The FDE Arms Race: AI Companies Spend $10B on Forward-Deployed Engineering
AI companies have quietly committed ~$10B in 12 months to forward-deployed engineering (FDE), embedding engineers inside customer organizations to install and integrate AI. The tension is stark: MIT’s “GenAI Divide” report found that 95% of enterprise GenAI pilots deliver no measurable P&L impact, despite companies spending ~$684B on AI in 2025. The bottleneck has shifted from model capability to deployment—GPT-4, Claude, and Gemini are powerful enough, but enterprises cannot configure and operate them without dedicated engineering teams on-site. FDEs are the implementation layer that turns model access into business outcomes, and capital is flooding into deployment faster than model development.
Three structural models are emerging. The internal army (Microsoft, Amazon, Salesforce) funds FDE from balance sheets, reusing existing employees for speed and control but unable to isolate P&L. The PE-backed JV (OpenAI‘s Deployment Company at $4B with a $14B post-money valuation, Anthropic‘s JV at $1.5B from Blackstone and others) enables scale without diluting the parent, but risks misaligned incentives between PE return floors and maximum deployment. The original is Palantir, which invented FDE and maintains ~400-500 engineers (~12% of headcount) as a core product. The talent math doesn’t work: FDE postings surged 42x from 2023 to 2025, while candidate pools grew only ~50% year-over-year. Senior FDE compensation at labs ranges from $350K to $550K, far above Palantir‘s $215K median. The single forward-deployed engineer label is fragmenting into five distinct sub-roles.
The real strategic question is whether FDE investment is a moat or a toll booth. OpenAI‘s acquisition of Tomoro creates switching costs: once engineers build custom workflows on OpenAI APIs, switching to Anthropic or Google means rebuilding. Anthropic‘s PE structure gives it access to 275 Blackstone portfolio companies as captive Claude customers. If model quality converges, the company owning the deployment relationship captures the customer. Venture-backed FDE startups like Riplo and Xavier AI face existential threat from labs that can outbid and absorb them. The companies that win enterprise AI will not be those with the best models, but those with the most engineers inside customer offices. However, scaling Palantir‘s proven model 10x may break the economics—the talent war, compensation inflation, and misaligned incentives make the path uncertain.
