determine what’s happening around the vehicle in an extremely short time. A general-purpose GPU is overkill and slow on this pipeline. The new model from Waymo’s recent partnership with Geely’s Zeekr is being switched over to this chip. Bloomberg reports that the in-house move also reduces supply dependence on Nvidia and AMD.

[COST] In the per-vehicle cost of a robotaxi, the compute platform is one of the few big-ticket items that hasn’t yet been driven down. By bringing this in-house, Waymo has reset the floor for per-kilometer operating costs. For peers still using third-party solutions, the constraint on expansion speed shifts from “can we get chips” to “are we willing to spend a few years making one.”

▪ SIGNAL Autonomous-driving companies making chips aren’t chasing peak compute—they’re after holding both latency and per-vehicle cost in their own hands.

❯ Unitree Robotics Falls 18.7% on Second Trading Day, Market Cap Breaks Below 300 Billion Yuan

[PULLBACK] Exchange data showed the stock opened 6.51% lower at 790 yuan on its second trading day, at one point falling more than 17% intraday and successively breaking below the 800-yuan and 700-yuan round-number levels. It closed at 687 yuan, down 18.7% for the day. Turnover reached 9.072 billion yuan, with a turnover rate of 41.98%. Total market capitalization pulled back to about 277.8 billion yuan, wiping out over 60 billion yuan from the previous day.

[DAY ONE] The gap stems from the first day’s extreme pricing. On the prior trading day — August 19, the listing day — the stock opened at a high of 1,100 yuan, up 629% from the 150.80 yuan IPO price. It closed at 845 yuan, up 460%, with full-day turnover of 24.5 billion yuan and a turnover rate of 85.28%. Over the two sessions, market capitalization has shrunk by more than 160 billion yuan from the peak at the day-one open. Founder Wang Xingxing himself acknowledged in public the day after listing that many of the carefully polished robot demos in recent years used pre-trained actions, and responses to real-time voice commands still have a delay of several seconds.

[PRICING] With the turnover rate above 40% for two consecutive days, the stock has yet to form a stable shareholder base, with the shares basically circulating among short-term funds. Embodied AI companies lack comparable revenue multiples, so the market can only fall back on order scale and demo videos for valuation — and the founder himself just undercut the credibility of those demos. The only thing that can support the stock price going forward is batch delivery order numbers; nothing else will be enough.

▪ SIGNAL What evaporated 160 billion in two days was not fundamentals — it was the imagination packed into that 629% opening price on day one.

❯ Hangzhou Puts Autonomous Traffic-Directing Robot on Streets, Works 7:00–18:00 with No Remote Control

[ON DUTY] Hangzhou has put a traffic-directing robot into service at a busy intersection, Reuters reports. Developed by local technology company Supcon Information, the T2 robot stands 1.88 meters tall and weighs 98 kg. It works autonomously from 7:00 to 18:00 each day, moving between preset points, requiring no remote control at any point; traffic police monitor and dispatch it via a central control platform.

[CAPABILITIES] Reuters reported on August 20 that the 98 kg robot uses cameras, radar, and onboard computing to identify four types of violations: riders without helmets, e-bikes carrying passengers, motor vehicles stopping across the line, and pedestrians running red lights. It then issues voice warnings and waves its mechanical arm in sync with the traffic signals to direct traffic. It has no enforcement or penalty authority; its role is to take over the most repetitive part of traffic officers’ work. Supcon Information’s core business has been industrial automation systems, and this pilot—with traffic police monitoring and dispatching via a unified platform—marks the first step from demonstration to scheduled duty.

[DEPLOYMENT SIGNAL] The value of this deployment lies not in technical difficulty, but in a robot being written into a city management duty roster for the first time. For peers in the industry, the metrics that matter are no longer lab success rates, but the failure rate across 11 consecutive hours of outdoor operation, battery endurance under summer road heat, and whether frontline traffic officers are willing to hand over the intersection to it. Only after a full quarter of these three datasets is accumulated can procurement by a second city even be discussed.

▪ SIGNAL The real threshold for a robot was never the moves—it’s whether it can show up at its post by 7 a.m. every day.

❯ DeepSeek Harness Ships rc.8, Adding Image Input and Installable Sub-Agents

[RELEASE] DeepSeek Harness released v0.1.0-rc.8 on August 19, with 14 changes total, centered on rounding out multimodal input: the model adaptation layer now supports native image requests, core commands like /goal and /plan accept mixed image-and-text input for the first time, and the @ menu adds references to local files and past sessions.

[DETAILS] The sub-agent system was refactored into optional installable packages, so both Claude Code and Codex can be installed and invoked. More noteworthy is the “tool vision” layer built for text-only models — it converts images into structured information via OCR, color statistics, and pixel scanning, letting models without vision capabilities reason over screenshots. The release also fixes long-standing issues with image requests, streaming generation, and custom gateways, and improves the Windows terminal experience. Per the official changelog, plugin support was strengthened too, adding concurrent web_search queries.

[USAGE] Screenshots, design mockups, and error screens can now be dropped straight into the command line, saving developers the step of describing interfaces in words — a step that was both time-consuming and prone to information loss. This tool-layer vision approach also shows that multimodality doesn’t have to wait for the model itself to upgrade: with OCR and pixel scanning paving the way, text-only models can take on image tasks too.

▪ SIGNAL Converting images to text and feeding the result to text-only models paves the way with engineering before model capabilities catch up.

❯ Game Science Unveils 15-Minute Gameplay Demo for Black Myth: Zhong Kui, Release Date Still Unannounced

[FIRST LOOK] Game Science released a 15-minute gameplay demo for Black Myth: Zhong Kui on August 20. It is the series’ second title, and since its first reveal at last year’s Gamescom, this is the first time full combat mechanics and story segments have been shown publicly, including a complete boss fight against “Great Clam Lord.” According to tallies by multiple domestic game media outlets, the video hit trending searches within ten minutes of release, with concurrent viewers briefly reaching around 100,000.

[ORIGIN] The protagonist draws from folk legend Zhong Kui, with the tiger-riding look inspired by the Ming Dynasty painting Zhong Kui Riding a Tiger and a dream by art director Yang Qi. The title was first shown at Gamescom in 2025, and only a trailer appeared over the following year. According to earlier official statements, after Black Myth: Wukong shipped in 2024, Game Science adjusted its plans and abandoned the originally planned DLC, shifting the entire team to this new project in order to create more distance from its predecessor and address the level design criticized in the earlier game. The art and rendering in this demo are more realism-oriented than Wukong’s.

[ASSESSMENT] The project is still in early development, with no release date set. Releasing a 15-minute demo without a launch window usually means the team is confident about the combat system’s completeness but not about the overall schedule. The real hurdle for China’s single-player studios lies ahead: Wukong proved a single title can sell explosively, but whether it can become a consistently producing series—one that gives the single-player business predictable revenue—is the variable to watch after this demo.

▪ SIGNAL The second installment consumed the DLC resources entirely, betting that those who bought Wukong will be willing to buy a new story again.

❯ Musk Reported to Pour $100M–$200M into Midterms, Betting on Texas

[FUND FLOW] According to Wired, Trump-aligned figures expect Musk to spend $100 million to $200 million in the November midterms, operating through his super PAC America PAC, with the focus on voter mobilization in key districts — the primary battlefield being his home state of Texas. Earlier reports also put his commitment at up to roughly $120 million across eight states.

[WHY TEXAS] The Texas Senate race is tight: Republican candidate and state attorney general Paxton is clearly trailing in fundraising behind Democratic rival Talarico. Texas is also home to Tesla, SpaceX, and xAI, and the state with the largest queue of AI data centers in the country — this week, the state’s grid regulator just froze grid-connection approvals for all data centers pending a queue audit. According to SemiAnalysis, the queue stands at roughly 474 gigawatts, nearly five times the state’s historical peak load.

[CONVERGENCE] Energy permitting and political money are both escalating in the same state, and the practical impact on the AI industry is more direct than the election outcome itself. In data-center siting decisions, the calculus used to be electricity prices, land costs, and fiber. Now, the political environment around grid interconnection has to be priced in separately — a single permitting freeze can stall multi-billion-dollar projects halfway. The first to feel the pressure are developers who have already secured land but not grid access.

▪ SIGNAL Once compute expansion hits the grid bottleneck, who makes the permitting calls becomes a question worth $200 million to influence.