Newsletter · · Ashutosh Agarwal

An Operator Calls Per Seat SaaS Broken as Hidden AI Debt Mounts - Is SaaS Broken? - Week of August 28, 2026

Is SaaS Broken? for the week of August 21 to 28, 2026. Podcast synthesis on a Cloudforce operator saying the per-user-per-month model cannot survive the cost of inference, a chorus of episodes quantifying roughly 2 trillion dollars of off-balance-sheet AI financing, falling token prices and model routing cutting AI bills 30 to 40 percent, and Anthropic's IPO math ahead of an S-1 that would show a frontier lab's real gross margin.

Is SaaS Broken?

Week of August 28, 2026: An Operator Calls Per Seat SaaS Broken as Hidden AI Debt Mounts

For weeks this newsletter has been hunting for one thing: a person who actually runs a software business saying, out loud, that the old "pay us per user, per month" model cannot survive the cost of AI. This week we got it, not from a pundit or a short-seller, but from a founder who sells AI software to hospitals, universities and governments for a living. Meanwhile, the loudest story in the podcast world moved somewhere unexpected: not to software margins, but to the mountain of hidden debt now propping up the whole AI build-out. That matters for our seven names too. But first, the quote we've been waiting for.

TL;DR

  • An operator said the quiet part out loud. The founder and CEO of Cloudforce, a company that builds AI software for regulated industries, flatly said the traditional per-seat software model "doesn't really work in the age of AI at all," because a per-user, per-month price "was never designed to really consider the cost of inference." His fix is the exact re-rate this newsletter tracks: consumption or outcome-based pricing.
  • The bubble talk has moved from software margins to hidden debt, and that is a bigger risk to your software book than any single earnings miss. A wave of podcasts this week put hard numbers on roughly $2 trillion of off-balance-sheet AI financing, GPUs used as loan collateral, and neoclouds that half of one investor expects to disappear within three years. If that story cracks, software multiples fall regardless of how any of our seven names actually perform.
  • Our seven names produced no fresh datapoint for a third straight week, with no Adobe, Salesforce, Datadog, Atlassian, HubSpot, Asana or Monday disclosure. Ten-plus weeks in, still no in-scope company has printed an AI-feature gross margin or a fresh net-retention number. The debate stays unsettled because the companies at the center of it keep saying nothing.

What's new

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

1. A software operator said the per-seat model "doesn't really work in the age of AI at all"

This is the datapoint the whole newsletter is built to catch. On The AI Files, "A conversation with Husein Sharaf, Founder and CEO of Cloudforce" (Aug 23), Husein Sharaf, who runs a Microsoft-backed company selling AI software into healthcare, education and government, was asked what the industry is underestimating. His answer:

The SaaS model, the traditional software model, doesn't really work in the age of AI at all. And it's because this… per-user-per-month license was never designed to really consider the cost of inference.

He then walked straight into the mechanism this newsletter exists to track, the reason cheaper AI does not automatically fix the problem:

The cost of inference is expensive… and it's going to continue to be expensive because as GPU costs come down, the number of tokens you eat goes up. So the models are becoming cheaper and more efficient, but we're using more tokens at a faster rate and we're spending more.

His conclusion is the exact re-rate the bull case depends on: "most SaaS companies are not ready for that… everybody needs to kind of think about consumption-based or outcome-based pricing for software." (Plain English: inference is the computing cost of running an AI model every time someone uses it. A flat monthly seat price doesn't flex with that cost; a consumption price does.)

Why it moves numbers: Every one of our seven names sells, at its core, a per-seat subscription with AI now bolted on. Sharaf is describing their exact problem from the inside: the flat monthly fee was designed for a world where serving one more user cost almost nothing, and AI breaks that assumption. His prescription (move to consumption or outcome pricing) is precisely what Salesforce's Agentforce flex-credits, Adobe's Firefly credits and Asana's AI Studio are groping toward. The bull reading: operators see the problem and are already re-pricing. The bear reading: an operator this candid about the seat model being obsolete is a warning that the transition is messy, unfinished, and margin-dilutive while it happens.

Two more useful nuggets from the same conversation. On China: Sharaf said customers are now "privately asking… why am I paying OpenAI or Anthropic prices if DeepSeek is close enough?" and while he won't put a Chinese model into a hospital, he argues the mere existence of cheap open models "is forcing everybody to really justify their premium price." And on the plumbing beneath our vendors: he confirmed the Microsoft and OpenAI exclusive arrangement has ended, so OpenAI's models are coming to Amazon's cloud (Amazon announced a limited preview), while Anthropic's models are now on Microsoft's Azure and his customers are "leaning much more heavily" on them. The model layer is being commoditized in real time, good for app-vendor input costs, bad for the labs' pricing power.

2. The bubble conversation moved to hidden debt, and this is the real threat to software multiples

Last week the bear case was about software margins. This week it became about the financial plumbing under the entire AI build-out, and several podcasts put real numbers on it. This is the single biggest swing factor for a software book right now, because it de-rates the whole sector at once.

The most detailed teardown came from Deep Values, "Follow the Money: Big Tech's Two Trillion Dollar Missing AI Debt" (Aug 21). The claim: the five biggest tech firms (Amazon, Microsoft, Google, Meta, Oracle) have about $1.65 trillion in issued debt, but the true underlying obligation, once you add signed-but-unbooked leases and off-balance-sheet vehicles, is roughly $2.13 trillion. The reason they hide it, per the episode, is narrative: "if Meta or Microsoft suddenly dumps hundreds of billions of dollars of raw debt onto their balance sheet… they transform from high-margin software companies into capital-intensive utility companies, and growth funds will dump the stock." Concrete examples cited: Meta's Louisiana "Project Hyperion" data center is $27.3 billion of financing but shows only about $2.37 billion on the balance sheet; Microsoft's capital-lease obligations jumped from $27 billion to nearly $63 billion; S&P has cut Oracle's debt rating to BB−, one notch above junk, and Oracle's cost of default insurance spiked to 203 basis points, "the highest level of risk premium… since the depths of the 2008 financial crisis."

The part that matters directly for "Is SaaS broken?": the same episode explained that private-credit giants (Blackstone, KKR, Blue Owl) used to pour money into leveraged buyouts of SaaS companies precisely because recurring subscription revenue was such reliable collateral, but "generative AI fundamentally threatened the SaaS model… why buy software if you can just use an AI agent to build it internally?" So the smart money that used to finance software buyouts has redirected its "capital fire hose" toward AI infrastructure instead. When the people who lend against recurring software revenue start treating that revenue as no longer safe, that is a bear-case datapoint hiding inside a macro story.

Other voices pounded the same drum:

  • On Eurodollar University, "Nvidia's Hidden Debt Risk Could Crash the ENTIRE AI Bubble" (Aug 21), Jeff Snider detailed how Nvidia offers "residual-value support," effectively promising to buy back GPUs used as loan collateral if their value collapses, and doesn't carry it on the balance sheet "because in Nvidia's modeling there's almost no chance" it's triggered, which he compared directly to credit-default swaps before 2008. His historical rhyme: Nortel and Lucent lent their own customers the money to buy their gear during the dot-com build-out, were right about the technology and wrong about the timing, and both went bankrupt.
  • On The SharePickers Podcast, "If an AI Crash Happens, It Will Be 5x Worse Than Dot-Com" (Aug 22), Justin Waite used CoreWeave as the poster child of "circular" financing: market cap about $50 billion, enterprise value about $100 billion because of debt, free cash flow of −$13.66 billion over the last twelve months; Nvidia invested over $2.1 billion of equity and granted priority chip access, and CoreWeave then borrows against those chips to buy more chips. He noted Alphabet went cash-flow negative "first time in 20 years" on AI spending, and argued any crash would be "three to five times bigger" than the dot-com bust because the market is far larger and index funds are far more concentrated in these names.
  • On The Twenty Minute VC, "The AI Bubble Will Burst: Half the Neoclouds Will Die" (Aug 22), the guest investor's line: "there's a whole bunch of [neoclouds]. I think at least half of them go away within 36 months. And if there's an economic disruption, a lot of them will go away right away." His warning was about "complacency," and "tremendous red lights… on the credit markets" that nobody is pricing.
  • And on The Enterprise AI Show, "NVIDIA's Pivot from Chipmaker to Financier" (Aug 23), the hosts made the blunt point that "nowhere in any of these equations is anybody highlighting profitability… we're now touching trillions of dollars of investment," and that companies keep spending "based on who's going to be the last one around when a whole bunch of these start collapsing." Their refrain: "the maths don't math."

Why it moves numbers: None of this is about Adobe or Salesforce directly. But software is a "risk-on" trade. If the market decides the AI capital structure is a house of cards, with hidden leverage, GPUs treated as if they hold value like real estate, and neoclouds surviving on vendor financing, it will sell the entire complex, and already-de-rated software names get pulled down with it regardless of their own fundamentals. This is the tail risk that dominates positioning right now.

3. The price relief is real and arriving, open models plus smart routing are cutting AI bills

The counterweight to the bear case is that the cost of the AI feeding every software product is genuinely falling, fast, and there's now hard evidence enterprises are capturing it. On The Information's TITV, "AT&T's Open-Source Pivot, Stripe to Buy OpenRouter" (Aug 21), the reporters said AT&T now processes about 45 billion tokens a day internally; using only frontier models from OpenAI and Anthropic "would cost in the hundreds of millions of dollars annually," but AT&T keeps its spend "comfortably less than that by relying on open source for nearly 40%" of usage, customizing Nvidia, Meta and Google open models and running them on its own data centers. (It is studying, but not yet using, Chinese models like DeepSeek and Moonshot's Kimi, citing political risk.)

The pricing itself keeps dropping. On The AI Files #113 (Aug 22), the hosts flagged OpenAI's cheapest model, "Luna," now at roughly "$1 per million tokens input and $6 output," cheap enough that Replit built a free coding mode on it that only escalates to a pricier model when a task needs "more horsepower." And on ThursdAI (Aug 21), the panel noted the expense-card company Ramp bought "router.com," which claims to cut AI costs "40% in seconds," and that Ramp says such routers have cut its own internal AI bills about 30%. OpenRouter, the routing marketplace Stripe is buying for somewhere between $7.5 billion and $8 billion (up from a $1.3 billion valuation in May), is now pushing about 88 trillion tokens with about 9% weekly growth and about 4 million users. Stripe CEO Patrick Collison's framing, per ThursdAI: "every business will have to manage both revenue flows and token flows," because token spend "now matches employee salaries and maybe will outgrow" them.

Why it moves numbers: This is the mechanism by which "cheaper models exist" becomes "your software vendor's AI actually got cheaper to run." Every point of gross margin a software company can claw back on its AI feature comes from exactly this: cheaper models plus automatic routing of routine tasks to the cheapest one that works. It is the strongest argument that the seat-plus-AI margin problem is self-healing over time. The catch, flagged on TITV: enterprises are starting to "sign contracts on a token basis," and "those prices themselves can be volatile," so vendors passing token costs through are taking on a new kind of price risk.

4. Anthropic's IPO math got bigger, and its "profit" claim got contested

The upstream story kept building. On Bloomberg Intelligence, "Broadcom Seeks More Than $60 Billion in Latest AI Debt Deal" (Aug 21), Mandeep Singh said Anthropic went "from almost $10 billion in ARR to $65 billion in a span of eight months" and expects "to close the year at around $100 billion in ARR, almost 10x." Reporter Anthony Hughes said Anthropic filed confidentially on June 1, could file publicly "as early as late next week," and would then likely list "late September… early October." On The Curve (Aug 23), the hosts put the growth at "80 times in six months" (from an about $9 billion run-rate last December), pegged the likely IPO at "$1 trillion… the largest tech IPO in history," and said Anthropic is "on track to out-earn every publicly listed software company except Microsoft," while stressing "they're not profitable yet." They cited valuation comps of Palantir at about 50x revenue and SpaceX at about 41x revenue and asked the obvious question: how do you value a company whose revenue "just keeps growing exponentially"?

Here's the delta worth flagging: last week the story was that Anthropic had turned its first profit. This week that claim got pushback. The Enterprise AI Show described Anthropic as entering "the phase of, I'm not really going to tell you what my revenues are… but none of these guys are profitable," and The Curve said plainly "they're not profitable yet." Separately, Ramp's data (via The AI Files #113) suggests OpenAI has started growing faster than Anthropic again on business bills, reversing last week's "Anthropic passed OpenAI" headline. Take the lab-leadership league table as noisy and oscillating, not settled.

Why it moves numbers: For our seven, the labs are the upstream supplier. The single most important thing the coming S-1 will reveal is a real gross margin for a frontier lab. If it's healthy and durable, the labs have less need to raise prices on their app-layer customers (mildly good for the seven). If the filing shows thin margins, huge losses and off-balance-sheet compute commitments, expect price increases after the IPO lands (bad). Either way, this is the first hard number that could actually settle the debate, and it's now weeks, not months, away.

5. The demand is not slowing, AI revenue is ramping faster than any tech wave in history

One more datapoint to keep the bear case honest. On TechStuff, "Everyone Is Saying Something Different About the AI Economy" (Aug 21), Azeem Azhar (summarizing an Exponential View research report) said AI industry revenue hit $126 billion in the year to July 2026, growing roughly 200% a year, "the fastest revenue ramp in any technology wave… probably of any sector in history," about three times faster than mobile advertising, cloud, or the internet. Crucially, the share of that revenue coming from AI companies spending on themselves has dropped from about 60% eighteen months ago to about 20% today; about 80% of the run-rate is now real enterprise demand, roughly $10 billion a month.

Why it moves numbers: The bear thesis that AI revenue is mostly circular (AI companies paying each other) is getting weaker, not stronger, the money is increasingly coming from ordinary enterprises. That's the demand tailwind under every consumption-priced vendor (hello, Datadog) and the reason the "it's all a bubble" crowd may be early even if they're eventually right.

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

The bear case (per-seat SaaS is breaking). This week produced the cleanest statement of it yet, and from an operator, not a critic. When the CEO of an AI software company says the per-user-per-month model "doesn't really work in the age of AI at all" because it "was never designed to consider the cost of inference" (The AI Files, Aug 23), that is the thesis speaking in the first person. Bolt a token-hungry AI onto a flat seat price and you've turned a near-zero-marginal-cost business into one with a real, rising cost of goods behind every click. The people who lend against software revenue apparently agree: private credit has redirected money away from SaaS buyouts because "why buy software if you can just use an AI agent to build it internally" (Deep Values, Aug 21). And the whole complex sits on a financing structure that a chorus of podcasts this week called a bubble waiting to pop, which, if it does, takes software multiples down with it no matter how well any individual vendor executes.

The bull case (incumbents re-rate and keep the margin). The cost of the AI itself is collapsing and enterprises are visibly capturing it: AT&T runs 40% of its 45-billion-tokens-a-day on open-source models and keeps its bill "comfortably" below frontier pricing (The Information's TITV, Aug 21); routers now cut costs 30 to 40% automatically (ThursdAI, Aug 21); and the cheapest frontier tokens are down to about $1 in and $6 out (The AI Files #113, Aug 22). Underlying demand is ramping faster than any prior tech wave and is now 80% real enterprise spend (TechStuff, Aug 21). And even the operator making the bear case is prescribing the bull outcome, move to consumption or outcome pricing, not predicting extinction. Vendors that own the customer relationship can pass token costs through, route the boring work to cheap models, and defend the margin.

The swing factor. It hasn't changed, and this week is the tenth-plus in a row it stayed unresolved: not one of our seven has printed an AI-feature gross margin or a fresh net-retention number. The falling token prices say the input cost is improving; the operator candor says the transition is painful; the financing bubble says the whole sector could get re-rated before the fundamentals resolve either way. The one number that would cut through all of it, a real AI-feature gross margin from a Salesforce or an Adobe, still hasn't appeared. Until it does, this stays a debate. The first place it might finally get settled isn't even one of the seven: it's Anthropic's S-1, now weeks away, which will show a frontier lab's true gross margin for the first time.

Stocks in play

Source note: for the third week running, none of the seven produced a new podcast datapoint of its own. The reads below are driven by this week's operator commentary, the model-layer economics, the financing-bubble risk, and clearly-flagged prior coverage. Every number traces to a cited episode.

Adobe (ADBE), no new datapoint this week

  • Bull: The strongest re-rate logic of the week applies squarely to Adobe, cheaper models plus routing keep pulling its Firefly inference costs down, and a move toward credit/consumption pricing (which Adobe has already started with AI credits) is exactly the fix operators are calling for.
  • Bear: Its freemium AI-credit giveaway is still uncharged inference, a deliberate near-term margin drag, and the governance overhang (CEO retired with no named successor, CFO departed) remains unresolved. No new information this week either way.
  • Next catalyst: Fiscal Q3 print in September, the first real chance for any Firefly/AI-native gross-margin disclosure and a freemium-monetization update.

Salesforce (CRM), no financial datapoint, one indirect pundit mention

  • Bull: The only in-scope name to surface at all. A podcast host noted that Salesforce, after a year of "we're cutting quota and headcount" messaging, "yesterday" said it is "actually creating new jobs, a thousand new internships" (The AI Files, Aug 23, host paraphrase, treat as pundit not company data). A softer "AI-adds-jobs" tone is friendlier to the seat base than "AI-replaces-jobs." Its Agentforce flex-credit model is the clearest example of the consumption re-rate operators are demanding.
  • Bear: Still no Agentforce consumption number, and the message-whiplash (cut jobs, then create jobs) reads as a company still figuring out its own AI story. Broader budget-capping pressure threatens seat growth and pricing power.
  • Next catalyst: Q2 FY2027 print (expected around now or early September, it has not surfaced in podcasts yet), the key tell is any hard Agentforce/flex-credit consumption figure and whether an outcome-based metric actually launches.

Datadog (DDOG), no new datapoint this week

  • Bull: This week's evidence is a tailwind for the one usage-priced name in the group, enterprise AI spend is still ramping about 200% a year and now 80% real demand (TechStuff, Aug 21); more tokens flowing means more to observe, meter and secure.
  • Bear: If routing sends the bulk of tasks to cheap models and CFOs cap AI budgets, the consumption that feeds Datadog's meter could plateau. Still no economics datapoint of its own.
  • Next catalyst: Any consumption or net-retention figure tying revenue to AI-observability growth.

Atlassian (TEAM), no new datapoint this week

  • Bull: Last quarter's blowout (revenue up 28%, Rovo AI reportedly in 80%+ of the Fortune 500) says AI is landing as a tailwind, not a threat.
  • Bear: Adoption stats aren't dollars, the market still needs Rovo to convert into an attach or revenue number.
  • Next catalyst: First Rovo revenue or attach disclosure.

HubSpot (HUBS), no new datapoint this week

  • Bull: Embedding AI (Breeze) into an SMB workflow could follow the embedded-AI playbook where AI drives usage rather than replacing it; balance sheet remains solid.
  • Bear: SMB customers are the most price-sensitive to any AI-cost pass-through, and there's still no Breeze monetization/attach or net-retention figure.
  • Next catalyst: Any Breeze attach or net-retention datapoint.

Asana (ASAN), no new datapoint 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 operators are calling for, if it ever gets airtime.
  • Bear: Smallest and most seat-dependent name here, the 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), no new datapoint for multiple straight weeks

  • Bull: Fast-growing work-management platform with room to layer AI onto an expanding seat base.
  • 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 Cloudforce operator quote is aimed straight at these three: they are the purest per-seat, per-month businesses in the group, with the least pricing power to pass token costs through and the most exposure if AI compresses the number of seats a customer needs. The cautionary example this week was Chegg, a subscription homework tool the same host described as "practically dead," from "$200 per stock" to "99 cents" in two years (The AI Files, Aug 23). That's the extreme version of what happens when a cheap AI substitute eats a subscription product.
  • Model and inference vendors (OpenAI, Anthropic, AWS Bedrock, Azure OpenAI, Google DeepMind). The whole layer is being commoditized and repriced in public: the Microsoft and OpenAI exclusivity has ended (OpenAI models heading to Amazon's cloud, Anthropic's already on Azure), the cheapest frontier tokens are about $1 in and $6 out, and enterprises are routing about 40% of usage to open-source models. Anthropic's IPO, filing possibly within days and listing by early October, will finally expose a frontier lab's real gross margin. Watch whether the "first profit" claim survives contact with an actual S-1; two podcasts this week said Anthropic is not profitable.
  • Multiple de-rating risk (the whole complex). This is the loudest read-through of the week and it sits above every single-stock view. A chorus of podcasts quantified roughly $2 trillion of hidden AI financing (Deep Values, Aug 21), Nvidia backstopping GPU loans like pre-2008 credit-default swaps (Eurodollar University, Aug 21), a crash potentially "three to five times bigger" than dot-com (SharePickers, Aug 22), and "half the neoclouds" gone within three years (20VC, Aug 22). If sentiment on that structure cracks, software multiples, already de-rated, get dragged down regardless of any one vendor's fundamentals. The offsetting comfort: real AI demand is still ramping about 200% a year and increasingly comes from ordinary enterprises, so the bubble bears may be early even if eventually right.

What changed vs. last week

The center of gravity moved again, from software margins down to the debt financing the whole build-out.

  • New this week: the first genuine operator statement that per-seat SaaS "doesn't work in the age of AI" and needs consumption/outcome pricing (Cloudforce CEO), the most on-thesis datapoint this newsletter has captured; a full chorus of hidden-debt and bubble episodes quantifying about $2 trillion of off-balance-sheet AI financing, GPU-as-collateral lending, and "half the neoclouds die in 36 months"; and the SaaS-specific detail that private credit has abandoned software buyouts because AI threatens the model.
  • Escalated from last week: the concentration and circular-financing bear. Last week it was one 11-company valuation model; this week it's five or six independent deep-dives with SPV mechanics, Oracle's BB− rating and 203bp default-insurance spike, and Nvidia's residual-value guarantees. The open-source and routing price relief also escalated with a hard enterprise example (AT&T at 40% open source, 45 billion tokens a day). Anthropic's IPO math grew (now an about $100 billion year-end ARR target and an about $1 trillion listing within weeks).
  • Reversed or contested: last week's "Anthropic passed OpenAI and turned its first profit" is now in question. Ramp's data shows OpenAI growing faster again on business bills, and two podcasts this week said Anthropic is not profitable. Treat lab leadership as oscillating.
  • No follow-up: last week's named price-war gross-margin math (about 70% flipping negative) and the cost-per-completed-task framework got no fresh update; Microsoft's 30-million-Copilot-seat story did not move.
  • Still missing (now 10+ weeks): no explicit AI-feature gross-margin percentage from any of the seven, and no direct net-retention print. The "sales held, margin cut, token costs named" print at an in-scope vendor still hasn't appeared, the single highest-value signal we're waiting for.