Newsletter · · Ashutosh Agarwal

Salesforce Ends the SaaSpocalypse and Starts Charging by Outcome - Is SaaS Broken? - Week of September 4, 2026

Is SaaS Broken? for the week of August 28 to September 4, 2026: Salesforce printed a strong quarter, announced a Claudeforce partnership with Anthropic, had its best day in six years, and then used its own earnings call to start walking away from the per-seat model it invented. The one number that would settle the argument, an actual gross margin on AI features, still has not appeared.

Is SaaS Broken?

Week of September 4, 2026: Salesforce Ends the SaaSpocalypse and Starts Charging by Outcome

For three straight weeks the seven software companies this newsletter follows said nothing. The whole "is per-seat software dying?" argument was being fought over their heads, by short-sellers, model-lab founders and podcast pundits, while Adobe, Salesforce, Datadog, Atlassian, HubSpot, Asana and Monday.com stayed silent.

That ended this week, loudly. Salesforce reported a strong quarter, stood on stage next to the CEO of Anthropic to announce a product literally named after a competitor's AI, and its stock had its best day in six years. Half a dozen podcasts declared a "software comeback." But the part that matters most for the question in our title is quieter and more important than the stock pop: on its earnings call, the biggest name in enterprise software openly started walking away from the pay-per-user model it invented, and moving toward charging customers based on how much they use the AI, or whether it actually works. The thing this newsletter has spent months predicting stopped being a prediction and became a pricing page.

TL;DR

  • The "software is dead" trade blew up. Salesforce posted $11.35 billion in quarterly revenue (up 11%), announced a deep Anthropic partnership called "Claudeforce," and the stock jumped about 21%, its best day since 2020, and up roughly 40% in a month from a level where it traded at just nine times earnings. Multiple shows called it the week SaaS came back from the dead.
  • Incumbents are officially repricing away from per-seat. Salesforce CEO Marc Benioff said customers now "want to buy and want to price in different ways," and Salesforce is signing custom Agentforce deals that charge based on the revenue the AI generates or the costs it cuts. The Information reported the same shift industry-wide and named Adobe and HubSpot, two of our seven, as already charging for AI "only when it works." This is the exact re-rate the bull case needs.
  • The math still cuts both ways, and one number is still missing. Token prices keep falling (OpenAI cut its cheap model 80%), but Gartner says the cost of running a single AI "agent" workflow will still rise more than fivefold by 2028 as usage explodes, a real margin headwind hiding behind cheaper headline prices. And after roughly 11 weeks, not one of our seven has printed the number that would settle everything: an actual gross margin on its AI features.

What's New

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

1. Salesforce Ended the SaaSpocalypse, and the Stock Had Its Best Day in Six Years

This is the datapoint of the week, and it's a direct hit on a name we cover. On Bloomberg Tech, "Nvidia Sees 70% Growth as AI Boom Accelerates" (Aug 28), Bloomberg's Brady Ford summarized why Salesforce shares "jumped beyond 21 percent, the most in six years": the company gave a strong revenue outlook and announced it was deepening its partnership with Anthropic, integrating Salesforce's products with Anthropic's Claude. His read on the earnings itself was refreshingly blunt: "the numbers were good. They weren't amazing." What actually moved the stock was fear leaving the building. As he put it, the whole sector has been living under "this whole SaaSpocalypse fear that the model layer is going to eat the application layer," the worry that AI models get so good you'd just build your own CRM and stop paying Salesforce. Seeing Dario Amodei (Anthropic's CEO) and Marc Benioff on TV together signaled "a bit of a non-aggression treaty… The application layer is here to stay, at least for the foreseeable future." (In plain English: the "application layer" is the software you actually click on, like Salesforce; the "model layer" is the raw AI underneath, like Claude. The fear was that the model would make the app unnecessary.)

Ford named the single most important proof point for our thesis: "What's really important is they show that customers are not attritioning off the platform. That's been the big fear, that customers are going to go grab an [AI model] and build their own CRM. But at least for now, it is not happening." He described the revenue model plainly: "you sell AI products kind of at the margins… You get people to upgrade their subscriptions. You get them to buy some tokens to weave AI into their existing platform."

The hard numbers, as relayed on the Elon Musk Podcast, "Salesforce Stock Surges After Claudeforce Announcement" (Aug 28), a news-recap show, so treat these as its retelling of Salesforce's report rather than company testimony: revenue of $11.35 billion, free cash flow of $1.1 billion, and remaining performance obligation, the pile of already-signed contracts not yet counted as revenue, of $33.5 billion. The hosts made the fair point that a big chunk of the profit came from a roughly $2.6 billion gain on Salesforce's own investment stake in Anthropic, not from selling software: "the venture investing success" has to be separated "from core software sales." But the software momentum showed up too: they said annualized revenue from Agentforce (Salesforce's AI-agent product) reached about $1.5 billion, and AI workflows across Agentforce and Slack hit $6.9 billion, "which nearly doubled in a short window." Slack, they noted, delivered "its fastest new business growth since being acquired," driven by a spike in Slack-bot users.

On Squawk on the Street (Sep 3), Jim Cramer, a long-time Benioff bull, captured the reversal: Salesforce is "up 40% in a month," was "40% above the 50-day" average, and "traded at nine times earnings at the bottom." His thesis in one line: Agentforce is "the product that everyone laughed at [and it's] working. And it makes it so that you want to hire more people, and therefore do more work. It's a very agent-oriented product." That directly contradicts the "one AI agent replaces five human seats" fear. Cramer's claim is that the product is additive to headcount, not a replacement for it. He expects the upcoming Dreamforce conference to be "extraordinarily good."

Why it moves numbers: For weeks the swing factor in this debate has been the total silence of the companies at its center. That silence just broke with the biggest possible print. A blowout quarter, contracts locked in for years ($33.5B of backlog), customers not fleeing to build their own tools, and a stock re-rating 40% off the lows. This is the app layer directly rebutting the "software is dead" thesis with its own results. The one honest asterisk: a meaningful slice of the reported profit was an investment gain on Anthropic, not core software, so don't confuse the headline EPS with the operating story.

2. Salesforce Started Charging by Outcome, Not by Seat, and Adobe and HubSpot Are Already Doing It Too

This is the most on-thesis development the newsletter has captured, and it deserves its own item. Last week an operator (the CEO of Cloudforce) told us the per-seat model "doesn't work in the age of AI." This week the incumbents actually did something about it.

On Tech Brew Ride Home, "50% Toward AI Takeover?" (Aug 31), host Brian McCullough read out The Information's reporting on exactly the pivot this newsletter tracks: "As software firms sell more AI, they are shifting from subscription fees to charging based on how much customers use it and whether it actually helps their business." He then quoted Benioff directly from the investor call: "Customers want to buy and want to price in different ways. This is something I've learned really aggressively recently." Salesforce, per the report, is now letting businesses "negotiate custom contracts that charge businesses based on how much the AI either grows revenue by helping salespeople close more deals or cuts costs by automating more customer service interactions."

Three things in that same segment matter for our seven:

  • Two of our names are explicitly named as already there. Per The Information's reporting: "some older software firms, including Adobe Systems, HubSpot, and Zendesk, have already moved in the direction of charging for AI only when it works." That is the first concrete, in-scope datapoint on Adobe or HubSpot in weeks, and it's directly on-thesis.
  • The startups are forcing the issue. OpenAI "has started giving some major customers the option of paying only when its AI completes tasks." The customer-service AI companies Sierra and Finn, which Salesforce is buying for $3.6 billion, "charge customers only when the AI completes tasks without human intervention." Coding startup Cognition is "promising enterprise customers up to $10 million in credits if it fails to deliver engineering results worth at least what customers pay."
  • Benioff has done this before. The segment noted the irony: more than 25 years ago, Benioff led the original shift, away from buying software outright toward renting it per employee (the per-seat model). Now he's leading the shift away from it. The cautionary note buried in the same piece: when Splunk made a similar pricing transition, its revenue "temporarily fell." Repricing is rarely clean.

The HR-industry hosts on The Chad & Cheese Podcast, "HiBob Beast Mode & DOL Dismantles Worker Protections" (Sep 4) put it in plain language, describing the two fears that had driven software stocks down and why both look wrong now. Fear one, that companies would just "vibe-code" their own CRM with an intern, "was always silly." Fear two: "if your pricing is on a per-seat basis and companies are laying off and agents are doing work, how does the model work?" Their answer: "shocker, they're all changing their pricing models from a per-seat to a consumption or how many tokens are you using… We're not bound to the seat pricing model." Salesforce also led a $166 million investment into the HR software platform HiBob (valuing it at $3.2 billion), which the hosts framed as Salesforce buying up the employee data that makes its AI agents smarter.

Why it moves numbers: This is the bull case turning from theory into revenue mechanics. The core worry has always been that a flat monthly seat price can't survive an AI feature with a real per-use cost behind it. The answer the industry is now giving, out loud, on earnings calls, is to stop selling seats and start selling usage and outcomes. Salesforce (Agentforce flex credits), Adobe (AI credits) and HubSpot are all named doing it. The risk, flagged in the same reporting: pricing transitions can dent revenue while they happen, and outcome-based deals hand some of the pricing power, and the measurement headaches, to the customer.

3. The Price of AI Keeps Falling, but the Total Bill Is Set to Rise Anyway

The counterweight to any margin worry is that the raw cost of AI is collapsing. On The AI Daily Brief, "How to Navigate the Next Wave of AI Competition" (Aug 31), host Nathaniel Whittemore detailed OpenAI's latest cuts: its cheapest model dropped about 80% and its bigger models about 20% through the API. The effect on usage was dramatic. The marketplace OpenRouter reported daily usage of one model up 5.6x and another up 13.8x after the discount, and even after the discounts expired, "nearly a third of them" kept using the models at full price. He quoted Box CEO Aaron Levie's framing of why cheaper tokens matter so much: "Anytime we can lower the cost of tokens, we will see a disproportionate increase in consumption. Even a 50% drop in token prices could result in a 5x increase in tokens."

But that same dynamic is exactly why the total bill doesn't shrink. On Business of Tech, "'Pricing Not Disclosed' Becomes a Risk as AI Screens MSPs Out of Deals" (Aug 28), host Dave Sobel laid out what he called, via Gartner, the "inference paradox" (inference is the computing cost every time an AI model produces an answer). The headline prices are down: OpenAI's frontier developer price fell "$5 per million input tokens down to $4, $30 per million output down to $20"; Google cut Gemini 3.7 Flash 50%. And yet, per Gartner, "the inference cost of a single agentic workflow rises more than fivefold by 2028," not despite falling prices but because of them, since cheaper AI gets used for far more work. His most useful figure: a simple chatbot question costs "about a penny," but an AI agent doing a real task, planning it, fetching data, checking its answer, "costs up to $1.50. That is 150 times the price for the same job." Agents burn "5 to 30 times more tokens than a chatbot," and reasoning-heavy models cost "8 to 10 times" more per token than simple ones.

The efficiency race is real, though. On Daily Tech News Show (Sep 2), the panel covered Anthropic's new Fable 5.1, which the company says is "about 25% less" to run on most workloads and "45% more efficient for highly agentic work," partly by charging less to reuse context it has already processed. Their takeaway: "efficiency is becoming the new selling point for these [models]." And on They Might Be Self-Aware (Sep 1), the hosts (in a novelty segment "interviewing" the model itself) laid out the Chinese open-weight model GLM 5.3 Flash: "15 cents in, 50 cents out per million tokens, about a tenth of what the fancy models charge," with a benchmark score near the frontier, citing a VentureBeat estimate that "45% of the AI workloads in most companies probably should be routed to" a model that cheap.

Why it moves numbers: This is the single most important cross-current for our seven. The bull half, cheaper models plus cheap open-source options for routine work, is how a Salesforce or an Adobe claws back gross margin on its AI features. The bear half, Gartner's inference paradox, is why "the model got cheaper" doesn't automatically mean "our AI feature got cheaper to run." As customers move from simple chatbots to expensive multi-step agents, the per-task cost can rise 150-fold even as the per-token price falls. Which force wins inside any one of our vendors is still, frustratingly, undisclosed.

4. Anthropic's IPO Number Doubled to About $2 Trillion, and It's Pitching a $30 Trillion Market

The upstream supplier to our seven kept getting bigger and stranger. On The Artificial Intelligence Show, "#235" (Sep 1), the hosts relayed a Wall Street Journal report that Anthropic is preparing to tell IPO investors that AI represents "an annual revenue opportunity of more than $30 trillion," its estimate of the total market AI could eventually address, not its own forecast. The valuation talk on this IPO has itself roughly doubled: several shows this week (Elon Musk Podcast, Aug 28; 20VC with Aaron Katz, Aug 31) now peg Anthropic at a potential $2 trillion listing, up from the roughly $1 trillion figure circulating just last week. Anthropic also won a legal fight this week: on The AI Daily Brief (Aug 31), the host reported a federal judge ruled the Pentagon had "no basis for declaring them a supply chain risk" and ordered the blacklist rescinded.

Why it moves numbers: Anthropic is now welded to Salesforce's story (Claudeforce) and is the upstream cost input for anyone building on Claude. The single most valuable thing its eventual IPO filing will reveal is a real gross margin for a frontier AI lab, the first hard read on whether the labs make money selling intelligence. If the margin is healthy, the labs have less need to raise prices on app-layer customers like our seven (mildly good). If it shows big losses and huge off-balance-sheet compute commitments, expect post-IPO price hikes that flow straight into our vendors' cost of goods (bad). The ballooning valuation and the enormous market pitch don't answer that question. The filing will.

5. The Bubble Alarm Got Louder, but for the First Time Serious People Pushed Back

For weeks the "it's all a debt-fuelled bubble" case has been a one-sided chorus. This week it escalated and met resistance.

On the bear side, The Meb Faber Show, "Paul Kedrosky: AI is the First Bubble With Every Ingredient at Once" (Aug 31) gave investor Paul Kedrosky room for the sharpest bear framework yet. As of the first half of 2026, he said, more than half of the financing for hyperscaler data centers "no longer comes from internal cash flows"; it's external debt (asset-backed securities, private credit, sovereigns). He called tokens "the first hyper-deflationary commodity in the history of modern economies," deflating "around 70% to 80% a year," and drew out the brutal consequence for the AI labs: if your price falls 80% a year, "just to stand still" on revenue "I need to grow 400% year over year." He noted GPU usage at big data centers is running "only… around 35, 40%" despite all the scarcity talk, a sign of hoarding and double-ordering, and that a JPMorgan report pegged "15 to 18% of the investment grade [bond] marketplace" as now data-center related. On the Elon Musk Podcast, "Nvidia Becomes The AI Industrial Landlord" (Sep 2), the hosts detailed Anthropic committing "$80 billion in cloud deals essentially all at once" ($35B with Lambda, $45B with NScale) ahead of its IPO, and walked through the "circular financing" loop the Bank for International Settlements has flagged: "If I give you a loan so you can buy my product, I can report record sales this quarter." And on The Compound and Friends, "The Four Horsemen of the AI Apocalypse with Ed Zitron" (Aug 28), Zitron argued OpenAI and Anthropic "are not generating enough cash right now to even get close to covering their $1.1 [trillion] or more… in commitments," while taking up "90% of AI infrastructure," creating what he called an "illusory demand signal."

For the first time, though, named bulls pushed back hard. On Prof G Markets, "Tyler Cowen: The AI Bears Are Asking The Wrong Questions" (Sep 4), economist Tyler Cowen rejected the framing outright: "I don't like the word bubble." His analogy: "Were automobiles a bubble in the 1920s? A lot of the companies failed… but cars are a big, big thing and AI is too." On the circular-financing worry, he said the big players act as a "buyer of last resort," the labs themselves "are not very heavily financed by debt," and "revenue growth for the current AI companies looks much better than what we saw before the dot-com bubble burst." And on RiskReversal Pod, "Imran Khan Isn't Worried About Nvidia's 'Circular' Deals" (Aug 28), investor Imran Khan, no permabull ("I'm a bear on the allocation of resources"), argued the circularity is "already priced in": strip out every dollar of circular-flow earnings (assume a quarter of Nvidia's revenue is tied to it) and Nvidia "is still trading… less than 20 times [earnings]." His bottom line: "We know that AI is real. And we do know that AI will generate positive ROI for businesses."

Why it moves numbers: This remains the biggest tail risk to a software book, not any single earnings miss, but a sector-wide de-rating if the market decides the AI capital structure is a house of cards. If that sentiment cracks, risk-on software gets sold regardless of fundamentals, and our already-cheap seven get dragged along. The new development is that the debate is now genuinely two-sided: the bear math got scarier (Kedrosky's 400%-to-stand-still is a real problem for the labs), but credible bulls are arguing the demand is real and the risk is priced. Treat this as an ongoing overhang, not a settled verdict.

The Debate: Is Per-Seat SaaS Structurally Broken, or Just Re-rating to Consumption?

The bear case (per-seat SaaS is breaking). Bolt a token-hungry AI onto a flat monthly seat price and you turn a near-zero-marginal-cost business into one with a real, rising cost behind every click. This week's bear evidence is subtle but sharp: Gartner's "inference paradox" says the cost of running the AI agents everyone is racing to sell rises more than fivefold by 2028 even as token prices fall, because a real agent task costs 150 times a simple query (Business of Tech, Aug 28). One AI-infrastructure CEO, Aaron Katz of ClickHouse, put a number on the margin question the incumbents won't: AI-native companies like Fireworks are running "30%, 35%" gross margins and merely "hope to be more over time," well below the 75 to 85% that classic software enjoys (20VC, Aug 31). Katz also named the deeper danger: "the switching costs can be very low for agentic applications," so AI revenue may not be as durable, as sticky, as the subscription revenue it's replacing. And the whole complex still sits on a financing structure a chorus of podcasts calls a bubble waiting to pop.

The bull case (incumbents re-rate and keep the margin). This week handed the bulls their best evidence yet, from the incumbents themselves. Salesforce printed a strong quarter, kept its customers, locked in $33.5 billion of backlog, and its stock re-rated 40% (Bloomberg Tech, Aug 28). More importantly, the incumbents are visibly solving the margin problem by changing how they charge: Benioff is signing outcome-based Agentforce deals, and Adobe and HubSpot are already charging for AI "only when it works" (Tech Brew, Aug 31). The raw cost of AI is collapsing, OpenAI's cheap model down 80%, open-weight models at a tenth the price (AI Daily Brief, Aug 31), and the incumbents own the customer relationship and the proprietary data that make the AI useful. As the Bloomberg reporter put it, the model layer is not eating the application layer; the two just signed a "non-aggression treaty."

The swing factor. It finally moved, but only halfway. For the first time, an incumbent both spoke and acted: Salesforce confirmed the per-seat-to-consumption pivot is real and underway, and named peers (Adobe, HubSpot) doing the same. That's a genuine win for the bull case. But the single number that would end the argument still hasn't appeared: not one of the seven has printed an actual gross margin on its AI features. Salesforce gave us Agentforce revenue and backlog, but not the margin on it, and not a net-retention figure. The closest proxies this week all came from outside our group: 30 to 35% gross margins at AI-native Fireworks, and net-retention rates above 200% at ClickHouse, roughly 160% at Palantir and 126% at Snowflake (The Compound and Friends, Sep 1). Until one of our seven shows its AI-feature margin, the bull case rests on a strong quarter and a pricing promise, not on a proven margin.

Stocks in Play

This week the group finally produced real datapoints: one direct (Salesforce), two named in pricing reporting (Adobe, HubSpot), and the rest still dark. No figures are invented; every number traces to a cited podcast.

Salesforce (CRM), the story of the week. Bull: A strong Q2 FY2027 ($11.35B revenue, up 11%; $1.1B free cash flow; $33.5B of signed backlog), customers not leaving, a headline Anthropic "Claudeforce" partnership, Agentforce at about $1.5B annualized, and a stock up roughly 40% in a month off nine-times-earnings lows. Crucially, Benioff is actively repricing to consumption and outcome deals, the exact fix this newsletter tracks (Bloomberg Tech, Elon Musk Podcast, Tech Brew, Squawk on the Street). Bear: A large chunk of the profit was a roughly $2.6B gain on its Anthropic stake, not core software; the operating beat was "good, not amazing." Pricing transitions can dent revenue mid-shift (Splunk's did). And still no Agentforce gross margin or net-retention figure. Next catalyst: Dreamforce, where the tells are a hard Agentforce consumption metric, an outcome-based number (deals closed, cases resolved), and any net-retention or seat-count disclosure.

Adobe (ADBE), first in-scope datapoint in weeks. Bull: Named by The Information as already "charging for AI only when it works." Adobe's AI-credit model is exactly the consumption re-rate operators are calling for, and falling model costs plus routing keep pulling Firefly's inference bill down (Tech Brew, Aug 31). Bear: Its freemium AI-credit giveaway is still uncharged inference, a deliberate near-term margin drag, and the leadership overhang (prior CEO retirement, CFO gap) hasn't been updated in the podcasts. Next catalyst: Fiscal Q3 print in mid-September, the first real chance for any Firefly or AI-native gross margin and a freemium-monetization update. The near-term catalyst for the whole group.

HubSpot (HUBS), first in-scope datapoint in weeks. Bull: Also named as already moving to outcome-based AI pricing (Tech Brew, Aug 31), encouraging for a company whose SMB customers are the most price-sensitive; if it can charge for Breeze only when it delivers, it defuses the pass-through worry. Bear: SMBs are still the most exposed to any AI-cost increase, and there's no Breeze attach or net-retention figure yet. Next catalyst: Any Breeze attach or net-retention datapoint, or confirmation of how the outcome-based pricing is landing.

Datadog (DDOG), dark this week. Bull: This week's evidence is a tailwind for the one usage-priced name in the group. More AI usage (OpenRouter volumes up 5 to 14x on cheaper tokens; agents burning far more tokens) means more to observe, meter and secure. Bear: If routing sends routine work to cheap open-source models and CFOs cap AI budgets, the consumption feeding Datadog's meter could plateau. No datapoint of its own again. Next catalyst: Any consumption or net-retention figure tying revenue to AI-observability growth.

Atlassian (TEAM), effectively dark, only a security mention. Bull: Prior strength (last quarter's growth, Rovo reportedly in most of the Fortune 500) still says AI is landing as a tailwind. Bear: The only Rovo mention this week was a security vulnerability found by Varonis (Cybersecurity Today, Aug 29), not a revenue datapoint. Adoption stats still aren't dollars. Next catalyst: First Rovo revenue or attach 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-consumption shift now 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 total 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, like the rest, free to reprice to consumption as the industry standard shifts. 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 pricing pivot is now the industry standard, not a fringe idea. Salesforce, Adobe, HubSpot and even Workday are all named moving on it, and the Chad & Cheese hosts summed it up: "they're all changing their pricing models from a per-seat to a consumption or how many tokens are you using" (Chad & Cheese, Sep 4). That's good news for the purest per-seat names if they can execute the same shift, and a warning if they can't, since they have the least pricing power. The adjacent proof this week that data-rich incumbents endure: Workday beat with $2.65B revenue, 14% subscription growth, and AI making up more than 25% of new contract value (Chad & Cheese, Sep 4).
  • Model and inference vendors (OpenAI, Anthropic, AWS Bedrock, Azure OpenAI, Google DeepMind). The whole layer is being repriced downward in public: OpenAI's cheap model down 80%, Anthropic's Fable 5.1 down roughly 25% and more efficient, Google's Gemini Flash down 50%, open-weight GLM at a tenth the price. That is a direct input-cost tailwind for every app vendor. But watch two things: the inference paradox (cheaper tokens, bigger total bills as usage explodes), and Anthropic's coming IPO filing, which will expose a frontier lab's real gross margin for the first time and tell us whether price hikes are coming after the listing.
  • Multiple de-rating risk (the whole complex). This sits above every single-stock view. The bubble bears escalated (Kedrosky: token prices deflating 70 to 80% a year, labs needing 400% growth just to stand still, half of data-center financing now external debt; Anthropic committing $80B of compute at once), but for the first time credible bulls pushed back (Cowen: "the bears are asking the wrong questions"; Imran Khan: circularity "already priced in," Nvidia under 20x earnings even stripping it out). If sentiment on the AI capital structure cracks, software multiples fall regardless of fundamentals, but the demand underneath is real and the debate is now genuinely two-sided.

What Changed vs. Last Week

The single biggest change this newsletter has recorded: after three straight weeks of total silence, the app layer roared back, and it repriced.

  • Reversed: Last week all seven names were dark for a third straight week and the story was entirely at the model and financing layer. This week Salesforce printed a blockbuster quarter, announced Claudeforce, and drove a sector-wide software comeback. The SaaSpocalypse fear got directly rebutted with results (customers not leaving, $33.5B backlog).
  • Prediction became action: Last week an operator (Cloudforce's CEO) merely said per-seat was broken and prescribed consumption pricing. This week the biggest incumbent did it, with Benioff signing outcome-based Agentforce deals, and The Information named two of our own (Adobe, HubSpot) already charging for AI "only when it works." The core bull mechanism moved from theory to pricing page.
  • New quantification on both sides of the bubble: The bear case escalated with Kedrosky's framework (70 to 80% annual token deflation, 400%-to-stand-still, 35 to 40% GPU utilization, 15 to 18% of the investment-grade bond market now data-center-linked) and Anthropic's $80B compute spree. But, new this week, it drew serious pushback from Tyler Cowen and Imran Khan. Last week's chorus is now a debate.
  • Anthropic IPO math doubled: From roughly $1 trillion last week to roughly $2 trillion this week, plus a market pitch above $30 trillion and a courtroom win over the Pentagon. Still described as not profitable.
  • Inference pricing got hard numbers: OpenAI's cheap model down 80% (usage up about 14x), Anthropic Fable 5.1 down 25% with 45% better agentic efficiency, Gemini Flash down 50%, GLM Flash at 15 cents in and 50 cents out, plus Gartner's inference paradox quantifying why the total bill still climbs.
  • Still missing, now about 11 weeks: no explicit AI-feature gross-margin percentage from any of the seven, and no direct net-retention print. Salesforce gave Agentforce revenue and backlog but not the margin on it. The "sales held, margin cut, token costs named" print at an in-scope vendor still hasn't appeared. DDOG, ASAN and MNDY stayed fully dark; TEAM surfaced only via a security vulnerability.