Summer is here, unofficially, now that we have reached June. School is letting out, people are starting to take time off, and yet the markets are preparing for one of the more interesting IPO windows we have seen in years.

Anthropic confidentially submitted its draft S-1. SpaceX has filed its public S-1 prospectus, and the filing makes clear the company is no longer only a rocket and satellite story. OpenAI is widely reported to be preparing its own confidential filing, with CNBC reporting that the company has been working with banks on the process.

The analysis will be everywhere. Valuation, revenue, losses, capex, GPUs, model quality, competitive positioning. All of it matters.

But my goal this week is not to write another valuation take.

I want to talk about what these companies represent directionally, and what that means for FinOps now and over the next few years.

The IPOs are not the destination for AI. They are the foundation being priced.

Compute, models, infrastructure, data centers, distribution, and the capital required to keep scaling all of it. That is what public markets are preparing to value. The numbers are enormous because the foundation is enormous.

But a foundation is not the building.

The more important question for FinOps is what gets built on top of it. That is where the next layer of work happens. That is where agents operate. That is where requests will get routed. That is where model selection will happen. That is where cost either will be controlled before it happens or explained after it is too late.

That is where I think FinOps is going.

The headlines will make these companies sound different.

SpaceX will be treated as the rocket and satellite company. OpenAI will be treated as the platform and application company. Anthropic will be treated as the enterprise AI and frontier model company.

But directionally, they are part of the same story. Compute and models are becoming the foundation layer for how work gets done.

SpaceX is the easiest company to misread because the public story has always been rockets, launch, and Starlink. But the S-1 tells a broader story. SpaceX is now presenting itself as a company with three major pillars: space, connectivity, and AI. The Space segment remains the original foundation, generating $4.1 billion of revenue in 2025 while operating at a loss as the company continues to invest heavily in Starship and launch infrastructure. Connectivity, driven primarily by Starlink, is the current profit engine, generating $11.4 billion of revenue and $4.4 billion of operating income. AI is the newer and more capital-intensive bet, with AI compute, Grok, and X producing $3.2 billion of revenue while generating a $6.4 billion operating loss. In plain English: Space is the origin story, Starlink is the cash engine, and AI is the next major investment thesis.

OpenAI and Anthropic are cleaner versions of the same foundation-layer story. They are building the models, developer platforms, enterprise adoption, and ecosystem gravity that other companies will build on top of. Whether you view them as model companies, platform companies, infrastructure companies, or application companies depends on where you stand. But the market is going to value them because they sit close to the foundation.

That matters.

The public markets are getting ready to price the base layer of AI. But almost none of the value created by that layer will stay entirely inside those companies. The real value will show up in the products, workflows, agents, automations, and operating models built on top.

And that is where the FinOps question changes.

The old question was, “What did we spend?”

The new question is, “What should this unit of work cost before we run it?”

That question cannot be answered only from an invoice.

The loud version of the AI story is replacement. Agents take the jobs. Teams shrink. The org chart collapses. Work gets automated and people get removed.

I do not buy that as the dominant story.

I am sure some work will be automated away. That has always happened with major technology shifts. But the more interesting version is not humans replaced by agents. It is humans working with agents.

That may sound like a small distinction, but it changes the economics.

When a person uses software, the cost model is relatively familiar. You can look at salaries, SaaS licenses, cloud spend, support cost, and productivity. It is imperfect, but at least the categories are recognizable.

When a person works with an agent, the cost object starts to change.

A completed unit of work may now include human time, model calls, retrieval, memory, tool use, infrastructure, data movement, third-party APIs, orchestration, and retries. Some of that cost is fixed. Some of it is variable. Some of it is obvious. Some of it is hidden. Some of it happens because the work is valuable. Some of it happens because the agent made a bad decision, chose the wrong model, retrieved too much context, called too many tools, or retried a task that should have failed earlier.

That is not a normal software cost problem.

It is not just cloud cost. It is not just SaaS cost. It is not just AI cost. It is the cost of work being performed by humans and agents together.

That is why I think FinOps has to move closer to the work itself.

It is no longer enough to look at the bill after the fact and allocate spend to a team. That will still matter, but it will not be sufficient. If agents are part of how work gets done, then the economics of that work have to be measured and governed in the path where the work happens.

That path increasingly runs through gateways and proxies.

I have been experimenting with OpenClaw, the self-hosted, open-source agent that runs on my own hardware instead of someone else’s cloud.

The point of the experiment is not just the agent. It is the layer underneath it.

OpenClaw lets me select different models through a plugin called Manifest. Manifest sits between the agent and the model providers. Its routing documentation says it scores prompts across 23 dimensions, assigns them into tiers, and routes them to the cheapest model that can handle the job.

Simple request, cheaper model. Reasoning-heavy request, more capable model. Local model when appropriate. Frontier model when needed.

That is the interesting part.

The important thing is not whether I save exactly 30%, 50%, or 70%. I do not put too much weight on any generic savings claim without more real-world testing. The important thing is the mechanism.

The request is being evaluated before it becomes spend.

That is the control point.

Manifest is one example. OpenRouter is another. OpenRouter recently announced a $113 million Series B led by CapitalG and said its volume has grown to 25 trillion tokens per week across more than 400 models. LiteLLM, a self-hosted proxy, is another example, with spend tracking across keys, users, and teams, along with budgets and rate limits that can be applied at the proxy layer.

Some are commercial. Some are open source. Some are developer-first. Some are infrastructure-first. Some are built for routing. Some are built for governance. Some are built for observability. Some are built for all of it.

The common thread is where they sit.

They sit between the application or agent and the model provider. That means they see the request before the cost happens. They can route it. Price it. Log it. Block it. Budget it. Attribute it. Compare it. Evaluate it. Change it.

That is very different from a cloud invoice.

A cloud invoice tells you what happened. A gateway can influence what happens next.

For FinOps, that difference matters.

If a gateway can route a request from a frontier model to a mid-tier model with no meaningful change in user outcome, it may change the unit cost by an order of magnitude. If it can block a request that exceeds a policy, it prevents spend before it lands. If it can attribute model usage to a user, team, product, feature, customer, or workflow, it starts to create the unit economics that AI products need.

That is why this layer matters more than most of the IPO commentary.

The providers are the foundation. The gateways and proxies are where many of the actual cost decisions will happen.

So where does this leave FinOps?

It moves the discipline closer to the work.

Traditional FinOps was built around visibility, allocation, forecasting, optimization, and accountability. Those practices still matter. I do not think they disappear. Cloud bills still need to be allocated. Infrastructure still needs to be rightsized. Commitments still need to be managed. Waste still needs to be removed.

But AI changes the speed and location of the cost decision.

In cloud, a bad architecture decision might show up over days, weeks, or months. In AI, a bad routing decision can happen thousands or millions of times. The unit cost problem gets repeated at machine speed.

That means AI cost governance cannot rely only on quarterly reviews, monthly reports, dashboards, and after-the-fact recommendations.

Some of the work has to become real time.

A request comes in. The system evaluates what kind of work it is. It chooses a model. It checks budget. It applies policy. It captures metadata. It logs the cost. It ties the request back to a user, team, product, customer, or business process. It learns whether the output was good enough. It improves the routing decision next time.

That is FinOps moving from reporting to governance.

This is also where the human-plus-agent model becomes important. If agents are doing more work, then FinOps cannot only govern the infrastructure underneath them. It has to govern the economics of the work they perform.

That does not mean humans disappear from the process. The routine work can become more autonomous: anomaly detection, budget enforcement, recommendation generation, idle resource detection, commitment checks, model routing, and policy validation. But judgment still matters. Humans still decide acceptable tradeoffs. Humans still define business value. Humans still decide when a higher cost is worth it.

The future is not the agent replacing the FinOps practitioner.

It is the FinOps practitioner governing a system where agents create, consume, and optimize spend continuously.

That requires better instrumentation than most companies have today.

This is where the vendor community has to be honest with itself.

Most FinOps tooling was built for a world of cloud invoices. You ingest the bill, normalize the data, allocate the spend, show the dashboards, generate recommendations, and help teams take action.

That is still useful. But it is downstream.

AI cost governance is moving upstream.

If the cost decision happens in the request path, then a platform built primarily around invoice ingestion may not be standing in the right place. Adding a token dashboard may help with visibility, but visibility is not the same as control.

The harder question is whether the product can influence cost before the request is executed.

Can it sit in the flow of work? Can it help route requests? Can it enforce budgets before spend happens? Can it connect a model call to a workflow, customer, product, or business outcome? Can it expose one interface for the agents doing the work and another for the humans governing it?

That is a different product problem.

I am not saying current FinOps vendors are irrelevant. That would be lazy, and I do not believe it. Many have customer trust, billing expertise, allocation logic, commitment management, executive reporting, and integration depth that new AI-native tools do not have.

But I am saying the center of gravity is shifting.

The most interesting signal right now is not only coming from incumbents adding an AI tab. It is coming from the layers forming around AI work itself: routing, observability, agent orchestration, policy enforcement, evaluation, budget controls, and probably categories we have not named yet. Some tools will decide which model handles the request. Some will monitor agent behavior. Some will measure quality, latency, retries, tool calls, and cost. Some will help teams determine whether a completed unit of work was worth what it consumed. The common thread is proximity. These tools are closer to the work, closer to the request, and closer to the economic decision. That does not mean they become the entire FinOps platform, but it does mean they may become the control points that traditional FinOps platforms have to integrate with, partner with, or eventually compete against.

The IPOs are dominating the headlines because they price the foundation. That is fair. The foundation matters.

But the next several years of FinOps will be shaped above that foundation, in the layers where AI work is requested, routed, observed, governed, measured, and improved. Some of that will happen in gateways and proxies. Some of it will happen in observability platforms, agent orchestration systems, evaluation tools, product workflows, and categories we have not named yet. The common thread is that FinOps moves closer to where work happens and closer to the moment cost is created.

The old question was, “What did we spend?”

The new question is, “What should this unit of work cost before we run it?”

That is a very different operating model.

It forces engineering leaders to think about model routing as architecture. It forces product leaders to think about AI features in terms of unit economics. It forces finance leaders to think beyond provider bills. It forces FinOps practitioners to move closer to the systems where work actually happens.

And it forces vendors to answer a harder question.

Are you helping companies explain the bill after the economics have already been decided, or are you helping them govern the decision before it becomes spend?

That is where I think AI cost governance is going.

Not just better reports. Not just more dashboards. Not just another tab labeled AI.

Closer to the work. Closer to the request. Closer to the moment cost is created.

If you would like your company included in the Spotlight, contact the CloudXray AI Team

Company: CloudXray AI

Category: Managed Services & Consulting

What They Do: FinOps Consulting & Advisory Services; Owners & maintainers of the single largest FinOps company directory (finops.cloudxray.ai)

Why It Matters: Companies still need guidance on implementing FinOps and understanding the landscape of companies that exist