Newsletter · · Ashutosh Agarwal

A Software Vendor Cut Its Margin Guide and Blamed AI Token Costs - Is SaaS Broken? - Week of August 14, 2026

For the week of August 7 to 14, 2026, software podcasts finally got a named margin casualty of AI token costs, watched Salesforce spell out its seat-to-outcome pricing pivot, and debated whether Adobe at 8x forward earnings is the cheapest quality software around or the next Kodak.

Is SaaS Broken?

Week of August 7–14, 2026: A Software Vendor Cut Its Margin Guide and Blamed AI Token Costs

For months this newsletter has chased one number: the moment an AI-feature product's cost of running the model shows up as a real dent in a real company's profit guidance. Everyone has argued about it in theory. This week a listed software vendor finally put it on a slide: cut its full-year margin forecast, held its sales forecast flat, and named AI token costs as the reason. That is the whole thesis in one earnings release. Meanwhile Salesforce's own team walked through, out loud, the exact pricing journey we keep saying incumbents will have to make, and Adobe got taken apart by two value investors who can't decide whether it's the cheapest quality software in the market or the next Kodak.

Quick vocabulary, since we try to keep this readable: a "seat" is a per-person software license (10 employees, 10 seats, 10x the bill). "Consumption" or "usage-based" pricing charges for what you actually use: here, the AI work measured in "tokens" (the little chunks of text an AI model reads and writes; you pay per million). "Gross margin" is what's left of a dollar of revenue after the direct cost of delivering it, and the fear all year has been that AI features carry a much thinner gross margin than old software because every answer costs real money to generate. "NRR" (net revenue retention) is how much more, or less, existing customers spend a year later. Now the news.

TL;DR

  • The margin hit is no longer hypothetical. Trend Micro's AI security product grew recurring revenue 49%, but its cloud bill nearly doubled (the company literally named AI token costs), operating income fell 54%, and it cut its full-year operating-margin guide from 19% to 15% while leaving sales guidance untouched (Business of Tech). The product sold as planned; the cost of running it ate the profit.
  • Salesforce is the third named operator to spell out the seat → consumption → outcome path, and it's telling on itself: flex credits today, a "leads worked" outcome metric "pretty soon," and a candid "in the future it's going to flip... more and more outcome based" (The GTMnow Podcast). Same week, the first named large-enterprise Salesforce budget cut surfaced (the USDA).
  • The de-rated names are getting bid back. Atlassian printed +28% revenue and popped 30%+ in a day (Rovo now in 80%+ of the Fortune 500), the software ETF is up 30% off its lows, and investors are hunting bargains in the cohort: Adobe at ~8x forward earnings, HubSpot near its floor. The "SaaSpocalypse" is being called over. Whether that's right is the debate.

What's new

Ranked by what actually moves a book.

1. The AI-margin thesis finally has a real, named victim: Trend Micro

[Operator/company data, relayed by an analyst]: Dave Sobel, Business of Tech: Automation's Cost Curve (Aug 14)

This is the datapoint we've been waiting eight-plus weeks for. Trend Micro's Q2: total net sales up 13%, and its AI security product Vision One grew annual recurring revenue 49%. That's the slide a vendor loves to show. The next slide: cloud costs nearly doubled in a year, from 6.8 billion yen to 12.7 billion, and, in Sobel's words, "the company named AI Token Costs directly." Operating income fell 54%. Management then cut the full-year profit forecast by 12 billion yen, taking the expected operating margin from 19% down to 15%, while leaving net sales guidance completely unchanged. As Sobel put it: "So the AI product is selling exactly as planned. And the cost of running it is eating the margin underneath... the thing customers want most now carries a cost that scales with how much they use it."

Why it moves numbers: for a year the bear case has been abstract: "AI features are lower gross margin than legacy SaaS." This is the first clean, public instance of it hitting the guide, and the shape is exactly what the thesis predicts: revenue fine, margin cut, cost that scales with usage. Sobel framed the deeper economics better than most: "Software has always had one economic property... the next copy costs nothing... Inference breaks that. When a product answers a question using a model, that answer costs money. Every time, for every customer, forever." Trend Micro isn't in scope for this letter, but it's the template: watch for the same fingerprint (sales guide held, margin guide cut, token costs named) at any of our seven.

2. Salesforce narrates the exact pricing pivot we keep forecasting, and it's flipping to outcomes

[Operator]: Kris Billmaier, Salesforce, The GTMnow Podcast: Will Salesforce Win the AI Era Like It Won Cloud? (Aug 11)

The single most useful in-scope operator quote of the week. Billmaier is candid that Sales Cloud "traditionally has just been a seat-based offering. And a lot of customers still want to buy seat-based from us... it's predictable, it's easy." But he then lays out the three-step ladder this newsletter has argued incumbents must climb: seats today; a consumption option where you "buy a bunch of flex credits and experiment"; and, the tell, "we're going to introduce a leads worked metric here pretty soon. So think about it more outcome-based pricing for sales solutions." On the mix: revenue is "still largely seat based... with the consumption more and more growing." And the punchline: "In the future it's going to flip and I think we'll be more and more outcome based and that will be consumptive in nature. I think that's just the necessary evolution and we're on that path."

Why it moves numbers: this makes Salesforce the third named operator (after Sift/Bret Taylor and Workday in prior weeks) to publicly commit to outcome-based pricing, but the first of our seven to do it. It's a double-edged read. Bull: a seat-based incumbent that successfully re-prices to outcomes protects revenue as seats erode. Bear: "leads worked" and flex credits are, mechanically, a move away from the predictable, high-margin seat annuity toward usage revenue whose cost of goods includes real inference. Billmaier waved off the doom: "I don't think we're in a SaaS apocalypse at all... anyone really in the trenches building knows and sees the revenue numbers." Note he's the salesman here, but this is the clearest operator roadmap yet of how a mega-cap seat business intends to survive.

3. The first named large-enterprise Salesforce budget cut

[Analyst/reporter]: Laura Bratton, The Information, TITV: Inside Kalshi's $4B+ in Annualized Revenue (Aug 12)

The counterweight to item 2. The U.S. Department of Agriculture is trimming Salesforce spend and moving a task (correspondence tracking) to C3 AI. The economics: C3 will cost more in absolute dollars (~$9M/year, started in January) than the ~$4M Salesforce license plus ~$2M in outside consultants would have, so it wasn't purely price. But the USDA's largest Salesforce enterprise agreement is shrinking: up to $277M for 2025-2030 versus ~$294M for 2020-2025. Bratton's key line: "This is the first example we're seeing" of a large enterprise pulling back Salesforce spend; previously she'd only seen small companies "vibe code their own solutions." Her read on the stock: "Investors are scared... that could threaten Salesforce's pricing power, which is why the stock is down so far this year." (For balance: Salesforce still has a large USDA footprint, and C3's own revenue fell more than half last quarter, with its CEO calling the results "unacceptable.")

Why it moves numbers: pricing power is the entire Salesforce bull case. One government contract isn't a trend, but it's the first concrete, named crack in the "decades-long hold on its customers," exactly the kind of anecdote that, repeated, compresses the multiple.

4. Adobe gets a full teardown: cheapest quality software in the market, or the next Kodak?

[Analyst]: Daniel Mahncke & Shawn O'Malley, The Investor's Podcast (We Study Billionaires): TIP837 (Aug 13)

Two value investors who own the stock spent a long segment on Adobe, and it's the richest in-scope discussion of the week. The setup: the multiple has collapsed from 22x earnings to 11x, and Adobe now trades around 8x forward earnings, while the business "has grown at about 10% per year over the last four years" with "no meaningful signs of disruption yet" in the numbers. The AI evidence they cite is genuinely two-sided:

  • Bull data: AI-native revenue "tripled year over year," and Firefly (Adobe's AI creative app) is "nearing $300 million in ARR." Adobe is also, in their words, choosing the long game on the low end: a "strategic shift to focusing more on freemium users, which comes with incurring more costs in the short term, because you're providing a lot of AI credits that you're not charging for." Notably, Adobe pushed back against Wall Street's expectation that it would raise freemium prices sooner, protecting the top of the funnel over next quarter's earnings.
  • Bear data: the risk isn't the Hollywood studios, it's everyone else. As one host put it, ad agencies "don't have the same sort of love for the creative process. And if they can use AI for a fraction of the cost, that's what they will do." Plus the threat of losing young professionals, the trained-on-Adobe funnel. They quoted Jeff Cantissaria of Value Forge Capital, who sold: "for now, generative AI actually helps demand for Adobe... Over the long term, however, we see a future where Adobe is rendered obsolete, except for the very high end use cases."

Why it moves numbers: two things travel beyond Adobe. First, the freemium AI-credit giveaway is the seat-SaaS version of the Trend Micro problem: you're eating inference cost you're explicitly not billing for, a near-term margin drag taken deliberately to defend the funnel. Second, one host proposed a leading indicator worth stealing for the whole group: watch the divergence between total paid corporate seat licenses and the actual volume of creative output: if agencies quietly route work to outside AI tools, churn shows up there long before it shows up in revenue, because long contracts mask it. And a governance flag they kept circling: Adobe's CEO is retiring with no named successor and the CFO left weeks later, with no insider buying, "the canary in the coal mine."

5. Volume is beating the price cuts, the token war quantified

[Operator]: Alex Atallah, OpenRouter, 20VC: Is the Routing Layer Becoming a Commodity (Aug 10)

The cleanest real-world measurement yet of the dynamic that decides whether cheap tokens help or hurt the AI-feature margin. Atallah says token prices have fallen "90%... in like 18 months." Then the case study: OpenAI's GPT-5.6 "Luna" model dropped 5x, then another 2x in coordination with OpenRouter, a 10x price cut in two weeks. Usage response? "Guess how much usage has grown? 13x." He calls it "a close to perfect Jevons paradox story," the classic economics idea that when something gets cheaper, people use so much more of it that total spend rises. He also notes the top five models by volume on his platform are Chinese, and that a plurality of enterprises now run several models at once.

Why it moves numbers: this is the mechanism behind item 1. Per-token prices keep falling, but total AI bills keep rising because usage rises faster, which is precisely why Trend Micro's cloud bill doubled in a year the unit prices it paid went down. For any of our seven that has moved to consumption or given away AI credits, cheaper tokens are not the rescue the bulls assume; they invite more usage, and the bill (and the COGS) climbs with it. Corroborated all week: a cybersecurity-adjacent panel warned pricing is "getting out of control... not sustainable for many businesses" and that tools like Claude still don't reliably report token counts they bill for (Software Engineering Daily, Aug 11), and the model price sheet kept dropping: Grok 4.6 at $2/$6 per million in/out (The AI Daily Brief, Aug 13) and Google's Gemini 3.7 Flash at $0.75/$3.75, half the price of the prior Flash (Everyday AI, Aug 14).

The debate: is per-seat SaaS structurally broken, or are the incumbents re-rating?

Bull case: the incumbents adapt, and you're buying the panic. The financials of the in-scope names simply aren't breaking. Adobe has grown ~10% a year for four years with "no meaningful signs of disruption yet" in the numbers, yet the multiple fell from 22x to ~8x forward, "the only thing that changed is the multiple the market is willing to give" (The Investor's Podcast). Atlassian just printed +28% revenue and swung to a $139M profit, with Rovo in 80%+ of the Fortune 500 (The Rundown). The distribution moat is arguably stronger in the AI era: Adobe and Salesforce ship AI into an installed base a startup can't reach. Salesforce is actively re-pricing to outcomes rather than dying on seats (The GTMnow Podcast). And the market is voting: the software ETF is up 30% off its lows, the beaten-down names are ripping on earnings, and the "SaaSpocalypse" is being called "overblown."

Bear case: the margin structure, the pricing power, and the demand base are all quietly eroding. Trend Micro is the proof of concept that AI features carry a cost that scales with usage and eats the guide (Business of Tech). Adobe is voluntarily eating uncharged inference to hold the freemium funnel (The Investor's Podcast), a margin drag by choice. The first named large enterprise is trimming Salesforce (The Information's TITV). Cheaper tokens don't save you because usage outruns the price cut (20VC). And the whole layer these incumbents depend on may be built on sand: analysts cited by Better Offline (Aug 7) estimate 70%+ of Microsoft/Google/Amazon AI revenue comes from just OpenAI and Anthropic, two companies that "have yet to prove any possible way of becoming profitable." Scott Galloway put a number on the reckoning: to justify current hyperscaler capex you need ~$2.5 trillion of incremental AI revenue against ~$150 billion today, a 15x gap (Prof G Markets, Aug 10).

The swing factor. It comes down to whether the incumbents can re-price to consumption/outcomes without importing a lower-margin, usage-scaled cost base that behaves like Trend Micro's. If they raise prices to cover inference, they protect margin but invite churn and cheaper agent-native competitors; if they eat the inference (Adobe's freemium credits), they protect the funnel but compress margin. The seat annuity was high-margin because the marginal cost of a copy was zero. Every step toward consumption chips at that. The bulls are right that the panic overshot; the bears are right that the business model is genuinely changing shape. Both can be true, which is why the group can rally hard and still re-rate lower a year out.

Stocks in play

No in-scope company disclosed an explicit AI-feature gross margin or a direct NRR figure this week, eight-plus weeks running. Catalysts below are the next scheduled prints/events; treat them as the moments these debates get resolved.

Adobe (ADBE). Bull: ~8x forward earnings for a ~10%-grower with no financial deterioration yet; AI-native revenue tripled YoY, Firefly nearing ~$300M ARR; distribution moat into creative enterprises. Bear: deliberately eating uncharged AI credits on freemium (margin drag); non-studio/agency customers may defect to fractional-cost AI; top-of-funnel young-creative risk; CEO retiring with no successor and CFO gone, no insider buying (The Investor's Podcast). Next catalyst: fiscal Q3 earnings (September): watch for any Firefly/AI-native margin disclosure and freemium-monetization commentary.

Salesforce (CRM). Bull: actively re-pricing seats → flex credits → an outcome-based "leads worked" metric, defending revenue as seats erode; management flatly rejects the SaaSpocalypse and cites accelerating revenue (The GTMnow Podcast). Bear: first named large-enterprise budget cut (USDA shifting work to C3), a direct shot at pricing power; the pivot to consumption/outcomes trades a predictable high-margin annuity for usage revenue with real inference COGS (The Information's TITV). Next catalyst: Q2 FY2027 earnings (late August/early September): watch for a hard Agentforce/flex-credit consumption number and any "leads worked" pricing launch.

Datadog (DDOG). Bull: still the reference agentic-security operator; its CISO detailed genuinely advanced agent governance (role-based access, an LLM "judge" catching malicious skills) it's productizing (The a16z Show), and it's the consumption-priced name most levered to more AI workloads to monitor. Bear: zero consumption/pricing/margin datapoint again; as a usage-priced vendor it's the most exposed if enterprises "tokenmaxx" (route to cheap/small models to cut bills), denting the very consumption its revenue rides on. Next catalyst: watch the next quarterly print for net-new consumption and any AI-observability attach metric.

Atlassian (TEAM). Bull: blowout quarter: revenue +28% to $1.77B, $139M profit, biggest one-day gain since the 2015 IPO; Rovo in 80%+ of the Fortune 500 (The Rundown); founder-CEO Cannon-Brookes reframing AI as tailwind not threat, having pivoted the company to profitability (The Money Puzzle). Bear: Jira/Confluence are squarely in the "AI eats the seat" crosshairs (Claude as substitute); "Rovo adoption" is not the same as Rovo revenue, no monetization/attach figure disclosed; still ~a quarter of its Nasdaq peak. Next catalyst: first Rovo revenue/attach or per-seat pricing detail, so far it's a usage stat, not a dollar.

HubSpot (HUBS). Bull: re-emerged as a value candidate: ~$2B of cash/FCF, buybacks shrinking the share count from ~52-53M to ~49.5M, ROE ~7.5% and rising, "a pretty difficult ecosystem to replicate," stock ~$215 and finding a floor with buyers on the dip; in one firm's "top ten software names" to buy (InvestTalk). Bear: still no Breeze AI monetization/attach or NRR datapoint (multiple weeks); SMB customer base is among the most exposed to agent-native substitutes and downgrades. Next catalyst: any Breeze consumption/attach metric or an NRR print on the next call.

Asana (ASAN). Bull: n/a this week. Bear: fully dark: zero podcast coverage across two distinct search angles. After last week's product-vision operator appearance (no economics), silence. For a small-cap seat-based work-management tool squarely in the "one agent replaces five seats" bullseye, no news is not good news, it means no evidence of AI monetization traction to offset the seat-erosion narrative. Next catalyst: AI Studio consumption/pricing or an NRR figure, none has surfaced.

Monday.com (MNDY). Bull: n/a this week. Bear: fully dark: only a passing "I use it for project management" mention. Multiple consecutive weeks with no AI-credit, pricing, or NRR datapoint. Same structural exposure as Asana. Next catalyst: any monday AI credit/consumption or attach metric, still absent.

Read-throughs

  • Seat-heavy SaaS (HUBS, ASAN, MNDY): the "one agent replaces five seats" risk got fresh operator color this week, though from outside the group. Grindr's CEO said revenue per employee (~$2.5-2.75M) will rise this year because "we are hiring way less people than we had intended... the team we had can do all the stuff," with engineers no longer hand-writing code and outside legal/audit bills forced down (The Pomp Podcast, Aug 13). A smaller founder claimed she went from 47 employees to 4 using AI agents (Dawn of a New Era, Aug 12). Every headcount that doesn't get hired is a seat that doesn't get sold, this is the demand-side erosion the SMB-heavy names (HUBS, MNDY, ASAN) feel first, and none of them showed monetization offsetting it this week.
  • Model/inference vendors (OpenAI, Anthropic, Bedrock/Trainium, Azure OpenAI, DeepMind): the price war and the run-rate arms race both accelerated. OpenAI's annualized run rate topped $40B, roughly double end-2025, but still trails Anthropic's ~$47B (May), which the FT reports could reach $100-120B by year-end; Anthropic is reportedly buying a startup for ~$6B to improve model efficiency, and enterprises are firmly multi-model (Bloomberg Intelligence, Aug 14). Anthropic is eyeing a September IPO at a rumored ~$1T (The Rundown, Aug 11) even as it borrowed $36B from Blackstone (Prof G Markets, Aug 10). For our seven, the read-through is mixed: relentless model price cuts help the input cost line, but Jevons usage growth and multi-model complexity push total bills up, the net effect on AI-feature gross margin still hasn't been disclosed by any incumbent.
  • Multiple de-rating risk (the whole group): this is the two-sided one. The near-term tape is a tailwind: the software ETF is up 30% off its lows, Atlassian and peers are ripping on earnings, and money is rotating back into the beaten-down cohort. But the structural bear (demand concentration in two unprofitable model labs, inference cost that scales with usage, first pricing-power cracks) argues the re-rating is a mean-reversion bounce, not a re-rating to a durably higher multiple. If the "AI trade" keeps wobbling, the de-rated SaaS names are the natural landing spot for rotation, but a Trend-Micro-style margin print from any in-scope vendor would reset the multiple lower fast.

What changed vs. last week

Last week (Aug 7 issue) was a pure thesis week: the seven were mostly dark and the substance came from model-layer economics. This week the tickers came back, and two of last week's explicit watch items were hit:

  • The inference-COGS margin cut finally showed up (watch item #4, hit). Last week we flagged the per-token price reversal and asked whether it would surface as a margin-guide cut at an AI-feature-heavy vendor. This week Trend Micro did exactly that: margin guide 19% → 15%, token costs named. First hard corroboration.
  • CRM re-emerged with the outcome-pricing move (watch items #2 and #3, hit). After going quiet last week, Salesforce not only re-surfaced but became the third named operator on the outcome-based path, with a "leads worked" metric coming and an explicit "it's going to flip." Simultaneously, the first named large-enterprise Salesforce cut (USDA) landed, a new bear datapoint on pricing power.
  • ADBE went from read-through to full deep-dive. Last week Adobe appeared only as a disruption-risk mention; this week it got a rich, two-sided teardown (8x forward earnings, Firefly ~$300M ARR, freemium AI-credit giveaway, C-suite exodus).
  • TEAM and HUBS came out of the dark. Atlassian's blowout quarter (+28% revenue, Rovo in 80%+ of the F500) and HubSpot's value case (buybacks, ~$2B cash, ~$215 floor) both surfaced after being fully dark last week.
  • The software rotation continued (constructive, as flagged). Last week's "rotation into the de-rated names" signal extended: IGV +30% off lows, earnings pops, "SaaSpocalypse is over" chatter.
  • Still missing, still the central hole: no explicit AI-feature gross margin % and no direct NRR print from any of the seven, now eight-plus weeks. ASAN and MNDY went fully dark (ASAN across two query angles), and DDOG surfaced only via security/governance color with no economics. The demand-concentration bear case escalated a notch: 70%+ of hyperscaler AI revenue tied to two unprofitable labs, and Galloway's $2.5T-to-justify-capex framing.