Anthropic’s Digital Labor Tax
The "Metalayer" could redefine AI economics. The better it works, the faster it strains already limited compute.
Anthropic has inserted a new layer into the AI stack with Claude Managed Agents. We term this managed runtime for autonomous agents the “Metalayer.” It is priced by the session-hour, bundles compute, intelligence, and execution into a single taxable unit. By capturing the Agent Unit, Anthropic positions itself as the essential utility for digital labor: every synthetic colleague, regardless of who builds the last mile, pays the Anthropic “tax”. The Metalayer has the potential to decouple revenue from the linear compute-to-megawatt equation, create software-like stickiness, and accelerate the sorting of enterprise value from technical ability to strategic position. But the product designed to accelerate agentic AI adoption intensifies the very compute bottleneck on which it depends - and the ground beneath it is fragile.
We are entering an Agentic Era paradox.
Agentic AI is moving into production and accelerating faster than even the most aggressive forecasts from a year ago. At the same time, the system it depends on is hitting a hard constraint: there is not enough compute to sustain that growth.
This is not a marginal risk. It is a single point of failure. Without sufficient compute, none of this works. Not the models, not the applications, not the emerging architecture of autonomous agents. The entire trajectory of the Agentic Era is bounded by the availability of a scarce, contested, and unevenly distributed resource.
Into this constraint, Anthropic has just poured accelerant.
On April 8, Anthropic launched Claude Managed Agents. This is a new way to deploy agentic AI in production that dramatically lowers the friction of building, running, and scaling autonomous systems. On the surface, it looks like just another product launch. Structurally, it represents something far more significant that I call the “Metalayer”: a new layer inserted into the AI stack in the form of a managed runtime for agents that could significantly reduce the complexity of orchestration, memory, tool use, and coordination.
The Metalayer also introduces a major economic twist by changing the unit of value in AI. Instead of pricing intelligence through tokens, it bundles compute, reasoning, and execution into what I define as an “Agent Unit.” This is a synthetic colleague, or digital worker, billed by the hour. Every agent becomes a metered entity. Every synthetic colleague becomes a revenue stream. And because Anthropic controls both the intelligence and the runtime, it positions itself as the economic intermediary for digital labor, regardless of who builds the final application. With Managed Agents, Anthropic is attempting to capture value at the level of digital labor (Agent Units), not just intelligence (tokens). By owning both the model and the runtime, Anthropic would position itself as the tax collector of digital labor. This is where the paradox emerges.
The Metalayer is designed to accelerate the adoption of agentic AI by removing friction. Every unit of adoption creates and consumes more compute. The more successful it is, the faster demand scales. And the faster demand scales, the harder it runs into the constraint that defines the system. Anthropic is not just building on top of this constraint. It is amplifying it.
That makes Anthropic’s position fundamentally fragile. The company has already faced whispers about whether it has secured sufficient compute to deliver the services at the rate it appears to be scaling. Those whispers burst into the public this week via a leaked memo from an executive at rival OpenAI, who insisted that Anthropic did not have the compute it needed. Whether or not that’s true, Anthropic announced a series of major compute deals and appears to be looking for more. Certainly, the company understands that the layer it is creating and the economics it is trying to capture depend entirely on a resource that, if constrained, can halt the system altogether.
Anthropic has a path through this, but it is narrow. To understand why this agentic paradox matters and where it leads, we need to unpack what Anthropic has actually built.
A new layer in the stack
Consider the AI stack as it stood until last week:
At the bottom, energy. Today’s clear bottleneck.
Then, hardware such as GPUs, TPUs, custom ASICs, Nvidia’s kingdom.
Above it, cloud infrastructure. This is the hyperscalers’ domain.
Above that, the model layer, frontier labs serving foundation models via API.
Above that, applications, whether delivered by software vendors or built by the model user.
What was missing was an operating environment purpose-built for agents. Not a model. Not a cloud VM. Not an application. Something in between.
Enter Claude Managed Agents.
Using a suite of APIs, Claude Manage Agents removes the need to build agent infrastructure from scratch and allows users to create and deploy cloud-hosted agents with all the necessary production infrastructure.
You define tasks and tools, and Claude handles the running, coordinating, and iterating agents in a managed runtime that replaces the need to build backend infrastructure.
Alongside this, the advisor tool introduces an architectural innovation by pairing Opus, a strategic brain, with Sonnet or Haiku as the execution engine. The executor drives the task end-to-end and escalates to Opus only when it hits decisions beyond its capability.
Anthropic’s benchmarks show near-Opus-level intelligence at costs below those of Sonnet alone. That’s a 2.7% point improvement on SWE-bench Multilingual with 11.9% lower cost per task. Haiku with an Opus advisor more than doubles its solo BrowseComp score at 85% less cost. The pattern works not by making models cheaper, but by making the allocation of intelligence dynamic. It knows when to think hard and when to act fast.
What Anthropic has done is vertically integrate the model and the runtime into a single managed service. This is structurally different from anything that existed before.
As I continue to refine the principles of Orchestration Economics, new concepts require precise definitions. Here’s how I define “Metalayer”: the managed runtime for persistent, long-running, multi-agent workloads that absorbs the overhead of dense infrastructure. That includes sandboxing, state persistence, tool orchestration, multi-agent coordination, retries, and error recovery. It can separate a chatbot demo from a production agent that runs for hours, delegates to sub-agents, and recovers from failures without human intervention.
The Metalayer now exists. Anthropic built it, creating a new stratum in the AI value chain that captures value that previously leaked into the enterprise’s infrastructure budget as engineering overhead.
OpenAI has its Assistants API and Responses API, but not a managed runtime. Google has Vertex AI agents, but the model and runtime are not as tightly integrated. AWS Bedrock Agents provides infrastructure, but not the model intelligence.
Anthropic is the only company that owns both the brain and the operating environment purpose-built for that brain. By owning the fifth layer and dominating the fourth year, it is the only company to occupy both. It is simultaneously captain of the Metalayer and the leading tenant of the model layer. It orchestrates other agents while remaining the primary intelligence provider.
This dual position is the moat.
The agent unit, the tax, and the last mile
Now that we have a clear picture of Metalayer's structure, let’s turn to understanding the new unit of economic value it creates.
Before last week, the AI economy was priced in tokens. These are the atomic units of model consumption. Tokens represent intelligence. They are a tax on the model provider, which in turn pays a tax to the GPU provider. The value chain was linear: energy to hardware to the cloud to the model to the application.
However, just as notable as the introduction of the new layer was Anthropic’s pricing for the Metalayer. Users pay per token usage PLUS $0.08 per active session-hour. In other words, you’re not just paying for intelligence.
The Metalayer introduces a bundled intermediate unit I will call the Agent Unit. It combines three components:
the raw compute (GPU)
the intelligence (tokens)
the execution (runtime - the session-hour)
The Agent Unit is the cost of running an autonomous agent. It is not, by itself, what it costs to produce an autonomous business outcome. For that, you need the last mile.
The last mile takes two forms. Think about when software sits on top of the Agent Unit, such as Salesforce embedding Claude agents into Agentforce or ServiceNow orchestrating ITSM workflows. The last mile is application value: the UI (or the GUI), the domain logic, the compliance rules, the customer-facing experience. The software vendor adds context and workflow control; the Agent Unit provides the brain and the nervous system.
When there is no software intermediary, such as an enterprise building custom internal agents directly on the runtime, then the last mile is the infrastructure cost of connecting the enterprise’s own context (data, systems, permissions) to Anthropic’s intelligence. In both cases, the Agent Unit sits underneath. And in both cases, there is a new tax to Anthropic.
This tax is consequential because of what sits at the end of the value chain: synthetic colleagues.
As I have argued across my Agentic Era series, we are observing a transition from applications that enhance human productivity to applications that replace entire organizational functions. Traditional AI applications compete for software budgets that typically account for 2–5% of enterprise spending. Synthetic colleagues compete for labor budgets: the salaries, benefits, and operational costs of entire teams.
When an agent can perform the functions of a marketing department or a customer service division, the addressable market extends from IT spend toward the broader costs of organizational cognitive work and the estimated $60 trillion global labor market.
Every synthetic colleague, regardless of who builds the last mile, runs on Agent Units when using Claude. Every Agent Unit pays the Anthropic tax. The Metalayer makes Anthropic the essential utility for digital labor, regardless of who builds the final interface.
The session-hour is not a software metric. It is a headcount metric. It is the cost of an autonomous worker, billed hourly.
This architecture allows Anthropic to play in three positions simultaneously. If it moves one layer up by launching verticals or expanding Cowork into domain-specific workflows, then it captures the entire value chain from intelligence to outcome, becoming a software company that owns its own silicon-to-solution stack.
When a software company builds on top of the Metalayer instead, Anthropic becomes the engine it cannot easily swap out. Every agent running inside Salesforce is a session-hour billed to Anthropic, every advisor-tool escalation is token revenue at Opus rates, and the Agent Unit remains Anthropic’s high-margin product while the software company provides the UI and the domain expertise.
For the most sophisticated companies, there is no last mile at all. They build custom internal agents directly on the runtime, a hedge fund’s proprietary research agents or a pharmaceutical company’s clinical trial workflows, and Anthropic sells digital headcount directly to the C-suite, bypassing traditional software vendors entirely.
In all three scenarios, the tax flows to Anthropic.
The revenue transformation
The Metalayer has the potential (potential, as there is no execution proof yet) to transform Anthropic’s position in two distinct ways.
The first concerns revenue quality. Before Managed Agents, Anthropic’s revenue was predominantly token-based: consumption of model intelligence, volatile, difficult to predict, and, as OpenAI’s Sarah Friar equation demonstrated, linearly correlated with compute. Friar framed OpenAI’s growth as a near-linear function of megawatts: 0.2 GW yielded $2B, 0.6 GW yielded $6B, 1.9 GW yielded $20B. I deconstructed that equation earlier this year and showed its fragility. But the deeper problem with the Friar Equation is strategic: if revenue scales linearly with compute, then revenue is structurally capped by compute availability. You cannot grow faster than your infrastructure.
Session-hour billing has the potential to break this linearity. The Metalayer adds a runtime fee that is not purely token-denominated. It is closer to software-like revenue: recurring, predictable, correlated with enterprise headcount equivalents rather than with prompt length.
More importantly, the advisor pattern shows that comparable intelligence can be delivered at lower token intensity with Haiku executing and Opus advising. That means the revenue per session-hour can be maintained while the compute cost per session-hour declines.
If this pattern scales, Anthropic begins to decouple revenue from the megawatt equation. That would be a structural break from the industrial economics that define the rest of the AI industry.
Session-hour billing also creates stickiness that token-level access does not. An enterprise that has deployed dozens of persistent agents with accumulated state, configured tools, and established multi-agent coordination patterns faces meaningful switching costs. This is not a model you can swap by changing an API endpoint. It is an operating environment.
The closest precedent for the session-hour is Databricks’ DBU. This is the consumption unit that turned a data platform into a trajectory toward a $300 billion valuation. Both are metered and usage-based, and they scale with enterprise workload volume. Both create stickiness through accumulated platform state.
But both also face the same open question I raised earlier this year: will consumption-based growth hold up as enterprises realize the extent of the spend at stake? The DBU’s answer is still being written. The session hour’s chapter has not yet begun.
The second way the Metalayer could transform Anthropic’s economics concerns the quality of growth.
The tax structure creates the conditions for network effects. Every software vendor that builds on the Metalayer expands the installed base of Agent Units. Every enterprise that deploys custom agents creates session-hours. Every synthetic colleague running on the runtime generates demand that funds a better runtime. All of that removes more friction, which drives more adoption.
The flywheel is structurally identical to what AWS did to cloud adoption in the 2010s. EC2’s per-hour cost was real, but the removal of friction was so valuable that it accelerated the very adoption it was taxing. More applications deployed, more compute consumed, more revenue, better primitives, more friction removed. Fifteen years of compounding.
Whether Anthropic can replicate that trajectory is an open question. EC2 succeeded because AWS owned its own infrastructure. Anthropic’s Metalayer runs on rented compute.
The flywheel is the same. The ownership structure is not.
The sorting
For the broader software landscape, the Metalayer accelerates a sorting that was already underway. But it also changes its basis from technical ability to strategic position.
Before Managed Agents, building agentic capability required months of backend engineering. That engineering tax was itself a moat: only the most technically sophisticated vendors could absorb it.
The Metalayer collapses that tax to a configuration exercise. Any ISV with a Claude API key can now deploy persistent, long-running agents. The barrier is that you can no longer build the infrastructure. The barrier is, do you have a strategic position worth defending?
This is already how I view a moat within my ARAF framework. The factors that determine survival are not technical; they are strategic: proximity to user intent, depth of proprietary context, and control of the workflow.
The Metalayer does not change these criteria. It accelerates their application by removing the engineering barrier that previously gave technically capable but strategically weak companies time to maneuver. It does not kill thousands of startups or incumbents. It kills the wrapper whose value proposition was technical and not strategic. These are the companies whose differentiation was “we built the agent infrastructure,” because Anthropic is now giving that infrastructure away at $0.08 per session-hour.
What survives is context depth, regulatory embeddedness, and proximity to the workflow that produces the business outcome.
The same sorting applies beyond software to the real world. The Metalayer enables enterprises to accelerate the adoption of agentic AI without building their own runtime. That means the gap between adopters and non-adopters widens faster. Companies that move now build operational intelligence that compounds session by session, exactly as the orchestration graph architecture predicted. Companies that wait will face the same architecture at a lower cost in six months, but without the accumulated understanding that early movers have already embedded in their agent configurations.
The flywheel on rented ground
Picture a flywheel mounted on rented ground. The flywheel is Anthropic’s Metalayer: the managed runtime, the session-hour meter, the adoption-acceleration loop.
Every revolution removes more friction from enterprise agentic AI, which drives more adoption, which generates more session-hours, which funds a better runtime, which removes more friction. The machine is elegant. It is working. And every revolution makes the ground beneath it more important and less secure simultaneously because the ground is rented compute, leased from hyperscalers who operate competing runtimes and have their own commercial reasons to prioritize their own workloads when capacity binds.
A virtual cloud’s control plane is sovereign. Its substrate is not.
This is the Single Point of Failure I identified in mid-2025. What has changed is not the nature of the vulnerability. It is exactly what it was a year ago. What has changed is the urgency and the scale of what the constraint now binds. When I first wrote the analysis, the constraint was on a model API business with a $4 billion run rate. Today, it is on a horizontal platform whose flywheel is designed to accelerate the adoption of agentic AI across the entire enterprise software economy.
The better the Metalayer works, the faster the demand grows. The faster the demand grows, the harder the substrate strains. Success and vulnerability are the same mechanism viewed from opposite sides.
While the situation is not necessarily fatal for Anthropic, it also leaves it with little maneuvering room. And signs of those strains began to show over the course of five days.
Anthropic has previously acknowledged in public that it walks a fine line between overspending and underspending on Capex for compute. While it hasn’t made any official statements regarding those overall plans, one can see how it’s rethinking existing resources.
On April 4, Anthropic announced subscription plans would no longer cover OpenClaw usage. The reason: capacity triage. The open-source agentic tool that went viral earlier this year remains wildly popular. Alas, perhaps a bit too popular. As Claude Code creator Boris Cherny put it: “Capacity is a resource we manage thoughtfully.”
Individual OpenClaw agents were consuming $1,000 to $5,000 in API-equivalent costs per day on $200 monthly subscriptions. Five-hour session limits during business hours followed. On April 6, Claude.ai partially failed.
Seen in this context, the advisor strategy announced on April 9 may have been positioned as a developer feature, yes, but it can also be read as a defensive response to the compute crunch: routing the bulk of execution to cheaper models, reserving Opus for the moments that demand it.
When you are capacity-constrained, the most valuable innovation is the one that delivers comparable intelligence at lower intensity.
Then came the substrate scramble via two big partnerships: Google–Broadcom (" our most significant compute commitment to date,” 3.5 gigawatts of TPU capacity from 2027), and CoreWeave (multi-year agreement for Nvidia capacity from a neocloud that does not run a competing runtime). Then Reuters reported that Anthropic is exploring custom silicon.
This is the cadence of a flywheel spinning faster than its ground can absorb.
Of course, the compute constraint is not unique to Anthropic. OpenAI CFO Sarah Friar told colleagues she did not believe the company would be ready for a 2026 IPO, citing risks from spending commitments.
But OpenAI’s structural exposure differs. Its relationship with Microsoft gives it preferential access to Azure capacity, a landlord that is also a $13 billion investor. The circularity is complex, but the alignment is real: Microsoft’s commercial interest in OpenAI’s success directly motivates compute allocation.
OpenAI senses a vulnerability for Anthropic here. In an internal memo that was leaked, Chief Revenue Officer Denise Dresser took a few shots at Anthropic: “Their strategic misstep to not acquire enough compute is showing up in the product. Customers feel it through throttling, weaker availability, and a less reliable experience. We saw the exponential compute curve earlier, acted on it faster, and now have a real structural advantage.”
Anthropic’s dual dependence on Google (TPUs for inference) and Amazon (Trainium for training) creates a more fragile geometry than OpenAI. Both landlords operate their own agent runtimes: Google’s Vertex AI agents, Amazon’s Bedrock Agents. And both have commercial reasons to prioritize their own workloads when capacity binds. Anthropic is the tenant whose success drives demand for infrastructure that the landlord could, at the margin, redirect.
The CoreWeave deal is the most revealing response. CoreWeave operates data centers with hundreds of thousands of Nvidia GPUs but does not build competing AI models or agent runtimes. It is the landlord with no reason to deprioritize the tenant. But CoreWeave is capital-constrained. It is funding expansion through aggressive debt issuance, with leverage that has raised analyst concerns. Custom silicon is years from production. The Broadcom SEC filing includes a clause noting that the new TPU capacity is contingent on “Anthropic’s continued commercial success.” The landlord put that clause there for a reason.
The company has built a product whose purpose is to accelerate the consumption of a resource it cannot scale fast enough.
Claude Code’s session autonomy has doubled from 25 to 45 minutes. Managed Agents create consumption patterns that drive hours-long sessions, multi-agent teams, and persistent state, all of which are categorically more compute-intensive than the API calls they replace. The advisor tool multiplies the number of model calls per task. Each product individually increases compute intensity.
Together, they represent a combinatorial explosion of inference demand.
Anthropic’s IPO, reportedly targeted for October 2026, will need to price this contradiction. With a $380 billion private valuation and a $30 billion run-rate in revenue, the market assigns roughly a 12.7× trailing revenue multiple. That premium reflects both the extraordinary trajectory and the orchestration authority the Metalayer is accumulating.
But the market will also need to price the risk that compute constraints cap the growth curve before the infrastructure catches up. This is the Durable Growth question: can the flywheel sustain its rotation, or does the rented ground give way under the centrifugal force?
The flywheel is real. The problem it solves is real. The ground it stands on is rented.