A decade ago, FAANG (Facebook, Amazon, Apple, Netflix, and Google) was shorthand for the firms that defined the consumer internet and mobile‑cloud era. Today, MANGOS has replaced it in investor shorthand as the roster of companies that define the AI era.
We are entering the MANGOS decade - Meta, Anthropic, Nvidia, Google, OpenAI, and SpaceX - and the agentic enterprises that depend on them will increasingly run on borrowed governance rather than their own.
The timing is not accidental. SpaceX has just raised roughly seventy‑five billion dollars in what is widely described as the largest IPO in history and quickly traded above a two‑and‑a‑half‑trillion‑dollar market value. Days later it agreed to acquire Anysphere, the company behind the Cursor coding agent, in a roughly sixty‑billion‑dollar all‑stock deal, the largest acquisition of a venture‑backed startup on record. OpenAI has filed confidentially for an IPO that could value it near eight hundred fifty billion dollars, while Anthropic, last valued near a trillion, is preparing to follow.
For Wall Street, MANGOS is a theme to trade. ETF issuers are already filing products to track the basket. For public‑company boards, the acronym should signal something different: a structural dependency map. The same companies investors are racing to own are the companies your enterprise increasingly runs on.
Every AI agent you deploy and every workflow you automate sits on someone’s model, someone’s chips, and someone’s runtime. When that “someone” is Meta, Anthropic, Nvidia, Google, OpenAI, or SpaceX, you inherit more than capability. You inherit their effective constitution: the technical and policy defaults that determine what your agents can do, what they must refuse, and how they behave under stress.
And the acronym understates the reality. Microsoft, Amazon, and Apple do not appear in MANGOS, yet Azure is the primary host and reseller for OpenAI, AWS underpins Anthropic and much of the AI cloud market, and Apple sits atop the device layer and on‑device inference path for more than a billion users. Acronyms are market language. Dependency maps are governance instruments. A realistic one includes at least these nine firms.
Cheap intelligence, infinite appetite
Why will this dependency deepen rather than plateau? Because AI is subject to the same Jevons paradox economists observed with coal: efficiency gains in using a resource often increase, rather than decrease, total consumption.
Over the last few years, the cost of a unit of machine intelligence—measured in inference cost per token or per task - has fallen sharply as models have improved and inference infrastructure has matured. Analysis from AI economics and sustainability researchers describes an “AI Jevons paradox”: AI adoption reduces per‑unit resource use but increases total usage as firms do more things with the technology.
Inside companies, that dynamic looks like this:
- A legal review that takes seconds and costs pennies is run on every contract, not just a subset.
- A pricing agent that can run thousands of micro‑experiments per day is left running continuously.
- A coding agent that can read and refactor a million‑line codebase turns what used to be quarterly rewrites into a daily background process.
The result is the agentic enterprise: not just employees with copilots, but standing populations of software agents that perceive, decide, and act across thousands of workflows. Many of these agents operate with some autonomy—routing tasks, calling tools, and coordinating with other agents—without a human in the loop on any single decision.
These agents are not “superintelligent.” They are simply always on. And almost none of them run on infrastructure the enterprise controls.
The MANGOS pattern
Beneath each MANGOS firm is an AI‑native stack with the same basic architecture:
- A frontier‑class model family.
- A coding agent or development environment.
- An orchestration harness or runtime for agents.
- A primary agent surface where humans interact with those systems.
They are not merely selling models or APIs. They are assembling the rails on which your agents will live.
Anthropic centers its stack on the Claude model family (Opus, Sonnet, Haiku, and successors), designed for long‑context reasoning, tool use, and “constitutional” alignment. Claude Code, its coding agent, integrates into developers’ workflows and uses a sophisticated tool‑orchestration layer to navigate files, refactor code, and run tasks. Technical disclosures and a recent source‑code exposure showed that Claude Code’s orchestrator - a TypeScript “harness” - instructs Claude on how to call tools, manage state, and enforce safety policies, and that Claude can generate dynamic execution harnesses to coordinate multiple agents on complex tasks. The real product is Claude‑inside‑Anthropic’s harness, not raw weights.
OpenAI is turning ChatGPT into an agentic super app. In 2026 it reorganized product leadership so that president Greg Brockman oversees all product strategy, promoted former Codex lead Thibault Sottiaux to run core product and platform, and began merging ChatGPT and Codex into “one unified agentic experience” ahead of a possible trillion‑dollar IPO. At VivaTech in Paris, Sottiaux described the ambition as a “personal AGI”: a single, highly capable agent that remembers what users care about and can handle everything from quick questions to planning complex trips, booking travel, and spinning up custom tutoring apps for their children - all in one interface. Codex, he said, has evolved into a generalized engine beyond code and will power this super app.
To support that pivot, OpenAI has shut down side projects, including its Sora video product, and pulled back from expensive infrastructure commitments such as the proposed Stargate data centers in the U.K. and Norway. The company wants to persuade users and IPO investors that it is more than a chatbot vendor; it is an agentic platform with a unified consumer, enterprise, and developer surface.
Nvidia is the inference fabric. Its AI Enterprise stack includes drivers, CUDA libraries, orchestration tools, and reference architectures for running AI in production. NIM—NVIDIA Inference Microservices - wraps leading models into containerized microservices with consistent APIs and GPU‑optimized performance, so they can be deployed across clouds, on‑premises data centers, and edge environments. For many enterprises, this means that regardless of which model they choose - Anthropic, OpenAI, Meta, proprietary - the actual inference often runs on Nvidia infrastructure. Nvidia’s “app” is not a chat interface; it is standardized, metered intelligence at scale.
Meta has shifted from an open‑weights Llama strategy toward a vertically integrated assistant. Its Muse Spark model, launched in April 2026 under the Meta Superintelligence Labs, is a proprietary multimodal system that now powers Meta AI across Facebook, Instagram, WhatsApp, and hardware such as smart glasses. Llama continues as an open‑weight line, but Muse Spark is the flagship behind Meta’s first‑party assistant. That means the alignment choices, refusal behavior, and data practices baked into an assistant used by billions are now governed by Meta’s internal processes, not open‑source debate.
Google has built an agent‑native cloud around Gemini. At I/O 2026 it announced Gemini 3.5 models, the Antigravity agent runtime, and Gemini Spark - a 24/7 personal agent that runs on dedicated Google Cloud virtual machines and stays active even when a user’s devices are offline. Antigravity 2.0 provides a developer‑facing desktop app and CLI to provision remote Linux sandboxes where agents can plan, execute code, manage files, and browse in isolation. The Gemini Enterprise Agent Platform ties these pieces together with managed agents, observability, and governance for enterprise deployments. Google is not merely attaching a model to a generic cloud; it is designing the cloud itself around agent workloads.
SpaceX/xAI is assembling a stack from orbit to IDE. SpaceX folded xAI into its AI division, giving it the Grok model family, and is buying Anysphere, the company behind Cursor, in a roughly sixty‑billion‑dollar all‑stock deal to gain a popular coding agent environment and developer user base. Beneath this sits Starlink, an AI‑heavy satellite network that already uses machine learning for trajectory planning, autonomous landing, and constellation management. At the same time, xAI has seen all eleven of its original co‑founders depart and has faced criticism over rushed releases and safety lapses. For enterprises whose developers begin to rely on Cursor backed by Grok, that governance history becomes part of their software supply chain.
The same pattern is emerging outside MANGOS. Microsoft is working on its own one‑stop agentic app that will connect GitHub Copilot, Copilot chat, its Cowork productivity tools, and a new agentic workflow engine into a single experience, while Elon Musk has long expressed ambitions to turn X into a WeChat‑style app for communication, media, and commerce.
The common thread is simple: the leading AI firms do not just want to sell models or APIs. They want to own the agent surface where users live and work, and the harness beneath it that decides how those agents behave.
Super apps and the enterprise
The push toward super apps is not just about consumer experience; it is about the enterprise fight.
In the U.S. business market, OpenAI no longer clearly dominates. A May 2026 AI Index tracking actual corporate spend across thousands of companies showed Anthropic at roughly 34.4% of U.S. business adoption versus OpenAI at 32.3%, the first time Anthropic pulled ahead. Analysts attributed the shift largely to Claude Code’s traction as an autonomous coding tool. ChatGPT, launched in 2022 as a conversational chatbot, has been perceived as more consumer‑focused - a perception OpenAI now wants to change.
OpenAI’s bet, articulated by Sottiaux, is that there is no fundamental divide between consumer and enterprise use cases for a unified agent. The same system, they argue, can serve both if it can “connect the right context, connect the right tools, and then take actions in a way that minimizes risk in order to deliver value.” The question is whether large enterprise buyers—historically wary of horizontal platforms that promise to do everything—will accept a single, personal‑AGI‑style surface for both work and life, or prefer vertical, domain‑specific agents embedded directly into existing systems.
Either way, the underlying physics is the same: more agents, more inference, more dependence on the same small group of infrastructure providers.
Inheriting a constitution you did not write
This brings the argument back to governance.
In constitutional economics, a constitution is the set of basic rules that constrain and channel behavior. In the AI economy, the “constitutional layer” is the combination of model behavior, safety policies, tools, and interfaces that determine what agents can and cannot do by default. When an enterprise builds agents on top of MANGOS stacks, it is effectively accepting those defaults as part of its own operating environment.
That has practical consequences:
- A safety policy change upstream can narrow what your customer agents are allowed to say overnight.
- A model deprecation can break a workflow that underpins a product or process.
- An export restriction on certain chips can limit your capacity to run agents at the scale your operations assume.
None of these decisions are made in your boardroom. All of them can affect your customers, your regulators, and your financials.
The more agents you run, the more these upstream choices matter, because each one touches not a single application but dozens of embedded workflows. This is strategic subordination by architecture. Traditional vendor management frameworks—focused on contracts, SLAs, and periodic assessments—do not fully capture it.
Until this year, three of the six MANGOS firms - Anthropic, OpenAI, and SpaceX - were private, with limited disclosure about their governance, safety processes, or board oversight. Their move toward public markets will increase transparency, but it does not guarantee robust governance. Governance quality varies widely, and the most capable model is not always the best‑governed one.
Your inherited constitution is only as sound as the governance of the company you inherited it from.
Continuous governance as counter‑architecture
Static governance cannot manage a dependency that mutates in real time. Annual reviews cannot govern systems that act in milliseconds. Committees designed for human chains of command do not map cleanly onto autonomous agents.
What the agentic enterprise requires is continuous governance: an always‑on set of controls and telemetry that inserts your company’s own constitution between its agents and the external infrastructure they run on.
Concretely, that means:
- Runtime constraints – Policies expressed as executable guardrails: whitelisted tools and actions, autonomy thresholds tied to risk appetite, mandatory human approval for high‑impact operations, and kill switches for misbehaving agents.
- Behavioral telemetry – Fine‑grained logs and dashboards showing what agents actually did in production: which systems they touched, what they changed, which exceptions they generated, how often humans intervened, and where.
- Decision traceability – The ability to reconstruct the chain behind a decision: which agents and models were involved, which data they saw, which humans approved, and how that maps to existing accountability structures.
This is counter‑architecture. It is where a board encodes its values, fiduciary obligations, and risk tolerance as constraints and incentives on agent behavior, even when the underlying models and runtimes are controlled elsewhere.
The regulatory clock
The timing is not theoretical. On August 2, 2026, the EU AI Act’s penalty framework, transparency requirements, and general‑purpose model provisions take effect, with sanctions reaching up to thirty‑five million euros or seven percent of global turnover. The detailed high‑risk system obligations follow on a longer timeline (currently extending toward late 2027 after provisional adjustments), but enforcement begins this summer.
Regulators will not accept “we relied on our provider” as a defense. Accountability remains with the enterprise that deploys AI systems in its products and processes, not with the model vendor.
What boards and executives can do now
For the board:
- Demand a dependency map. Ask management for a clear view of which consequential workflows - those that touch revenue, customers, or capital -depend on which AI providers and stacks. If it does not exist, that absence is itself a finding.
- Assign ownership explicitly. Place AI vendor concentration and systemic dependence on the risk committee’s standing agenda, and agent controls and assurance within the audit committee’s remit. Make this explicit in charters, not just informal practice.
- Look beyond cost and capability. Request management’s assessment of each major provider’s documented safety posture, governance transparency, and incident history - not just pricing and performance.
- Assess director readiness. If no committee has a director who can interrogate an AI risk dashboard or an agent’s decision chain, treat that as a composition gap to be addressed.
For the executive team:
- Inventory the agents. Catalogue agents already in production, the providers they rely on, and the types of decisions they touch. Use this as the foundation for both regulatory readiness and internal governance.
- Instrument before expansion. Introduce runtime guardrails, telemetry, and traceability first in your highest‑exposure workflows. Continuous governance is built incrementally; it cannot be bolted on after an incident.
- Diversify or consciously concentrate. Where a single provider underpins a revenue‑critical path, either develop a credible alternative or document for the board why the concentration is accepted and how the associated risk is mitigated.
- Map to the EU AI Act and other regimes. Identify systems likely to fall into regulated categories and ensure you can evidence both technical controls and governance processes before supervisory authorities ask.
So what's next?
If the board asks management a single question this quarter, it could be this:
What share of our consequential decisions - the ones that touch revenue, customers, or capital - is now executed by agents running on a single AI provider this board has never evaluated?
Not the cloud bill. That share.
If management cannot produce a credible answer quickly, the board has just measured the size of its governance gap.
Six companies are writing much of the constitution your enterprise runs on, and three of them have just asked public markets to fund the next chapter. You do not get to amend their constitution. You do get to write your own - and to govern to it continuously - because in the agentic enterprise that capability is not a compliance cost. It is the difference between running your business and renting it.
