The Assembly and the Assembler

On May 4, Anthropic announced a standalone enterprise services firm with Blackstone, Hellman & Friedman, and Goldman Sachs as founding partners, backed by roughly $1.5 billion in committed capital. Hours earlier, Bloomberg reported OpenAI raising $4 billion for a venture with the same shape. The model makers, and some of the most disciplined capital in the world, are building consultancies.

The number underneath the move explains it. As Fortune put it in its coverage of the launch, for every dollar companies spend on software, they spend six on services. Six to one. That ratio built the modern consulting industry, and it is now arriving in AI on schedule.

Read it plainly. The constraint in enterprise AI is no longer intelligence. Models are capable and getting cheaper. The constraint is deployment: making models work inside a real business, on budget, inside policy, with evidence a regulator can read. The market just priced that layer at six times the software underneath it.

The model makers are not wrong about where the money is. They are choosing how to collect it.

Two ways to sell the same layer

There are exactly two ways to sell the make-it-work layer.

The first is hours. Staff the layer with people. Embed engineers, run the workshops, tune the deployment, bill the time. This is a real business. It is a multitrillion-dollar business. It also scales with headcount, walks out the door at the end of the engagement, and has to be bought again for the next project.

The second is a machine.

Consultants sell the assembly. APERION sells the assembler.

What the assembler is

The assembler is not a metaphor stretched past its limit. It is a specific piece of infrastructure with a specific address: the line between a company and every AI provider it uses.

Smartflow sits on that line as a transparent proxy inside the company's own perimeter. Every request passes through it, so the make-it-work functions stop being projects and become properties of the wire. Every call is metered and priced in real time, attributed to a verified person and rolled up the organization. Provider keys come off developer cards and into central custody. Each task checks a local cache first, at zero token cost, then routes to the lowest-cost model that completes it, escalating only on complexity or risk. Padding the model never needed is stripped before it is billed. Policy is enforced as the request happens, and the record it leaves behind is the one an examiner asks for.

None of that requires changing a line of application code. That is the point of a machine. It does the work the engagement would have done, and it is still there in the morning.

The pattern is older than AI

Every layer of computing infrastructure has taken this walk. Network security began as hardening engagements and became the firewall. Traffic management began as failover runbooks and became the load balancer. Identity began as directory consulting and became the identity provider. In each case the services wave was real, profitable, and early. In each case the layer ended up as a product, because enterprises do not want to rent the same fix twice.

Enterprise AI's make-it-work layer is at the start of that walk. The services ventures announced this spring will do well, and they will be busy for years. The six-to-one ratio guarantees it. It also guarantees the question every CFO eventually asks about a recurring engagement: what stays installed when the engagement ends?

An assembly leaves furniture and an invoice. An assembler stays on the line, metering every call, enforcing every policy, and reporting the savings from the customer's own traffic.

For every dollar of AI software, six now go to making it work. The durable share of those six belongs to whoever turns the work into a machine, and that is the question a buyer should put to everyone now selling this layer, including the model makers themselves: are you selling me the assembly, or the assembler?

Craig Alberino is the founder and CEO of APERION. APERION builds Smartflow, the on-premises runtime governance control plane for enterprises deploying AI, with metering, key custody, routing, and inline policy at every prompt, response, and tool call. Related coverage: Anthropic's services venture announcement (May 4, 2026), Fortune's reporting on the one-to-six software-to-services ratio, and Bloomberg's report on OpenAI's counterpart venture via TechCrunch.