The PM With Loaded Dice

A friend asked me last week how I’d want to be a product manager at OpenAI or Anthropic right now. I told him I’d want that job more than almost any other PM seat.

Here’s why.

Every product manager in history has had to guess. I remember once upon a time when I was a product manager at Applied Biosystems. You survey customers. You run interviews. You watch competitors. You read tea leaves. Then you pitch the higher ups on which feature to ship next, and your bonus rides on a market you’re estimating.

Now picture a different job. You open a dashboard. The dashboard shows you what every customer of your company is trying to do with your product, ranked by volume and by how fast each use is growing. It shows you which categories doubled in the last three months. It shows you which jobs in the economy have already shifted a chunk of their work onto your platform. Your bonus is tied to picking the right category to build into.

Both companies publish sanitized versions of that dashboard. Anthropic’s most recent Economic Index report, published March 24, 2026, looks at how Claude was being used in February. It names the most common tasks. It tracks which categories grew the fastest. It points out that two specific kinds of business automation, sales outreach and trading workflows, at least doubled in API volume between November 2025 and February 2026.

OpenAI runs its own version called Signals at openai.com/signals. Same idea. Privacy-preserving slices of what people are doing with ChatGPT, updated regularly, posted publicly.

The internal versions are even richer.

So, when Anthropic made Claude Code generally available in May 2025 and hit a $1 billion annualized revenue run rate just six months later in November, the people who read that as a surprise are missing the mechanics. Anthropic didn’t have to guess that developers were running coding tasks at scale. Their biggest API customer was Cursor, the AI coding tool. The second biggest was GitHub Copilot, also an AI coding tool. Both products were paying Anthropic millions to put Claude behind a coding interface, and in doing so, both were handing Anthropic a real-time picture of how developers used the model.

The dashboard wasn’t abstract. It had two specific customers on it generating a big chunk of Anthropic’s revenue and showing Anthropic the entire shape of the market.

On May 22, 2025, Anthropic launched two new models, Claude Sonnet 4 and Claude Opus 4. Both were optimized for coding. They came out the same day Claude Code went generally available. Eight days later, on May 30, Anthropic introduced a new Priority Tier pricing structure that effectively raised costs for big production customers like Cursor. According to billing data reported by Ed Zitron at Where’s Your Ed At, Cursor’s Anthropic-related spend on Amazon Web Services more than doubled in a single month, from $6.2 million in May to $12.6 million in June. Cursor had to restructure its consumer pricing in mid-June and lost a lot of customers in the process.

That’s not a coincidence. That’s a company watching its biggest customer’s business and deciding when to start competing with it directly.

Then look at Windsurf. Windsurf was another AI coding startup. In February 2025 it was raising money at a $2.85 billion valuation. In May, OpenAI offered to buy it for $3 billion. In early June, Anthropic cut off Windsurf’s access to Claude. Co-founder Jared Kaplan said it would be odd for Anthropic to sell Claude to OpenAI. The polite version. The blunt version is that the labs can see who their downstream customers are about to merge with, and they can pull the plug on the model access those companies depend on to operate.

The OpenAI deal collapsed in July. Windsurf’s CEO and senior researchers left for Google in a $2.4 billion deal a few days later. What was left of Windsurf got sold to a smaller AI company called Cognition.

If your company is a thin layer of software on top of one of these model APIs, the lab can see exactly what you’re doing. You see press releases. They see API calls.

There’s more than one lab running this play. OpenAI, Anthropic, Google, Meta, xAI. Five companies all watching what’s happening on their own platforms, all racing each other to ship into whatever category looks like it’s about to take off. Ramp, the corporate card company, published a study in May 2026 looking at how 50,000 U.S. businesses spend on AI. Anthropic just passed OpenAI for the first time in business adoption, at 34.4% versus 32.3%. A year before that, Anthropic was at roughly 8%.

The competition between the labs isn’t a slowing force. It’s an accelerant. Each one is racing the others into visible adjacent markets.

Now, why is any of this suddenly possible. For most of the last thirty years it wasn’t.

The honest answer is that the models themselves crossed a threshold somewhere between 2023 and 2025. They can now read, write, and reason in plain language well enough to do work that used to require hand-coded business logic. Before, if you wanted software to handle a billing question, you needed an engineer to write every rule. If the question didn’t match a rule, the software gave up. Now the model handles the question in language, the same way a human customer service rep would. That alone collapses a lot of what used to be expensive custom software. There is so much more that I can do now – by myself – without any technical assistance than I could even 6 months ago. (Just to remind people – I am not the technical guy)

The second piece is plumbing. In June 2023 OpenAI added something called function calling to GPT-4, which let the model trigger external actions instead of just generating text. In November 2024 Anthropic released an open standard called the Model Context Protocol, or MCP. By March 2025, OpenAI had adopted it too. By December 2025 it was donated to the Linux Foundation with Google, Microsoft, AWS, and Bloomberg backing. What MCP does in plain language: it gives the model a clean, standard way to reach into your databases, your CRM, your scheduling system, your billing system, and take action. Read a patient record. Update a claim. Look up a balance. Send a message. Before this, every connection between a model and an internal system was custom plumbing built by engineers. Now it’s something closer to plug and play. The same way USB-C made every charger work with every device.

And the coding tools themselves got good enough to build the application layer fast. According to Anthropic’s own published customer stories, Stripe used Claude Code to migrate 10,000 lines of code from one programming language to another in four days, work estimated at ten engineer-weeks by hand. The security firm Wiz migrated 50,000 lines from Python to Go in roughly 20 hours of active work, a project the team estimated at two to three months manually. Rakuten cut its average feature delivery time from 24 working days to 5.

The combination is what’s new. The model can understand what a user is trying to do. It can reach into the underlying systems and take action. And the cost of building the customer-facing app on top of all of it has dropped by something like an order of magnitude. The data has not changed. The customer relationships have not changed. The regulations have not changed. The only thing that’s changed is who can plausibly build the thing.

For thirty years the way industry data got to market looked like layers in a stack. The original holder, the insurer, the payer, the EHR vendor, the bank, sat at the bottom. A middle layer of companies licensed slices of that data, combined it with other data, built a customer-facing interface on top, and sold it back to the original holder or to their customers. Most of the margin lived in the middle layer because building the interface was expensive. You needed engineers, compliance work, product designers, sales reps.

That math has changed. A small in-house team can now ship what a third party used to charge millions a year to provide.

If you’re an insurer sitting on member data, the question you should be asking now is why you’re paying an outside vendor to build a patient communication tool you could ship in six months with a small team. Hint: this happened to a good friend with an amazing product.

Epic is the cleanest case. According to coverage of Epic’s user group meeting in Verona, Optum Advisory’s VP of provider technology services said Epic “is not simply an EHR company anymore, it is building the operating system of healthcare.” Epic now ships its own AI scribe called Art. Its own AI billing agent called Penny. Its own AI patient assistant called Emmie, inside MyChart, that handles scheduling, bill explanations, payment plans, and reimbursement statements. As of February 2026, 85% of Epic’s healthcare customers were live on these built-in tools. Epic told Becker’s it has more than 150 additional AI features in development for 2026. Every one of those features was a startup pitch deck three years ago.

The same logic applies to insurers. PEP Health reported that roughly 70% of health executives are planning significant digital platform investments, and the framing from analysts is that insurers no longer see themselves as just payers. They see themselves as the operator of the digital care experience their members touch. Building that themselves means not licensing the data out to a startup who repackages it and sells it back.

This is the model lab dynamic one layer up. Whoever owns the foundational asset, model weights for the labs, raw industry data for the insurers and EHRs, moves up the stack the moment the cost of building the user-facing layer collapses. The buyer becomes the builder. The third-party middle layer gets squeezed.

So where does that leave the rest of us.

If I’m writing checks as a venture investor today, I’m not investing in companies that are just a wrapper on top of a frontier model. The dashboard makes that category bait. I’m not investing in third-party communication startups whose customer is supposed to be the insurer or the hospital system either, because the data owner is now incentivized to build the same thing in-house. I’m looking for three different things.

Companies with proprietary data the labs can’t reach. Hospital workflow data. Clinical trial data. Regulated financial data. Industrial sensor data. The lab has no way to acquire that through API telemetry. The data is the product. The model is a commodity that sits next to it.

Companies where the moat is the integration, not the model. The deep connection into an ERP system. The regulated reporting hook. The medical device that’s already credentialed in fifty hospitals. The workforce system that’s been negotiated with the unions. The lab can ship a coding tool in six months because coding is generic. The lab cannot ship a sepsis prediction system that took eight years to clear FDA and credential through hospitals. (Although I reserve the right to change my mind on this based on the speed of change and the current FDA)

The picks and shovels. Companies that sell to the labs and the data owners themselves. Inference optimization. Evaluation infrastructure. Security tooling for AI-generated code. Governance platforms. Anthropic acquired Bun in December 2025, an open-source JavaScript runtime, to accelerate Claude Code. That’s the pattern. When the platform owner is your customer, you don’t get eaten by them.

What I’m not investing in is a Series A for a vertical AI agent whose only edge is a clever prompt and a workflow on top of GPT or Claude. That edge has a half-life measured in months.

Now flip the question. If I’m running a Fortune 500 enterprise (yes you are not as safe as you believe) and the model company likes my market, can I outspend them on compute. The answer is no. And most boards I talk to do not grasp the order of magnitude.

Goldman Sachs projects that the five biggest cloud companies, Microsoft, Google, Amazon, Meta, and Oracle, will spend a combined $527 billion in 2026 alone on data center capacity for AI. A single Nvidia training cluster of 100,000 H100 chips runs $3 to $5 billion in total cost depending on networking, power, and cooling. A frontier training run today costs $200 to $500 million. The next generation expected in 2027 is projected at $1 to $3 billion per run.

If you’re a typical Fortune 500 company outside of tech, you do not have a billion dollars a year to spend on training runs. You don’t have access to the chips even if you did. On May 6, 2026, Anthropic signed a deal with SpaceX to take 300+ megawatts of capacity and 220,000+ Nvidia GPUs at the Colossus 1 data center in Memphis, just to keep up with their own customer demand. SpaceX, of all the strange bedfellows. That’s the level of scarcity we’re talking about.

And the people who do have the compute are the same people who can see what’s growing in their platforms. Microsoft sees Azure traffic. AWS sees Bedrock traffic. Google sees Vertex traffic. They all see what categories are heating up in their own clouds. Those are precisely the categories most at risk of getting absorbed.

For most large enterprises, the honest answer isn’t to compete on compute. It’s to be uncompetable on the other axes. Proprietary data. Regulated workflows. Customer relationships built over decades. Distribution. Integration depth. Trust. Things the lab can’t reproduce by spinning up another model.

The warning signal here is Uber. The Information reported in April that Uber CTO Praveen Neppalli Naga said the company burned through its entire 2026 AI budget in four months. Claude Code adoption across Uber’s 5,000-engineer organization jumped from 32% in December 2025 to 84% in March 2026. Individual engineers were running $500 to $2,000 per month in API costs. Naga himself spent $1,200 in a two-hour personal demonstration. Roughly 70% of code committed at Uber now comes from AI.

That’s what happens when a single big company gets serious about using AI in a category the labs already own. The cost curve doesn’t bend in the customer’s favor. It bends in the supplier’s favor, because the supplier is also setting the meter. (Now to be fair to my programming friends I realize it isn’t as easy as pushing a button but let’s face it – it is moving things along much faster if you know what you are doing).

I don’t think boards have missed this. Dataiku, an AI governance company, ran a survey of 900 CEOs globally with Harris Poll between February and March 2026. 80% said their job will be at risk by the end of 2026 if they don’t deliver on AI. 62% reported direct board pressure to show measurable AI outcomes. 56% admitted their competitors have a stronger AI strategy than they do. Dataiku sells AI governance tools, so they have a horse in the race. The directional finding lines up with everything else though. The Uber numbers. The Ramp spend data. Boards are not patient anymore.

When I hear people talk about fast follower as an AI strategy, I think about what Tom Davenport told me on the show. Tom’s been studying analytics in enterprise for thirty years. The point he made was that this isn’t a market where waiting on the sidelines works. It requires too much data, too much learning, too many cycles of trial and error. If you skip the curve, you don’t catch up. You watch from a distance.

Complacency used to be cheap. It’s now a documented governance failure with a Harris Poll attached.

What I’d tell a board or a CEO asking about this.

You don’t need to outguess the labs. You can’t. What you can do is make sure someone in your strategy function reads the Economic Index and the OpenAI Signals page every quarter the same way someone reads earnings reports. If the answer to “what are the three task categories on the labs’ dashboards that are growing fastest right now” is silence around the table, you are not setting strategy. You’re hoping.

You also need to be honest about which axis you can compete on. Compute is not it. Foundation models are not it. The labs will outspend you and out-train you or out innovate you in these areas. Your edge must live in the things they can’t reach without your help. Your data. Your customers. Your regulators. Your workflows.

The PM at the lab already knows what their dashboard is telling them about your category. The only question is whether you do, and whether you’ve built something they can’t replicate even after they see it.