Newsletter · · Ashutosh Agarwal
ServiceNow Blowout and the SaaSpocalypse Debate as Kimi K3 Resets AI Pricing - Weekly SaaS Software Podcast Recap - Week of July 26, 2026
Weekly SaaS / Software Podcast Recap for the window of July 19 to 26, 2026. The tape argued about the SaaSpocalypse, a cheaper Chinese model called Kimi K3, the shift from per-seat to outcome pricing, ServiceNow's blowout quarter, and how far software valuations still have to fall.
Weekly SaaS / Software Podcast Recap
Week of July 26, 2026: ServiceNow Blowout and the SaaSpocalypse Debate as Kimi K3 Resets AI Pricing
Window: July 19 to 26, 2026.
Here is what the podcasts were actually arguing about in software this week. Everything below is drawn from episodes published in the last seven days, with links so you can listen to the exact moment.
Top of Mind This Week
Two things ran through almost every software conversation this week: one fear and one number.
The fear was the "SaaSpocalypse," the idea that AI is about to hollow out traditional software companies. It came up on founder shows, investor shows, and even the CNBC desks, usually phrased as some version of "is SaaS dead?"
The number was Kimi K3, a surprisingly good and much cheaper AI model out of China. One podcast cited an estimate that its release wiped roughly $314 billion off what investors think OpenAI and Anthropic are worth, and it reframed the biggest question in the sector: if great AI is getting cheap and open, who actually keeps the profit?
Tying those two together was a quieter but very concrete shift in how software gets priced, away from charging "per person" and toward charging "per usage" or "per result." Salesforce made real news here. And the single most-discussed public stock of the week, by a wide margin, was ServiceNow, which reported a blowout quarter and became the poster child for "good software can still win in the AI era." Security deal-making (Palo Alto buying Protect AI for $700 million) and a brutal read on software stock valuations rounded out the week.
1. Dominant Themes
Theme A: The "SaaSpocalypse," is AI Killing Software?
This was the week's loudest topic, and the takes ranged from doom to defiance.
The defiant end came from Rob Walling on Startups For the Rest of Us (Ep. 842, "What is the Future of SaaS in an AI World?", July 21). His core argument: the thing that made software valuable was never the code. "The moat was never the code. It was everything you added to the code... AI has just made the obvious part the lowest-risk part of building." He flips the usual panic on its head, the real danger, he says, is not for small startups but for bloated incumbents: "The real SaaS-pocalypse is the big companies, the big incumbents. They're the ones who have raised their prices so much... they have a customer base that's probably quietly resentful about the last 3 price hikes. Those are the companies that people are now motivated to replace, and AI just lowered the cost of building that replacement." On the popular claim that "agents will replace SaaS," his answer is that an agent is just software you pay for on a subscription, still SaaS, and that agents actually need SaaS underneath them: "An agent floating in space is useless... it needs a backend to act on. So agents aren't going to kill SaaS. They need SaaS." His verdict: "The internet was not dead in 2000, and SaaS is not dead in 2026." (Note: "SaaS" = software-as-a-service, the standard model of renting software by monthly or annual subscription.)
The gloomy end came from The Synopsis ("AI SaaS Risk (again), Atlassian, Duolingo, & Uber", July 23), which built a detailed bear case on Atlassian (see the Names section for the numbers). Their framework for which software is most at risk is worth repeating: software is in trouble when "if I had a full-time human sitting next to me who could solve certain problems for me, then that would be a problem," they use Chegg (homework help) as the classic victim. Software is safer when you still want to sit with the data yourself, they use QuickBooks (accounting) as the survivor. And they make a broader point that even survivors get riskier: "The distribution of outcomes has widened... in a negative way."
The operator's version of the same anxiety came from Futureproof Founder (Ep. 407, "The SaaSpocalypse", July 21). The guest's blunt read: thin "wrapper" apps (products that just pass your request to ChatGPT and hand back the answer) are dead, "The wrappers are done." He expects "a lot of shakeout of the established SaaS providers... for the next probably 18 months, two years" on whether they can adapt. What survives, in his view, is anything hard to copy: data, compliance, and trust. He flagged a genuinely new idea here, software vendors getting formal "trust scores" and even insurance for their AI features (he referenced a standard called AIUC-1, now backed by a consortium including MITRE): "You can't vibe-code your way into" security and compliance certifications. ("Vibe coding" = building an app quickly by describing it to an AI in plain language, rather than writing the code by hand.)
Theme B: The Great Pricing Reset, Seats to Usage to Outcomes
This was the most concrete, money-on-the-table theme of the week.
The clearest explainer came from the Chat GPT Podcast ("How AI Native Startups Hit Billions Faster", July 22), citing analyst Sean Kanungo. For 20 years, software was sold "per seat," 50 employees, 50 licenses. That math breaks with AI: "If an AI agent can autonomously route IT tickets, write marketing copy, and answer customer emails, literally doing the work of five humans, why would a company pay for human seats?" The problem is that a heavy AI user "is generating 100 times the compute costs, actively eroding your profit margin with every single click." So pricing is moving to usage-based ("kind of like a utility bill") or, more radically, outcome-based: "You pay for a successfully resolved customer support ticket. If the AI hallucinates and fails to resolve the issue, you pay nothing." The same episode noted the eye-popping efficiency of AI-native companies: 55% generate over $400,000 in revenue per employee, and startups are now hitting a billion-dollar valuation in about 3.5 years, half the roughly 7 years it used to take. (It also cited a RAND stat that "80% of AI projects still fail to deliver business value," so the winners are impressive, but the failure rate is high.)
The real-world headline here was Salesforce, dissected on Insights for IT Negotiations ("What Salesforce's Pay-Per-Resolution Model Really Means", July 23). Salesforce's Agentforce customer-service agent will now charge $2 per resolution (a resolved issue) instead of $2 per conversation, and crucially, "You don't pay Salesforce anything unless an issue is resolved... any type of escalation, you don't pay for that." Analyst Adam Mansfield's verdict was unusually positive: "good job, Salesforce." His read on why they did it is the tell for the whole sector, customers are scared of open-ended consumption bills: they "see the horror stories that are out there about OpenAI and Anthropic and consumption... companies that started using ChatGPT Enterprise and all of a sudden have a bill for $4 million that they weren't expecting." Consumption pricing, he argued, was actually slowing Salesforce's sales cycles because buyers couldn't forecast what they'd owe. The catch he stressed: there's fine print. There's still a "consumption-like" element (you prepay for packs of at least 1,000 resolutions), and the definition of "resolution" has real mechanics buried in it, "even though you're outcome-based... customers should still be going down that path of what's a resolution."
ServiceNow's product chief made the same macro point on The Product Podcast (Amit Zavery, July 22): every past tech shift changed the business model, and this one is moving pricing "from subscription... to a more consumption-driven pricing. When I use it, I want to pay you."
Theme C: Kimi K3 and the "Commoditization of Intelligence"
The single biggest market-moving story of the week got a full breakdown on The Rundown ("Did China Break the AI Trade Again?", July 25). Kimi K3, from a Beijing startup called Moonshot AI (roughly 300 employees, about $20B valuation, founded by 33-year-old Yang Zhilin, a Carnegie Mellon PhD who worked at Meta and Google Brain), is a 2.8-trillion-parameter open-weight model, the largest open model released so far. Independent benchmarks rank it #3 in the world, behind only Anthropic's Fable 5 and OpenAI's GPT-5.6, and it beats both on some coding tests. The kicker is price: Moonshot charges about $15 per million output tokens, versus roughly $30 for OpenAI's top model and $50 for Anthropic's Fable 5, a 50 to 70% discount. ("Open-weight" means companies can download the model and run it on their own hardware instead of paying the maker each time; "tokens" are the units of text AI processes, and the standard unit AI usage is billed in.)
Why it matters for software: enterprises can now run most of their work on a cheap model and save the expensive frontier models for only the hardest tasks. This is already happening: "DoorDash said that they use the Moonshot models for lower-level work, and they use Anthropic's models for their hardest cutting-edge tasks. Airbnb is also doing the same thing. Their customer service agents run mostly on Alibaba's Qwen model," and "even Microsoft is reportedly testing whether Kimi K3 could eventually power some of their features inside Copilot." One analyst estimated about $314B was shaved off OpenAI and Anthropic valuation estimates the day after Kimi dropped. The optimistic counter in the same episode: unlike the DeepSeek scare of early 2025 (which briefly cost Nvidia about $600B in market value on fears AI would need less hardware), Kimi is a huge model needing about 1.4 terabytes of memory to run, so via Jevons' Paradox (cheaper resources get used far more), it could actually be good for chip and memory makers and for software companies that can now afford to build AI into everything.
Theme D: Open-Source Models Quietly Taking Over Token Volume, and the "Router" as the New Control Point
Two of the week's best conversations were about how fast open models are eating share, and how enterprises are managing the switch.
Matan Grinberg, CEO of Factory, on Sequoia's Training Data ("The Coming 'Dark Factory' Where Software Builds Itself", July 21) gave the most vivid numbers. Open models went from "less than 1% of tokens" at the start of the year, to "a single-digit percent" in Q1, to "now crossed into being a double-digit percent." Internally at Factory, "half of our tokens are open" and he called the open model GLM 5.2 "incredible." His key insight: open models are usually about one generation behind, but "the question is, are the open models getting as good as frontier minus one? And the answer is unequivocally yes." He also coined the week's best description of AI waste, "token maxing," where companies push staff to use AI for everything: "there are banks that we are working with where they are spending literally hundreds of thousands of dollars a month on people asking things like... what is the weather." Factory's pitch is a "router" that automatically sends each task to the right model (cheap models for trivial work, frontier models for critical work), and, echoing the whole week, enterprises fear lock-in: "We cannot put our fate in any one of these model providers' hands."
Lin Qiao, CEO of Fireworks, on 20VC (July 20) gave the macro version. Her forecast: "10x cost reduction in the next three years. And this 10x cost reduction will drive 100x usage." And her thesis for why this doesn't collapse into one giant AI: "The future is not one AGI; it's millions of specialized models," companies tuning open models on their own data. She was refreshingly candid that this is a lower-margin business than classic software: inference margins "traditionally sit in the 30 to 40% range," versus the 80% gross margins investors expect from SaaS, though she framed that as a choice to prioritize hyper-growth over optimization. Fireworks itself has hit $1 billion in revenue in four years and processes about 40 trillion tokens a day, up from 15 trillion, and just hired George Hu, former president of Salesforce.
The private-equity operator's view came from Vista Equity Partners' Monti Saroya on Alt Goes Mainstream (July 22), with a concrete cost story. Vista built an AI "first notice of loss" feature into its insurance software (Duck Creek), ran it on a frontier model, and "the cost of hosting it was $11, $12 million for a small customer... it wasn't tenable." The fix: switch to open-source models running on cheaper SambaNova hardware. His broader bet: "everything goes open-source over some period of time... over 20 years." His analogy: open models are like Linux beating Unix. And his pointed line on why you don't need the smartest model for routine work: "Do you want geniuses running your claims process? ... Geniuses are expensive at the end of the day."
Theme E: Where Does the Money Actually Land, Infrastructure vs. Applications?
On the news-roundup 20VC (Harry Stebbings with Rory O'Driscoll of Scale Venture Partners and Jason Lemkin of SaaStr, July 23), Rory laid out a stark map of the AI economy in three buckets: the infrastructure layer is about $800 to 900 billion a year in spend; the two big foundation-model companies together are about $100 billion; and "rounding up every other apps company, you struggle to make 40 or 50 [billion]." His summary: "All the good investments sure seem to be in the infrastructure," and the application layer "is almost a rounding error." He noted Cursor alone (about $4B) is most of the app layer, and that spending on training data, Mercor (about $2B ARR), Surge (about $3B), rivals the entire app ecosystem outside Cursor. Jason Lemkin pushed back that the app revenue is there in the leaders and that "everyone wants your own model" once you reach scale. (ARR = annual recurring revenue; ACV = annual contract value.) The same episode noted Databricks is raising a $3 billion "Series M" at a $188 billion valuation, and both flagged the mega-story they almost forgot, Stripe buying PayPal.
Theme F: Software Stock Valuations, the Air is Coming Out
The bluntest valuation take came from Dan Rasmussen of Verdad on Planet MicroCap ("Private Equity Unwind", July 25). On the private-equity software rollups: "I would be willing to buy your entire software portfolio at 5 to 6 times EBITDA. But the problem is you paid 25 times EBITDA. So... you have 10 times EBITDA of debt." He'd stretch to "7 or 8 times because there's no CAPEX," but the gap between that and where PE marked these deals is enormous, "it's just fascinating to think about how far those private equity portfolios need to get marked down." (EBITDA = earnings before interest, taxes, depreciation and amortization, a rough proxy for cash profit; the multiple is the price paid per dollar of that profit.) He also, for the first time, called the semiconductor market "something that looks like a bubble... it is cyclical, and it will go back down."
The public-market echo came on Squawk on the Street (July 23): the five-year rolling return for software (the IGV software index) is "the worst since 2013," with many names down about 35%.
Theme G: Security Software Consolidation
Ed Sim of Bold Start Ventures on Resilient Cyber ("Why AI Security Is Getting Rebuilt From Scratch", July 23) walked through Palo Alto Networks' about $700 million acquisition of Protect AI, which he called "the first exit in AI security history." His read on deal timing: "the first one is usually the highest one," and Protect AI's exit "started a wave of 12 or 15 acquisitions." His candid industry take: "The world needs more cybersecurity. But we don't need all the cybersecurity companies that we have right now." He also floated a striking framing of the model labs as a competitive threat to services and security firms: "the biggest heist ever happening right now is that OpenAI and Anthropic created their own forward-deployed engineering companies."
2. Key Debates
Debate 1: Is SaaS dead? Bear side: AI erodes the moat, the user interface migrates to an "AI agent layer" that sits above your software, and bloated incumbents face an 18 to 24 month "adapt or die" shakeout (The Synopsis; Futureproof Founder). Bull side: SaaS isn't dead, agents need software as the underlying "system of record," code was never the real moat, and the pain lands on overpriced incumbents, not lean players (Startups For the Rest of Us). The wedge: Do you believe AI mostly replaces the software workflow, or mostly sits on top of it and increases demand for good underlying systems?
Debate 2: Per-seat vs. consumption vs. outcome pricing. Consumption camp: the default for AI-native products, because a heavy user burns real compute (Chat GPT Podcast). Outcome camp: Salesforce's $2-per-resolution model is the customer-friendly answer to surprise consumption bills (Insights for IT Negotiations). The wedge: Outcome pricing removes buyer fear but requires airtight definitions of what counts as a "result," and there's fine print.
Debate 3: Are OpenAI and Anthropic overvalued now that open models are "good enough"? Bear side: Lin Qiao argues most enterprise workflows can run on cheaper open or tuned models, so frontier usage "will not be as large as it was if it was needed for everything," and Kimi K3 is "a direct attack on that pricing power" (20VC/Fireworks; The Rundown). Bull side: on 20VC, "the only thing that matters is the OpenAI and Anthropic growth rate in [20]26 and [20]27. If you're growing 10X year on year and you have any kind of positive and improving gross margin, it just covers all the nut," and they still own the best models and the biggest ecosystems (ChatGPT is "closing in on 1 billion users"). The wedge: Does cheap open-source cap the frontier labs' revenue, or does exploding overall usage lift them anyway?
Debate 4: Infrastructure vs. application layer, where does value accrue? Infra wins now: Rory O'Driscoll, apps are "almost a rounding error" next to the roughly $800 to 900B infrastructure spend, and "I don't believe the application layer is here yet." Apps camp: Jason Lemkin, the revenue is real in the leaders, and every serious app will build its own tuned model or reasoning layer (20VC).
Debate 5: Can the big incumbents (ServiceNow, Salesforce) actually monetize agents and own the "orchestration layer"? Bull side: ServiceNow's Amit Zavery says they've been the orchestrator of business processes "for 20+ years," their AI business is now $1.5B, and, importantly, he rejects the winner-take-all view: "I think the industry thinks they're going to be one orchestrator, which is a fallacy. There are going to be multiple... orchestrated by different providers" (The Product Podcast). DA Davidson's Gil Luria backed the execution story on Closing Bell. Skeptic side: going into the print, The Exchange expected weakness, arguing enterprise spend is rotating away from "management platforms" toward cybersecurity and AI buildouts, a thesis the actual beat then rebutted.
Debate 6: Vendor lock-in, model labs vs. neutral platforms. Factory's Grinberg argues enterprises want "model independence" so no single AI lab controls their fate, but the honest follow-up ("am I now just locked into Factory?") got the honest answer: the automations and artifacts "stay in your code base" (Training Data).
3. Specific Names
Public Companies
ServiceNow (NOW), the week's clear winner. On Closing Bell, it beat on revenue ($3.99B vs. $3.93B expected) and adjusted EPS ($0.90 vs. $0.85), with the stock up about 6% ("which we're not used to saying about software"). ServiceNow's AI business "crossed that billion-dollar ACV run rate," and CPO Amit Zavery said on The Product Podcast the AI business will do $1.5 billion this year, up from an original $1 billion plan. On Squawk on the Street, CEO Bill McDermott raised the full-year guide "to $32 billion-plus" and took a direct shot at Salesforce and Benioff, saying ServiceNow is "coming after the CRM platform." Bull case (Gil Luria, DA Davidson, Buy): "This has little to do with AI... this really has to do with them executing well in their core businesses," helped by the Armis and Moveworks acquisitions; ServiceNow now joins Palantir, Datadog, and Snowflake as the rare software names with accelerating revenue growth, while most peers decelerate. Bear and watch items: pushback on seat-based pricing, and reports of customers signing two-year instead of four-year contracts; the stock notably did not rocket on the beat.
Salesforce (CRM). New Agentforce $2-per-resolution outcome pricing was the pricing story of the week (Insights for IT Negotiations). Separately, The Synopsis used Salesforce as the good example of a maturing software company, operating margins went "from 3% margins to now 20% margins" since 2022, with stock-based comp cut to "mid to high single digits."
Atlassian (TEAM), the week's detailed bear case (The Synopsis). The knock: a 24-year-old software company with a -4% operating margin, R&D at about 50% of revenue (with "customers... barely" noticing product change), stock-based comp at 25% of revenue, and customer growth that has slid from about 40% pre-COVID to "mid-teens." They argued its recent $600M+ browser acquisition is "pretty confident this is going to be a write-off." The AI risk: workflows starting in an AI assistant instead of in Jira. The pushback (from the co-host): still 80%+ gross margins, high net revenue retention (NRR, how much existing customers grow their spending year over year), and it's still winning new customers, "worse business after AI," but not a melting ice cube.
Adobe (ADBE). On Squawk on the Street, Cramer was harsh, a board that has failed to fill the CEO and CFO roles, competitive pressure from Figma and Canva, and the AI threat: "if you're with Claude, you can crush them all."
Datadog (DDOG). Named by Gil Luria as one of the few software names with accelerating growth (Closing Bell).
Snowflake (SNOW) and Palantir (PLTR). Also cited as the "direct AI exposure = accelerating growth" cohort (Closing Bell).
IBM (IBM) and Pegasystems (PEGA). Both cited as caught in the enterprise "crowding out," Gil Luria's memorable framing: "AI is taking the oxygen out of the room. It's not that we're replacing software with AI. It's that we're replacing spend on any other technology that's not AI with spend on tokens" (Closing Bell).
Palo Alto Networks (PANW). Bought Protect AI for about $700M, kicking off a wave of AI-security consolidation (Resilient Cyber).
CrowdStrike (CRWD). Cramer's aside on ServiceNow's cybersecurity push: "I would rather have CrowdStrike doing that" (Squawk on the Street).
Private Companies
Anthropic and OpenAI. The week's central characters. Both have reportedly filed to go public at trillion-dollar-plus valuations, and both are now exposed to the Kimi pricing threat on their enterprise business (The Rundown). Vista's Saroya cited Anthropic's revenue run-rate jumping "from 30 to 44 billion" in a matter of months (Alt Goes Mainstream). On Squawk, Cramer contrasted the two: "no one's saying anything bad about Anthropic because I think Anthropic is making a lot of money," while OpenAI faces "where all the money's going to come from" questions. The Elon Musk Podcast (July 23) ran with the headline that Anthropic has overtaken OpenAI as the most valuable startup.
Moonshot AI / Kimi. The disruptor of the week, see Theme C. 2.8 trillion parameters, about $20B valuation, about 300 employees, roughly 50 to 70% cheaper than the US frontier (The Rundown).
Databricks. Raising a $3B "Series M" at a $188B valuation; did about $5.4B of ARR in 2025 (20VC; Alt Goes Mainstream).
Fireworks. Inference leader, $1B ARR in four years, about 40 trillion tokens/day, fresh $1.5B round, hired ex-Salesforce president George Hu (20VC).
Cursor (Anysphere). Repeatedly named as the single biggest AI application by revenue (about $4B) and the exception to "the app layer is a rounding error" (20VC). One caution: a passing line in that same episode and on the Elon Musk Podcast claimed Cursor was "acquired by SpaceX," that surfaced only as an offhand, unverified aside and should be treated as podcast chatter, not a confirmed deal.
Mercor and Surge. Data-labeling businesses cited at about $2B and about $3B ARR respectively, a reminder of how much money flows to training data (20VC).
SambaNova. Vista-backed AI chip company; its air-cooled hardware is the basis for cheaper inference centers that reuse old telco data centers without needing net-new power (Alt Goes Mainstream).
GLM 5.2 (Zhipu) and Alibaba's Qwen. The open models actually being used in production, GLM called "incredible" by Factory (Training Data); Qwen running Airbnb's customer service (The Rundown).
Stripe / PayPal. Flagged as a "mega" deal of the week, Stripe buying PayPal, though the crew ran out of time to dig in (20VC).