The Investor

What Changed

On May 12, Anthropic converted every Claude subscription into a dollar-matched API credit pool, which is the polite way of saying a $200 monthly plan now buys exactly $200 of programmatic tokens at standard API rates and not a token more. The 70–90% discount that third-party harnesses (Cline, OpenCode, and a long tail of portfolio-stage tools) had been quietly arbitraging is gone. OpenAI answered within hours with two months of free Codex for enterprise switchers inside a 30-day window.

The same day, Ramp's April numbers showed Anthropic at 34.4% of business spend versus OpenAI at 32.3%, the first documented lead change. Pair that with a new CFO hire and an October 2026 IPO target and the read is fairly obvious: margin recovery dressed up as platform policy, timed to pre-IPO diligence.

Who Gets Hurt

Anyone whose COGS quietly assumed subscription-rate Claude tokens just lost 20–40% of effective runway, or rather, lost it on the next billing cycle. The squeeze runs in both directions. Anthropic meters from below while Notion's External Agents API (hosting Claude, Codex, Cursor, Decagon, Warp, and Devin inside one workspace) commoditizes the harness layer from above.

The coding-agent thesis is now a duopoly subsidy fight with a commoditized harness layer beneath it, and every portco priced on a Claude subscription arbitrage is worth less today than it was last Friday.

The Enterprise Share Flip in Context

Multiple sources point the same direction: Anthropic quadrupled business adoption year-over-year while OpenAI grew 0.3%. The caveat is real, since Ramp skews to US credit-card billing and understates invoiced enterprise contracts, but even after adjusting, the signal is the same: vendor stickiness in the LLM layer is effectively zero. Customers flip on each model release. That guts the single-vendor moat assumption baked into most AI application-layer marks.

The Counter-Thesis

OpenAI's 2-month Codex promo could reverse the Ramp numbers by July. The subscription arbitrage was always a feature of early-market pricing, not a permanent economic structure. And the IPO could slip if markets soften. All fair. None of it changes the fact that the June 15 change is live code rather than a memo, and portfolio COGS models need updating before it hits production billing.

The ServiceNow Data Point

ServiceNow, which is approximately the most sophisticated enterprise software buyer alive, exhausted its full-year Anthropic budget by May. Not because Claude underdelivered. Because Anthropic ships no granular per-user, per-tool usage telemetry and offers no SLAs, which is a problem when the buyer is trying to run a P&L rather than a science fair. National Life Group's CIO put the same point less politely: Anthropic is 'great for consumer usage but not great for companies.'

This is the company the market is valuing at over nine hundred billion dollars on the premise that enterprise revenue justifies the number. Consumer-grade plumbing does not, historically, support enterprise-grade multiples.

The FDE Consensus

Four organizations independently arrived at the same conclusion this quarter, which is the part worth paying attention to: deployment is the bottleneck, not model capability.

• Google Cloud — hiring hundreds of forward-deployed engineers
• OpenAI/Bain — stood up DeployCo, bought a consulting firm for 150 FDEs
• Salesforce and ServiceNow — staffing the same function internally

When four firms independently allocate headcount to deployment rather than model work, the margin is probably in deployment services. This is the Palantir playbook arriving at every frontier lab at roughly the same time, which means the labs are not spending those dollars on something else. That something else is the next model.

The Category Forming Underneath

Modal's four and a half billion dollar round and ServiceNow's AI Control Tower are two sides of the same trade: AI observability and FinOps is becoming a standalone category. Token-level cost attribution, per-user spend caps, SLA monitoring across model APIs — none of which exists as an independent product today. The first CDIO who watches the category form inside her own P&L and decides to be the vendor rather than the line item is already building it.

The people building the models are not the ones who get paid first. They are not even the ones who get paid second. They are the ones whose spend schedule everyone else is quietly trying to observe.

Investment Implications

This is probably wrong in at least one direction, but here is the view: for anyone holding LLM-layer or AI-wrapper ARR, enterprise AI spend is reversible. The counter-thesis is that switching costs build quietly through integration depth, and that is plausible. The data so far does not support it. No SLAs, no granular telemetry, no contractual lock-in — apply a 20–40% reversibility discount to any model-layer ARR multiple where those elements are absent.

The a16z Thesis, Decoded

a16z published its system-of-intelligence framework with a Stitch investment attached, which is to say the thesis has already written checks. The argument is that most of the next decade's GTM enterprise value migrates off the system-of-record layer (Salesforce at one hundred and forty billion dollars, HubSpot at nine billion) and onto the orchestration layer sitting on top of it. The record of what happened matters less than the thing deciding what to do about it.

The proof point is Lemkin's SaaStr anecdote, and it belongs in every IC memo: Salesforce cut from 10+ human seats to 2 humans + 1 API seat while spend rose eighty-three percent (twelve thousand dollars to twenty-two thousand) with more than twenty agents running underneath. The seat count collapsed. The bill went up.

Incumbents Moved First This Time

The cloud transition gave AWS time to define the infrastructure layer before the incumbents noticed. The ERPs are not making that mistake again:

• SAP — a one hundred million euro Autonomous Enterprise partner fund, with NVIDIA and Microsoft wired into the platform layer
• ServiceNow — Action Fabric decouples logic from UI and exposes workflows as headless APIs for agents
• Notion — a developer platform with Claude and Codex running as hosted teammates

Every agent-infrastructure startup just acquired large potential buyers and lost most of a moat in the same announcements. The buyer shows up once. The moat erodes continuously.

Where the Alpha Actually Sits

Vercel's production AI Gateway data has agent workloads at 59% of token volume. The spend-versus-volume split is the part worth staring at: Anthropic takes sixty-one percent of spend on expensive tool-using agentic calls, while Google takes thirty-eight percent of volume on cheap high-throughput work. Two different businesses are emerging inside the phrase 'foundation models.'

Value in GTM software is migrating from data gravity to orchestration gravity, and the incumbents' product cycles suggest a roughly twelve-to-eighteen-month gap before they ship intelligence layers of their own.

The Investable Wedge

What an AI-native GTM bet at Series A/B has to look like to be interesting: narrow high-frequency workflows with measurable outputs, institutional context that compounds with use, and consumption-based pricing capable of structurally exceeding one hundred and fifty percent NRR. The anti-pattern is the horizontal 'AI CRM' copilot, which walks straight into Salesforce's API-first counter-punch.

The counter-thesis, and it is not silly: Salesforce absorbs the intelligence layer through acquisition inside eighteen months, the wedge narrows sharply, and this turns out to be the Facebook-news-feed analogy with the incumbent winning rather than the challenger.