SaaS after seats
The last decade of software was paid for by the seat. Agents do the work; some incumbents are already selling tokens. The multiple moved first. Atlassian’s 10-Q then said paid seats were still expanding.
Nikesh Arora did not hedge. If you sell analytical software, he said, it is over. He can run a large language model against the data. The incremental modules that vendors spent a decade attaching to the system of record do not need to be sold.
If you're an analytical SaaS company, it's over. I can just go run an LLM against the data.
Nikesh Arora
The rest of his taxonomy is less terminal. Data-infrastructure software — the Databricks and Snowflake layer — is a winner, because enterprises will store more, not less. System-of-work names, Salesforce and Oracle and SAP, have to be re-engineered for a world where five people become one and the interface is an agent. That is a five-year rebuild, not a funeral.
One industry tally, single-sourced, has per-seat pricing falling from about 21 percent to about 15 percent of SaaS firms in a year. Gartner is cited as calling for at least 40 percent of enterprise SaaS spend to shift to usage, agent, or outcome pricing by 2030. The cross-section is cleaner than the causal arrow: seat-based net revenue retention sits around 95 percent (a 95–105 percent range) against 108 percent for usage-based and 115–130 percent for hybrid. That is a 13-to-25-point gap. The time series already slid from about 105 percent in 2021 to about 101 percent in 2024 — before agentic deployment. AI-feature gross margins run about 50–60 percent against traditional SaaS at about 80–90 percent. Selling the agent can defend revenue and still dilute the mix.
Beat, raise, fall
ServiceNow fell about 11 percent after a beat-and-raise fourth quarter and about 50 percent over the trailing year. A February 2026 SaaS selloff, put at about $285 billion in one tally and tied there to Anthropic’s Claude Cowork, took Salesforce and Monday.com down about 25 percent and Workday and Adobe 26–38 percent. Trade outlets had Atlassian reporting its first-ever decline in enterprise seat counts in March 2026, and Workday cutting 8.5 percent of its own workforce on AI. Atlassian’s later 10-Q walked the seat tell back. Subscription growth, the MD&A said, came from paid seat expansion and price increases. Mike Cannon-Brookes: “We are not seeing any signal of seat compression from customers. If anything, we are seeing the opposite.” The Workday 8.5 percent cut is a February 2025 filing — later restated around 7.5 percent in a 10-Q — not a March 2026 print. The multiple still moved. The operational tells that were supposed to confirm it did not survive the filings. ServiceNow’s third quarter is unprinted.
William McDermott’s answer is to sell the work, not the chair. On the first-quarter call he said 50 percent of net new business already comes from non-seat pricing, including tokens. By the second quarter, Now Assist had crossed $1 billion of AI annual-contract value, agentic-AI production customers were up nine times in nine months, and he was asked the compression question directly. “Are we worried about seat compression? Not at all.” The company raised full-year subscription guidance. Seat erosion specific to ServiceNow is still absent in the numbers. The bear case is forward-looking. Ben Thompson’s version is a category move, not a ServiceNow idiosyncrasy: Microsoft’s E7, he said, is usage-based pricing that “breaks the seat model.”
Marc Benioff said Salesforce will spend $300 million on Anthropic tokens in 2026. Anthropic itself, according to a late-2025 contract rewrite, replaced fixed-seat bundles with a low base fee plus token consumption at API rates. Buyers are cutting licenses for cost as much as for substitution: Microsoft terminated internal Claude Code seats after per-engineer bills hit $500–$2,000 a month; Uber capped agentic coding tools at $1,500 per employee per month after burning a year’s AI budget in four months. IBM’s June 8-K described clients shifting quarterly capex toward servers, storage, and memory. Josh Brown’s read: enterprise SaaS is not the priority for the rest of the summer. That is a shortage-driven reallocation, not a counted seat replacement.
John De Fucci, the Guggenheim analyst, has been upgrading traditional software — including ServiceNow and Workday — on the view that these companies will at least persist, and are trading as if they will not. Brian Belski cuts the other way on Salesforce and Adobe: do you need that stuff? He does not think so. Josh Brown added a price-action tell, not a step change: Silver Lake talks to acquire Workday, the stock up 19 percent, “a possible sign the AI driven SAS apocalypse is ebbing” — at least in prices. That sits next to De Fucci. It does not re-rate ServiceNow.
The All-In table, first on an incomplete late-August clip and then on the speaker-labeled full episode, drew the same line Arora did and then narrowed it. Horizontal systems of record — Salesforce, Workday, Oracle, SAP — were the names they called safer. Vertical SaaS was still at risk. Calacanis read a Salesforce print as $11.3 billion of revenue, up 11 percent, adjusted earnings of $5.90 against $3.27, a $46 billion full-year figure, and the stock up 20 percent that Thursday. Those are host numbers, not a fetched 10-Q. Chamath said the monoliths have net-dollar-retention and negative churn, and claimed a larger bounce from the bottom than Calacanis’s one-day print. David Friedberg said he had tried a weekend Cloud Code and Cursor CRM and switched back after Benioff’s persistence — a story he said Benioff retold on the earnings call with Dario Amodei in the room. David Sacks, now attributed on the full tape: the SaaS-apocalypse narrative was “totally overdone”; core CRM is “not gonna get ripped and replaced with something vibe coded.” The Anthropic deal makes Claude the front end and the risk is disintermediation. A Leopold short of Salesforce was, in Sacks’s telling, wrong, and in Chamath’s a leverage or prime-broker artifact. It is not a 13-F. The wiki did not mint a Salesforce page. It did not flip a step. ServiceNow’s printed second-quarter numbers were not overwritten.
A week later, on The Compound, Michael Batnick and Josh Brown put the same consumption pivot inside a peer data platform. Snowflake, they said, named “over 9,000 corporate customers utilizing agentic workflows directly inside of Snowflake’s data lake layer.” Brown’s cut: an inflection where “AI is not going to be all things to all people.” Operator color, not a ServiceNow 10-Q, and not a step flip. An unlabeled third speaker was skipped.
IBM said the quiet part in an 8-K
On July 14 IBM pre-announced a weak quarter and the market treated it as sector news. The stock lost about 25 percent — roughly sixty-seven billion dollars of market cap on a six-hundred-sixty-million-dollar revenue miss. Josh Brown’s read: disproportionate because IBM was first to say aloud what enterprise buyers were already doing.
The load-bearing sentence is in IBM’s own SEC letter, quoted on The Compound: in the first weeks of June, clients shifted quarterly capex toward servers, storage, and memory purchases to secure supply-constrained infrastructure ahead of expected price increases. Deals did not close because customers were buying hardware and, in Brown’s inference, cybersecurity — not because demand vanished. Same-day tape matched the split: Dell up, CrowdStrike to a record high, Palo Alto ripped. The filing is buyer-side confirmation of the memory shortage the wiki had built from seller guidance — a customer breaking its own quarter is stronger evidence than a vendor claiming scarcity.
The chain is messier than the headline. IBM’s distributed infrastructure business grew 37 percent — its best quarter in reported history — while software missed. Money moved within IBM, not only away from it. The letter is silent on how long the squeeze lasts; slipped deals may return. Michael Batnick noted software put in a bullish candle and IGV closed green — rebuttal one: IGV is now mostly cybersecurity weights, so green does not mean ServiceNow won. ServiceNow, Workday, and Salesforce Q2 prints are the dated falsifiers Brown and Batnick both named. Starbucks telling Wall Street it is building its own AI ops tools is a separate insourcing squeeze on the same budget line. Brown scoped the SaaS deprioritization to the rest of the summer, not forever.
The legal sidecar
Legal research is a thinner story — one podcast, plus a later filing pass. Max Junestrand, who runs Legora, puts legal services at about a trillion dollars a year and legal-tech software at about $40 billion — 4 percent software, 96 percent service. He also says the early AI bet that whoever owned the data would win is failing, and, in the same conversation, that you cannot build legal research without the corpus. Westlaw, he says, has a monopoly with the American government to report cases. RELX’s legal AI skin, Lexis+ with Protégé, is showing double-digit growth in law-firm and corporate legal. The equity tell is a multiple: Relx at about 17 times 2027 earnings, down from a peak around 31 times. That is a derate on the tape, not a cited revenue collapse. Whether disruption hits law-firm headcount before it hits the research vendors is still open.