前沿AI市场正在分化成一个个封闭阵营,各实验室纷纷挑选合作伙伴、排挤竞争对手,并限制其最强模型的获取渠道。
Salesforce 选择 Anthropic 作为其专属 AI 合作伙伴,让 Claude 成为全球最大 CRM 及 Slack 内部的默认模型。¹ 在 SpaceX 收购 Cursor 后,OpenAI 于 11 月 12 日切断了 Cursor 的 API 访问权限,理由是此前存在合同违约。² 就连开放权重模型如今也设置了门槛:Z.ai 于 8 月 26 日以 MIT 许可证发布了 GLM-5.3-Flash,两天后又将其旗舰模型置于 100 亿美元主机营收审查门槛之后。³
Anthropic 今年夏天率先开启了这一模式。其最强模型 Mythos 5 现在仅通过 Project Glasswing 计划提供,该计划将部分能力限量开放给受信任的公司。⁴
其姊妹模型 Fable 在美国境外下线了三周,随后回归并仅限美国境内推理。⁴
OpenAI 做出了镜像选择,在公开 API 上线之前,先向一小批受信任的合作伙伴推出了面向政府的 GPT-5.6 变体。⁵ 各国政府也在参与其中,决定谁才有资格运行前沿模型。
Salesforce 将 Claude 硬编码为默认模型已成定局;SaaS 应用自行挑选组件,并以单一应用的形式交付给客户——要么接受,要么放弃。
可插拔、与模型无关的理想主要存在于编码工具中,开发者期望在那里可以随意切换后端。
企业买家正面临日益严峻的治理挑战:严格的零数据保留政策、数据主权合规要求,以及对通过提示词将自身知识产权外泄的日益警惕。当平台硬编码单一供应商的模型时,买家必须通过谈判争取到能够更换模型的主动权。
总而言之,前沿AI不再是一种按 token 计费的公共事业。实验室正在建立白名单和黑名单。
就连宽松的开源权重时代也在收窄。模型创建者正在引入免费转付费的转化门槛——这是一种产品驱动增长的策略——既允许开发者自由实验,又要求大规模云服务商在营收达到触发条件时接受安全审查并获取商业许可。开源如今意味着:在你规模化之前是开源的。
各国政府正将前沿 AI 视为关键的主权基础设施,通过出口管制、国籍审查和分级模型体系来管控谁有资格运行前沿模型。
与这种封闭趋势相抗衡的,是一股 5 万亿美元的反向力量:Nvidia 凭借其市值优势,积极投资开放生态系统。而且它们动作频频:向 Hugging Face 投入 130 亿美元,为 Poolside 的模型工厂投入 70 亿美元,并对 Nemotron 作出 260 亿美元的承诺。这份清单还在继续。
开放权重是 Nvidia 针对自身云客户的结构性防御手段,确保基础模型不会被专有实验室或主权限制完全封锁。
在前沿领域,稀缺的不再是价格,而是访问权限。
-
Salesforce 与 Anthropic 联合推出 Claudeforce(2026 年 8 月下旬):Claude 成为 Agentforce 和 Slack 的默认模型,同时推出 Salesforce in Claude 插件,该插件内置 37 个预置销售技能,目前正处于试点阶段。Salesforce。 ↩︎
-
OpenAI 向 Cursor 发出终止通知,自 11 月 12 日起生效,理由是“与埃隆·马斯克旗下公司违反合同的经验有关”;Cursor 回应称,OpenAI 约占其流量的 5%,而 Anthropic 在同一天提高了 Claude 的配额。Business Insider、Bloomberg Law。 ↩︎
-
Z.ai 于 2026 年 8 月 26 日以纯 MIT 许可证发布了 GLM-5.3-Flash(320B),两天后又以定制许可证发布了其旗舰模型 GLM-5.3:任何在连续 12 个月内总营收超过 100 亿美元的模型即服务(MaaS)提供商,在商业使用前必须通过 Z.AI 的安全审查。Digital Applied、The New Stack。 ↩︎
-
在美国商务部发布要求外国国籍验证的出口令后,Fable 在美国境外停服三周(6 月中旬至 7 月 1 日);恢复后仅限美国境内推理,并加收 1.1 倍附加费。Mythos 5 仅限获批的 Project Glasswing 合作伙伴使用。Anthropic、Reuters。 ↩︎ ↩︎
-
应政府要求,GPT-5.6 的 Sol、Terra 和 Luna 变体已向一小部分可信合作伙伴推出。OpenAI。 ↩︎
-
Nvidia 承诺在 5 年内投入 260 亿美元用于开放权重 AI 模型开发和基础设施建设,该消息已在监管文件中披露,并得到高管确认。WIRED(2026 年 3 月)。 ↩︎
The frontier AI market is sorting itself into closed camps, as labs pick partners, cut rivals, & ration access to their strongest models.
Salesforce chose Anthropic as its dedicated AI partner, making Claude the default model inside the world’s largest CRM & Slack. 1 OpenAI cut Cursor’s API access on November 12 after SpaceX bought the company, citing prior contract breaches. 2 Even the open weights models now impose gates : Z.ai released GLM-5.3-Flash under an MIT license on August 26 & put its flagship behind a $10b host-revenue review two days later. 3
Anthropic started the pattern this summer. Mythos 5, its strongest model, now ships only through Project Glasswing, a program that rations certain capabilities to trusted companies. 4
Fable, its sibling, went dark outside the United States for three weeks, then returned to US-only inference. 4
OpenAI made the mirrored choice, launching its government-facing GPT-5.6 variants to a small group of trusted partners before any public API. 5 Governments are getting in on the act, deciding who may run the frontier at all.
Salesforce hardcoding Claude is de rigueur ; SaaS apps pick their components & pass them on a single app to the customer : take-it-or-leave-it.
The pluggable, model-agnostic ideal lived primarily in coding tools where developers expected to swap backends at will.
Enterprise buyers face an escalating governance challenge : strict Zero Data Retention policies, data sovereignty mandates, & an increasing wariness of sending their IP out the door in prompts. When platforms hardcode a single provider’s models, buyers must negotiate for leverage to be able to change it.
All of this to say, frontier AI is no longer a utility sold by the token. Labs are developing whitelists & blacklists.
Even the era of permissive open-source weights is narrowing. Model creators are introducing free-to-pay conversion thresholds, a form of product-led growth, allowing open experimentation for developers while requiring revenue-triggered security reviews & commercial licensing for large cloud hosts. Open source now means open until you scale.
Governments are treating frontier AI as critical sovereign infrastructure, using export restrictions, nationality screening, & specialized model tiers to police who runs the frontier.
Pushing back against this closure is a $5t countervailing force : Nvidia, leveraging its market cap to invest in open ecosystems. & they’ve been active : $13b for Hugging Face, $7b for Poolside’s model factory, & a $26b commitment to Nemotron. 6 & the list continues.
Open weights are its structural defense against its own cloud customers, ensuring foundation models cannot be entirely gated by proprietary labs or sovereign restrictions.
Access, not price, is the new scarcity at the frontier.
-
Salesforce & Anthropic launch Claudeforce (late August 2026) : Claude default across Agentforce & Slack, & Salesforce in Claude, a plugin with 37 prebuilt sales skills in pilot now. Salesforce. ↩︎
-
OpenAI termination notice to Cursor effective November 12, citing ’experience with Elon Musk’s companies violating contracts’; Cursor response, OpenAI near 5% of traffic & Anthropic raising Claude’s allocation the same day. Business Insider, Bloomberg Law. ↩︎
-
Z.ai shipped GLM-5.3-Flash (320B) under a plain MIT license on August 26, 2026 & its GLM-5.3 flagship two days later under a bespoke license : model-as-a-service hosts with more than $10b in aggregate revenue over any consecutive 12 months must pass a Z.AI security review before commercial use. Digital Applied, The New Stack. ↩︎
-
Fable dark outside the US for three weeks (mid-June to July 1) after a Commerce Department export order requiring foreign-national verification; US-only inference at a 1.1x surcharge on return. Mythos 5 limited to approved Project Glasswing partners. Anthropic, Reuters. ↩︎ ↩︎
-
GPT-5.6 Sol, Terra, & Luna variants launched to a small group of trusted partners at the government’s request. OpenAI. ↩︎
-
Nvidia 5-year $26 billion commitment to open-weight AI model development and infrastructure, disclosed in regulatory filings and confirmed by executives. WIRED (March 2026). ↩︎