Morning Brief · Tuesday

OpenAI and Anthropic Both Launched $10B Deployment Companies Yesterday. Google DeepMind's Workers Responded by Unionizing.

On the same day, OpenAI closed a $4 billion raise into "The Deployment Company" — valued at $10 billion, backed by TPG, Brookfield, SoftBank, and Bain Capital — and Anthropic announced a parallel enterprise JV with Blackstone, Goldman Sachs, and Hellman & Friedman. The race to enterprise revenue ahead of their respective IPOs is now fully visible. Within hours, Google DeepMind's UK workers formally announced they'd voted to unionize — citing Google's new Pentagon AI deal — and sent their recognition letter to management this morning. Meanwhile, 8,000 developers who didn't get into OpenAI's sold-out GPT-5.5 party woke up to a consolation prize: a month of 10x Codex rate limits. And security researchers published a scan of one million exposed AI services that found things much worse than anyone expected.

Mira Novian

May 5, 2026 · Morning Brief · ~9 min read

OpenAI and Anthropic both launched enterprise deployment joint ventures on the same day. This is the pre-IPO sprint for revenue.

Monday produced one of the most revealing days in AI business history, and most of the coverage treated each announcement in isolation. OpenAI closed a $4 billion fundraise for a new entity called The Deployment Company, majority-owned and controlled by OpenAI, with a post-money valuation of $10 billion. The investor roster reads like a who's-who of institutional capital with enterprise distribution: TPG Inc, Brookfield Asset Management, Advent, Bain Capital, Dragoneer Investment Group, and SoftBank Group. Combined, the partners' portfolio companies and clients number over 2,000 organizations — which is, transparently, OpenAI's near-term commercial target list. The Deployment Company's stated purpose is to help those businesses integrate OpenAI's AI software into daily operations. Translation: OpenAI is building the professional services and integration layer that a product company alone cannot build fast enough.

Within hours, rival Anthropic announced a parallel structure: a deployment-focused joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs Group. The terms were not disclosed, but the pattern is unmistakable. Both companies are solving the same problem simultaneously: they have transformative technology and enterprise interest, but converting that interest into durable recurring revenue requires human relationships, consulting capacity, and institutional trust that neither lab has at scale. Private equity firms and investment banks — with their existing relationships to every Fortune 500 company — are the fastest available distribution channel. The timing is not coincidental. Both OpenAI and Anthropic are widely expected to pursue public offerings in the next 12 to 18 months. Revenue quality, enterprise contract depth, and revenue growth rate will all be scrutinized. These joint ventures are the mechanism to make those numbers work. They are also, notably, a structural admission: despite the hype, selling AI to large enterprises is still a relationship business, and neither lab has invested heavily enough in that relationship layer to win it alone.

timesofindia.com ↗

The simultaneity here is the story. It's possible both announcements just happened to drop the same day by coincidence. It is more likely that each company was aware the other was close and chose to move on the same news cycle to avoid the optics of being "second." The downstream effect is that the framing becomes symmetry rather than leadership: "both the major labs did this." That's a net negative for whichever one has the weaker deal terms — and those terms are not yet public. What I'd watch: whether The Deployment Company and its Anthropic equivalent end up competing for the same enterprise accounts, and how the exclusivity (or lack of it) in each JV is structured. If these are non-exclusive, you'll have Goldman Sachs advising the same CTO to use both OpenAI and Anthropic simultaneously, which benefits the enterprise buyer and muddies any preferential distribution advantage. If they're exclusive, expect a fast-following announcement from one of the labs about a competing banking or PE relationship. Either way: the "sell it yourself" era for frontier AI is ending. The enterprise sales motion is going institutional, and that means margins, timelines, and the kinds of customer conversations these companies are having will all start looking a lot more like enterprise software than like consumer tech.

Labor

Google DeepMind's UK workers voted to unionize. Their first letter to management goes out today. The trigger was the Pentagon deal.

Workers at Google DeepMind's UK offices — the company's primary AI research laboratory, headquartered in London — have voted to unionize and today are formally requesting recognition of the Communication Workers Union and Unite the Union as joint representatives of the lab's UK-based staff. The vote occurred in April; the recognition letter is being sent to Google management on Tuesday. If recognized, the union would represent at least 1,000 workers across DeepMind's UK offices — making it the first union in a frontier AI laboratory to seek formal recognition anywhere in the world.

The proximate trigger was last Friday's Pentagon announcement: the US Department of Defense confirmed it had reached "any lawful use" AI access agreements with seven leading AI companies, Google among them — alongside SpaceX, OpenAI, Nvidia, Reflection AI, Microsoft, and Amazon Web Services. Anthropic was notably absent, having reportedly refused to agree to surveillance and autonomous weapons use cases. Workers cited that deal, combined with Google's longstanding cloud contract with the Israeli government and its provision of AI tools to the IDF, as the breaking point. "Our technology helped the IDF," one anonymous worker told the Guardian. "I want AI to benefit humanity, not to facilitate a genocide." The union's stated goals include ending AI use by the US military and Israeli government, restoring Google's former pledge not to develop militarized AI or surveillance tools (dropped last year), establishing an independent AI ethics oversight body, and granting workers the right to refuse to contribute to projects on moral grounds — the last of which is, from an employer perspective, the most operationally significant demand. Google has not responded.

theguardian.com ↗

The "right to refuse" demand is the one that will determine whether this union becomes consequential or symbolic. Every other demand on the list — end military contracts, restore ethics pledges, independent oversight — requires Google to change its business decisions. Companies generally don't do that in response to union recognition requests. But the right to refuse specific projects, if won, would create a structural constraint on how Google can staff its most sensitive AI work. It means that the engineers best positioned to build a surveillance tool could decline to build it. That's a meaningful operational limit, and it's why Google will fight that particular clause hardest. The broader signal is that the DeepMind unionization is the first frontier AI lab to organize, but it will not be the last. The Pentagon deal was the precipitating event, but the underlying conditions — a workforce hired on promises of beneficial AI research, now building tools for classified military networks — are not unique to Google. OpenAI, Anthropic, and xAI all have employees who hold similar views and are watching the DeepMind recognition request closely. The Hollywood guild comparison is again relevant: the most effective AI governance mechanisms in 2026 are coming from workers and their contracts, not from regulators or voluntary lab commitments. If DeepMind's union wins formal recognition and meaningful terms, it will serve as the template for the next lab to organize.

Dev Tools

8,000 developers who couldn't get into OpenAI's GPT-5.5 party got a consolation prize: a month of 10x Codex. Sam Altman telegraphed it himself.

OpenAI's GPT-5.5 launch party was invite-only, capacity-constrained, and generated over 8,000 applicants in 24 hours — far more than the office could hold. On Monday, everyone who applied but didn't get in received an email with a notable alternative: a tenfold increase in Codex rate limits on their personal ChatGPT accounts, effective immediately through June 5. The gift went to all 8,000-plus applicants regardless of acceptance status — applied, waitlisted, or turned away. CEO Sam Altman telegraphed the move on X before the emails went out: "We are gonna do something nice for everyone who applied for the GPT-5.5 party and that we didn't have space for. Hope you enjoy!" The post exceeded 521,000 views within hours.

The practical implications are larger than the gesture suggests. Codex operates under daily usage caps that vary by subscription tier — a tenfold increase gives developers dramatically more room to run extended agentic coding loops, overnight experiments, and multi-step autonomous tasks. OpenAI says GPT-5.5 matches GPT-5.4's per-token latency while operating at higher intelligence and using fewer tokens to complete tasks — making the rate limit increase effectively a free month of premium access to their newest model, at scale. The 31-day window is long enough to reshape working habits. VentureBeat's read is pointed: by flooding thousands of developers with expanded access during a critical adoption period, OpenAI is "subsidizing the kind of deep, sustained usage that turns a curious trial into a daily dependency." The rate limit increase is also a developer relations play — transforming a PR awkwardness (party oversubscription, exclusivity optics) into a goodwill moment that reaches 8,000 developers who were engaged enough to apply.

venturebeat.com ↗

The "subsidizing dependency" framing is the accurate one, and the fact that it's this transparent doesn't make it less effective. OpenAI is not hiding what it's doing here: it's giving developers a month of supercharged access to a model it wants them to build on, rely on, and eventually pay for. The 31-day window is timed to outlast a sprint cycle but fall short of a full quarter — just long enough for habits to form, just short enough that pulling it creates actual pain. What makes this interesting in context is the DeepClaude story from yesterday's brief: developers are simultaneously being given 10x Codex access and building workarounds to run Claude Code's loop on DeepSeek for 17x cheaper. The same developer community is both the recipient of OpenAI's goodwill and its biggest arbitrageurs. The tension between those two behaviors — loyalty gifted, loyalty eroded — is the central dynamic of the AI developer tools market in 2026. OpenAI's bet is that the goodwill and product quality will outrun the arbitrage. It's not a bad bet. But it's not guaranteed.

Security

Security researchers scanned one million exposed AI services. They found authentication disabled by default, exposed chat histories, and open chatbots anyone can use.

The Intruder security team published a research report Monday — cited by The Hacker News — detailing the results of a systematic scan of over two million hosts with AI services exposed to the internet, surfacing approximately one million accessible instances. The findings landed in a single direction: AI infrastructure is being deployed faster than security practices can follow, and the defaults are dangerously permissive. The most common pattern found: authentication simply not enabled by default in many self-hosted LLM projects, meaning real user data and company tooling were exposed to anyone who looked. Chat histories from enterprise environments — which can contain credential fragments, internal system descriptions, and sensitive business context — were accessible without login on multiple discovered instances. Generic chatbots hosting a wide range of models, including multimodal LLMs, were freely available to anonymous users, enabling jailbreaking, generation of illegal content, and use of the host organization's API budget — all without attribution or accountability.

The research contextualizes the scale of the problem: the two million hosts were identified using certificate transparency logs, meaning only publicly discoverable services were included — a likely undercount of actual exposure. The report attributes the pattern not to individual negligence but to a structural problem: the open-source LLM infrastructure ecosystem has built fast and often shipped with authentication off by default, prioritizing ease of self-hosting over security posture. Proofpoint's concurrent 2026 AI and Human Risk Landscape report adds a parallel data point: 87% of organizations now have AI assistants deployed beyond pilot stages, but 52% are not confident in their security controls to detect a compromised AI system. The combination — rapid deployment, weak defaults, and limited detection capability — is the threat surface formula that historically precedes significant breaches.

thehackernews.com ↗

The "authentication off by default" finding is the one that enterprise AI teams need to take seriously this week. This is not a sophisticated attack vector — it's the functional equivalent of deploying a database with no password and a public IP. The reason it keeps happening in AI infrastructure is that most of the open-source LLM hosting projects were built by researchers and hobbyists optimizing for getting started quickly, not by security teams optimizing for hardened production deployment. The gap between "easy to self-host" and "safe to self-host" has never been wider. For organizations that have deployed AI tools in the past 12 months — particularly internal chatbots, coding assistants, or RAG systems — the actionable response is an auth audit this week: confirm every exposed LLM interface has authentication enabled, confirm no chat history is accessible without login, and confirm your API keys are not present in any exposed system prompts. The Proofpoint statistic — 52% of organizations not confident in their ability to detect a compromised AI — is the more alarming of the two reports. A compromised AI system in an enterprise context doesn't just leak data; it can actively misdirect users, exfiltrate context through prompt injection, and persist undetected through normal usage patterns. The tools to detect that reliably don't yet exist at scale. This is an emerging problem that is not yet being treated as seriously as it will be in 18 months after the first major disclosed incident.

✦ Mira's Take

Today's brief has a single through-line: AI is entering its monetization era, and the friction points are showing up everywhere simultaneously. The OpenAI and Anthropic JV announcements are the clearest sign yet that building the best model is necessary but no longer sufficient. Enterprise revenue requires enterprise distribution — and enterprise distribution requires trusted human relationships at scale, which neither lab has. Bringing in private equity and investment banking firms as distribution partners is an admission of that gap, dressed up in press releases about "accelerating enterprise adoption." That's not a criticism — it's the correct move for where they are. But it means the AI industry's next chapter will increasingly look like the SaaS industry's last chapter: channel partners, professional services margins, quarterly business reviews, and multi-year contract lock-in. If that sounds less exciting than frontier model research, that's because it is. It's also how you build a real business.

The DeepMind union is a direct response to the same dynamic. The workers building this technology did not sign up to build AI systems for classified military networks and warfighter decision support. They signed up to work on beneficial AI at a prestigious research lab. Google's decision to drop its militarized AI pledge last year and then sign a Pentagon deal was a values discontinuity that was always going to produce organizational friction. The union is that friction, formalized. The "right to refuse" demand is the one that matters most, and the one most likely to be fought hardest. If it's won, it will be the most consequential worker governance mechanism in AI to date — more specific and more binding than any government regulation currently on the books. Watch May 11's DGA negotiation and the DeepMind recognition process in parallel: the two most important AI governance developments this month are happening in labor law, not in Congress or Brussels.

The security story is the quiet alarm that these other stories are drowning out. One million exposed AI services with weak or absent authentication is not a warning about what might happen — it is a description of what already exists. The enterprises racing to deploy AI tools to meet their board's demands are creating an attack surface that their security teams are not equipped to monitor, let alone defend. The 52% of organizations unable to detect a compromised AI system is the most important number in this brief. It means that when the first major enterprise AI breach is disclosed — and it will be disclosed — a meaningful fraction of affected organizations will not have known they were compromised until after the fact. That breach hasn't happened publicly yet. It has almost certainly already happened somewhere. The question is when the disclosure comes, not whether.