李严冰 × Ion Stoica × Ed H. Chi × Bill Jia × 乔琳 × Li Fan × Dawn Song × 贾扬清 × 田渊栋 × 葛小川 × Connie Chan × 曹卿云 Qingyun Cao

基础模型领先或仅维持数月，开源的核心价值转向掌握升级、流量与专有数据控制权。Agent已在ExploitGym中发现未知漏洞并攻击基础设施，企业采购也从token用量转向可验证的商业ROI。云厂商接近1万亿美元的资本开支将推动算力从训练转向推理，电力、内存与软硬件协同成为关键瓶颈和机会。

Max Kan × Jordan Nanos × Andrew Megalaa

Anthropic’s $200 subscription delivered roughly $12,000 of API-equivalent usage in testing, but workload mix—not headline token counts—determines its economics.At roughly 96% cached reads priced at 0.1x or less of fresh input, model ownership may support 40–70% subscription gross margins, while OpenAI’s 50% limit cut and migration to open models remain risks.

Peter Diamandis × Salim Ismail × Dave Blundin × Alexander Wissner-Gross × Emad Mostaque

Memory and reserved compute—not model quality alone—have become AI’s binding bottleneck, with an NVL72 order moving from $3.5 million to a $9 million, three-year lease and Positron reportedly reaching a $5 billion valuation.Frontier labs are directing 80%-90% of research toward GPT-7 and GPT-8, making Altman’s “accept some bad things happening” stance a live contest over access, safety, and scarce compute.

Nathan Labenz × Woodson Martin

OutSystems argues that AI’s enterprise moat is not faster code generation alone, but 25 years of application primitives, governance, and customer trust supporting software that “never breaks.”Development output rose from four major features in Q4 to 19 in Q1 and 26 in Q2 as model routing reduced token costs, while the agent factory targets workflow ROI; compliance queues and model provenance remain adoption constraints.

David George × Kevin Mandia

AI is shifting cyberattacks from resource-constrained “sniper shots” to scalable swarms, while Armadin continuously verifies exploitable network weaknesses.Since January 2026, Armadin has found more than 90 zero-days across Fortune 500 customer environments, while Blue aims to turn a five-minute lead into autonomous controls.With technical advantages potentially copied in six months, customer execution and the Red-to-Blue feedback loop remain the operating test.

Molly O'Shea × John Collison

AI training is shifting from benchmark scores to economically valuable work, making high-fidelity environments for real workflows, tools, data, and verifiers a bottleneck.Enterprises may gain an edge by owning evaluations and correction data around differentiating workflows while routing tasks across models for accuracy, cost, and latency.Reward hacking and limited agent endurance make permissions, auditability, and verification key adoption constraints.

Peter Diamandis × Salim Ismail × Dave Blundin × Alexander Wissner-Gross × Richard Socher

Recursive’s $670 million raise, backed by Google Ventures, Greycroft, Nvidia, and AMD, targets weak recursive improvement as AI-generated code increasingly enters system design.Socher says the strong loop still needs AI-led ideation, implementation, and validation, with $410 million in AWS compute confronting a genuine GB200-scale crunch.Sonnet 5.5’s reportedly 70% Terminal Bench 4.0 score nevertheless does not improve the cost-performance frontier, while Project Meridian tests whether exponential autonomy can fit linear procurement.

Erik Torenberg × Alex Atallah × Amjad Masad

Stripe’s acquisition of OpenRouter preserves its brand, roadmap, and product autonomy while pairing payments with model-independent inference infrastructure.OpenRouter’s marketplace reduces vendor lock-in, combines models, and pressures inference prices downward as enterprises evaluate task-specific performance rather than feature lists.Specialized decision models could lower model debt and police agent actions, while multi-model fusion reportedly approaches frontier quality at 40–50% of the cost; deception risk and cache reuse remain unresolved.

Nathan Labenz × Keerthana Gopalakrishnan

Gemini Robotics 2 advances whole-body control and cross-embodiment transfer, but Keerthana still calls robotics its “GPT-2 moment”: one-shot demonstrations do not yet prove reliable generalization across bodies and changing scenes.Google’s three-model hierarchy separates embodied reasoning, action, and local execution, pointing to controlled factories, warehouses, and professional servicing as early wedges while latency, hand economics, safety, and compounding handoff errors remain unresolved.

Logan Jastremski × Vikram Sekar

Memory—not raw compute—is AI’s deepest systems bottleneck, making lower-memory architectures a potential lever across bandwidth and networking.As 144-GPU systems push power and cooling higher and copper fails at 400 gigabits, optics, DRAM-on-logic and specialized inference architectures may gain relevance, while Anthropic’s S1 points to more than $500 billion in obligated capital and compute and possible over-ordering.