Last Monday, Anthropic dropped an announcement that had nothing to do with a new model, a benchmark, or a safety paper. Instead, the company unveiled a $1.5 billion enterprise AI services joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs, with Apollo Global Management joining as a founding partner shortly after. Within hours, OpenAI announced its own competing vehicle, "The Deployment Company," a $10 billion venture majority-owned by OpenAI and backed by TPG, Brookfield, SoftBank, and others.
Two major AI labs. Two blockbuster Wall Street partnerships. One very clear signal about where the battle in enterprise AI is headed.
The Real Fight Is Now Over Implementation, Not Models
For the past three years, the AI conversation in every boardroom has revolved around models. Which one is smarter? Which scores better on MMLU? GPT-4 vs. Claude 3 vs. Gemini — the benchmarks flew, the comparisons piled up, and differentiation between frontier models gradually compressed toward zero. The models got good fast. Really fast.
In enterprise software, when the model layer commoditizes, value always migrates to implementation.
The Anthropic-Goldman-Blackstone venture is the institutional acknowledgment of exactly that reality. Anthropic's stated rationale for creating the firm is telling: "demand for Claude is outpacing any single delivery model." Companies want to use Claude. They just can't figure out how to make it actually work inside their organizations. So Anthropic is sending in engineers and bringing in capital-backed partners to solve the last-mile problem at scale.
This isn't a software sales play. It's a services and transformation play. The two are very different businesses.
The Hard Part Was Never the Model
Every experienced enterprise technology consultant will tell you that the hardest part of any major software implementation — SAP, Salesforce, you name it — is never the software itself. It's the organizational plumbing: the messy data, the legacy integrations, the change management, the politics of getting 14 different departments to agree on a workflow.
AI amplifies this by an order of magnitude.
A 2025 MIT NANDA report on the state of AI in business found that 95% of enterprise generative AI pilots fail to deliver measurable P&L impact. Not because the models don't work. Because they never make it into the actual workflows that drive business outcomes. Gartner predicts that through 2026, organizations will abandon 60% of AI projects that lack AI-ready data. Meanwhile, 63% of organizations either don't have or aren't sure they have the right data management practices for AI in the first place.
The barriers are almost never technical. They're organizational: fragmented data, siloed teams, no executive sponsor, no clear business case tying AI output to a revenue or cost line. A survey of over 1,000 enterprises found that the rate of abandoned AI initiatives jumped from 17% in 2024 to 42% in 2025, more than doubling in a single year.
That 42% abandonment figure is the implementation gap made concrete, and it's exactly what Anthropic's new venture is designed to address. The new firm will embed small teams of applied AI engineers directly inside client organizations. They'll identify where Claude can have the most impact, co-build custom solutions, and stay engaged as models and use cases evolve. It's Palantir's forward-deployed engineer model, proven at scale for high-stakes deployments, now applied to the broader commercial market.
This is what AI services actually looks like. Not slide decks and strategy workshops. Engineers embedded with your team, building systems that run in production.
The Strategic Logic for Blackstone, Goldman, and Apollo
The financial partners here aren't passive investors looking for a return. Their involvement is structurally strategic.
Blackstone, Apollo, and Goldman collectively control stakes in hundreds of portfolio companies. Private equity-backed CFOs are under enormous pressure to demonstrate AI-driven efficiency before exits (85% of PE buyers now factor AI capabilities into their valuations). By co-founding an AI services firm built around Claude, these firms gain a captive, deeply capable implementation partner they can deploy across their portfolio at scale. They become both investors in and customers of the venture.
The PE firms get a competitive advantage across dozens of portfolio companies at once. Anthropic, in turn, gets a distribution channel pre-loaded with motivated clients whose investors own a piece of the outcome. Few enterprise software arrangements are structured quite this neatly.
The Sovereign Wealth Fund Institute called this "less a routine joint venture than a deliberate institutional response" to the deployment problem. The distinction matters: a routine JV is passive capital. This is active infrastructure, a purpose-built pipeline for getting Claude into core operations at scale.
Who Gets Left Out
Anthropic's framing is nominally about "mid-sized companies." But mid-sized here means companies held in the portfolios of multi-billion-dollar private equity firms. GIC, Sequoia, TPG, Brookfield, and Bain Capital are all additional backers. The portfolio network this firm will serve is enormous, well-capitalized, and firmly in the large-enterprise tier by any reasonable definition.
If your company doesn't have a PE parent or Goldman Sachs on your board, you won't be getting an Anthropic applied AI team embedded with your engineering team anytime soon. The economics don't work at that scale.
This creates a notable divide. The companies that most need structured AI implementation support — growth-stage businesses, independent mid-market firms, companies in sectors without heavy PE concentration — are precisely the ones this new venture isn't built to serve. Meanwhile, the existing systems integrators that Anthropic cites as part of its Claude Partner Network — Accenture, Deloitte, and PwC — operate at price points and timelines (6-12 months, seven-figure engagements) that are equally out of reach for most mid-market organizations.
The MIT NANDA report put it clearly: there's a GenAI Divide where just 5% of enterprise AI initiatives are extracting meaningful value while the vast majority remain stuck. The divide doesn't run along industry or geography lines. It runs along capability and capital lines.
The new Anthropic-Goldman-Blackstone venture will almost certainly help the 5% get further ahead. The question is what it signals for everyone else.
What to Actually Do About It
The following is what I'd prioritize if I were running a mid-market company right now.
Own Your AI Roadmap Before Anyone Else Does
The biggest risk the new mega-ventures introduce for mid-market companies isn't being priced out. It's being rushed into decisions without a clear strategy. As model providers expand into services, the vendor relationships formed in the next 12-18 months will be extraordinarily hard to unwind. Implementation creates switching costs that compound every month your data and workflows get more entangled with a specific provider's ecosystem.
Before engaging any external AI partner, nail down your own use-case priorities. Which workflows, if improved by 30%, would move a meaningful needle on revenue or cost? What data do you actually have, and what state is it in? Answer these questions before an enterprise sales team answers them for you.
Judge Partners on Track Record, Not Brand Names
The announcement that Anthropic has Goldman and Blackstone behind it is impressive branding. For a mid-market company, what actually matters is whether a partner has shipped AI into production environments like yours — similar industry, similar data complexity, similar organizational size. Ask for references. Ask what failed and why.
Boutique AI engineering firms and technically deep specialists can often deliver production results in 6-8 weeks at a fraction of the cost of large consulting engagements. The right partner is less likely to be the biggest name in the room and more likely to be the one asking the most pointed questions about your data and your team's existing capabilities.
Fix Your Data Before Anything Else
This one deserves more weight than anything else on this list. Gartner's finding that 63% of organizations lack proper AI data practices isn't a peripheral concern. It's the single variable that most consistently separates successful implementations from abandoned ones.
Before spending a dollar on an AI services firm, audit your data. Where does it live? Can you get it into an AI system without rebuilding three legacy integrations first? Is it clean enough to be useful? Organizations that treat data readiness as a prerequisite see dramatically better ROI outcomes than those who approach it as an afterthought. The best AI implementation partner in the world can't fix bad data. No one can.
Watch the Vendor Lock-In Risk
Anthropic is now building both the model layer and the services layer, which means the firm is structurally incentivized to embed Claude as deeply as possible into client operations. Every custom integration built on Claude-specific APIs, every RAG architecture that assumes Claude's context window, is a switching cost. If a better or cheaper model emerges in 18 months, you want to be able to move.
Demand that any AI implementation partner design for model portability. Abstract the model layer through an orchestration framework. Keep your data pipelines and business logic separate from whatever model you're running today.
Build Internal Muscle in Parallel
The Palantir FDE model produces the best outcomes when the client's own team learns alongside the embedded engineers. Whatever external partners you bring in, insist on genuine knowledge transfer. Build internal AI literacy. Hire or develop at least one person who deeply understands AI workflows, data pipelines, and implementation-level engineering, not just prompt writing.
Companies that treat AI as a pure outsourcing problem stay permanently dependent. The goal is to build enough internal capability so that when an engagement ends, you can operate and evolve what was built.
The Bigger Picture
The Anthropic-Goldman-Blackstone venture, alongside OpenAI's competing Deployment Company, marks a clear inflection point in how AI will actually reach enterprise users. Both lab founders have concluded, correctly, that they can't just hand companies an API and wish them luck. The gap between "we have access to Claude" and "Claude is running in our core operations" is wide, expensive, and full of organizational complexity. Services close that gap.
Goldman Sachs won't build your AI strategy for you. But the companies that do that work themselves, before the first vendor call, are the ones who won't need Goldman Sachs to.