AI 大厂转行做咨询：40 亿部署公司与 FDE 时代（2026）
AI Labs Become Consulting Firms: The $4B Deployment Race (2026)
> 📌 TL;DR
> 2026 年 5 月，AI 行业发生了一件比任何新模型都重要的事：头部大厂集体「转行」做咨询。5 月 4 日 Anthropic 联手高盛、黑石成立约 15 亿美元的服务公司；5 月 11 日 OpenAI 砸 40 亿美元设立「部署公司」并收购咨询团队；半个月内 PwC、KPMG 把 30 万员工搬上 Claude。信号很清楚：当模型本身越来越便宜、越来越同质化，真正的护城河变成了「把 AI 装进企业」的能力。 而扛起这件事的，是一个叫 Forward Deployed Engineer（前线部署工程师） 的岗位——2026 年最贵、最抢手的 AI 工作。
三周时间线：AI 大厂同时「下凡」
先把事实摆出来。下面每一条都有官方公告或一线媒体（CNBC、Fortune、TechCrunch、路透）佐证，括号里是发生时间：
| 日期 | 事件 | 关键数字 |
|------|------|----------|
| 5/4 | Anthropic 联合黑石、Hellman & Friedman、高盛成立企业 AI 服务公司 | 约 15 亿美元，瞄准中型企业 |
| 5/5 | 路透报道 OpenAI 已在三笔服务公司收购的「后期阶段」 | 3 笔交易 |
| 5/11 | OpenAI 正式成立「OpenAI Deployment Company」，并收购咨询公司 Tomoro | 40 亿美元+，TPG 领投，约 150 名工程师 |
| 5/14 | PwC 扩大与 Anthropic 联盟，认证 3 万名员工使用 Claude | 3 万人，新设「CFO 办公室」实践 |
| 5/19 | KPMG 上线「Digital Gateway Powered by Claude」 | 27.6 万员工，覆盖 138 国 |
| 5/21 | 微软以约 10 亿美元多年承诺留住安永（EY） | 约 10 亿美元 |
短短三周，三件事同时发生：AI 实验室自己下场做服务、四大会计师事务所选边站队、华尔街资本大举入局。 截至 5 月下旬，Anthropic 已在不到 8 个月里拿下四大里的三家（PwC、KPMG、德勤），微软则靠重金守住了安永。
这不是巧合，是一次集体转向。
为什么：模型在贬值，部署才是护城河
要理解这波操作，得先承认一个残酷现实：纯粹「卖模型」这门生意，正在变得不性感。
- 价格在崩。 Google 在 5 月 19 日 GA 的 Gemini 3.5 Flash，定价低到每百万 token 输入仅 1.5 美元；开源阵营（DeepSeek、GLM、Qwen 等）一路把前沿能力的价格打到地板。
- 能力在趋同。 三大厂的高管现在都公开承认前沿是「势均力敌」，差别更多在成本、速度和取舍，而不是谁聪明一个数量级。
- 真正的瓶颈在别处。 企业买了最强的模型，照样跑不起来——数据是乱的、流程是旧的、没人知道该把 Agent 接到哪。模型能力 ≠ 业务价值，中间隔着一整条「部署」的鸿沟。
于是逻辑就通了：如果模型本身赚不到溢价，那就往下游走，去赚「帮你把模型真正用起来」的钱。Counterpoint 分析师 Neil Shah 一句话点破：「控制应用层和服务层，才能把企业客户锁死。」 服务不只是新收入，更是护城河和黏性。
这就是 AI 行业的「成人礼」——从卖智能（selling intelligence）转向卖结果（selling outcomes）。
Forward Deployed Engineer：从 Palantir 怪癖到行业标配
扛起「卖结果」这件事的，是一个突然爆火的岗位：Forward Deployed Engineer（FDE，前线部署工程师）。
这个概念并不新——它大约 20 年前诞生于 Palantir，指的是直接坐进客户公司里、把能跑起来的生产系统做出来的工程师（交付的是工作代码，不是 PPT）。FDE 站在工程和咨询的交叉点：吃透客户的业务问题，用自家平台搭定制 Agent，反复迭代到真正在生产环境跑通为止。
过去一年，这个「Palantir 的怪癖」变成了全行业的默认配置：
- OpenAI 的部署公司、Anthropic 的合资公司，核心都是成建制地养 FDE 团队；
- Google Cloud 2026 年计划招约 59 名 FDE，Salesforce 则公开承诺组建 1000 人的 FDE 队伍；
- 收入也很夸张：据 2026 年的多份薪酬报告，FDE 平均总包约 23.8 万美元，中级起步 30 万美元+，顶级实验室的资深 FDE 能拿到 60 万美元+。
一句话：当 AI 落地的瓶颈从「模型不够强」变成「没人会装」，会装的人就成了最稀缺的资源。 如果你是工程师，这是 2026 年最值得关注的职业方向之一。
四大的两难：被颠覆，还是当渠道？
这波转向里最尴尬的，是埃森哲、四大这些传统咨询/集成商。Fortune 直接点出，四大面对 AI 有「两个噩梦」：AI 可能替代它们卖人天的核心生意，同时它们又不得不押注某个模型厂当渠道。
KPMG 的选择最典型——它把 Claude 铺给 27.6 万员工，还成了 Anthropic 在私募股权（PE）领域的「首选咨询伙伴」，帮 PE 把 AI 装进被投公司。这既是防守（别被 AI 取代），也是进攻（抢做 AI 落地的渠道）。
但风险也恰恰在这里：当 OpenAI、Anthropic 自己养起几百上千号 FDE，它们和昔日的「渠道伙伴」就成了正面竞争对手。 四大今天是合作方，明天可能就是被绕开的中间商。
这对不同的人意味着什么
- 如果你是买 AI 的企业：好消息是终于有人帮你「装到底」了；坏消息是要警惕厂商把应用层和服务层一起锁死——签约前先想清楚数据、流程、退出成本掌握在谁手里。
- 如果你是工程师：FDE 是当下最稀缺、最高薪、最不内卷的方向之一，技能栈（RAG、评测、Agent 开发、生产可观测性）值得现在就补。
- 如果你在传统咨询/集成商：要么成为模型厂的深度渠道，要么被它们的 FDE 军团绕开，中间地带正在消失。
- 如果你是创业者：垂直行业的「最后一公里部署」依然有大量空白——大厂的 FDE 军团再大，也覆盖不完所有细分场景。
> ⚠️ 注意
> 「服务化」是一把双刃剑。当你的模型供应商同时也是你的实施方，锁定风险（vendor lock-in）会指数级上升——模型、应用、服务、数据全在一家手里。企业在享受「一站式落地」便利的同时，务必在合同里守住三条底线：数据所有权、模型可迁移性、退出条款。
写在最后
2026 年最重要的 AI 故事，可能不是哪个模型又在榜单上涨了几分，而是这个：当所有人都能买到聪明的模型，能不能把它真正用起来，才是分水岭。 AI 大厂们用 40 亿、15 亿真金白银投了票，告诉了我们答案——未来几年，决定胜负的不是参数，是部署。
> ✨ 金句
> 卖模型的时代正在结束，卖结果的时代刚刚开始。下一个十年最值钱的 AI 能力，也许不是「训练出最强的脑子」，而是「让这个脑子在你乱糟糟的真实业务里，真的开始干活」。
---
本文数据来源截至 2026 年 5 月底，主要依据 OpenAI、Anthropic、PwC、KPMG、黑石的官方公告，以及 CNBC、Fortune、TechCrunch、路透等公开报道交叉核实。
> 📌 TL;DR
> In May 2026, something more important than any new model happened in AI: the top labs started turning themselves into consulting firms. On May 4, Anthropic teamed with Goldman Sachs and Blackstone on a ~$1.5B services venture; on May 11, OpenAI launched a $4B "Deployment Company" and bought a consulting team; within two weeks, PwC and KPMG put 300,000+ staff on Claude. The signal is clear: as the models themselves get cheaper and more interchangeable, the real moat becomes the ability to actually get AI into a business. And the people carrying that load have a job title — Forward Deployed Engineer — now the hottest, highest-paid role in AI.
A three-week timeline: the labs come down to earth
Let's start with the facts. Every line below is backed by an official announcement or first-tier reporting (CNBC, Fortune, TechCrunch, Reuters); dates in parentheses:
| Date | Event | Key number |
|------|-------|-----------|
| May 4 | Anthropic launches an enterprise AI services firm with Blackstone, Hellman & Friedman, Goldman Sachs | ~$1.5B, targeting mid-sized companies |
| May 5 | Reuters: OpenAI in "advanced stages" on three services acquisitions | 3 deals |
| May 11 | OpenAI launches the "OpenAI Deployment Company," acquires consultancy Tomoro | $4B+, TPG-led, ~150 engineers |
| May 14 | PwC expands its Anthropic alliance, certifies 30,000 staff on Claude | 30,000 people, new "Office of the CFO" |
| May 19 | KPMG launches "Digital Gateway Powered by Claude" | 276,000 staff across 138 countries |
| May 21 | Microsoft retains EY via a ~$1B multi-year commitment | ~$1B |
In just three weeks, three things happened at once: the AI labs went into the services business themselves, the Big Four picked sides, and Wall Street capital poured in. By late May, Anthropic had locked in three of the Big Four (PwC, KPMG, and Deloitte) in under eight months; Microsoft spent heavily to keep EY.
This isn't a coincidence. It's a collective pivot.
Why: models are depreciating, deployment is the moat
To understand the move, you have to accept an uncomfortable truth: the pure "sell the model" business is getting un-sexy.
- Prices are collapsing. Google's Gemini 3.5 Flash, GA on May 19, is priced as low as $1.50 per million input tokens; the open-source camp (DeepSeek, GLM, Qwen, and others) keeps dragging frontier-grade capability toward the floor.
- Capabilities are converging. Executives at all three big labs now openly call the frontier "neck-and-neck" — the differences are in cost, speed, and tradeoffs, not in who is an order of magnitude smarter.
- The real bottleneck is elsewhere. A company can buy the best model on earth and still get nowhere: the data is messy, the workflows are old, and nobody knows where to wire the agents in. Model capability ≠ business value — and the gap between them is the whole of "deployment."
So the logic clicks into place: if the model itself can't command a premium, move downstream and charge for making it actually work. Counterpoint analyst Neil Shah put it bluntly: "Controlling the application and services layer allows them to lock in enterprises." Services aren't just new revenue — they're a moat and a source of stickiness.
This is AI's coming-of-age: a shift from selling intelligence to selling outcomes.
Forward Deployed Engineers: from a Palantir quirk to an industry default
The people carrying the "sell outcomes" load have a suddenly-hot job title: the Forward Deployed Engineer (FDE).
The idea isn't new — it was born at Palantir roughly 20 years ago, describing engineers who sit inside the customer's company and ship a working production system (working code, not slide decks). An FDE lives at the intersection of engineering and consulting: understand the customer's domain problem, build a custom agent on your platform, and iterate until it actually runs in production.
Over the past year, this "Palantir quirk" has become the industry default:
- OpenAI's Deployment Company and Anthropic's joint venture are both, at their core, ways to staff FDE teams at scale.
- Google Cloud plans to hire ~59 FDEs in 2026; Salesforce has publicly committed to a 1,000-person FDE force.
- The pay is eye-watering: per 2026 compensation reports, FDE total comp averages ~$238K, mid-level starts at $300K+, and senior FDEs at top labs clear $600K+.
In one line: when the bottleneck shifts from "the model isn't good enough" to "nobody knows how to install it," the people who can install it become the scarcest resource. If you're an engineer, this is one of the most important career directions of 2026.
The Big Four's dilemma: get disrupted, or become a channel?
The most awkward players in this pivot are the traditional consultancies and integrators — Accenture and the Big Four. Fortune framed it neatly: the Big Four face "two AI nightmares" — AI may replace their sell-the-hours core business, while they're also forced to bet on a model vendor as their channel.
KPMG's move is the template: it put Claude in front of 276,000 employees and became Anthropic's preferred consulting partner for private equity, helping PE firms deploy AI inside portfolio companies. That's both defense (don't get replaced by AI) and offense (own the channel for AI rollout).
But here's the risk: once OpenAI and Anthropic staff up hundreds or thousands of their own FDEs, they and their former "channel partners" become direct competitors. Today's collaborator is tomorrow's disintermediated middleman.
What this means for you
- If you're a company buying AI: the good news is someone will finally install it end-to-end; the bad news is you must watch for a vendor locking up the application and services layers at once. Before signing, be clear on who controls your data, your workflows, and your exit costs.
- If you're an engineer: FDE is among the scarcest, highest-paid, least-saturated paths right now. The skill stack — RAG, evals, agent development, production observability — is worth building today.
- If you're at a legacy consultancy or integrator: either become a deep channel for a model vendor, or get routed around by its FDE army. The middle ground is disappearing.
- If you're a founder: vertical "last-mile deployment" is still wide open — no FDE army, however large, can cover every niche.
> ⚠️ A word of caution
> "Services-ization" is a double-edged sword. When your model vendor is also your implementation partner, lock-in risk rises exponentially — model, application, services, and data all sit with one company. Enjoy the convenience of one-stop deployment, but hold three lines in the contract: data ownership, model portability, and exit terms.
The bottom line
The most important AI story of 2026 may not be which model gained a few points on a benchmark. It's this: when everyone can buy a smart model, whether you can actually put it to work becomes the dividing line. The labs just voted with $4B and $1.5B of real money — and the message is that the next few years will be decided not by parameters, but by deployment.
> ✨ The line to remember
> The era of selling models is ending; the era of selling outcomes has just begun. The most valuable AI capability of the next decade may not be "training the smartest brain," but "getting that brain to actually start working inside your messy, real-world business."
---
Data in this article is current as of late May 2026, drawn from official announcements by OpenAI, Anthropic, PwC, KPMG, and Blackstone, and cross-checked against public reporting from CNBC, Fortune, TechCrunch, and Reuters.