On May 4, 2026, the two largest frontier AI labs stopped pretending they were selling models. OpenAI announced "The Development Company" — a separately capitalized enterprise services venture valued at $10 billion, raising $4 billion from 19 investors led by TPG, Brookfield Asset Management, Advent, and Bain Capital. Hours later, Anthropic unveiled its own joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs as founding partners, valued at $1.5 billion with each anchor committing $300 million.
The two announcements were not coordinated. They were inevitable.
For enterprise buyers, this is the moment the API era officially ended. The two labs that hold roughly 67% of all enterprise LLM API spend (Menlo Ventures data, late 2025) just told the market that selling tokens is not enough — the real margin is in the engineers who deploy those tokens inside Fortune 500 workflows. That changes how every CIO, CFO, and head of AI engineering should think about budget, vendor selection, and lock-in risk for the next 24 months.
What Changed
Both ventures use the same playbook: the forward-deployed engineer (FDE) model, a Palantir invention now being copied by Anthropic, OpenAI, and Google. An FDE is an embedded engineer who sits inside the customer's organization, writes integration code, configures data pipelines, redesigns workflows around AI capabilities, and stays until the system runs in production. It is not consulting. It is closer to staff augmentation with deep model expertise — a hybrid role between SaaS account engineering and big-four delivery.
OpenAI's Development Company raised $4 billion at a $10 billion valuation. Investors include TPG, Brookfield Asset Management, Advent, and Bain Capital — all alternative asset managers, none overlapping with Anthropic's roster. The unit absorbed OpenAI's acquisition of Tomoro, an AI consulting firm that brings roughly 150 deployment specialists into the new entity. OpenAI's framing: "Real impact comes from helping people and organizations use those systems safely, effectively, and at scale" (TechCrunch, May 4, 2026).
Anthropic's joint venture is smaller but better positioned for the mid-market. Blackstone President and COO Jon Gray stated the venture aims to break down "one of the most significant bottlenecks to enterprise AI adoption" — namely, the scarcity of engineers who can implement frontier AI at speed. Beyond the $900 million anchored by the three founders, the venture pulled in Apollo Global Management, General Atlantic, GIC, Leonard Green, and Sequoia Capital. Anthropic's Applied AI team will work alongside the new company to identify use cases and build custom agents.
The two ventures operate at meaningfully different scale: OpenAI's vehicle is roughly 6.7x the valuation and 4x the announced commitment. Yet Anthropic enters with a stronger enterprise position. Menlo Ventures and Ramp data both show Anthropic now captures about 40% of enterprise LLM API spend versus OpenAI's 27%, with Claude winning approximately 70% of head-to-head evaluations among first-time enterprise buyers (analysis here). Over 70% of Fortune 100 companies have already integrated Claude-powered tools, and Claude processes more than 25 billion API calls per month with 45% coming from enterprise users.
The strategic context: Anthropic's ARR grew from $9 billion to roughly $44 billion across 2026 — doubling approximately every six weeks. OpenAI raised $122 billion in March 2026 at an $852 billion valuation, and Anthropic is closing a $50 billion round at a $900 billion-plus valuation. The race to monetize the enterprise installed base — not just to win it — has begun.
Why This Matters
Technical Implications (CIOs, CTOs, Heads of AI Engineering)
Three things change for your architecture playbook starting now:
-
The pilot-to-production gap is being priced explicitly. MIT research cited across enterprise AI literature found that 95% of AI pilots deliver zero measurable P&L impact. S&P Global reported that 42% of companies abandoned most of their AI projects in 2025. IBM put the share of initiatives delivering expected ROI at 25%. OpenAI and Anthropic are now selling the engineering capacity to close that gap — but at premium rates and with vendor lock-in attached.
-
Forward-deployed engineers will become a recurring line item. Until now, the labs sold infrastructure (tokens, fine-tuning, RAG endpoints). The new model bundles infrastructure with embedded labor. Justin Greis, cited in CIO.com's analysis, notes that "a few hundred thousand dollars to get something into production" is not the problem. The problem is vendor dependency afterward — the FDE who built the agent is the only person who can extend it.
-
Multi-LLM architectures are no longer optional. Gartner predicts that by 2028, 70% of organizations building multi-LLM applications will use AI gateway capabilities, up from less than 5% in 2024. The same firm warns separately that "70% of enterprises will be forced to abandon agentic AI solutions from FDE-led engagements because of high vendor costs and lack of internal skills to evolve them independently" (Alex Coqueiro, Gartner senior director analyst). Translation: if your AI strategy is single-vendor with embedded FDEs, you are statistically on track to abandon it within 24 months.
Business Implications (CFOs, COOs, Heads of Strategy)
The 2026 enterprise LLM budget is no longer just a software line. According to Menlo Ventures, average enterprise spending on LLMs jumped from approximately $2.5 million in 2024 to $7 million in 2025 — a 180% increase. Surveyed companies expect another 65% jump to $11.6 million in 2026. Add embedded FDE services on top of that and total enterprise AI spend approaches the range of a mid-sized ERP rollout.
Three business risks deserve board-level attention:
- Concentration risk. Goldman Sachs publicly committed to a single-vendor Anthropic dependency for its OneGS 3.0 platform. ServiceNow took the opposite position in January 2026 and signed multi-year deals with both OpenAI and Anthropic explicitly to give customers "model optionality." Both are defensible. Picking one without a written governance rationale is not.
- Access volatility. Anthropic's Safeguards Team revoked an entire customer organization's access in April 2026. In June 2025 it cut Windsurf's direct access to Claude with less than five days' notice. In August 2025 Anthropic revoked OpenAI's own API access over a terms-of-service dispute. Embedded FDE deployments are not immune to these decisions — they compound the disruption.
- Forward labor cost. AI strategist rates range from $250 to $500 per hour, ML architects $200-$350/hr, LLM specialists $175-$300/hr, and AI engineers $140-$220/hr (market data). An FDE-heavy engagement of 4 specialists at average rates over 6 months runs $1.5-$2.4 million before any model inference cost.
Market Context
The FDE model did not appear from nowhere. Palantir invented it and used it to compound from a $19 IPO in 2021, through a $6 low in 2022, to delivering roughly 640% returns over five years. Q1 2026 earnings showed 85% year-over-year revenue growth, with government revenue up 84% quarter-over-quarter. Palantir's pitch was always: the customer's engineers know the business, our engineers know how to make AI work — embed both, and outcomes happen.
What is new in 2026 is the scale and capital structure of the imitators. Anthropic and OpenAI are not slowly building FDE practices internally. They are spinning out separately capitalized vehicles backed by private equity. That decision matters because PE-backed services entities are structurally optimized to grow headcount fast — and to bill recurring services revenue at high margins.
Three analyst perspectives frame the implications:
- Faisal Kawoosa (Techarc): "IT deployments in the enterprise domain have been consultations or advisory-driven." AI labs want to remain "in the driver's seat" rather than become generic IT vendors. The FDE model is how they keep that position.
- Deepika Giri (IDC): "AI model providers are moving beyond being platform vendors to actively shaping the entire AI value chain." This is a strategic move away from horizontal infrastructure and toward vertical solution delivery.
- Tulika Sheel (Kadence International): Direct AI services "reduce deployment risk" but create "deeper dependency across the stack," increasing long-term vendor lock-in.
For competitive context: traditional Big Four consulting firms (Accenture, Deloitte) charge $150-$300/hr blended for AI engagements, with partner-level rates of $400-$600/hr. Engagements typically run $500K to $5M+. Palantir's own engagements range $500K to $5M+ as multi-year platform commitments. The new FDE model from OpenAI and Anthropic will compete in that same price band but with a different value proposition — closer model expertise, weaker domain depth, faster time to first production deployment.
Framework #1: The OpenAI vs Anthropic vs Big Four vs Hybrid Decision Matrix
Before signing any FDE engagement, score your situation against these four paths. Each is defensible. The wrong choice is the one you make without explicit criteria.
Path A: OpenAI Development Company (FDE-led)
Best fit when:
- You need broad coverage across multiple use cases (sales, support, knowledge, code).
- You are already standardized on GPT-class models and the OpenAI ecosystem.
- You have a $2M+ annual AI engineering budget and can absorb premium FDE rates.
- Your highest-value use case is a generalist agent that touches multiple departments.
- You want a long-term strategic partner aligned with OpenAI's roadmap.
Watch out for:
- Lock-in compounds at the agentic layer faster than at the model layer. Model architecture decisions are 24-month commitments minimum.
- The Development Company has investor incentives to grow services revenue — engagement scope tends to expand.
Path B: Anthropic + Blackstone Joint Venture (FDE-led)
Best fit when:
- You are a mid-market enterprise ($500M-$5B revenue) where the new venture is explicitly targeted.
- Your priority workloads are in financial services, legal, customer operations, or other text-heavy regulated environments where Claude has demonstrated advantage.
- You value Anthropic's safety positioning for compliance or board-level risk posture.
- You want vertical-specific solutions — the Anthropic venture has emphasized financial services and legal.
Watch out for:
- Anthropic's access revocation history (three documented incidents in 13 months) is a real continuity risk.
- The smaller venture size means fewer FDEs available — engagement queues may extend.
Path C: Big Four / Palantir (Traditional consulting + AI specialists)
Best fit when:
- You need deep domain expertise (regulated industries, complex change management, legacy integration).
- You want model-agnostic delivery — the consultant is happy to deploy any LLM you choose.
- Your project requires significant business process redesign, not just AI integration.
- You have an existing relationship that lowers vendor risk and contracting overhead.
Watch out for:
- Slower time-to-production than FDE-led engagements.
- Less depth on cutting-edge model behavior (prompt engineering, agentic patterns, evals).
Path D: Hybrid (AI Gateway + Internal Build + Selective FDE)
Best fit when:
- You have an in-house AI engineering team of 5+ engineers.
- Your strategy explicitly requires multi-model portability (the Gartner-predicted majority by 2028).
- You can afford to absorb the 6-12 month ramp on internal capability.
- Your CFO has flagged vendor concentration as a board-level risk.
Watch out for:
- Build cost and time-to-first-value are highest in this path.
- Requires senior leadership (VP of AI Engineering or equivalent) to be effective.
Quick Scoring
| Criterion | OpenAI Dev Co | Anthropic JV | Big Four / Palantir | Hybrid |
|---|---|---|---|---|
| Time to first production | 3-6 months | 3-6 months | 6-12 months | 9-18 months |
| Domain depth | Medium | Medium | High | Variable |
| Model depth | High | High | Medium | High (if hired) |
| Lock-in risk | High | High | Medium | Low |
| Annual cost (typical) | $2-5M | $1-3M | $500K-$5M | $1.5-4M |
| Multi-model portability | Low | Low | High | High |
Decision rule: if your AI strategy spans more than 18 months and crosses more than two business lines, the math almost always favors Path D (Hybrid) or Path C (model-agnostic Big Four) over single-vendor FDE lock-in.
Framework #2: The 90-Day Enterprise AI Vendor Lock-In Audit
Before you sign or extend any FDE-led engagement, run this audit. It exists for one reason: Gartner's prediction that 70% of FDE-led agentic AI projects will be abandoned by 2028 because of high vendor costs and the inability to evolve them independently. You do not want to be in that 70%.
Week 1: Exit Cost Inventory
- List every system the FDE engagement will touch (data sources, identity, observability, internal APIs).
- For each, document whether the integration uses a vendor-specific construct (e.g., OpenAI Assistants API, Claude Computer Use, Anthropic memory primitives) or an open standard (REST + JSON, OpenAPI specs).
- Estimate replacement cost if you needed to swap models in 12 months. If you cannot estimate it within 30%, your dependency is opaque — fix that before signing.
Week 2: Knowledge Transfer Plan
- Require a written knowledge-transfer artifact for every FDE deliverable: prompt library, eval suite, agent architecture diagram, integration runbook.
- Negotiate a minimum number of pair-programming hours between FDEs and your internal engineers. Anthropic's existing FIS engagement included this — make it standard, not optional.
- Identify the two internal engineers who will own the system after FDEs depart. Their names go in the SOW.
Week 3: Multi-Model Optionality
- Sit any new agent or workflow behind an AI gateway (LiteLLM, Portkey, Langfuse, or equivalent). Gartner expects 70% of multi-LLM applications to do this by 2028; doing it now is cheap.
- Define an objective evaluation harness independent of the vendor — 50+ task examples, scored automatically.
- Run the same harness against a second model every quarter. The discipline alone changes vendor behavior.
Week 4: Contract Levers
- Include a "right to extend" clause: your internal engineers can modify any FDE-built artifact without voiding warranty or SLA.
- Cap services billing at a fixed percentage of inference spend. Without a cap, services revenue compounds independently of value delivered.
- Include access continuity language. Anthropic's three documented revocations and OpenAI's capacity-tier shifts are precedent. Demand notice periods of 90+ days for production workloads.
Common Challenges and Solutions
- Challenge: FDE scope expands beyond original engagement. Solution: tie scope to a fixed number of agent "use cases," each with a written success criterion. Net-new use cases require change orders.
- Challenge: Internal engineers cannot read FDE-built code. Solution: mandate code review on every PR from the start. If FDEs push without internal review, you bought labor, not capability.
- Challenge: Eval coverage is verbal only. Solution: require a versioned eval suite checked into your repo before any system goes live. No suite, no go-live.
- Challenge: Model upgrades break production. Solution: pin model versions in production with explicit upgrade windows. Don't auto-adopt.
- Challenge: Renewal pricing surprises. Solution: lock multi-year pricing during initial engagement when leverage is highest.
Real-World Example: Anthropic + FIS + Bank of Montreal
Anthropic's first major FDE engagement is already in production. FIS, the financial services platform provider, embedded Anthropic Applied AI engineers to co-design its Financial Crimes AI Agent. The agent compresses anti-money-laundering investigations from hours to minutes and is now in initial deployment with Bank of Montreal (BMO) and Amalgamated Bank.
Three things are worth highlighting about how this engagement was structured.
First, knowledge transfer was explicit in scope. Anthropic stated publicly that the engagement would "transfer knowledge so FIS can build and scale additional agents independently over time." This is the right model — and it is the exception, not the norm. Most FDE engagements do not write explicit knowledge-transfer obligations into the SOW.
Second, the financial impact is concentrated and measurable. AML investigations are a high-cost, high-volume, regulated workload. Compressing investigation time from hours to minutes maps to a specific cost-per-investigation savings that the bank's finance team can attest to. If your candidate FDE engagement does not have a similarly precise economic anchor, the project is more likely to drift.
Third, the integration footprint is contained. The agent runs inside FIS's existing platform with connectors and "ready-to-run" templates. The customer (BMO, Amalgamated) is not absorbing Anthropic-specific architecture into its core systems — the dependency lives at the FIS layer. That is a defensible position.
The lesson for buyers: when evaluating FDE engagements, ask whether the system being built lives inside your core stack or inside the vendor's platform. The latter contains the lock-in. The former multiplies it.
What to Do About It
For CIOs and CTOs
- Stand up an AI gateway in the next 60 days. Whether you use LiteLLM, Portkey, or a homegrown router does not matter as much as having a routing layer between your applications and your model providers. This is the cheapest insurance you can buy against access volatility.
- Create a written FDE engagement playbook. Every FDE SOW should be reviewed against the same internal checklist — scope, exit criteria, knowledge transfer, contract levers. Make sure no team can sign one without it.
- Mandate dual-vendor evaluation. For any new agentic system, require an objective comparison between at least two models. ServiceNow's "model optionality" framing is the right CIO posture in 2026.
For CFOs
- Budget FDE labor as a multi-year line. A pilot starting at $300K is likely to extend to $1.5M+ once scope expands. Plan for it, do not be surprised by it.
- Cap services-to-inference ratio. A reasonable ceiling is 1.5x inference spend in year one, dropping to 0.5x by year three. If services revenue grows independent of inference, you are paying for capability transfer that is not happening.
- Demand quarterly ROI evidence. Every FDE-built system should produce a quarterly impact memo with specific cost-saved or revenue-generated numbers. No memo, no renewal.
For Heads of AI Engineering and Business Leaders
- Identify your internal "future owners" before the FDE arrives. The two engineers who will run the system after FDEs leave should be named in the SOW and present in every working session.
- Establish governance early. Who approves new use cases? Who owns evals? Who decides on model upgrades? Decide these questions before the engagement starts, not at month 9 when the first conflict appears.
- Treat the engagement as capability transfer, not a finished product. If you finish the engagement with a working agent but no internal engineers who can extend it, you bought a hostage situation, not an asset.
The two-vendor announcement on May 4 is the clearest signal yet that the enterprise AI market has matured past the "buy tokens, hire consultants separately" phase. The labs that built the models now want to deliver the labor that turns those models into outcomes. That is a rational move for them — and a manageable risk for you, if you treat the engagement as a capability-transfer project, not a turnkey service.
The buyers who get the next 24 months right will be the ones who said yes to FDE labor and no to FDE dependency. Everyone else will discover, around 2028, that they are in Gartner's 70%.
