Stripe 70 亿美元收购 OpenRouter，AI 路由层价值重估

TheInformation had the scoop last month, but OpenRouter’s acquisition by Stripe for $7B was seems all but closed this weekend, 90 days after their $1.3B Series B. Their last revenue number out there was $140m annualized, so this represents a “standard” 50x multiple for a top tier AI company. What’s incredible is the profitability:

Although much smaller than Cursor, OpenRouter likely has better economics. Its costs to serve its model-routing product were recently about $40 million on an annualized basis, or 28.5% of its revenue, meaning it was generating $100 million in annualized gross profit. With a roughly 70% gross profit margin, OpenRouter was near the level of high-performing, publicly traded software firms in that regard….

… Overall, OpenRouter is facilitating AI model usage at a rate of 250 trillion tokens per month, up from 50 trillion tokens per month in February.

A 70x P/E ratio is possibly cheap for a high growth (5x in 6 months) startup with a broad (8 million developers) base. Certainly a good outcome for new billionaire Alex Atallah, and good for fellow router startups, but certainly there are a lot of implications on Stripe’s AI strategy and where value accrues in AI infra (much less GPU infra, much less Agent Labs, much less Frontier Model Labs).

You can catch Alex’s last public appearance on the AIE State of Model Routing panel.

AI News for 8/15/2026-8/17/2026. We checked 12 subreddits, 544 Twitters and no further Discords. AINews’ website lets you search all past issues. As a reminder, AINews is now a section of Latent Space. You can opt in/out of email frequencies!

AI Twitter Recap

AI Infrastructure, Compute, and the Platform Stack

• OpenAI’s power-and-compute strategy is getting very literal: Two related posts suggest OpenAI is moving beyond “GPU supply” narratives into long-horizon control of the full infrastructure stack. @markchen90 described a 4+ GW NVIDIA capacity commitment; @kimmonismus added detail on an 8 GW Ohio campus, with SB Energy building and operating the site, NVIDIA backing the initial 4.25 GW, and a multi-year buildout through 2032. For infra engineers, the notable point is not just scale, but vertical coupling across power, data centers, chips, and long-dated access.
• The model access/routing layer is being repriced in real time: The reported Stripe–OpenRouter deal crystallizes how valuable the aggregation/routing API layer has become, but reaction from @kimmonismus also underscored how fragile that position could be if markup compresses to zero. In parallel, OpenRouter cut GPT-5.6 Sol pricing while Vercel did the same on AI Gateway, reinforcing that model brokerage is becoming a pricing battlefield rather than a stable tollbooth.

Developer Platforms, Coding Agents, and Agentic Tooling

• Cursor’s Origin points toward the AI-native IDE becoming the system of record: Origin’s launch is more than a GitHub competitor headline. It suggests Cursor wants first-party control over the full loop: repository, agent, review surface, and deployment hooks. @kimmonismus notes GitHub remains syncable and source-of-truth-compatible, but the strategic direction is clear: agentic coding products are trying to absorb the surrounding platform, not just autocomplete against it.
• Multi-agent orchestration is shifting from demoware toward operating patterns: Several posts converged on the same motif. @tonbistudio showed Hermes Desktop bots self-assigning game-dev work based on inferred specialties; @Teknium formally reintroduced Bot Mode, where agents maintain distinct memory, skills, tools, and inter-bot communication; and @omarsar0 recommended material on orchestrating multiple agents in Codex. The common thread is specialization plus persistent context, not generic “agents talking to agents.”
• Evaluation and harness work remains the real leverage point: Hamel Husain’s updated eval-skills plugin adds an error-discovery workflow that turns model outputs/traces into annotated failure modes and clustered review surfaces. That pairs well with Agent Arena’s new cost-per-task and category filters, which are based on 1.7M+ real-world sessions. The field is slowly moving from model-level evals to harness-level measurement: routing, decomposition, memory, verifier loops, and total completion cost.
• Computer-use and sandboxing are getting productized: Vanta’s new computer-use capability for its TrustVanta agent addresses a real enterprise workflow gap: screenshot evidence capture when there is no API surface. Likewise, LangChain’s monday.com case study highlights isolated workspaces via LangSmith Sandboxes for agents doing iterative work like CSV analysis or map generation. “Agent” product quality is increasingly about permissioning and execution isolation, not just reasoning quality.

Model Efficiency, Post-Training, and Small/Open Model Progress

• Open models continue to compress the capability frontier: The strongest signal here was @cline’s note that Qwen3.8-27B now scores at DeepSeek V4-Pro / GPT-5.6 Luna territory on the Artificial Analysis Intelligence Index, described as the first time a local model has reached that capability tier. Ollama immediately positioned deployment paths for local users, and anecdotal reports like @rishdotblog’s suggest the model is already practical for long-context local coding setups.
• Inference efficiency is becoming architecture-level, not just quantization-level: @cwolferesearch’s discussion of Nemotron 3.5 Lightning is a good example: a 30B MoE with 3B active, trained for high-throughput agent execution, with multi-token prediction support for speculative decoding and additional drafters/quantized checkpoints. Similarly, @PandaAshwinee reported RL for large MoEs with zero train-infer mismatch, highlighting open ablations around post-training sparse models.
• Latent reasoning and memory are emerging as a separate scaling track: The BDH-CQ writeup shared by @TheTuringPost is notable less for raw benchmark strength than for the recipe: a 150M model doing latent-space reasoning with temporary memory, hitting 29.5% pass@2 on ARC-AGI-1 at around $0.0007 per task. In parallel, OpenAI Devs reported that with retained reasoning and compaction, GPT-5.6 Sol improved from 13.3% to 38.3% on ARC-AGI-3 while using roughly 6× fewer output tokens. The shared idea is that memory/compaction strategy is now a first-class capability multiplier.

Retrieval, Skills, Memory, and Research Tooling

• Search/retrieval people are questioning the “retrieve more, rerank more” reflex: The Weaviate podcast episode with Mathew Jacob revisits “Drowning in Documents”, phantom hits, listwise reranking, and ranking cascades. The practical implication for RAG systems is that naively increasing retrieved set size can degrade final quality, and future systems likely need per-query effort prediction and smarter scoring cascades rather than brute-force retrieval volume.
• Agent skills are being demystified and operationalized: @omarsar0’s summary of “Demystifying Agent Skills” is useful because it quantifies a common intuition: skills help mostly through procedural anchoring (65.7%), not factual knowledge injection (4.5%). Precision also collapses as skill pools expand. Related posts on the “skills” paper and GitSkills dataset mining ~3.8M SKILL.md files point to a maturing ecosystem around discoverability, packaging, and trigger management for agent skill libraries.
• Native memory is becoming a research object, not just a product feature: Engram Lab’s first research blog frames a future where agents are trained with native memory, while @jxmnop emphasizes the hard parts: memory calibration, self-generated training data, and getting models to actually exploit remembered information efficiently. This lines up with the broader move from stateless prompt engineering toward persistent internal/external memory systems.

Multimodal Models: Video, Audio, and Speech