Building for trillions
of digital workers.

Samuel Ekpe came to this problem through earlier startups, consulting, and running a talent marketplace. Matching someone to a job was only the beginning. Onboarding, coordination, exceptions, and delivery still needed people. Software organized the work; a person carried responsibility for getting it finished.

Our central bet is that this responsibility can become a reusable capability. A Worker should carry a mission across the software a business already uses, without losing its purpose at each handoff. The screen is the deployment surface. The opportunity is the work that crosses it, especially the last mile where an exception can make an otherwise successful workflow unusable.

History offers two useful precedents. Research on U.S. manufacturing from 1890 to 1940 finds that electrification brought productivity gains alongside investment in equipment and changes in organization. Our inference is that improving a capability and redesigning how work uses it belong together. [1]

Cloud computing changed how capability could be accessed. Amazon's 2006 EC2 launch let developers obtain resizable computing capacity through a web service and pay for what they used. Teams could provision capacity as a service, reducing the need to assemble the underlying infrastructure themselves. [2]

Our thesis brings these two ideas to digital work: redesign execution around capable models, then make that execution available on demand. The analogy has a limit. Provisioning a server supplies capacity; delegating a mission requires a correct result under real constraints. Permission, judgment, recovery, and evidence are central to the infrastructure we must build. History suggests a pattern, not a guarantee that we will reproduce its outcome.

Stanford's 2026 AI Index reports an OSWorld result of 66.3% by late 2025, with roughly one-third of benchmark tasks still unfinished. That is a snapshot of benchmark performance, not a guarantee of production reliability. Our reading is that useful execution is becoming possible while recovery from failure remains essential. [3]

The economics are moving too. Stanford's 2025 report recorded a more than 280-fold fall in inference cost at GPT-3.5-level performance between November 2022 and October 2024. Cheaper reasoning creates room to spend on checking and recovery. It does not, by itself, make a complete Worker cheap or reliable. [4]

Together, these changes open a specific research and engineering opportunity: turn improving models into dependable work, and measure the full cost of getting the result right.

The infrastructure opportunity is to let people and companies delegate completed work without rebuilding execution for every assignment. Customers need an accepted result, whether they buy it directly or build it into their own product. The value of the tools underneath depends on how reliably and economically they deliver that result.

Work carries responsibility. Errors, missed deadlines, and unauthorized actions have consequences. The useful economic measure is cost per accepted outcome, including execution, verification, retries, external resources, and human intervention. A cheap run that needs expensive rescue is expensive work.

Our ambition is a Worker more capable on defined work than a remote worker, at one-tenth the total cost of comparable skilled labor in the world's lowest-cost regions. That target must be tested at equivalent quality, with the cost of failed attempts included.

Consider a supplier renewal. A Worker may need to read a contract, reconcile an invoice, check a portal, resolve a discrepancy, and request approval before a commitment. The job crosses systems, but its context, permissions, and definition of completion must remain intact. That is the continuity we are engineering.

Different jobs repeatedly need the same foundations: identity, authorized access, persistent state, coordination, recovery, and evidence. Our infrastructure bet is that these foundations can be reused across Workers and roles. A developer should be able to provision a Worker with a mission and boundaries, then inspect progress and results through a shared API.

Revi, a food-ordering startup, asked us for an API. Similar requests came from a consulting firm and YC startups. Those requests shaped our decision to build Worker infrastructure alongside a desktop product. They are early signals of demand for a shared foundation, not proof of a repeatable market.

We start with repeated operational work for founders and teams, with completion criteria a person can review. Momentum provides the direct Worker experience; the Workers API exposes the execution foundation. The path is to deepen reliable missions, extend continuity across environments, and coordinate teams. Each expansion must preserve quality while reducing the attention a person supplies.

Our intended advantage is cumulative operational knowledge: evaluations that catch real failures, reusable recovery behavior, and deployment practices that make the next mission easier to complete. This requires permissioned learning and isolation of customer information. The value should compound through better execution without depending on unrestricted access to customers' data.

That is a hypothesis we must earn. If each new workflow requires the same amount of bespoke engineering or human rescue, the infrastructure thesis weakens. The tests are accepted outcomes, deadline reliability, intervention time, deployment effort, and full delivery cost. Our own results must establish that these improve as the system takes on more work.

A person may direct roughly ten Workers for personal life, work, and business. An enterprise may coordinate swarms across many customers and processes. A product may provision Workers when a request arrives. Demand can therefore expand with the number of useful assignments, beyond the number of people buying a software seat.

Using the UN's 2024 population estimate, 8.2 billion people × roughly ten Workers gives about 82 billion Workers. This is an illustrative universal-adoption scenario, not a market forecast. Business swarms and product-directed work motivate our trillion-Worker ambition; they do not yet establish demand or simultaneous usage at that scale. [6]

If dependable digital work becomes programmable, a person can delegate a mission, an enterprise can provision capacity, and a founder can build an autonomous business on the same foundation. The long-term ambition is infrastructure for trillions of Workers, with human authority and evidence of completion built into how they operate.

One trusted Worker is the beginning. Infrastructure that powers the world's work is the company we are building.

The GrupaAI Pioneer AI Lab Journal shares our research questions, design convictions, and what we are learning as we build. These essays describe work in progress. They are not independent validation of our performance.