Goldman Called Claude a Security Risk. Then It Helped Sell It.
Anthropic, Blackstone, Hellman & Friedman, and Goldman Sachs launched a $1.5 billion joint venture on May 4 to sell Claude to hundreds of mid-market companies — and on April 29, Goldman Sachs had barred its own Hong Kong bankers from using the same model.
The venture Blackstone confirmed the partnership in a May 4 press release. Anthropic, Blackstone, and Hellman & Friedman each committed $300 million to the new enterprise AI services firm, with General Atlantic, Leonard Green, Apollo Global Management, GIC, and Sequoia Capital also in the investor consortium. The firm targets mid-size companies across healthcare, manufacturing, financial services, retail, real estate, and infrastructure — sectors where Anthropic's blog post said enterprise demand for Claude was significantly outpacing any single delivery model.
On April 29, Goldman Sachs removed Claude access for its Hong Kong bankers, citing data security concerns. The ban was specific to Claude — Google's Gemini and OpenAI's ChatGPT remained available on the same internal platform. The bank has not said what changed between April 29 and May 4 that made the same model acceptable as an investment vehicle when it was not acceptable as an internal tool.
Goldman Sachs declined to comment on why a model it considered too risky to run internally was an appropriate vehicle to sell to enterprise customers five days later. Anthropic pointed to its May 4 announcement and did not address the April 29 restriction.
The venture structure offers a partial explanation. The joint venture is a separate legal entity staffed by implementation partners, not a direct Goldman deployment. Mid-market customers buying through the channel would work with systems integrators and resellers rather than Goldman bankers — a different governance surface area than an internal bank platform. But the venture structure does not fully close the logical gap. The question is not whether a separate entity has different governance terms — it does — but whether the data security concern that triggered the internal ban was a firm-specific technical judgment about Claude's data handling, or a broader posture that would apply across any Claude deployment. If the concern was specific to how Claude handles data in a Hong Kong banking context, that concern does not disappear when the same model is deployed at a mid-market manufacturer through a different legal entity. If the concern was about Goldman's internal compliance infrastructure rather than Claude itself, the investment makes more sense — but Goldman has not said which it was.
The narrowness of the ban is the only public signal available. Goldman's internal platform blocked only Claude while leaving Gemini and ChatGPT accessible — a distinction that suggests the concern was specific to Anthropic's model or its data handling posture for that market, not AI systems in general. No other bank has been publicly identified as implementing a similar Claude-specific restriction in Hong Kong, which isolates the ban to Goldman specifically rather than confirming it as a sector-wide compliance posture.
The five-day window between the ban and the announcement is short for a policy reversal of this kind — the venture would have required due diligence, term negotiation, and legal review that typically takes longer. One reading: the investment was under development before the restriction was issued, and the ban reflected a compliance decision made independently of the investment track. Another reading: the restriction and the investment reflect separate decision-making chains that did not coordinate. Both readings remain open without comment from Goldman on the due diligence process.
The venture structure points to a broader industry pattern. OpenAI launched a parallel enterprise services joint venture — now reported by TechCrunch under its updated name, The Deployment Company — suggesting the channel-based mid-market strategy is an industry-wide approach rather than an Anthropic-specific bet. That convergence means the question of whether the Goldman contradiction is idiosyncratic or representative matters for the broader enterprise AI distribution model.
What to watch: whether other financial institutions show the same pattern, and whether the venture's implementation partners provide technical documentation addressing the specific concerns that triggered the Goldman ban. Enterprise buyers evaluating the new channel should ask directly whether the data security concerns behind Goldman's internal restriction apply to their use case — and whether the channel's governance structure addresses those concerns substantively or merely legally.