Stripe Eyes Seven Billion Dollar OpenRouter Acquisition as AI Industry Shifts to Agent Infrastructure
English version above, Chinese version below.
Trend One: The race moves from model capability to agent infrastructure
On August 17, DeepSeek's peak/off-peak API pricing took effect. Peak hours (9:00–12:00, 14:00–18:00 Beijing time) are priced at twice the off-peak rate; the v4-flash model's peak output price sits at 9.0 yuan per million tokens. IT Home reported that in the five hours after the price rise, OpenCode Go's flash-model call volume dropped 94%. Before the adjustment, DeepSeek V4 Pro's cached-hit input price was 0.025 yuan per million tokens, uncached input 3 yuan, and output 6 yuan.
The same day, DeepSeek released Harness, an MIT-licensed, plugin-based agent runtime built on the Cordis meta-framework. Its core design is "everything is a plugin": the model, tools, skills, sessions, sandbox, agent loop, and even the Web UI are all swappable modules. Four preset modes—standard, PTC (TypeScript-composed tool calls), minimal, and creative—assemble different tool and skill sets on top of the same runtime. The project is in Developer Preview, with the official note that breaking changes may follow.
On August 16, Bloomberg reported Stripe is close to buying OpenRouter for over $7 billion. OpenRouter has 8 million users and offers access to more than 400 models; the price is a more than five-fold premium over its last funding round.
These three items move on the same axis. Two years ago the AI story was which lab released the biggest model; now the focus is on who owns the layer between the model and the work it does—pricing levers, runtime frameworks, and model gateways.
Trend Two: Frontier labs restructure as they head to IPO
OpenAI dissolved its Preparedness risk-evaluation team at the end of July, the Financial Times reported on August 17. The team's responsibilities—assessing major model risks and designing mitigations—have been split by subfield into biosafety, cybersecurity, and other operating teams. This follows the earlier disbanding of OpenAI's AGI-readiness and superalignment teams. Ethics lead Chloe Bakalar, chief futurist Josh Achiam, and safety head Johannes Heidecke left recently. Dylan Scandinaro, who joined from Anthropic in February to lead Preparedness, now focuses on recursive self-improvement.
On August 17, IT Home also reported that Anthropic plans a September IPO. CEO Dario Amodei published a rare long post addressing "trust crisis" criticism, predicting AI will cure most diseases within five to ten years, and anchoring it to his father's 2006 death from hepatitis C before cure rates jumped to 95% in 2010. Fidelity investor Gavin Baker said Amodei once suggested Anthropic could become the world's only private company; Anthropic's Sholto Douglas called that claim nonsense, and Amodei argued the "regulation = regulatory capture" framing is a false dichotomy.
Nvidia scaled back its guarantee for OpenAI's Ohio data-center financing. Reuters, citing WSJ, reported on August 14 that the initial guarantee is now expected below $120 billion, down from a previously discussed $250 billion, driven by investor concern over risk exposure.
Alibaba sold its gaming unit Lingxi Interactive Entertainment to Trustar Capital for at least $1.5 billion, IT Home reported on August 17, accelerating its shift toward AI.
Trend Three: Safety, watermarking, and synthetic-media disputes
Anthropic began embedding SynthID-Text watermarks in Claude-generated text for models released from August 2 onward. The watermark is machine-readable, survives copy-paste, and resists light editing. A paper disclosed alongside it describes "mind viruses" that spread self-modifying SOUL.md files between AI agents; four test payloads passed a 20-hop stress test.
On August 17, Shenzhen's internet information office confirmed that a widely shared video of "large-scale road collapse with a overturned vehicle" was AI-generated and spliced. The real event, on August 4, was a pothole of over 10 cubic meters on a Binhai Avenue arterial road, with no vehicle rollover.
McAfee Labs found AI can identify the location of travel photos with 87%–91% accuracy using visual cues—buildings, shop signs, road markings, vegetation, landmarks—even after GPS and EXIF data are removed. Researchers flagged the combination with phishing as a rising personalized-fraud risk.
On August 17, Claude suffered a major outage starting around 5:58 Beijing time, affecting Claude.ai, Claude Code, and Claude Cowork; Anthropic marked them as a major outage while Console and API stayed operational, and service was restored the same morning.
Trend Four: China's compute supply chain and social acceptance
On August 17, Biren Technology filed a Hong Kong exchange announcement guiding first-half 2026 revenue of 1.15–1.3 billion yuan, up 1,852%–2,107% from 58.9 million yuan a year earlier. Adjusted loss narrowed 35%–47%, and the company attributes growth to GPGPU demand from AI coding, long-horizon agent tasks, and generative AI.
Stanford's AI Index 2026, released August 16, found 84% of Chinese respondents feel excited about AI versus 38% in the United States; trust runs 72% to 32% in Edelman's survey. The report also notes the US–China performance gap has largely closed since early 2025, with models alternating the lead.
On August 17, First Financial reported Guangdong launched a "Token Loan" product, using enterprise token output/consumption and compute-service contract value as credit basis. Zhangjiagang Rural Commercial Bank placed its first 1-million-yuan "compute loan" in May, with base credit up to 10 million yuan and up to 50 million with compute-voucher enhancement.