Newsletter · · Ashutosh Agarwal

AI Agents Break the Per-Seat Model and Nobody Agrees Who Pays - Weekly SaaS / Software Podcast Recap - Week of July 12, 2026

Weekly SaaS and software podcast recap for the week of July 12, 2026. Nearly every show circled the same question as AI agents erode per-seat pricing, with Salesforce, Intuit, and the incumbents cast as most exposed.

Weekly SaaS / Software Podcast Recap

Week of July 12, 2026: AI Agents Break the Per-Seat Model and Nobody Agrees Who Pays

This week's podcasts were dominated by one uncomfortable question: if AI can do the work, who still pays for the software? Almost every software conversation, from tiny operator shows to the big VC and markets podcasts, circled back to the same idea. The old way software was sold (a fixed monthly fee for each person who logs in) is being pulled apart by AI "agents" (software robots that do tasks on their own), and nobody yet agrees on what replaces it or who gets hurt.

A few quick definitions so the rest reads easily:

  • SaaS = "software as a service," i.e. software you rent monthly instead of buying once (Salesforce, HubSpot, etc.).
  • Seat-based pricing = you pay per person ("per seat") who uses the app.
  • Consumption / usage pricing = you pay for how much you actually use (often measured in "tokens," the little chunks of text an AI reads and writes).
  • Capex = the huge amounts Big Tech spends building data centers.
  • Hyperscaler = the giant cloud landlords: Amazon, Microsoft, Google.

Here is what people actually said.

1. Dominant Themes

Theme 1, The death of the "seat." This was the loudest topic of the week. The clearest version came from Dave Sobel on Business of Tech: Daily 10-Minute IT Services Insights (July 7). His argument: for 30 years software companies could charge per user because a human sat inside the app to do a job. AI agents break that link, an agent "completes a task by reaching across multiple systems," so "the task gets done without a person sitting in any one application." His line: "In plain terms, the value never lived in the software. It lived in the job the software helped a person do... Agents just unbolted them." He leaned on a Gartner estimate that agentic AI "could affect $234 billion of SaaS spending by 2030", roughly 20% of everything businesses spend on software subscriptions. As proof the sellers themselves believe this, he pointed to Notion shutting down its own working email product (because more than half its users never opened the inbox, the agent handled it), and to the giants pivoting to human services: Microsoft's new "Frontier" unit ($2.5 billion, ~6,000 consultants embedded at customer sites), Amazon's own ~$1 billion embedded-consultant commitment two days earlier, and OpenAI's goal of "300,000 certified consultants by the end of the year."

Theme 2, Usage-based pricing is winning, but the honest model is a hybrid. Stripe's head of AI monetization laid out the mechanics on The MAD Podcast with Matt Turck (July 9). Old SaaS had "beautiful and simple" economics because serving one more customer "costs you basically nothing", so flat subscriptions and per-seat licenses worked. AI breaks that because "every prompt and every API call and every task has a real marginal cost." His read of the market: "I see very few scaling or scaled AI companies that are still exclusively subscriptions or seat-based," because a few heavy users "cost you a ton" and you can't tell "the sheep from the goats" without a usage meter. The pattern he sees everywhere is a hybrid: a familiar fixed subscription to get people comfortable, then usage charges once they cross a threshold, he cited Lovable and ElevenLabs both starting on subscriptions and then moving to pay-as-you-go on top. And for agents specifically, he expects "real-time metering" and real-time billing, because "agents can consume at machine speed" and could otherwise "rack up a bunch of spend and then go dark."

A vivid, real example of pricing "up the value stack" came from Higgsfield CEO Alex Mashrabov on The Official SaaStr Podcast (July 10). His AI-video company went "from per-token costs to charging by outcome, by video," and now roughly 40% of usage is higher-level "Cinema Studio / Marketing Studio" workflows rather than just picking a model. The eye-opener: the average Higgsfield customer spends "around $1,000 a year," versus about $200 for Canva, 5x the revenue per customer, and about 70% of his stated ~$300 million revenue run-rate comes from creative agencies using it to replace slow, expensive production.

Theme 3, The AI bill has become a boardroom problem. Jordan Wilson devoted a whole Everyday AI episode (Ep 813, July 7) to "AI cost control." The trigger: the industry is quietly killing unlimited plans, GitHub Copilot replaced unlimited "premium requests" with token credits, Microsoft's Copilot "Cowork" moved to task-level credits, and even xAI's Grok started charging by credit. He put real numbers on how expensive heavy use is: he burns "about 2-ish billion tokens a week" on a $200/month OpenAI Codex plan, which at API prices "is a $200,000 a month" bill, a reminder of how heavily today's subscriptions are subsidized. He cited concrete belt-tightening: "Uber reportedly burned through its 2026 AI coding budget in just four months," "Tesla capped employee AI tool spend to just $200 a week," and a UBS survey finding "60% of interviewed enterprises are throttling AI spend already."

Theme 4, "Is SaaS dead?" (the "SaaSpocalypse"). This label, coined after roughly $300 billion was wiped off software valuations earlier in 2026, came up again and again. Most serious voices landed on "no, but…" (see Debates below). A useful framing came from The Information's TITV (July 6, "Is Claude Enabling a SaaSpocalypse?"), which reported real examples of companies ripping software out: French drugmaker Sanofi is "aiming to take 80% of their workloads out of ServiceNow and other software apps" into a custom in-house platform (called Concierge) built with Cursor and Claude Code, worth "at least tens of millions of dollars annually." The reporter's caveat mattered though: big companies are "largely not replacing their biggest software providers," and consultants say keeping systems like Salesforce as a "system of record" and building AI on top is usually smarter than the disruptive rip-out. Salesforce itself is leaning in, having launched "Headless 360" and declared "the user interface is dead."

Theme 5, AI infrastructure spending: still accelerating, but the returns question is getting louder. On Closing Bell (July 7), Deepwater's Gene Munster read Amazon's new $25 billion bond raise (on top of $37 billion four months earlier) as proof "the AI buildout is in its early innings", Street expectations for hyperscaler capex growth next year have jumped from 17% to 23%, and recent Amazon/Google moves suggest the real number "could be up around 37%." The counterweight all week: who is actually driving that demand, and does it pay off? On Catalyst with Shayle Kann (July 9), a SemiAnalysis guest argued the frontier AI labs are the real engine: of the gigawatts Amazon and Microsoft are building, "half of that" combined is going to OpenAI and Anthropic, "these guys are basically proxies increasingly for OpenAI and Anthropic." Anthropic reportedly wants to go from 1.5 gigawatts of capacity at end-2025 to "over 10 gigs" by 2027, "you're building a Google in 2 years."

Theme 6, The great model price war moved down into enterprise software. OpenAI launched GPT-5.6 (three tiers: Sol, Terra, Luna) plus "ChatGPT Work," a desktop agent aimed squarely at Anthropic's enterprise lead. On Tech Brew Ride Home (July 10), the numbers OpenAI is marketing: on the Artificial Analysis Coding Agent Index, Sol "sets a new state-of-the-art at 80, 2.8 points above Fable 5, while using less than half the output tokens, taking less than half the time, and costing about one-third less." Meanwhile Tech Brew's "China (AI) Rising" (July 8) noted Chinese models (DeepSeek, Z.ai) captured "over 30% of US token usage (peaking at 46%)," undercutting the US labs on price. The relevance to software investors: the "engine" that infra software runs on is getting cheaper and more commoditized fast.

2. Key Debates

Debate 1, Is traditional SaaS being killed by AI, or just forced to evolve?

  • "It's overblown" side: Dame Anne Glover of Amadeus Capital Partners, on Private Equity Spotlight (July 8), called the SaaSpocalypse "an overreaction", AI mainly makes SaaS companies "more efficient at building great products." Geoff McQueen on Spark of Ages (July 10) put it bluntly: "SaaSpocalypse is very overrated," because "vibe-coded" demos (apps built quickly by non-engineers with AI) lack the "security, governance, scalability, and enterprise-grade rigor" real products need, and because most people simply don't want to build their own software. His gardening analogy: only ~20% of people actually enjoy it; giving everyone the tools doesn't make them want to.
  • "But some die" side: Glover's own warning is that SaaS firms will become "dinosaurs", not because their engineers get replaced, but "because they're not understanding the customer need and someone inside the organization can do it faster and better" now that in-house teams can build with AI. Justin Watt on Futureproof Founder (July 7) was specific about who's at risk: "SaaS companies that will disappear... are the ones who are very vertical, who aren't flexible," while those that pivot to a "headless" data-and-API approach thrive.
  • The wedge: does easier coding mostly help incumbents build better products, or does it hand their customers the ability to replace them? Everyone agrees the pricing model and upsell change; they disagree on whether the need for the vendor survives.

Debate 2, What comes after per-seat pricing: pure usage, or something else?

  • Usage/consumption: Stripe's AI chief (above) sees usage-based billing as "really critical for AI companies."
  • Outcome/ROI-based: Higgsfield charges "by outcome, by video." A founder on INspired INsider (July 9) deliberately avoids both seat and consumption pricing in favor of ROI-based pricing tied to the work the agent completes.
  • The wedge: consumption pricing aligns your revenue with your costs, but it also punishes customers for using your product more, the very behavior you want. That tension is why the honest answer this week was "hybrid," not "pick one."

Debate 3, Is OpenAI in real financial trouble, or just repositioning?

  • Bear: On Prof G Markets (July 10), author Sebastian Mallaby held to his prediction that OpenAI could "run out of money," describing an unsustainable burn, weak consumer monetization ("like 5% of the retail consumers were actually paying," and its biggest user markets, India, Brazil, Indonesia, "are not rich consumers"), and being "squeezed between Anthropic" (better at enterprise coding, cybersecurity, agents) and Google's Gemini (better at consumer reach and ad monetization). He even questioned a headline $122 billion raise where "about two-thirds" was future promises or payment-in-kind.
  • Bull: On All-In (July 11), Brad Gerstner argued OpenAI has "got its swagger and mojo back," with revenue reportedly ticking back to "$70 billion" run-rate and GPT-6 rumored within 30 days, a plausible trillion-dollar-plus IPO candidate alongside Anthropic (rumored trending over $100 billion in revenue). The template both are following is the SpaceX IPO ($75 billion raised at $1.75 trillion).
  • The wedge: is the AI sector healthy while one badly-managed company struggles (Mallaby's view), or is OpenAI's momentum real and the doubters early?

Debate 4, Are we near an AI-spending "reckoning"? On All-In (July 11), Chamath relayed an enterprise customer whose "token costs are doubling every 45 days" while downstream productivity gains were "maybe 5% max", because "you need to use a lot more tokens to get to this next iteration of improvement" now that model quality has "effectively already asymptoted" (flattened out). His conclusion: every big company will hit this reckoning "in the next 3 or 4 years." The bull rebuttal (Munster, above) is that spending accelerating is exactly what you'd want if the growth is real. Notably, closed models are gaining enterprise share even as costs rise, one figure cited had open-source falling "from 19% to 11%" of enterprise spend.

Debate 5, Will the foundation-model labs (OpenAI, Anthropic) eat the application layer, or will startups/incumbents hold it?

  • On 20VC (July 6), USV's Mike Mignano argued the labs won't win the app layer: regulatory moats and specialization give startups durable advantages (e.g., Abridge's 8-year healthcare build), and he expects a fragmented market where the leader takes only ~30% share. Glean co-founder Arvind Jain made a similar case on 20VC (July 11): "Why OpenAI and Anthropic Won't Win the App Layer."
  • The counter-tension, voiced repeatedly, is that value is commoditizing at the model level and accruing "at the application level", which is exactly why Palantir's Alex Karp is pitching data defense (see Names).

3. Specific Names, bull/bear as articulated on the podcasts

Intuit (INTU), bearish tilt, but "priced for disruption" debate. The most-discussed public software stock of the week, down ~58% year-to-date. On Stock Club (July 9): a "genuine miss operationally", TurboTax revenue "well down," triggering "a 20% one-day drop", plus multiple securities-fraud investigations into pricing disclosures. Goldman Sachs downgraded it to sell, seeing growth slowing "from about 14%... to 5% to 10%" as AI-native competitors emerge and Claude/ChatGPT erode the "guided software" moat. Intuit responded with a "17% workforce cut and an $8 billion stock buyback." The bull sliver: "10% revenue growth, 55% EBITDA margins" haven't deteriorated, the brand (QuickBooks) is strong, and 27 of 34 analysts still rate it a buy, "thrashed... but they're buying back $8 billion." On Motley Fool Hidden Gems (July 10), the hosts leaned bearish/"value trap": existential risk that Claude "can handle all my taxes," a B2B business ("automating accounting, managing payroll... driving automated email marketing") that "sounds very, very replaceable with AI" and "gets disrupted first," plus a worry that Intuit keeps making stretch acquisitions (MailChimp, Mint) that "add goodwill... can't get the returns."

HubSpot (HUBS), bearish, self-inflicted wound. On The Information's TITV (July 8), enterprise reporter Kevin McLaughlin described HubSpot changing its terms of service to use customer CRM data for a new AI sales-lead feature, then "backtrack[ing] so forcefully just four days after" a customer backlash. He called HubSpot "one of the hardest hit companies by the SaaSpocalypse," down "75%... over the last 18 months," and structurally exposed because its small-and-medium-business customers can switch CRMs relatively easily. Some customers "were already looking to get off HubSpot" over costs and feature-gating. Bull case is thin: if it can ship "a hit AI feature" it could reverse sentiment, and it's a plausible acquisition target, but "trust is so hard to build with customers and so easy to lose."

Shopify (SHOP), bullish on the core, unresolved on the next act. On The Watson Weekly (July 6): its "strongest quarter in years", over "$100 billion of GMV in a single quarter" for the first time, revenue up 34% to ~$3.2 billion, ~90% from merchants on the platform over a year. ("GMV" = gross merchandise value, the total dollars of goods sold; "take rate" = the cut Shopify keeps.) The bull: ShopPay, its one-tap checkout, is "growing more than 60% a year" and runs "close to 40% of eligible checkout"; merchant-solutions revenue up 39%. The bear/risk: "agentic commerce... could quietly kill the thing Shopify is known best for, the storefront", if buying moves inside ChatGPT or Google's AI, "the merchant store could go dark." Shopify's answer (a catalog/indexing business with Google, opened even to non-Shopify merchants) is "nowhere near proven." Also flagged: rising "transaction and loan losses" at Shopify Capital.

Figma, bearish data point. On Big Digital Energy (July 9), Palantir's Alex Karp reportedly cited Figma as a cautionary tale, an Anthropic employee on Figma's board resigning before "Claude Design" launched, "causing Figma's stock to drop 80%." Separately, designer Patricia Reiners on Future of UX (July 8) described Figma sliding "from the center of design practice to a backup tool," with one client rebuilding an entire "700-component" design system as code "in one day."

Palantir (PLTR), bull thesis is "data defense." Also on Big Digital Energy (July 9): Karp's pitch is that the foundation-model companies are "stealing enterprise data and competitive advantage" (his Cursor example: it competed with Claude Code "after being Claude's biggest customer"), so enterprises "need Palantir's data security to defend against this threat", the hosts summarized it as the Jack Nicholson "you need me on that wall" argument. The same hosts noted the honest tension: value "is at that application level" and "the foundational models are getting pretty commoditized." Karp's broader "steamroll SaaS" framing was also revisited on The Artificial Intelligence Show (July 7, #224).

Salesforce (CRM) & ServiceNow (NOW), the perennial "first on the chopping block" names. On TITV (July 6), the discussion kept returning to Salesforce because Marc Benioff "pioneered SaaS," and small companies buy its CRM (easier to switch than enterprise). But the takeaway was measured: big customers are "largely not replacing their biggest software providers," and the smarter play is AI on top of these systems as a "system of record." Salesforce's own "Headless 360 / the UI is dead" launch signals it's trying to lead the shift rather than be disrupted by it. ServiceNow appeared mainly as a system enterprises (like Sanofi) are trying to move workloads out of.

Cursor (Anysphere, private), the week's hottest infra name. On Revenue Builders (July 5), Cursor's President of Global Revenue Brian McCarthy (ex-Rubrik CRO) explained its edge: Cursor is "agnostic to what model you use", you can run Claude, OpenAI's Codex, or Cursor's own "Composer" model inside one "harness", positioning it as the "picks and shovels" of software development (his Snowflake/Databricks analogy). Cursor grew via 3 million self-serve engineers with "very little competition... because it was done at the task level." The mega-story: on Limitless (July 10), Cursor was acquired by Elon Musk's SpaceX AI for a reported "$60 billion" as a "data moat" that helped train the coding-focused Grok 4.5 (said to be "17 times cheaper than Opus 4.8 for coding").

Snowflake (SNOW), internal-adoption bull signal. On The Engineering Leadership Podcast (July 7), Snowflake VP of Engineering Vivek Raghunathan said "95% of his 2,500-person engineering team actively uses coding agents weekly," calling AI-tool adoption "the highest-leverage organizational priority." A read-through for how fast even infra incumbents are internalizing AI to do more with smaller teams.

Rubrik (RBRK), repositioning into agent security. On Eye On A.I. (July 7), CEO Bipul Sinha's colleague Devvret Rishi positioned "Rubrik Agent Cloud" as a governance/security layer over AI agents, arguing the real enterprise bottleneck "is not model capability but managing agent risk." Rubrik acquired Rishi's prior startup Predibase to combine data/identity security with AI platform capabilities, targeting Global 2000 enterprises.

Neo4j (private), infrastructure bull. On Invisible Machines (July 9), President/CPO Sudhir Hasbe argued graph databases are "essential infrastructure" for AI agents (structured, relational knowledge), positioning Neo4j against vector databases for agentic decision-making.

Vercel (private) & AWS, the AI-app plumbing. On the AWS for Software Companies Podcast (July 7), Vercel and AWS detailed "self-driving infrastructure" cutting Aurora serverless database creation "from minutes to seconds," letting developers ship production apps in hours via Vercel's v0 AI builder, a concrete example of AI collapsing the cost and time of building software.

Amazon / AWS (AMZN), capex and cloud reacceleration. Via The Watson Weekly (July 6): AWS "up 28%, its fastest in 15 quarters," but CapEx of "$44.2 billion this quarter, almost double the year before," with training capacity committed to OpenAI and Anthropic, and trailing free cash flow collapsing "to $1.2 billion" from "nearly $26" billion a year earlier, the AI buildout showing up in cash flow even as the income statement holds. Advertising (up 24% to over $17 billion in the quarter) remains the high-margin bright spot.

Meta (META), new cloud competitor. On The Rundown (July 11), Meta is reportedly entering cloud computing to rent out excess AI capacity, competing with CoreWeave/Nebius and undercutting the labs on model prices, giving Wall Street a clearer revenue path from its infrastructure spend.

CoStar Group (CSGP), bearish, cash-burn story (non-software adjacent, flagged for context). On Stock Club (July 9): down 55%, dragged by Homes.com, whose path to profitability was pushed out "to 2030" with "$550 million net investment in 2026 alone," plus an activist proxy fight and Nasdaq-100 removal forcing index selling.

Palo Alto Networks (PANW), CEO on AI pricing. CEO Nikesh Arora appeared on Squawk on the Street (July 9) alongside Sam Altman, discussing AI pricing and enterprise adoption (and separately noting "90% of employees aren't ready" for AI, favoring hackathon-style training over mass layoffs). No fresh bull/bear investment thesis was articulated beyond his commentary.