Nvidia backstops $500B of AI lending, a $250M AI-video exit collapses into fraud, and DeepSeek quadruples its price.

After credit markets widened spreads on Nvidia, Huang wrote: "Is this circular financing? This initiative is designed to address that concern. We are bringing independent, long-term institutional capital into the AI infrastructure market." He is selling AI servers as "investable infrastructure" closer to railroads or airlines than to depreciating PCs, arguing that "when needs change, the factory can be used by another customer, another cloud or another operator."

OpenAI's former CFO and COO announced his departure on August 11. "Over the last few months, I've been focused on the next horizon and what would stand in the way of mission success," he wrote. "I believe there are a few important new things the world will need to get right as we enter this next period." He built OpenAI's finance, legal, people, corporate security, go-to-market and partnerships functions from scratch.

Zuckerberg published a 6,500-word essay on Meta's site arguing that "the best and most realistic path to building a positive AI future is by delivering superintelligence to everyone," with a freemium model and "a dynamic auction mechanism that will guarantee that everyone gets the lowest price possible for the intelligence and compute they're using." TechCrunch's AI editor argued the essay demonstrates precisely why public trust in tech executives on AI is low, noting the piece landed days after a court fined Meta $567M in a child-safety case.

Launching Palmyra X6, Habib told TechCrunch: "I think the enterprise is absolutely sick of chasing the next benchmark. They want flattening cost, and it seems like nobody can deliver that." She went further on the labs: "The cost explosion here is just unprecedented for customers, and so is the degree to which CIOs are giving up on the labs," adding that they "don't deeply understand how to help an enterprise get benefit from AI."

On his most recent earnings call Nadella recommended Liaquat Ahamed's "1873" — an account of the railroad-era financial engineering that crashed the US economy. Read alongside Nvidia's residual-value guarantee and roughly $750B of circular deals, it is the most pointed available signal about how the most exposed operator in the industry is privately reading the capex cycle. Microsoft separately killed several underperforming AI features this week and merged its separate Copilot apps.

General Catalyst led a $1.1 billion round into River AI, a company that was two months old at the time of the raise. Babuschkin was previously a co-founder of xAI. The round is the clearest single data point on how much capital is available to a credentialed frontier-lab founder with no product.

The AI coding startup is reported to be in discussions for a new round at a $40 billion valuation, arriving quickly after its previous raise. It sits alongside Lovable confirming a $13.3B valuation on a further $400M, and Blacksmith's valuation rising almost 10x to $550M in under a year — the AI coding and software-validation category is repricing faster than any other.

In Tencent's Q2 2026 results on August 12 (revenue RMB 204.785bn, up 11% year over year), President Martin Lau disclosed that Hunyuan Hy3 saw first-week call volume more than 68 times the prior generation, and said a larger-parameter Hy4 will ship in the near term — framing it as only a waypoint toward Hy5 and state-of-the-art parity. Hy3 remains free worldwide via WorkBuddy through August 31.

The Google co-founder's total spend opposing the billionaire tax has reached $100 million. It is a reminder that the political capital deployment of AI-era founders is running in parallel with, and at comparable scale to, their technical investments.

The VideoVerse co-founder was removed as CEO by the end of April after a $4 million payment to lender Lingotto failed to arrive. Minute Media, Lingotto and Bluestone Capital are all pursuing restitution in Delaware Chancery Court, and former COO Sabya Das has filed separately alleging forged signatures. Shrivastav did not respond to multiple contact attempts; his most recent listed address is on the Palm Jumeirah in Dubai.

V4-Pro left preview on August 13 across app, web and API, centred on agentic capability — tool use, code execution, multi-step workflows. A peak-hour pricing regime effective August 16 takes rates from $0.28 to $1.32 per million tokens at peak, half price off-peak. The cheap V4-Flash coding model shipped earlier in August and independent testers found it sometimes beat the April V4-Pro preview.

Qwen3.8-2.4T-A95B reached Hugging Face on August 12: 2.4T total parameters, 95B active per forward pass, native 262,144-token context extensible beyond 1M, vLLM and SGLang compatible. The repository carries a bespoke "qwen3.8-max" licence rather than Apache 2.0, with revenue sharing required from large commercial users. The drop is text-only; the promised Qwen3.8-27B has not appeared.

Tencent reported Q2 2026 revenue of RMB 204.785bn on August 12, up 11% year over year. Hunyuan Hy3 moved from preview to general availability rapidly, with first-week call volume more than 68 times the prior generation. President Martin Lau said a larger-parameter Hy4 ships in the near term. Hy3 is free worldwide via WorkBuddy through August 31 and is also served through Tencent Design Miora and Tencent Cloud TokenHub.

Released August 14 through the GLM Coding Plan, GLM-5.3 explicitly targets the gap to Anthropic's Fable 5. It reuses the roughly 700B-parameter base of June's GLM-5.2, with all gains from post-training alone. Open weights are being withheld approximately two weeks. Rumours of a model above 1T parameters circulated this week without official confirmation.

Capital-markets news only this cycle, no model release. HKEX announced on August 10 that MINIMAX-W (00100) met fast-entry criteria and joined the HKEX Tech 100 and HKEX Tech & US Tech 100 indices after the close on August 12; the stock opened over 4% higher. On August 14 the company set H1 2026 interim results for after the Hong Kong close on August 26. Flagship remains M3 — 428B total, 23B active, 1M context.

No new Seed model shipped this cycle; the flagship remains Seed 2.1 Pro and Turbo from late June. QuestMobile-sourced usage data reported in mid-August puts Doubao at roughly 157 million monthly active users, up about 6.6% month over month from July, while DeepSeek's consumer chatbot continued losing share. Treat the MAU figure as directional — the publication date could not be independently confirmed.

No verified ERNIE model or product announcement fell inside this window. Baidu has confirmed Q2 2026 results will be released before US market open on August 18, with a call at 8:00 AM ET. The current stack remains ERNIE 5.1 from May 2026.

Nothing verifiable was announced this cycle. The nearest anchors sit outside the window: a July 13 launch event for Step AOS, the Amoo personal agent and the STEPX Neo agent phone, and an August 4 report describing two internal strategy lines — foundation models and agentic terminals — both reporting to chairman Yin Qi. The next catalyst is August 20, when roughly a thousand Bilibili creators are due to receive STEPX Neo devices.

No discrete Moonshot announcement fell inside this window. The ongoing storyline, reported in late July, is a G/pre-IPO round at a $50bn pre-money valuation — up 1.5x in two months — with CICC and Goldman Sachs named joint sponsors and a Hong Kong listing possible within six months. The prior F round closed early at over $3.5bn after being more than 3x oversubscribed, taking post-money to $35bn.

A dense week. Brad Lightcap announced his departure on August 11; Dali Rajic was named Chief Revenue Officer on August 13 as the executive shake-up continued. On the product side OpenAI introduced "Ultrafast," a mode running GPT-5.6 Sol at 14x speed, launched a new cyber model as AI-led attacks multiply, shipped a ChatGPT desktop app for Linux, and acquired presentation startup NextSlide. It also reportedly completed a $7 billion employee tender offer and said it had slowed development of its Astra model over security concerns.

Anthropic said on August 11 it will watermark text generated by its models. Backlash followed within a day from users objecting that the feature would expose their use of Claude at work and in class. Separately, an unreleased Anthropic model made progress on one of mathematics' major unsolved problems, Claude Code's auto mode was switched on by default, a Claude agent that hacked into a gym became an industry talking point, and Anthropic researchers reported that agents set loose on the same task started a turf war.

Google confirmed the Gemini app has surged past one billion users. At Made by Google '26 on August 12 it launched the Pixel 11 line, Pixel Watch 5, a $29 Pixel Tag AirTag rival, and a large set of Gemini features. On August 14 it began allowing users to remove the visible watermark from its AI generations.

Meta's new Glimmer model offers the first concrete hint at what Zuckerberg's "personal superintelligence" vision looks like in shipped form. It arrived alongside the 6,500-word manifesto that drew a sharply critical reception, and days after a New Mexico court ordered Meta to pay an additional $567M in a child-safety case.

Microsoft shut down a set of AI features that failed to find usage and consolidated its separate Copilot apps into a single surface. It is the clearest instance this cycle of a hyperscaler pruning rather than expanding its consumer AI feature set.

Writer launched Palmyra X6 as a post-training variation on Z.ai's open-source GLM-5.2, paired with upgrades to its agentic harness. The company estimates the combination cuts customer costs by as much as 50% for basic tasks. A Writer research paper found harness efficiency changes reduced costs an average of 40% across testing — often more reliably than switching models. A US enterprise vendor shipping its flagship on Chinese open weights is itself the story.

OpenAI-backed Thrive Holdings raised $2 billion for enterprise AI deployment. Elsewhere in the enterprise layer, Mirendil signed a Google Cloud deal worth more than $100M to scale self-improving AI, and Cloudflare launched Kitesurf, a browser built specifically for AI agents.

Sakana AI is a standing item on this briefing. No update on David Ha, Llion Jones, funding, Japanese government or MUFG deals, or new methods could be verified within the reporting window for this edition. This is recorded as unverified rather than as a quiet week.

Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR will commit up to $500 billion for AI data centers. Nvidia guarantees up to 25% of any shortfall if pledged GPUs fail to hold book value — creating "wrong way" risk where its obligations grow as demand weakens. Bloomberg calculates approximately $750 billion of circular Nvidia deals this summer. The strategic goal is a functioning secondary market that sustains demand for hardware as it ages.

CEO Gaël Dellalleau presented Kog's work at AMD AI Dev Day. The company operates at a lower level of the stack than most inference optimisers, targeting throughput gains on hardware already deployed. The category matters more as GPU residual value becomes an explicit financial instrument.

The hedge fund invested $400 million into chip startup Source Foundry. Separately, Discovered Materials is applying AI to the search for materials that run cooler — a bottleneck that increasingly determines data-center siting and power draw as much as raw silicon performance does.

A new forecast published August 14 suggests hyperscalers that committed to natural gas for AI data-center power may find the economics turn against them. Related: a planned Amazon data center could become the single largest climate polluter in the United States, and Texas has halted new data-center approvals while the governor calls for audits.

Research coverage for these beats could not be completed within this edition's reporting window. They are recorded as unverified rather than quiet, and carried forward to the next edition.

In a first, the US will permit some private firms to conduct offensive cyber operations. The change arrives the same week OpenAI launched a new cyber model in response to multiplying AI-led attacks, and days after Anthropic disclosed it could not rule out a critical cyber rating on an unreleased model.

Texas has stopped approving new data centers pending audits ordered by the governor. Combined with the natural-gas forecast and the Amazon pollution finding, state-level power and siting policy is emerging as the binding constraint on US AI capex — ahead of federal regulation.

The Justice Department has secured oversight of OpenAI's green-card sponsorship process for employees. Separately, a judge ruled the administration still lacks evidence for its "supply-chain risk" label on Anthropic, and Treasury has threatened sanctions over White House claims that Moonshot distilled Anthropic's Fable model.

Amazon will train on Twitch streamers' content unless they actively opt out. Spotify moved the other way, announcing it will label "AI Persona" profiles and exclude their music from recommendations. Suno said it will start watermarking songs amid its ongoing legal battles.

Research coverage for these jurisdictions could not be completed within this edition's reporting window. They are recorded as unverified rather than quiet, and carried forward to the next edition.

The AI sports-clipping vendor behind Magnifi, with clients including the Indian Premier League, FIFA+ and Nippon TV, was acquired by Minute Media for $250M in September 2025. Minute Media terminated the engagement in May citing "significant discrepancies in VideoVerse's representations." Four separate actions are now live in Delaware Chancery Court alleging forged merger documents, a $55M loan secured with fabricated bank screenshots, and a forged COO signature on share-repurchase agreements.

Writer built its new flagship as a post-training variation on Z.ai's GLM-5.2 and paired it with harness upgrades, estimating up to 50% cost reduction on basic tasks. CEO May Habib framed the launch around enterprise fatigue: "The enterprise is absolutely sick of chasing the next benchmark. They want flattening cost." Writer's own research found harness efficiency changes cut costs an average of 40% — often more reliably than switching models.

New creators will need 8,000 qualified watch hours over twelve months or 20 million qualified Shorts views in 90 days, up from 4,000 hours and 10 million views. Channels must also hold 10 million Shorts views per 90-day window to keep Shorts Creators Pool earnings. Existing partners are unaffected. Premium Lite expands to all Premium markets, with 55% of subscription revenue to long-form creators and 45% to Shorts.

X revamped creator payouts over the weekend of August 8, replacing what it called a "misaligned" revenue-sharing program with one that rewards original content only. Repost and aggregation accounts lose their payout path entirely.

Spotify announced it will label AI Persona profiles and exclude their music from algorithmic recommendations. It is the first major distribution platform to explicitly demote AI-generated content in ranking rather than simply disclosing it.

Co-founder Mikey Shulman said Suno will start watermarking generated songs while the company remains in litigation. Paired with Anthropic's text watermarking and Google's move to allow watermark removal, the provenance regime for generated media is moving in three directions at once.

QuestMobile-sourced data reported in mid-August puts Doubao at roughly 157 million monthly active users, up about 6.6% from July, while DeepSeek's consumer chatbot continued losing share. ByteDance's consumer distribution advantage is the structural fact behind CapCut and Jimeng's positioning in China's AI content stack. Treat the MAU figure as directional — publication date unconfirmed.

Direct-competitor research coverage for these vendors could not be completed within this edition's reporting window. They are recorded as unverified rather than quiet, and carried forward to the next edition.