OpenAI、Anthropic、Amazon 以及现在的 Microsoft：为何一些最大的科技公司要派遣数千名员工到客户办公室‘驻场’

Big Tech has found a new obsession, and it isn't another chatbot. Over the past few months, five of the world's most powerful technology companies have quietly reached the same conclusion: selling AI software isn't enough anymore. Amazon, Microsoft, Meta, OpenAI and Anthropic have all announced plans to physically send thousands of their own engineers into client offices, embedding them inside corporate teams to build and ship AI systems on-site. Between them, these bets add up to roughly $9 billion and tens of thousands of engineers, all chasing the same goal: making AI actually work in the messy reality of a real company, not just in a demo.The latest and biggest of these bets comes with Microsoft’s new “Microsoft Frontier Company,” backed by $2.5 billion and 6,000 engineers and industry experts. It arrived just days after Amazon Web Services committed $1 billion to a similar unit, and weeks after OpenAI and Anthropic launched competing ventures. Meta, too, quietly began building its own version around the same time. What started as a niche staffing model has, in the space of a single quarter, become the defining strategy of the entire AI industry.

Why "forward-deployed engineers" are suddenly everywhere

The job title behind all this is the forward-deployed engineer, or FDE—a technical employee who moves into a client's office to build and ship AI systems on the spot, instead of handing over software and walking away. Palantir invented the role over a decade ago, but it's become one of the hottest jobs in tech over the past year. LinkedIn data shows demand for FDE-type roles grew 42-fold between 2023 and 2025, and Box CEO Aaron Levie has called it one of the most in-demand jobs in the industry.The reason is simple: model access alone stopped closing enterprise deals a while ago. A widely cited MIT study found that around 95% of enterprise generative AI pilots showed no measurable profit impact, with the failure traced to weak integration rather than weak models. Getting AI to run inside legacy systems, compliance rules and internal office politics takes humans on the ground—not just an API key.

Before Microsoft, there was OpenAI, Anthropic and Amazon

Anthropic and OpenAI kicked off the current wave, both launching standalone deployment ventures backed by outside capital. Anthropic partnered with Blackstone, Goldman Sachs and Hellman & Friedman on a venture reportedly valued above $1.5 billion, aimed at mid-sized businesses that lack in-house AI talent. OpenAI went bigger, launching the OpenAI Deployment Company with more than $4 billion from TPG, Advent, Bain Capital, Brookfield and others, while acquiring Tomoro to add roughly 150 deployment engineers.Meta moved almost in parallel, building an Enterprise Solutions unit under head of product Naomi Gleit that embeds engineers to push its AI agents into everyday business use.Amazon joined soon after, when AWS committed $1 billion toward its own Forward Deployed Engineering unit. Unlike the joint ventures from OpenAI and Anthropic, Amazon's stays entirely in-house—run by Francessca Vasquez, AWS vice president of frontier AI engineering and services. The plan is five to six engineer "pods" at a time, each spending 45 days embedded with a client and working alongside AI agents. Early customers named include the NBA and electronics maker Ricoh, with Vasquez framing the effort around speed.

Microsoft is betting bigger, and pitching itself as neutral

Microsoft's answer, led by commercial business CEO Judson Althoff, is designed to dwarf all of them. Its Frontier Company brings together existing FDEs, consultants and support staff under Rodrigo Kede Lima, who previously led Microsoft's Asia business. Unlike Anthropic, OpenAI and Meta, which push their own models, Microsoft is selling itself as vendor-neutral: customers can run OpenAI, Anthropic, Microsoft or open-source models, and keep full control of their own data throughout.Early Frontier clients include the London Stock Exchange Group, Unilever, Land O'Lakes and Novo Nordisk, with consultancies like Accenture, EY, KPMG and PwC helping scale the effort worldwide.Microsoft has been making a version of this argument for months: that the model itself is becoming a commodity, and the real value lies in how well it's woven into a company's workflows. Althoff has said the mistake Microsoft made three years ago was binding Copilot to a single model, when customers actually wanted the flexibility to swap models without losing what they'd built. That's the pitch behind Frontier Company: not a better model, but people on the ground who plug in whichever model fits, and stay long enough to make it stick.