The Product Desk

The Distribution Shift No GTM Team Is Staffed For

A head of sales at a mid-market SaaS company took a call last Tuesday from a portfolio operator at the PE firm that just acquired their third-largest customer. The operator wanted to know, by Friday, which parts of the product had AI and which didn't. The customer's CISO had not been told. Neither had procurement. The mandate came out of the investment committee, landed on the CEO's desk as a one-page memo, and was pushed to operators with 90-day deadlines to show progress.

This is the shape of the channel now, not a one-off. OpenAI finalized a $10B venture with TPG, Brookfield, and Advent. Anthropic launched a $1.5B venture with Blackstone, Goldman Sachs, and Hellman & Friedman. These get pitched as capital raises. What they actually do is hand model vendors a GTM channel across thousands of portfolio companies, where one conversation with an operating partner deploys a product fifty times.

How Decisions Get Made in This Channel

Here is what operators actually do. They do not run demos. They look at vendors already deployed inside portfolio companies, ask which ones shipped an AI feature in the last six months, and standardize across the portfolio. The decision is made on install base plus shipped capabilities. Not on RFPs. If a product is in 5 of 50 companies and shipped something credible in Q3, it wins the other 45 by default. If it is in 2 and shipped nothing, it is not in the conversation.

The competitor you spent two quarters cultivating a champion against does not matter. The decision was made in a room your GTM team never entered, by three people who read twenty of these memos this year.

Anthropic's Mid-Market Play Confirms the Segmentation

Anthropic launched a separate company specifically for mid-market Claude deployment. OpenAI raised $4B+ for enterprise deployment. The AI platform market is disaggregating into model providers, deployment companies, fine-tuning infrastructure, and domain-specific platforms. PE operators pick from these categories, not from feature comparisons. The product team question this quarter is which layer the operator thinks it lives on.

The 2x2 That Determines Your Next Year

One axis: is the product already deployed in a top-10 PE portfolio in the category. Other axis: has the team shipped an AI capability a portfolio operator can cite in a board deck. The 'deployed plus shipped' cell is where renewals expand into portfolio-wide deals at 3-5x. The 'not deployed plus not shipped' cell is where pipeline quietly disappears. The other two cells are where the work is this quarter. Pick the cell before picking the roadmap.

The Churn Event That Won't Show Up For Two Quarters

Ben, who runs Ben's Bites, paid for Superhuman for years. He did not leave for a competitor. He spent one week with AI coding agents and built his own full-featured replacement: split inboxes, shortcuts, command palette, reply/compose, 20-second undo send, one-click unsubscribe, search, and rules. Marginal cost: $100/month for Factory. Roughly two months of Superhuman.

The comfort story product teams tell themselves is that Ben is an edge case, an AI newsletter operator who can vibe-code. What users are actually doing looks different. Airbnb reports 60% of new code is now AI-generated, and Factory is marketed at non-coders running serious projects. The "they can't code" moat is dissolving.

The competitor is not another SaaS. The competitor is a weekend and an agent. The churn will not show up for two quarters because the seats stay paid while procurement figures out what happened.

The Diagnostic: Which Cell Are You In?

Two axes. First: how personalized is the workflow, generic or deeply idiosyncratic to the user. Second: how much value is the interface vs. data gravity, network, or integrations behind it? Products in the "idiosyncratic workflow plus thin interface moat" cell get rebuilt first. Email clients sit squarely there. CRMs with real data gravity do not. Yet.

Agent-Native Architecture Is the Response

The detail worth copying: Ben added hidden selectors, state, and debug endpoints so AI agents could operate his email app without any visible UI change. Same product, dual interface layer. That is what agent-native means in practice — a surface an agent can drive end-to-end. The recommendation follows. If the power user is going to build a thin client anyway, make sure they build it on top of your API surface. You lose them as a UI user and keep them as infrastructure. The tradeoff is that metered API revenue is smaller than a seat, and you have to decide this quarter whether smaller-and-sticky beats larger-and-gone.

The Coding Tool Pipeline (May 2026)

What Survives the DIY Wave

One tell from Ben's build: he chose not to use AI for email categorization, because deterministic rules worked better. The parts of SaaS that survive are not features. They are sync, security, updates, multi-device, and support — the operational work DIY builders underestimate until month three. Price against that maintenance burden, not the feature matrix.

The Numbers Behind the Squeeze

A product lead opened the AWS cost dashboard on Monday and the inference line had doubled since August. Nothing in the product changed. The rate card did. Amazon, Microsoft, Meta, and Alphabet averaged $45 billion per quarter in free cash flow since the pandemic. Q3 2026 is projected at $4 billion. That is a 91% decline, and it is entirely AI capex. These four companies are the substrate under most modern products. When free cash flow collapses at the substrate layer, list prices rise and free tiers quietly disappear.

Three signals point the same direction:

• SoftBank cut an OpenAI-backed loan from $10B to $6B. Lenders are balking at private AI valuations.
• Anthropic is compute-constrained, buying data center access from Musk's Colossus despite the adversarial history.
• Google's Demis Hassabis admits internal compute contention. Search, Cloud, and DeepMind are fighting for the same racks.

The Contradiction: Costs May Rise Before Oversupply Materializes

Last week's take here was that GPU oversupply is building through 2026-2027 — CoreWeave's $35B capex, orbital data centers, rural buildout. That still looks right. But the supply arrives in 2027-2028. In the interim, providers under FCF pressure raise prices to defend margin. The planning assumption should be: inference costs stay high or rise through late 2027, with relief possible after. I could be wrong if Azure eats margin to hold share. Nothing in the last two earnings calls suggests they will.

A product priced on today's OpenAI, Azure, or AWS rates is priced on a melting ice sheet. Model break-even at 2-3x current costs before the renewal conversation forces it.

Google Cloud as Counterweight

Google Cloud hit $20B revenue with 63% growth, outpacing AWS and Azure. OpenAI broke Microsoft Azure exclusivity to run on AWS, Google Cloud, and Oracle at the same time. The pitch is "multi-provider resilience." What model providers are actually doing is buying capacity wherever they can find it. Anthropic running on Musk's infrastructure is the clearest tell that no single provider has enough.

What This Means for Your Cost Model

The forcing function for this sprint: stress-test unit economics at 2-3x current API pricing. Split the feature list into two cells. Survives at 3x, or goes margin-negative. The margin-negative features move behind the enterprise tier or get usage caps. A planning doc that does not name multi-provider failover, usage caps, and tiered inference routing is missing the constraint that decides 2027.