The 2026 SaaS Selloff: IGV Fell 36%, Then Rose 45%. Revenue Grew 11% to 37% the Whole Time

I sell compute to the companies that are supposed to be eating software. The agent products that run on my microVMs (PandaStack is my company, so read what follows knowing that) are the ones the market decided in February would hollow out Salesforce, Workday and Thomson Reuters. So I had a professional interest in the "SaaSpocalypse", and an engineer's suspicion of any story in which the price moves before the data does. This post puts the two side by side: what software stocks did between October 2025 and this week, and what the same companies filed during the drawdown. Prices are from Yahoo Finance's chart endpoint and revenue from the SEC's XBRL API, both pulled by me; the script is at the end. The short version is that the operating data did not move the way the prices did, in either direction, and that the metric the selloff was really about, retention, has not yet appeared in a single quarterly report.

The timeline, by close

The selloff did not start with Anthropic. The iShares Expanded Tech-Software ETF, IGV, 106 holdings and $14.3 billion of assets per iShares, peaked at 117.08 on 27 October 2025 and had lost 20 percent by 29 January 2026, when ServiceNow fell 9.9 percent on results and CNBC declared software in a bear market. On Friday 30 January Anthropic published Cowork plugins for legal, sales, marketing and data-analysis work. On Tuesday 3 February Thomson Reuters had the worst day in its history, down about 18 percent at the low per Reuters and 15.7 percent close to close in my data; LegalZoom lost 19.7 percent, RELX 14, Wolters Kluwer 13, FactSet 10.5, IGV 4.6. Bloomberg counted $285 billion of market value gone across software, financial services and asset management that day; by Wednesday Reuters put the week's loss for software and services at nearly a trillion dollars. On Thursday Anthropic shipped Claude Opus 4.6 with a PowerPoint plugin, and Gartner told Fortune that predictions of the death of SaaS were premature.

The second leg was not plugins. On 7 April Anthropic announced Project Glasswing, a preview of Claude Mythos, a model it called far ahead of anything else in cyber capability, for nine named partners and about forty other organisations, with no general release planned (Fortune). Software fell again on 9 April (Reuters); ServiceNow lost 7.9 percent that day and 16 percent for the month, which The Motley Fool attributed to Mythos, a UBS downgrade and a gross margin down from 79 to 75 percent as revenue shifted from seats to Now Assist. IGV bottomed on 10 April at 74.67, 36.2 percent below the peak, then rose 40 percent by 2 June and 45 percent by 2 October, to 108.43: 7.4 percent below the peak, 2.6 percent up on the year. The individual names fell far further than the index, from Salesforce's 43.6 percent to monday.com's 71.3; the table below has each one.

What the companies reported while this was happening

Every company below reported at least two quarters during the drawdown; the revenue lines come from SEC exhibits, not coverage. None reported a decline.

Read the growth column against the right-hand column and the correlation is not what the story implies. The XBRL series shows ServiceNow accelerating from 22.4 percent growth in the June 2025 quarter to 24.0 in June 2026, Datadog from 28.1 to 35.6, Snowflake from 31.8 to 35.1, Salesforce from 9.8 to 10.8 with Informatica's help. Workday printed 12.6, 12.6, 12.6, 13.5 and 12.8 percent for five consecutive quarters. HubSpot's 20 percent is 14 percent more customers and 4 percent more revenue per customer: the seat-expansion engine is slowing, not reversing.

Then the legal and tax names, where the February story actually lives. Thomson Reuters, whose worst day named the episode, reported 10 percent growth in Legal Professionals the following quarter and raised its outlook. LegalZoom grew subscriptions 11 percent. Intuit is the one case where the operating data moved: fiscal 2027 guidance of 9 to 10 percent against 14 delivered, TurboTax guided to 2 to 3 percent with US units down 2 percent, and a 17 percent workforce reduction in a May 8-K carrying $300 to 340 million of charges. That is a real deceleration in a consumer tax product with a 63 percent stock decline attached. It is also the only one in the table. The market's problem is that retention is a lagging metric with a twelve-month denominator: a customer who decides in March 2026 to shrink a seat count shows up as a renewal in 2027. Nothing in these filings can refute the bear case yet, and nothing supports it either.

Multiples: 4.9x to 3.0x and back to 4.2x

Jamin Ball's Clouded Judgement publishes the median EV to next-twelve-months revenue for the public cloud basket every Friday, the cleanest series I know for what the market pays for a dollar of expected software revenue.

The median compressed 39 percent from December to April with no change in median growth and retention up a point, so the entire move was in what investors would pay for an unchanged stream: a terminal-value argument, not an earnings argument. The recovery has been lopsided: the high-growth bucket roughly doubled from 9.9x to 19.4x while the low-growth bucket, where most of the basket sits, went from 2.3x to 3.4x and remains below December. The 2 October issue puts the sector at 76 percent gross margin and 4 percent operating margin. ICONIQ's September State of Scaling, across 137 private companies, found the same split: high-growth names at 9.2x in the second quarter against 2.7x for everyone else, and a Pacesetter cohort at 120 to 146 percent net retention against 99 to 102 for the rest. "AI eats software" has so far been priced as "AI eats slow software". The market sorted by growth rate, not by category.

Seat compression: announced versus measured

The clearest evidence that agents remove seats comes from the vendors' own payrolls. Marc Benioff said on the Logan Bartlett Show in September 2025 that Salesforce's support organisation had gone from 9,000 people to about 5,000 "because I need less heads", with support costs down 17 percent and interactions split half to agents, half to humans (Fortune). Salesforce cut fewer than a thousand more in February (Reuters), and Intuit's 17 percent is in the 8-K above.

Klarna has the reversal on the record. Its agent was doing the work of 700 full-time staff at the start of 2025 and 853 by the third quarter, saving $60 million, per its CEO on the November call reported by CX Dive; in the same quarter customer service and operations cost rose from $42 million to $50 million. In May 2025 it went back to offering human support as a premium tier, recruiting its own customers as contract agents (Charter in Time). Headcount went from over 7,000 in 2022 to under 3,000, with a target of under 2,000 by 2030 through attrition (Fortune). The seats went away. The cost line did not, and the humans came back at a higher price.

Those are suppliers' headcounts. The buyer-side measurement I trust most is Battery Ventures' June 2026 survey of 100 technology leaders controlling $66 billion of annual spend: 49 percent were deploying agentic AI, up from 33 six months earlier, 81 percent planned to increase technology spending and none planned to cut AI. But only 29 percent were signing contracts of 25 months or longer, against a historical 64 percent, only 16 percent saw positive returns on more than half their AI projects, and 94 percent had no consistent way to measure AI return. Shrinking contract duration is the measurable form of seat compression in 2026: not cancellation, but a refusal to commit for three years to a seat count you expect to be wrong. Microsoft said the vendor-side version aloud in April, as Ball covered: per-user businesses become per-user plus usage, and Dynamics bookings saw weaker renewals while customers paused to evaluate. I covered the pricing mechanics in the repricing post; the point here is that the pricing change makes retention unreadable. A seat count that falls while consumption rises prints as flat net retention, which is exactly what the medians show.

The other side of the ledger

The AI-native run-rates are why the fear is not irrational. Anthropic's annualised revenue reached $65 billion in July according to a single source cited by Reuters; it was $30 billion in May when I last wrote it down, and Axios reported a $100 billion pace in September, which I could not verify beyond the headline. OpenAI's annual recurring revenue was nearing $70 billion at the end of September, per Axios via Reuters. Cursor doubled to $2 billion of annualised revenue in the three months to March (Bloomberg via TechCrunch). Against that, Salesforce guides to $46.1 to 46.4 billion for fiscal 2027, ServiceNow booked $13.3 billion in calendar 2025 and Workday $9.6 billion in its fiscal 2026, both from the XBRL. If the run-rates are honest, Anthropic alone is already larger than Salesforce.

Two caveats, both operational. Run-rate is a month times twelve, not recognised revenue, and the labs do not file. And margins: ICONIQ's July State of AI, from about 300 executives at software companies shipping AI products, puts gross margin on AI products at 45 percent in 2025 and a projected 53 in 2026, against the 76 percent sector median; consumption pricing has risen to 42 percent of respondents from 35, outcome pricing to 23 from 18, and almost half say their agents need human intervention on more than 30 percent of tasks. ServiceNow's slide from 79 to 75 percent is the incumbent version of the same fact. Software's cost of goods now includes inference, and the gap between 45 and 76 points of gross margin is not a model problem. It is metering, caching, routing, scale-to-zero and the evals that stop a human stepping into a third of tasks. That is platform work, and it is the part of this story I sell, so discount my enthusiasm accordingly.

Who bought, and at what price

The buyers of record are private equity, and they bought before the fall. Thoma Bravo agreed to take Dayforce private on 21 August 2025 at $70 a share, a $12.3 billion enterprise value and a 32 percent premium, with a minority cheque from the Abu Dhabi Investment Authority (Dayforce 8-K); it closed on 4 February 2026, the morning after Thomson Reuters' worst day (Alternatives Watch). Four days later it agreed to buy Verint at $20.50 a share, a $2 billion enterprise value and an 18 percent premium (Verint 8-K); that closed on 26 November 2025. I could not verify a 2026 Vista Equity take-private from a primary source and have left Vista out. On the strategic side the incumbents bought the outcome-priced agent companies: Salesforce completed its roughly $3.6 billion purchase of Fin on 10 September. So private equity bought systems of record at 2025 prices that look generous against a 3x median, while the incumbents bought the per-resolution businesses meant to replace them, the same shape I found in the no-code post.

Systems of record, workflows, clearinghouses, and what the prices say

The theses have been argued most carefully in the newsletter whose multiples I used. In December Ball argued that agents raise the standard for systems of record rather than replacing them; in February his bear case was that classic software could still be reduced to middleware with agents capturing the incremental value on top. In June the thesis became a clearinghouse model: systems of record won the SaaS era by storing governed data and owning where workflows started; in the agents era the prize goes to whoever controls memory, context, execution and governance across agents from many vendors, with Microsoft's Agent 365, Salesforce, Snowflake and Databricks the named contenders. A week later the companion piece argued the moat moves from where the data sits to where the work is orchestrated.

The prices test these roughly as follows. The pure "systems of record survive, workflow apps die" version is not what the market paid for: the two archetypal systems of record, Salesforce and Workday, are down 11 and 13 percent on the year, the data platforms the clearinghouse argument favours, Snowflake and Datadog, are up 56 and 104, and the workflow tier split, Atlassian up 16, monday.com and HubSpot down 47 each. The "vertical SaaS plus agents" thesis, that domain data plus a domain agent is the defensible combination, got the harshest test: Intuit, Thomson Reuters and LegalZoom are the three largest decliners in my table, down 58, 26 and 44 percent, despite Thomson Reuters growing legal 10 percent and LegalZoom growing subscriptions 11. The fear there is specific and not silly: a frontier lab publishes a plugin that does the work your vertical sells, priced in tokens, and the consumer end of the vertical, tax preparation, is where units actually fell. For the enterprise names it is a 2027 renewal question. For Intuit the numbers have already moved.

Pull the numbers yourself

The growth figures came from the SEC's XBRL company-facts API, the only way to put eight companies with four fiscal year-ends on one axis. This is what I ran; it needs only Python 3.12's standard library, and the SEC requires a User-Agent with contact details.

# saas_growth.py: year-over-year quarterly revenue growth from SEC XBRL company facts.
# Run: python3 saas_growth.py   (Python 3.12, standard library only; set your own User-Agent)
import json, urllib.request

UA = {"User-Agent": "your name your@email.example"}          # SEC asks for this; anonymous calls get throttled
CIKS = {"Salesforce": 1108524, "ServiceNow": 1373715, "Workday": 1327811, "Atlassian": 1650372,
        "Datadog": 1561550, "Snowflake": 1640147, "HubSpot": 1404655}
TAG = "RevenueFromContractWithCustomerExcludingAssessedTax"  # the ASC 606 revenue tag every one of these uses

def facts(cik):
    url = f"https://data.sec.gov/api/xbrl/companyfacts/CIK{cik:010d}.json"
    with urllib.request.urlopen(urllib.request.Request(url, headers=UA), timeout=60) as r:
        return json.load(r)

for name, cik in CIKS.items():
    units = facts(cik)["facts"]["us-gaap"][TAG]["units"]["USD"]
    # 'frame' (e.g. CY2026Q2) is the SEC's de-duplicated calendar bucket; it lines up companies
    # with July, January and June year-ends on the same quarter and drops restated duplicates.
    q = {x["frame"]: x for x in units if "frame" in x and "Q" in x["frame"]}
    rows = []
    for frame in sorted(q):
        prior = f"CY{int(frame[2:6]) - 1}{frame[6:]}"
        if prior in q:
            rows.append((frame, q[frame]["end"], q[frame]["val"] / 1e9, (q[frame]["val"] / q[prior]["val"] - 1) * 100))
    print(name)
    for frame, end, bn, yoy in rows[-5:]:
        print(f"  {frame} ending {end}: ${bn:6.2f}B  {yoy:+5.1f}% y/y")

On 3 October this printed 22.4, 21.8, 22.1 and 24.0 percent for ServiceNow's last four frames and 28.1, 28.4, 32.2 and 35.6 for Datadog. Atlassian's June quarter is absent because it files a 10-K for that period; subtract the three 10-Q quarters from the fiscal-year value. For prices, Yahoo's chart endpoint at query1.finance.yahoo.com/v8/finance/chart/IGV?range=1y&interval=1d returns daily closes as JSON given a browser User-Agent. Neither needs a key.

What I take from it. The February selloff was a repricing of terminal value, not a response to anything in a filing: median growth and retention did not move while the median multiple fell 39 percent, and six months of reports since have shown acceleration at the data platforms, stability at the systems of record and one genuine deceleration, at Intuit. The bear case is unrefuted because it cannot be refuted yet; the renewal cycle that would show seat compression runs into 2027, and the shift to consumption pricing means it may never appear as a clean retention number at all. Meanwhile the companies doing the eating are at $65 and $70 billion of run-rate on 45 to 53 percent gross margins with humans in a third of their tasks. For my own quarter that is the plan: the agent companies are my customers, their gross margin is partly my bill, and every idle millisecond I remove is a point of margin that decides whether the trillion-dollar story was right. I would spend the quarter on metering and evals, not on the chart.

Related: Seats Are Dying. What the 2026 SaaS Repricing Looks Like From the Compute Layer, No-Code Didn't Die. It Sold for 2.7x ARR While Its Replacement Raised at 27x and FinOps for AI Token Costs.