Newsletter · · Ashutosh Agarwal

Adobe Beats and Raises but Keeps Its AI Margin Hidden - Is SaaS Broken? - Week of September 11, 2026

Is SaaS Broken? for the week of September 4 to September 11, 2026: Adobe beat on profit and revenue and raised guidance but missed on backlog and again withheld the gross margin on its AI features, Databricks and ThoughtSpot operators declared per-seat pricing finished, new benchmarking found 49% of AI-enabled software firms had to re-price on margin pain, and Anthropic's march toward a $2 trillion listing revealed it has cut its own gross-margin forecast.

Is SaaS Broken?

Week of September 11, 2026: Adobe Beats and Raises but Keeps Its AI Margin Hidden

Last week the app layer roared back to life: Salesforce printed a blockbuster quarter, half a dozen podcasts declared a "software comeback," and the stock had its best day in six years. This newsletter said the near-term test would be Adobe's fiscal third-quarter report, due in mid-September. It arrived Wednesday afternoon.

The verdict is a shrug, which, for a stock the market had left for dead, counts as a small win. Adobe beat on profit and revenue, nudged up its full-year outlook, and said its AI business is now growing more than 150% a year. And yet the shares slipped, because the pile of already-signed contracts came in light and because Wall Street is nervous about the new boss taking over in December. Most importantly for the question in our title: after roughly twelve straight weeks of waiting, Adobe still did not tell us the one number that would actually settle whether AI is quietly wrecking software margins, the gross margin on its AI features. It gave us a headline growth rate and kept the cost side in the dark.

Off-stage, though, two things got a lot more interesting. The people who actually run software companies kept confirming, out loud, that the old "charge per user" model is dying. And Anthropic, the AI lab now wired into Salesforce and half the industry, edged closer to the largest stock-market debut in history while quietly admitting its own costs are rising as fast as its sales. That second point is the closest thing we have ever gotten to a real read on AI margins. Let's get into it.

TL;DR

  • Adobe finally reported, a quiet beat, not a breakout. Third-quarter earnings of $6.13 per share (versus $6.08 expected) and revenue of $6.76 billion (versus $6.70 billion) both edged past estimates, and Adobe raised its full-year guidance slightly. Its "AI-first" annual recurring revenue is now above $650 million and growing more than 150% a year. But the backlog of signed contracts missed, the stock dipped about 1.3%, a new CEO takes over December 1 (the market wanted an outside hire and was underwhelmed), and, crucially, Adobe disclosed an AI revenue number but not an AI margin number.
  • The people who run software keep burying the per-seat model. Databricks' revenue chief said flatly that "user-based pricing is a thing of the past" and that firms still charging per employee "are under siege" because customers are adding AI agents, not headcount. And new research says 49% of software companies that bolted on AI had to re-price because the cost was eating their gross margins, with a quarter of them killing an AI project outright over cost overruns. That is the exact squeeze this newsletter tracks, finally put to a number.
  • Anthropic's IPO got bigger, and let a margin secret slip. The AI lab is heading for a debut that could value it above $2 trillion, the largest ever. Buried in the coverage: Anthropic has lowered its own gross-margin forecasts because compute costs are rising as fast as, or faster than, revenue. That is the first hard hint of what an AI-heavy income statement really looks like, and it is not the 80%-margin software business investors are used to.

What's new

Ranked by what actually moves a book, the most relevant first.

1. Adobe reported: a beat, a backlog miss, and still no AI margin

This is the datapoint of the week and a direct hit on a name we cover. On Schwab Network, "EARNINGS PANEL: ORCL, ADBE" (Sep 10), hosts Marley Caden and Kevin Green walked through Adobe's fiscal third-quarter numbers live as they crossed the wire. The headline figures beat: "Third quarter EPS coming in at $6.13. Estimate was looking for $6.08… Revenue for their third quarter, $6.76 billion. The estimate was for $6.7. So a slight beat there as well." (EPS, earnings per share, is company profit divided by shares; a "beat" means it came in above what analysts predicted.)

The guidance moved up, barely. Adobe raised its full-year adjusted profit target "to $24.45 to $24.50 from a previous range of $24.35 to $24.45" and lifted full-year revenue guidance "to $26.58 to $26.63 billion." The next quarter's profit guide, $6.30 to $6.35, was "right in line" with estimates but "at the low end of that range."

The reason the stock fell despite the beat: the backlog. "The remaining performance obligations, $22.16 billion. That missed the estimate. And that is likely why we're seeing this move to the downside," Green said, the shares were "one and a third percent lower." ("Remaining performance obligations," or RPO, is the dollar value of contracts a company has signed but not yet booked as revenue, a read on future demand. A miss there spooks investors more than a small revenue beat reassures them.)

The AI number Adobe did give was a growth rate, not a margin: "Their AI-first ARR grew more than 150 percent year over year," now "exceeding 650 million," against monthly active users of "$1 billion across creativity and productivity solutions." (ARR, annual recurring revenue, is the annualized run-rate of subscription sales.) The hosts were honest about how much that actually helps: "if you're looking at it from a nominal standpoint, it's… not a really big mover right now on the overall revenue front. So it's… a little bit more of an icing on the cake." In plain terms: $650 million of AI revenue growing fast is nice, but it is small next to Adobe's roughly $27 billion of total sales, and, the point that matters here, Adobe still didn't say what it costs to deliver.

Two more things from the same segment matter for the thesis. First, the hosts named Adobe as "arguably one of the [companies] most caught up in the… SaaSpocalypse narratives," the fear that free or cheap AI tools make Adobe's paid software unnecessary. ("SaaSpocalypse" is the nickname for the worry that AI breaks the software-as-a-service business model; SaaS just means software you rent by subscription.) Second, on what management needs to prove: "They need to talk about the fact that they are not seeing a loss in subscriptions or seats, especially on the enterprise front. And in fact, they're actually seeing that reaccelerate." The bear worry, spelled out: if the economy slows and free AI competition bites, "you're going to have less seat usage. And that's… going to have an impact on the top line, kind of similar to what we saw for Microsoft about three to four quarters ago." Add a leadership question mark, a new CEO starts December 1, and "the street was looking for someone… from the outside… Did not get a great response."

Why it moves numbers: This was the near-term catalyst the whole group was waiting on, and it landed as a "prove-it, but not yet a disaster" print. The good: a real beat, raised guidance, AI revenue compounding at 150%, and no evidence yet of the dreaded seat collapse. The bad: a soft backlog, a leadership overhang, and, twelve weeks into this watch, still no AI-feature gross margin from any company we cover. Adobe told us its AI is selling; it pointedly did not tell us whether that AI makes money. Until it does, the single most important number in this debate remains a blank.

2. The people who run software keep killing the per-seat model, and now there's a margin number to go with it

This is the most on-thesis development of the week, and it came from operators, not pundits. The core question of this newsletter is whether charging a flat monthly fee per user can survive AI features that cost real money every time they run. Two company insiders this week said the answer is no, the industry is moving to charging by usage, and a piece of research finally put a number on the pain.

On The Peel with Turner Novak, "How Databricks Went $1M to $7B+ ARR in 10 Years | Ron Gabrisko, CRO" (Sep 10), Databricks' chief revenue officer Ron Gabrisko was blunt about why per-user pricing is doomed. When Databricks charged per user, "people would be like restricting people using the product. Which then also drives down usage… So I was like, okay, let's get rid of user pricing and all the users are free. And all of a sudden usage took off." His verdict on the model half the software industry still runs on: "I think user-based pricing is a thing of the past, honestly." And the reason is exactly the fear at the heart of this newsletter: "the companies that are doing user-based pricing are… under siege because people aren't growing employees… they might be growing agents. So maybe you're charging agents. But… I think everyone is or will move to usage-based pricing." He pointed at the AI coding tools as the tell: "Even the coding tools, like early days, they were user-based. I… was like, you guys need to go usage-based and now that market is just blown up with usage-based pricing." (Gabrisko put Databricks' revenue at "6.9 billion" and rising, context for how big a business is making this bet.)

If Gabrisko explains why the model is breaking, new benchmarking research explains how much it hurts. On AI to ROI, "Why every AI conversation should be an ROI conversation, with Ketan Karkhanis, CEO ThoughtSpot" (Sep 10), host and analyst Ray Rike shared the sharpest margin datapoint of the week: "I just did some benchmarking research on AI cost management. 49% of software companies that integrate generative AI into their product had to re-price because their gross margins were taking too big of a hit… And 25% actually said we had to stop an AI initiative because of the cost overruns." That is the thesis in one sentence: for roughly half the industry, adding AI hurt gross margins enough to force a pricing change, and for a quarter of them, the math simply didn't work.

The same episode's guest, ThoughtSpot CEO Ketan Karkhanis, laid out the fix in plain language. His company charges by "credits," "those credits are basically query costs. So how much ever you use the system," and he described a real cost problem he calls "token anxiety… it's actually causing people to quarantine themselves. They're like stopping using AI, which is terrible." (A "token" is the unit of text an AI model processes; you pay per token, so heavy use runs up a bill.) His engineering answer is what he calls "LLM thinning" versus rivals' "LLM fattening": instead of dumping a whole question on the expensive AI model, ThoughtSpot uses the model only to figure out intent, then answers with cheaper conventional software. The result, he says, is "controlled predictable cost," the thing a customer's finance chief can actually budget for.

Why it moves numbers: For months the bull case rested on a promise, that incumbents would re-price from per-seat to usage and protect their margins. This week two operators confirmed the shift is real and accelerating, and, for the first time, research quantified the damage that makes it necessary: half of AI-enabled software companies re-pricing, a quarter abandoning projects on cost. For our seven, that cuts both ways. It validates the pivot (good, the ones that move to usage/credits can defend margins) and it confirms the underlying squeeze is severe (bad, the ones that can't re-price, or whose customers resist, get hurt). The engineering angle matters too: whoever spends the least on the AI model per answer wins, and "thin" architectures like ThoughtSpot's are a template our vendors will be judged against.

3. Anthropic's IPO ballooned toward $2 trillion, and let slip that its margins are getting worse

The AI lab that now sits inside Salesforce (via last week's "Claudeforce" deal) and behind a growing share of the industry's features moved closer to going public, and the coverage contained the single most important margin clue we've seen. On Rich Habits Podcast, "Trump's $5K Bribe, Anthropic's $2T IPO, & Meta's Muse" (Sep 11), Austin Hankwitz relayed the mechanics: Anthropic "is now targeting mid-October at the earliest to start marketing what could be the largest IPO in history," with its S-1 prospectus, the official document a company files before going public, "expected… late September," a slip from an earlier plan to file "as early as last week." The valuation talk: "north of $2 trillion… more than double the [$965 billion] valuation within months, making it the single largest public listing ever." He walked through the growth that number demands: a revenue run-rate of "$47 billion" in May, which "insiders now saying… is now going to be $120 billion by the end of this year," on the way to a company target of "$200 billion… by the end of 2028." Even on those figures, "$2 trillion works out to be about 16 times revenue," or "10 times forward revenue" against the 2028 goal. Anthropic is also "finalizing a $15 billion revolving credit facility" before it even sits with the banks.

Here is the part that matters for software margins. Co-host Robert Croak noted Anthropic "just posted its first ever quarterly operating profit in Q2 of this year, about $559 million on $10.9 billion of revenue," but then the kicker: "In January, the Information reported Anthropic actually lowered its gross margin projections, even as revenue skyrocketed. And that's a sign… compute costs are growing just as fast or faster than the revenue that's supposed to be justifying these lofty valuations." In other words, the upstream AI supplier is telling us, indirectly, that scale is not automatically fixing AI economics, the cost of running the models is keeping pace with the money coming in.

One honest caveat on the numbers: a second show, AI to ROI, "The OpenAI vs Anthropic Battle for the Enterprise" (Sep 9), reported different Q2 figures, Anthropic revenue of "$11.6 billion" and a smaller operating profit "of about $300 million," while noting Anthropic's quarterly revenue "surpassed OpenAI's for the first time" ($6.7 billion for OpenAI, whose operating loss "widened… to $12.3 billion"). The two shows don't agree on the exact profit, and neither cited a filing, so treat both as competing estimates until the S-1 actually lands. The direction of travel, though, is consistent: Anthropic is closer to breakeven than OpenAI, but its own margin outlook is going the wrong way.

Why it moves numbers: Anthropic's eventual S-1 will be the first time the world sees a frontier AI lab's real gross margin, the truest read yet on whether selling intelligence is a good business. This week we got a preview, and it leaned bearish for the whole stack: Anthropic has cut its own margin forecast because compute costs won't come down as fast as revenue goes up. If a lab operating at Anthropic's scale can't expand margins, the app-layer companies buying its tokens, including Salesforce, which this week was reported to be spending $300 million a year on Anthropic, have less room than bulls hope. Watch for the S-1 in late September; it is the most important document this sector will see this quarter.

4. The price of AI keeps falling, and the bills keep rising anyway

The cross-current that decides everything for our seven got fresh numbers. On the price side, competition is brutal and deflationary. On AI to ROI (Sep 9), analysts laid out the token-price ladder: Anthropic's Claude "Sonnet 5" at about "$10 per million tokens," "Opus 5… about 2.5 times more expensive, $25 per million," OpenAI's coding model at "$30 per million," and China's Moonshot "KimiK3… about $15 per million." The most striking stat: cheap open-source models "now account for 72% of all the tokens processed in OpenRouter… up from 3% a year ago." (OpenRouter is a marketplace that routes AI requests to whichever model a developer picks; a jump from 3% to 72% means developers are fleeing to cheaper models en masse.) On What the AI?! (Sep 8), Jeff Keltner added the newest frontier prices, including Google's "Gemini at $0.75 [input] and $3.75 [output] per million tokens," a fraction of the premium models.

But cheaper tokens are not translating into cheaper AI bills, and this week produced the clearest evidence yet of why. In the same AI to ROI episode, analysts cited data from Weave, which "tracks AI spending across more than 200 enterprise customers," showing that the median monthly cost of one popular AI coding tool "tripled from $69 a month in January… to $219 in June." Companies are not cutting back despite the rising cost: "Rubrik mandated Claude Code for its 1,000-person engineering team… holding the line despite their costs going up a lot." This is the "inference paradox" in one datapoint: the per-token price fell, and the per-user bill still tripled, because people used far more of it.

The operator view backs this up. On Tech Disruptors, "AWS on Managing AI Costs and Enterprise ROI" (Sep 10), AWS enterprise-finance strategist Christian Weedbrook confirmed the pattern from inside thousands of corporate budgets: companies are "burning their annual token budgets within months," and most aren't even shopping for the cheapest option, "maybe 75% of customers that I [meet] with at enterprise level are still leaning in with a single model provider," and around "95% of enterprises still run on the most expensive model." He also flagged the macro squeeze: global IT spending is set to grow "around 14% year over year this year," while "S&P 500 revenue… is only expected to grow about 8%," meaning companies are spending on AI faster than their own sales are growing, and finance chiefs are "yearning for value delivery." His two big cost levers, caching (not paying twice for the same answer) and intelligent routing (sending easy questions to cheap models), are precisely the tools our seven will need to protect their AI margins.

Why it moves numbers: This is the tension that will decide whether SaaS is broken. The bull half, collapsing token prices and cheap open-source options, is how a well-run vendor claws back margin on its AI features. The bear half, the tripling bills, the burned budgets, the 95% still paying for the priciest model, is why "the model got cheaper" does not mean "our product got cheaper to run." The vendors that master caching and routing (the AWS playbook) and thin architectures (the ThoughtSpot playbook) can win. The ones that just pass expensive tokens straight through to a flat subscription price are the ones this newsletter is worried about.

The debate: is per-seat SaaS structurally broken, or just re-rating to consumption?

The bear case (per-seat SaaS is breaking). Attach a token-hungry AI to a flat monthly per-user price and you convert a near-zero-marginal-cost business into one with a real, rising cost behind every click. This week's evidence is the strongest yet for that worry: 49% of software companies that added generative AI "had to re-price because their gross margins were taking too big of a hit," and 25% killed an AI project on cost overruns (AI to ROI, Sep 10). The bills are not falling even as token prices do, one coding tool's median enterprise cost tripled in five months (AI to ROI, Sep 9), and enterprises are "burning their annual token budgets within months" (Tech Disruptors, Sep 10). Even the upstream supplier is signaling trouble: Anthropic lowered its gross-margin forecast because compute costs are rising as fast as revenue (Rich Habits, Sep 11). And per-seat pricing itself is being called a relic, "a thing of the past," with the firms still using it "under siege" (The Peel, Sep 10).

The bull case (incumbents re-rate and keep the margin). The same evidence, read the other way, is a roadmap for survival. Software companies are already doing exactly what the moment demands, moving from per-seat to usage and credit pricing (Databricks, ThoughtSpot this week; Salesforce, Adobe and HubSpot in prior weeks), which realigns what they charge with what the AI costs. The raw price of AI is collapsing (open-source models are now 72% of tokens on OpenRouter, and Google's Gemini runs at $0.75 per million input tokens), and the winners are the ones who engineer around cost with caching, routing and "thin" model usage. Adobe's print showed the demand side holding: a beat, raised guidance, AI revenue up 150%, and no sign yet of the seat collapse bears feared (Schwab Network, Sep 10). The incumbents own the customer, the data and the workflow; the model is just an input, and inputs are getting cheaper.

The swing factor. It is still the same missing number. Twelve weeks in, not one of our seven has disclosed an actual gross margin on its AI features. Adobe had the perfect chance this week and instead gave us an AI revenue figure ($650 million, +150%) with the cost side blank. The closest thing to a real AI margin read this week came from outside the group, and it was a warning: Anthropic cutting its own gross-margin outlook. So the bull case still rests on a pricing pivot and a demand beat, while the bear case now has a hard number (half the industry re-pricing on margin pain) but still no smoking gun inside our seven. The argument won't end until one of them shows the margin.

Stocks in play

This week produced one direct print (Adobe), two indirect datapoints (Salesforce's token spend, HubSpot's pricing model), and four names that stayed dark or near-dark. No figures are invented; every number traces to a cited podcast.

Adobe (ADBE), direct coverage, the print of the week

  • Bull: A clean beat ($6.13 EPS vs $6.08; $6.76B revenue vs $6.70B), full-year guidance raised, AI-first ARR above $650 million and growing more than 150% year over year, and about 1 billion monthly active users, with no evidence yet of the feared seat collapse (Schwab Network, Sep 10).
  • Bear: The signed-contract backlog (RPO) missed at $22.16 billion and the stock fell about 1.3%; the AI ARR is still small next to roughly $27B of total revenue ("icing on the cake"); a new CEO starts December 1 to a lukewarm reception; and, the big one, no Firefly/AI gross margin was disclosed, so we still can't see whether the AI growth is profitable. Commentators flagged an economic-slowdown-plus-free-AI scenario that could cut seat usage "similar to… Microsoft about three to four quarters ago."
  • Next catalyst: The earnings call detail and any follow-up on freemium monetization and, the prize, a Firefly/AI gross margin or an enterprise net-retention figure.

Salesforce (CRM), indirect datapoint

  • Bull: Momentum from last week's blockbuster print carries; this week it was reported to be spending "$300 million on Anthropic tokens" and making Claude the default across Agentforce and Slack (AI to ROI, Sep 9), evidence it is investing hard in the agent layer that drove the re-rating.
  • Bear: That same $300 million is a concrete, rising cost of goods behind the AI features, the very margin pressure this newsletter tracks, and there's still no Agentforce gross margin or net-retention number. No Dreamforce reporting surfaced this week despite the event being close.
  • Next catalyst: Dreamforce, watch for a hard Agentforce consumption metric, an outcome-based number ("deals closed," "cases resolved"), and any net-retention or seat disclosure.

HubSpot (HUBS), indirect and practitioner color

  • Bull: Practitioners on HubShots (Sep 10) describe the Breeze Assistant as genuinely good and, tellingly, still free, "not consuming credits yet," while other Breeze features do consume credits. HubSpot is running a freemium AI motion to drive adoption before it monetizes, with its UNBOUND conference imminent.
  • Bear: "Free… not consuming credits yet" is, by definition, uncharged inference, a deliberate near-term margin drag on the most price-sensitive (small-business) customer base, and the practitioners openly wondered "if they'll change that model." No Breeze attach rate or net-retention figure.
  • Next catalyst: HubSpot's UNBOUND conference, watch for Breeze pricing/credit changes, an attach or monetization metric, and any net-retention datapoint.

Datadog (DDOG), dark this week

  • Bull: The week's evidence is a tailwind for the one usage-priced name in the group, surging AI usage (open-source models jumping to 72% of OpenRouter tokens; agents burning ever more tokens) means more to observe, meter and secure.
  • Bear: If routing sends routine work to cheap open-source models and finance chiefs cap AI budgets (enterprises "burning their annual token budgets within months"), the consumption feeding Datadog's meter could plateau. No datapoint of its own again.
  • Next catalyst: Any consumption or net-retention figure tying its revenue to AI-observability growth.

Atlassian (TEAM), dark this week

  • Bull: Prior strength (last quarter's growth, Rovo reportedly across much of the Fortune 500) still says AI is landing as a tailwind.
  • Bear: No revenue or attach datapoint surfaced this week, only a tangential mention in an unrelated episode. Adoption stats still aren't dollars.
  • Next catalyst: First Rovo revenue or attach-rate disclosure.

Asana (ASAN), dark for multiple straight weeks

  • Bull: AI Studio is a genuine consumption-priced product bolted onto a seat base, the cleanest test in the group of the exact per-seat-to-usage shift Databricks' CRO says is sweeping the industry, if it ever gets airtime.
  • Bear: Smallest and most seat-dependent name here, most exposed to "one agent replaces five seats," and continued podcast silence is itself a mild negative on mind-share.
  • Next catalyst: Any AI Studio consumption or net-retention datapoint.

Monday.com (MNDY), dark for multiple straight weeks

  • Bull: Fast-growing work-management platform with room to layer AI onto an expanding seat base, and free, like the rest, to re-price to usage as that becomes the industry standard.
  • Bear: Same structural seat-erosion exposure as Asana, and equally invisible in the conversation.
  • Next catalyst: Any AI monetization or net-retention datapoint.

Read-throughs

  • Adjacent seat-heavy SaaS (HUBS, ASAN, MNDY). The per-seat-to-usage pivot is now the consensus view of operators, not a fringe prediction, Databricks' revenue chief called user pricing "a thing of the past" and said the holdouts are "under siege" (The Peel, Sep 10). That is good for the purest per-seat names if they can execute the shift, and a real warning if they can't, they have the least pricing power and the most seat-erosion exposure. HubSpot's free-Breeze-Assistant approach is the freemium version of that bet: buy adoption now, charge for usage later.
  • Model and inference vendors (OpenAI, Anthropic, AWS Bedrock, Azure OpenAI, Google DeepMind). The layer is being repriced downward in public (Gemini at $0.75/$3.75 per million tokens; open-source models now 72% of OpenRouter volume), which is a direct input-cost tailwind for every app vendor. But two cautions: the total bill keeps rising anyway (one coding tool's median enterprise cost tripled to $219/month), and Anthropic's coming S-1, now expected late September, will reveal a frontier lab's real gross margin for the first time. This week's preview (Anthropic lowering its margin forecast) suggests price relief for app vendors may be slower than the falling headline prices imply. Azure OpenAI and AWS Bedrock produced no pricing datapoints this week; Bedrock surfaced only via the AWS cost-management discussion.
  • Multiple de-rating risk (the whole complex). Quieter this week than last, but not gone. The dominant "software comeback" euphoria of last week faded to near-silence, evidence it was a one-week relief rally, not a regime change. The overhang shifted to the AI labs themselves: Anthropic pushing toward a roughly $2 trillion IPO on a $15 billion revolver, with roughly 1% of businesses driving about 80% of the labs' revenue and big tech supplying 70% to 80% of it (Prof G Markets, Sep 10). If sentiment on that capital structure cracks, software multiples fall regardless of fundamentals.

What changed vs. last week

  • The flagged catalyst resolved, with a shrug. Last week's issue named Adobe's fiscal Q3 as the near-term catalyst for the whole group. It printed: a beat on profit and revenue, slightly raised guidance, AI ARR above $650 million (+150%), but a backlog miss, a roughly 1.3% dip, and a lukewarm response to the incoming CEO. The first in-scope earnings datapoint since Salesforce, and still no AI-feature gross margin.
  • The "software comeback" chorus thinned out. Last week half a dozen shows declared SaaS back from the dead after Salesforce's +21% day. This week that chorus all but vanished, a strong sign it was a relief rally, not a durable re-rating. Attention rotated to Adobe's individual print and Anthropic's IPO.
  • Anthropic's IPO escalated again, and leaked a margin signal. From roughly $2 trillion talk last week to a firmer figure above $2 trillion (largest ever) this week, with the S-1 slipping to late September and marketing to mid-October. New this week: a first-ever Q2 operating profit (reported as either about $559M on $10.9B revenue or about $300M on $11.6B, depending on the source, unreconciled), a $15 billion revolver, a run-rate path of $47B to $120B to $200B, and, most important, the disclosure that Anthropic lowered its gross-margin projections because compute costs are rising as fast as revenue. That is the closest we've come to a real AI-margin read.
  • The pricing pivot got more operator proof, and a hard number. Last week it was Benioff and The Information naming Adobe and HubSpot. This week two more operators (Databricks, ThoughtSpot) confirmed the death of per-seat pricing, and benchmarking research quantified the squeeze driving it: 49% of AI-enabled software firms re-pricing on gross-margin pain, 25% killing AI projects on cost.
  • The inference paradox got its cleanest datapoint. Cheaper tokens (open-source now 72% of OpenRouter, up from 3% a year ago; Gemini at $0.75/$3.75) did not lower the bill, one coding tool's median enterprise cost tripled from $69 to $219 a month, and enterprises are "burning their annual token budgets within months," with 95% still on the priciest model.
  • Still missing, now roughly 12 weeks in: no explicit AI-feature gross margin from any of the seven, Adobe gave AI revenue but not the margin on it, and no direct net-retention print. DDOG, ASAN and MNDY stayed dark; TEAM was tangential; HUBS surfaced only as practitioner color.