AI 场景连续洞察合伙人AI Scene Intelligence Partner

Menlo Ventures 对 600+ 企业的调查显示，企业将 89% 的 $700 万 AI 预算集中于有 SLA/厂商背书的闭源模型；开源模型企业直接采购份额从 19% 降至 11%，方向性逆转而非正常波动（来源：Menlo Ventures，2026 State of Gen AI）。A Menlo Ventures survey of 600+ enterprises shows 89% of $7M AI budgets go to SLA-backed closed-source models; open-source direct procurement fell from 19% to 11%, a directional reversal, not normal fluctuation (Source: Menlo Ventures, 2026 State of Gen AI).

DeepSeek V4 技术能力达到历史新高（256K 上下文、原生 Agent 架构、MIT 许可证），同期企业直接采购开源模型比例持续下滑。缺失的不是技术质量，而是"可问责性封装"。评估开源 AI 公司商业价值的正确指标不是"企业直接采购份额"，而是"谁在其权重上构建了有付费客户的可问责服务"。DeepSeek V4 achieves historic technical capability highs (256K context, native Agent architecture, MIT license), while enterprise direct procurement keeps declining. What's missing is not technical quality, but "accountability packaging." The correct metric for evaluating open-source AI companies is not "enterprise direct procurement share," but "who has built accountable services with paying customers on top of its weights."

训练（Prime Intellect 去中心化训练，成本降 30–50%）、推理（GPT-5.6 Terra 达 Sol 95% 性能但成本降 60%）、架构（Meta SDM 线性 RNN，内存效率提升 10 倍、推理速度提升 3 倍）三层同步出现。竞争逻辑从"谁的模型最强"转向"谁在同等智能下成本最低、可问责性最强"。Training (Prime Intellect decentralized training, -30~50% cost), inference (GPT-5.6 Terra reaches 95% of Sol performance at 60% cost reduction), and architecture (Meta SDM linear RNN, 10× memory efficiency, 3× inference speed) all synchronized. Competition logic shifts from "whose model is strongest" to "who delivers the lowest cost with strongest accountability at equivalent intelligence."

GPT-5.6 原生多模态 Agent 编排与 DeepSeek V4 原生 Agent 架构在同一周完成，两条独立路线同步内化"模型即 Agent"范式。LangChain/AutoGen 类中间件的差异化护城河已不再是"编排能力"，而必须依赖行业数据壁垒或企业级合规集成深度。GPT-5.6 native multimodal Agent orchestration and DeepSeek V4 native Agent architecture were completed in the same week — two independent paths simultaneously internalizing "model-as-Agent." Differentiation moats for LangChain/AutoGen-type middleware can no longer rely on "orchestration capability"; they must depend on industry data barriers or enterprise compliance integration depth.

SambaNova（硬件替代，生产就绪）+ Prime Intellect（软件优化，商业早期）+ Meta SDM（架构变革，研究阶段）构成三维验证三角。Nvidia CUDA 生态锁定在接下来 36 个月内首次面临多维联合侵蚀——这是中期配置逻辑，不是近期颠覆信号。SambaNova (hardware alternative, production-ready) + Prime Intellect (software optimization, early commercial) + Meta SDM (architectural change, research stage) form a three-dimensional validation triangle. Nvidia's CUDA ecosystem lock-in faces its first multi-dimensional joint erosion within the next 36 months — this is a mid-term positioning thesis, not a near-term disruption signal.

AWS $10 亿 Forward Deployed Engineering 组织的出现，表明业界主流应对是"用人力服务把成本变量转化为合同固定值"，而非修复消费计制本身（来源：SiliconANGLE，2026/06/30）。FDE 工程师嵌入企业内部，客户看到的是项目合同金额，不是 token 消费额度。The emergence of AWS's $1B Forward Deployed Engineering organization signals the industry's mainstream response: convert cost variables into fixed contract values through human services, rather than fixing the consumption model itself (Source: SiliconANGLE, 2026/06/30). FDE engineers embed within the enterprise; customers see a project contract amount, not token consumption quotas.

Together AI 完成 $8 亿 C 轮，估值 $83 亿，年度 Booking 突破 $11.5 亿，16 个月内估值 2.5×（来源：TechCrunch，2026/07/01）。Aramco Ventures 领投 + Nvidia 跟投的组合释放双重信号：主权能源资本认为 AI 算力基础设施是国家战略资产，AI 芯片制造商在押注自己的下游分发渠道。Together AI closed an $800M Series C at an $8.3B valuation, with annual bookings surpassing $1.15B and a 2.5× valuation increase in 16 months (Source: TechCrunch, 2026/07/01). The Aramco Ventures lead + Nvidia follow-on combination sends a dual signal: sovereign energy capital views AI compute infrastructure as a national strategic asset, and AI chipmakers are betting on their own downstream distribution channels.

美团发布 LongCat-2.0——1.6T 参数 MoE 模型，完全在 5 万张华为 Ascend 910 国产 ASIC 上训练，MIT 开源（来源：FelloAI，2026/06/30）。非 NVIDIA 算力路径需分裂为两条独立路径：西方路径（Cerebras/Trainium，推理加速层）与中国国产路径（Ascend 910，已达到前沿规模训练能力）。Meituan released LongCat-2.0 — a 1.6T parameter MoE model trained entirely on 50,000 Huawei Ascend 910 domestic ASICs, open-sourced under MIT license (Source: FelloAI, 2026/06/30). The non-NVIDIA compute path must now be split into two independent tracks: the Western track (Cerebras/Trainium, inference acceleration) and the China domestic track (Ascend 910, which has achieved frontier-scale training capability).

Warner AI AGENT Act 草案（7/1）聚焦 AI Agent 隐私保护、守门人反垄断、可治理性三个维度（来源：CompleteAITraining，2026/07/01）。联合上周 GAAIA 讨论稿，美国联邦 AI 立法在两周内完成了"覆盖→深化"的跳跃。The Warner AI AGENT Act draft (7/1) focuses on three dimensions: AI Agent privacy protection, gatekeeper antitrust, and governability (Source: CompleteAITraining, 2026/07/01). Combined with last week's GAAIA discussion draft, US federal AI legislation completed a "coverage → depth" leap within two weeks.

本周三条信号（AWS FDE 服务层、Together AI Neocloud 算力层、LongCat-2.0 训练层）+ 上周 Dream 主权 AI（$30 亿）+ Odyssey 物理 AI（$1.45B），五周内持续的信号密度表明：AI 基础设施正在从 VC 叙事升级为主权资本叙事。This week's three signals (AWS FDE service layer, Together AI Neocloud compute layer, LongCat-2.0 training layer) + last week's Dream sovereign AI ($3B) + Odyssey physical AI ($1.45B) — five weeks of sustained signal density indicate: AI infrastructure is upgrading from a VC narrative to a sovereign capital narrative.

96% 的已部署 Agentic AI 企业报告 ROI 达到或超出预期（其中 42% 超出），72% 企业员工满意度上升。核心问题已从"是否值得试"切换为"如何快速规模化，成本谁来管"。96% of enterprises with deployed Agentic AI report ROI meeting or exceeding expectations (42% exceeding), and employee satisfaction rose at 72% of companies. The core question has shifted from "Is it worth trying?" to "How do we scale quickly, and who manages the costs?"

Microsoft Copilot Cowork GA 全面切换消费计制，AI 交互成本自 2023 年以来跳涨 30 倍——从 $0.04 升至 $1.20 / 次交互。ROI 已验证（96%）与成本不可预测（30×）形成直接张力：CFO 层面的采购阻力将在 ROI 兑现之后、而非之前出现。Microsoft Copilot Cowork GA fully switched to consumption-based pricing, with AI interaction costs jumping 30× since 2023 — from $0.04 to $1.20 per interaction. The tension between validated ROI (96%) and unpredictable costs (30×) is direct: CFO-level procurement resistance will emerge after ROI is proven, not before.

Mistral 以 Apache 2.0 开源 675B MoE 模型 + $14.99 / 月 Pro 层参战。闭源走消费计制（成本随 Agent 复杂度放大），开源走订阅 + 自托管（成本上限可控）。企业供应商选择将越来越由成本结构而非能力差异驱动。Mistral entered the market with an open-source 675B MoE model under Apache 2.0 + a $14.99/month Pro tier. Closed-source follows consumption pricing (costs scale with Agent complexity); open-source follows subscription + self-hosting (cost ceiling is controllable). Enterprise vendor selection will increasingly be driven by cost structure rather than capability differences.

EU AI Act 综合修正案 423:57 票终审通过，2026/12/02 与 2027/12/02 两个截止日期写入法律文本，"不再有进一步推迟的政治空间"。弹性合规采购正转为刚性采购，6 个月内将出现可观察到的采购加速。The EU AI Act omnibus amendment passed its final vote 423:57, with December 2, 2026 and December 2, 2027 deadlines written into legal text — "no further political space for delays." Flexible compliance procurement is converting to mandatory procurement; observable procurement acceleration will emerge within 6 months.

Dream 主权 AI 平台（$30 亿估值）+ Odyssey 世界模型（$1.45B 估值）单周合计超 $5.7 亿融资。两个赛道各自独立，不依赖通用 Agentic AI 叙事——赛道分层信号已出现。Dream sovereign AI platform ($3B valuation) + Odyssey world model ($1.45B valuation) raised a combined $570M+ in a single week. Both tracks are independent, not reliant on the general Agentic AI narrative — sector stratification signals have emerged.