Newsletter · 2026-10-02 · Ashutosh Agarwal

Atlassian Says AI Agents Grow Seats as Anthropic's IPO Filing Leaks - Is SaaS Broken? - Week of October 2, 2026

Is SaaS Broken? for the week of October 2, 2026. Podcast synthesis on whether AI agents erode seat-based software: Atlassian's CEO says AI-heavy customers grow seats at least 5% faster and ARR at twice the rate, finance chiefs at Zuora and SAP shift budgets from seats to tokens, Microsoft defends a seat-plus-usage model, the AI gross-margin gap (a projected 52% versus mid-70s for traditional software), and Anthropic's leaked S-1 showing roughly $5 billion of 2025 revenue against an $8 billion operating loss.

Is SaaS Broken?

Week of October 2, 2026: Atlassian Says AI Agents Grow Seats as Anthropic's IPO Filing Leaks

TL;DR

• The strongest "seats are fine" evidence yet came from a CEO inside our coverage. Atlassian's Mike Cannon-Brookes said customers who plug AI agents into Atlassian grow their seats at least 5% faster and their annual recurring revenue at twice the rate of customers who don't, and that the company grew "30-odd percent last quarter on a $7 billion run rate." That is the opposite of "one agent replaces five seats," at least for now.
• But the people who buy software are moving their AI budgets from seats to tokens, and their AI bills are growing faster than the payoff. Zuora's finance chief said nearly all of his own AI tools have moved off per-seat pricing, and that next year's budget will have "a people budget and... a token budget." SAP's CEO said AI made some staff 20% more productive while costs went up 30%. Microsoft is splitting the difference: seats for the base, usage for the premium stuff.
• Anthropic's IPO filing leaked, and it is a cautionary tale for anyone paying model bills. Per Reuters' copy, discussed on several podcasts: roughly $4.6–5 billion of 2025 revenue, an operating loss of about $8 billion, $518 billion of compute commitments, and 47% of sales routed through Amazon and Google. Still no gross margin figure in what we heard. Meanwhile a widely cited industry benchmark puts AI-company gross margins at a projected 52% this year versus mid-70s for traditional software.

Quick vocabulary note: gross margin is what is left of each sales dollar after the direct cost of delivering the product (for AI features, mostly the computing bill). Traditional software keeps roughly 75–85 cents. A token is the unit AI models charge by, roughly three-quarters of a word. Seat-based pricing means you pay per user; consumption (usage-based) pricing means you pay for what you use. ARR is annual recurring revenue. NRR (net revenue retention) measures how much more, or less, existing customers spend a year later. MCP servers and CLIs are the plumbing that lets AI agents read and act inside an application without a human clicking around the screen. A "headless" world is one where agents use software only through that plumbing and nobody looks at the app itself.

What's new

Ranked by what matters most for positioning, most actionable first.

1. Atlassian's CEO: AI agents make customers buy more seats, not fewer

This is the most important datapoint of the week for anyone holding seat-priced software, because it comes from the CEO of one of our seven names, on the record, with numbers.

On Decoder with Nilay Patel, "The SaaSpocalypse that wasn't, with Atlassian's CEO" (Sept 28), Nilay Patel put the bear case to Mike Cannon-Brookes directly: if an AI coding agent can read your data straight out of the database, why does anyone need Atlassian's interface, and isn't it "like $200 a month and you can just token max your way into the future"?

Cannon-Brookes's answer, in full, because the numbers are the point:

"We have one of the most used MCP servers in the world... It is massively used on a daily basis by millions of people... However, and people who use our MCP server and our CLI, they actually grow faster. They create more issues in JIRA. They grow their seats faster. We've said this publicly on earnings calls. I think it's like at least 5% faster. They grow their ARR at twice the rate of customers who don't use our CLI and MCP servers."

In plain English: the customers who are most aggressive about letting AI agents into Atlassian are the ones adding seats fastest and spending more each year. He added that "more than 98% of people using our MCP server also use the user interface," which is a direct rebuttal of the "headless" idea. His one-liner on that: "headless is brainless... It is submission in slow motion if you're a software vendor."

He also gave a growth number: "We grew 30-odd percent last quarter on a $7 billion run rate, which is the fastest we've grown in about two years." And he made the long-run argument against seat erosion: "I would argue more knowledge workers, more developers in five to 10 years time than today, because the ability to do tasks to compete will go up." His reasoning: if a frontier model can run your entire business for you, "you have a relatively simple business... Most businesses aren't simple."

The caveats a careful reader should hold onto:

• The "5% faster seat growth" and "twice the ARR growth" are a comparison between AI-heavy customers and everyone else. That could simply mean Atlassian's most engaged, fastest-growing customers are also the ones that adopt AI first. It is evidence, not proof, that AI causes seat growth.
• Patel reminded listeners that Atlassian "laid off about 10% of the company" in March. Cannon-Brookes said that was about changing the mix of skills, not replacing people with AI, and that "our AI products and services specifically, and our enterprise sales areas are growing very, very fast." Put simply, Atlassian is cutting its own headcount while arguing its customers will grow theirs.
• He still gave no Rovo (Atlassian's AI assistant) revenue figure and no AI gross margin.

Why it moves numbers: For TEAM, this is a direct challenge to the bear thesis that AI agents shrink seat counts. If the twice-the-ARR-growth pattern holds across the base, it supports retention and expansion, which is exactly what investors have been unable to verify for 15 weeks because nobody reports NRR on podcasts. It also gives HubSpot, Asana and Monday.com a template argument, though they have not made it themselves.

2. The buyers are switching from seats to tokens, and the bill is outrunning the payoff

Three finance and operating leaders described the same shift from the customer's side of the table this week. Together they are the clearest picture we have had of how AI is changing the way companies pay for software.

Zuora's finance chief says his own AI tools are leaving seats. On Run the Numbers, "Zuora's COFO On Going Private and Pricing AI | Todd McElhatton" (Oct 1), Todd McElhatton, Zuora's chief operating and financial officer, said:

"We've gone this year from a standpoint of a lot of our original AI tools were seat-based. And everyone is moving away from those. We've got one that's got two years left on it. So we're kind of happy about that. But everything else is on a token."

Notice the "kind of happy" about a seat contract with two years left. A software buyer is treating seat-priced AI tools as something to escape.

He went further on budgets, which is the real seat-erosion mechanism: "maybe I had some attrition to organization, and I don't want to back those. I'd much rather have it in token costs. And that's one of the things we move to in the upcoming budget cycle is you'll have a people budget and you'll have a token budget." And: "if a leader tells me they can deliver better with more tokens than people, we'll do more tokens." On his collections team, he said AI means "we could probably double the size of our business without having to make any changes in the size of that group."

Every employee not replaced is a seat not bought. That is the bear case in one budget memo.

He also gave the cleanest description yet of the gross margin trap, as a worry he hears from fellow finance chiefs: "something that was a product that had, you know, a nice 80, 85% gross margin? And I added some function and feature or function and features to it. And now all of a sudden, you know, I've tanked my margins." And a striking anecdote about how lean AI-native companies are: one customer, which he described as AI-native, has "hit $40 billion of ARR and they've got 2,500 people and they have less than 100 people in their finance function. I mean, it's maybe 50 or 60." (He did not name it, and we won't guess.)

SAP's CEO says the token bill is growing faster than productivity. On Big Technology Podcast, "SAP CEO: AI Won't Kill Software, But It Will Change Your Job, With Christian Klein" (Sept 30), Christian Klein was blunt about SAP's own internal use: "the overall token spend is up, which is good, but... it doesn't help you if some of your employees is getting 20% more productive, if at the same time, the cost is going 30% more up." SAP has put token limits on developers, product managers, designers, finance and HR staff, and now requires that every agent it ships "needs to be tested with different models," including open-source ones, to keep costs down.

Klein also made the incumbent's case: "If you just use an LLM alone, there's no way that you can run a warehouse with it, that you can do a financial close with it." His example: a financial-close agent that is right 93% of the time is useless when auditors need 100%. Host Alex Kantrowitz noted that SAP's stock is "down 21%" this year "but it's been up 43% over the past two months."

Pegasystems deliberately refused to charge by the token. On CFO Thought Leader, "1218: Why Pegasystems Is Betting on Work, Not Tokens | Ken Stillwell, CFO, Pegasystems" (Sept 30), CFO Ken Stillwell argued the opposite of Zuora's buying habits: Pega controls AI costs for clients "by not charging them based on tokens," steering them to the cheapest model that does the job. He compared token-maxing to seeing "which household in the neighborhood can use the most electricity in the summer." Two figures worth noting: Pega's cloud gross margin went "from 30 to 45 to 50 to 55" on its way to roughly 80% today, which is a reminder that a low-margin new product can grow into software-like margins. And he made an uncomfortable point for the "tokens keep getting cheaper" crowd: "some of the frontier token costs are actually more now than what some of the earlier models were." He also said plainly: "I think ROI has been elusive around AI."

Seat budgets are turning into metered budgets. On Everyday AI Podcast, "Ep 872: AI Cost Control 101: Why Your Chatbot Bill Is Becoming a Board-Level Problem" (Sept 29), host Jordan Wilson captured the mood shift in corporate IT: "it used to be a couple of months ago, right, the message being pushed down was let's increase our seats, right? Hey, we're paying for, you know, 5,000 seats, but only 2,500 are being used... And now that has shifted because a lot of those seats... now have meter billing maybe on top of a very limited, you know, monthly or weekly quota." He listed the vendors doing it: Microsoft's GitHub Copilot "replaced their premium requests with token-based AI credits," Microsoft is moving Copilot Cowork to "task-level credit pricing," and Anthropic's top model, Fable 5, goes API-only at "$10 per million input tokens and $50 per million output tokens." His conclusion: "We are seeing actual real spend limits on the amount of AI that these employees can use."

Why it moves numbers: This is the seat-erosion thesis described from inside real budgets. If finance chiefs literally put "people" and "tokens" in competing budget lines, every software vendor that charges per person is competing for a shrinking line. The offset is that the same buyers are capping token spend too, which limits how fast consumption revenue can replace seat revenue. Neither side of that trade is easy for seat-heavy names like HubSpot, Asana and Monday.com.

3. Microsoft's answer: keep the seat, put the cheaper AI inside it, and charge usage for the premium

On Sources with Alex Heath, "Satya Nadella on Microsoft's agent bet and AI's trust problem" (Sept 25), host Alex Heath noted that "token cost has fallen roughly half every quarter since 2023" and that "we're maybe past the token maxing moment," then asked how pricing should work. Satya Nadella laid out the hybrid model:

"The approach we've taken in commercial segments is to have a combination of what I'll call seat-based pricing and usage-based pricing... seat-based pricing, it just is a much more convenient way for any customer to be able to budget and buy without surprises. At the fundamental level, seats are fixed price."

The interesting part is how he plans to use falling model prices: "the models are dropping in price, means every day I can add more value to the subscription... we can pass through the advances in AI and the drops in prices and keep adding more and more value to essentially their membership." Anything cutting-edge is usage-priced, but "over time, even what is today usage-based will be tomorrow in the subscription."

That is a specific answer to the gross-margin question. Microsoft's plan is to keep the seat price fixed and let cheaper models widen the margin inside it, rather than pass savings on as price cuts. It is also a quiet threat: if Microsoft keeps stuffing more AI into the seat customers already pay for, standalone seat-priced tools have to justify their own line item.

The upsell is real, at least by one estimate. On Bloomberg Intelligence, "Muse Powers Meta to Cusp of $2 Trillion in Best Month Since 2013" (Sept 25), Bloomberg Intelligence's Anurag Rana said Microsoft's new E7 suite, which bundles Copilot and agent features, is "priced almost 75% more than the last plan, which is called the E5," and that paid Copilot seats are "something around 30 million right now" (his words: "if my memory goes right") out of "about 300-plus million users." He expects "massive double-digit growth in Microsoft Office revenue over the next couple of years, even when user growth is not there." His take on who keeps the money: "the model companies are not going to take all the economics here."

Why it moves numbers: The biggest software seller in the world is betting on seats plus usage, not usage alone. That supports Salesforce's six-part pricing menu from last week and the hybrid that AI-native CRM founders described two weeks ago. It also shows what pricing power looks like: a 75% price increase on the top suite while the user count stays flat. Whether ADBE, HUBS, ASAN or MNDY can do the same is the question. Only about 30 million out of 300-plus million Microsoft users have taken the upgrade so far, about one in ten.

4. The AI margin problem, in hard numbers

We have tracked the "~52% AI gross margin" question for 15 weeks without a single in-scope company disclosing one. This week a podcast finally put sourced figures on the table, plus a founder who explained exactly where the margin goes.

The benchmark. On Topline, "If the AI Money Dries Up, Which Companies Burn?" (Sept 27), host Sam Jacobs of Pavilion cited research from Iconic: "scaling AI companies averaged roughly 41% gross margins in 2024, 45% in 2025, and a projected 52% in 2026... At the application layer, margins are even lower, 33%, 38%, and a projected 45%." For comparison, "the median SaaS company generated gross margins in the mid-70s while the best exceeded 85%." (That is where our ~52% figure comes from.) Two things stand out. AI margins are improving by about 4–7 points a year, and the companies that build on top of the models are worse off than the average.

The horror story: Jacobs said legal AI startup Harvey "doubled its revenue while its gross margin tanked" this year. A co-host's reaction: "you're at negative 50% margins, you don't have a business." Asad Zaman explained why application companies are exposed: if the model makers "decide to provide you a cheap model, like OpenAI is using pricing to capture market share, you're in a good spot... And if they feel like they don't need to do that in that particular moment, they can just like capitalize on price as much as they want. And you're screwed."

The fix, from a founder. On 20VC, "$1BN ARR in 18 Months; The Untold Story of Higgsfield..." (Sept 28), Alex Mashrabov, CEO of AI video company Higgsfield, gave the single most useful margin split we have heard: "The margin on own models and open weights models is over 80%... And for closed source models, it's probably between 20% and 30%." In other words, who owns the model decides whether an AI product has software margins or reseller margins. He also reported "NRR at month 12 is over 300%. It just never happens in B2B SaaS," and that his own staff average "over $10,000 a month" each on models, "over 4 million" a month in total. On Adobe, he was dismissive: Adobe and Canva "build the best software for the pixel first era... But that's clearly not how the world is going to work in the future."

Why it moves numbers: This is the margin bridge investors have been missing. A seat-priced incumbent that routes AI features to third-party frontier models is closer to the 20–30% end; one that routes to its own or open models can protect something close to 80%. That puts a premium on what SAP, Microsoft and Pega all described this week: model routing and switching to cheaper models. Atlassian's ad claim of "44% more accurate results with 48% less token usage" (it ran during the Nadella interview) is the same pitch. The open question for ADBE, CRM, HUBS, ASAN and MNDY is which end of that range their AI features sit at. None has said.

5. Anthropic's IPO filing leaks: huge growth, big losses, heavy concentration

Last week we said to watch for Anthropic's public filing and an audited gross margin. Half of that arrived: Reuters obtained the S-1 (the registration document a company files before an IPO), and several podcasts went through it. Note, as hosts on More or Less pointed out, "it was a leak, right? And it doesn't have up-to-date numbers."

On Pivot, "AI's Rocky Road to Wall Street, Hegseth's Macho Military, and Trump's AI Safety Theater" (Oct 2), Kara Swisher and Scott Galloway walked through it. Swisher: Anthropic "reported a net loss of $42 billion in 2025 and plans to spend $518 billion on computing infrastructure... 47% of its sales were routed through Amazon and Google." Galloway added context: "They generated about $5 billion in revenue in 2025. That's up 12x, but they lost nearly $8 billion. The headline $42 billion loss is a bit of a... big asterisk because $34 billion of it is accounting charges." He said the company wants "a valuation north of $2 trillion or roughly 435 times last year revenue," and that "one quarter of revenue came from two customers." His verdict: "It feels like a hostage note with a cap table."

On The Enterprise AI Show, "The Anthropic IPO?" (Sept 30), host Brian Grace Lee gave a slightly different set of reported figures: 2024 revenue "somewhere in the $380 to $400 million range," 2025 revenue "about $4.6 billion," an operating loss of "roughly $3 billion" in 2024 and "a little over $8 billion" in 2025, Q2 2026 revenue of "about $11.5 billion," and a claimed run rate "somewhere around $60 billion." He also flagged "about $71 billion in off balance sheet" commitments on top of the $518 billion, so "that future commitment is roughly $600 billion." (He extrapolated a 2026 operating loss of $15–20 billion. That is his own estimate, not a filed number.)

The growth worry: on More or Less, "Meta Muse vs Instinct vs OpenAI Dots: Do AI Agents Have Any Moat?" (Oct 2), one host described the online reaction: "as token prices have dropped, their revenue, shockingly, just kind of flatlines," and joked about "memes of SaaS investors welcoming Anthropic to the no growth table." Another host pushed back: "I don't think their revenue growth is zero."

And the price war continued. On All-In, "Anthropic IPO at Risk, Meta's Muse Pop, Token Prices Fall, Open Source Gains Share, Alignment Fails" (Sept 26), Jason Calacanis said Anthropic and OpenAI "both released models this week at 50% less token prices. And their IPOs are looking like they're both going to get delayed." A co-host made the point that matters most for our seven names: falling token value "will force open AI and Ant to go up the stack... they'll have to go and do cyber, they'll have to go and do law, they'll have to go and do customer support."

Why it moves numbers: Three read-throughs. (1) We still don't have Anthropic's gross margin from any podcast, so the "what does inference really cost" question stays open. (2) If token prices are halving and Anthropic's revenue is flattening, more of the AI value is flowing to whoever uses the tokens. That is good for software margins. (3) The All-In point is the real risk: the labs moving into applications like customer support puts them into direct competition with Salesforce, HubSpot and Atlassian's service products.

The debate

The bull case: seats are bending, not breaking.

• The only in-scope CEO who spoke this week said AI-using customers grow seats faster and ARR at twice the rate (Decoder, Sept 28). Atlassian is growing 30%-plus on $7 billion.
• Microsoft, the biggest seat seller in the world, chose seats plus usage over usage alone, and is getting a 75% price increase on its top suite without more users (Sources, Sept 25; Bloomberg Intelligence, Sept 25). That shows seat pricing can carry AI upsell.
• Falling model prices are a margin tailwind for anyone who controls routing: own or open models carry "over 80%" margin versus 20–30% for closed ones (20VC, Sept 28), and AI-company margins are improving about 4–7 points a year (Topline, Sept 27). Pega took a cloud product from under 30% gross margin to about 80% (CFO Thought Leader, Sept 30).
• Incumbents can still sell AI: a16z partners said ServiceNow has "more than a billion in AI ACV," and that cybersecurity and observability software "have really stood out" (The a16z Show, Sept 30). SAP's stock is up 43% in two months as investors decide "an LLM alone" won't run a business (Big Technology, Sept 30).
• And the labs themselves look shakier: Anthropic's leaked filing shows heavy losses and concentration, which weakens the "the model makers capture everything" story.

The bear case: the budget has already moved.

• The buyers are the ones changing. Zuora is happy to be leaving its last seat-priced AI contract and will set "a people budget and... a token budget" next cycle; every unbacked hire is a lost seat (Run the Numbers, Oct 1).
• Token bills are rising faster than the payoff: SAP saw 20% productivity against 30% higher costs (Big Technology, Sept 30); Pega's CFO says "ROI has been elusive" and frontier tokens can cost more than older models did (CFO Thought Leader, Sept 30). That means buyers will squeeze all software spend, not just seats.
• Atlassian's own stat may just be selection: its most engaged customers adopt AI first. And it cut 10% of staff in March, so "AI grows headcount" is not what it practised internally.
• The margin numbers are still ugly: a projected 52% for AI companies, 45% at the application layer, versus mid-70s for SaaS; Harvey fell to negative 50% (Topline, Sept 27). None of our seven has shown its AI-feature margin.
• The labs are moving into applications because selling tokens is getting harder (All-In, Sept 26), and Microsoft keeps cramming more AI into a seat customers already own (Sources, Sept 25). Both squeeze standalone seat-priced tools.

Our read: For the first time this week, the evidence on seat counts for incumbents with strong platforms tilted bullish (Atlassian). But the evidence on how buyers budget tilted bearish (Zuora, SAP, Everyday AI). Those can both be true: strong platforms gain share of a seat budget that is shrinking and being cut into a token budget. That favours large platforms (MSFT, CRM, TEAM) over point tools (ASAN, MNDY), which is consistent with what we have seen for months.

Stocks in play

No company in our coverage changed financial guidance on a podcast this week. The only new operator figures from the seven came from Atlassian.

Read-throughs

Adjacent seat-heavy SaaS (HUBS, ASAN, MNDY). This week's split verdict, platforms winning but budgets moving to tokens, is worst for focused, seat-priced tools. Atlassian's argument works because it is a platform that agents plug into ("We want to be Union Square," as Cannon-Brookes put it). A smaller tool that agents can bypass, and that buyers can drop when they consolidate, has a harder time making that case. Wilson's "5,000 seats, but only 2,500 are being used" line (Everyday AI, Sept 29) is exactly the kind of seat audit that hits second-tier tools first. Mostly Technical's hosts noted that "everyone is raising prices" for AI features (Sept 29), but price increases in a capped budget can speed up consolidation.

Model and inference vendors. The price war that began last week kept going ("50% less token prices," All-In, Sept 26), but the picture is more mixed than "everything gets cheaper":

• At the very top, prices are going up: Anthropic's Fable 5 at $10/$50 per million tokens (Everyday AI, Sept 29) compares with the $4/$20 Opus 5.5 price we reported last week, and Pega's CFO says some frontier models cost more than earlier ones (CFO Thought Leader, Sept 30).
• Buyers are responding by routing: SAP to open-source models, Microsoft through "auto mode," Higgsfield to its own models. a16z partners cited "14x growth in agent token usage on OpenRouter" and a fine-tuned smaller model that was "60% cheaper" (a16z, Sept 30).
• The leaked Anthropic filing shows how much the labs depend on a few big partners: 47% of sales through Amazon and Google, and a quarter of revenue from two customers (Pivot, Oct 2).
• Net for our seven: whoever controls routing and owns the customer keeps the margin. The risk is the labs coming "up the stack" (All-In, Sept 26).
• We have no new podcast detail on AWS Bedrock, Azure OpenAI or Google Gemini enterprise pricing. The only Google item was Wilson's note that Gemini's $20 plan started enforcing stricter limits after Google's May developer conference (Everyday AI, Sept 29).

Multiple de-rating risk. One new macro input: on Bloomberg Intelligence (Sept 25), Bloomberg rates reporter Michael McKenzie described "nominal growth running in excess of 6%" and said "we've since saw the Fed raise rates for the first time in three years last week." Higher interest rates make future profits worth less today, which hits long-duration growth stocks like software hardest. On the other side, the software sell-off may already be reversing for platforms: SAP "down 21%" for the year but "up 43% over the past two months" (Big Technology, Sept 30), and a16z partners noted vertical software and observability "held up much better than the horizontal applications" (a16z, Sept 30). Our read: the de-rating risk is shifting from "AI kills software" to "rates and AI cost both squeeze growth," and it falls hardest on horizontal, seat-priced names.

What changed vs. last week

New this week

• First in-scope operator datapoint on seats vs. AI in months: TEAM's 5%-faster seat growth and 2x ARR growth for MCP/CLI users, plus 30%+ growth on a $7B run rate (Decoder). Last week TEAM was only an ad.
• Buyer-side seat-to-token shift, with names: Zuora (Run the Numbers), SAP (Big Technology), Pega (CFO Thought Leader), Everyday AI. A new theme: the "people budget vs. token budget."
• Microsoft's pricing framework on the record: Nadella's seat-plus-usage model, and passing cheaper models into the seat (Sources). E7 at ~75% above E5, ~30M paid Copilot seats (Bloomberg Intelligence; last week Dan Niles also cited 30M Copilot).
• The first sourced AI gross margin benchmark in this series: 41% → 45% → projected 52% for AI companies; 33% → 38% → 45% at the application layer; Harvey at negative 50% (Topline). Own or open models >80% vs. closed 20–30% (20VC).
• Anthropic S-1 leaked (watch item from last week): about $4.6–5B 2025 revenue, about $8B operating loss, $42B net loss incl. $34B accounting charges, $518B compute commitments plus ~$71B off-balance-sheet, 47% of sales via Amazon and Google (Pivot, Enterprise AI Show).
• Rates entered the story: Fed hike, 6%-plus nominal growth (Bloomberg Intelligence).

Gone quiet

• Salesforce: after dominating last week, no new operator commentary. Only Dan Ives's view and a price-increase mention.
• Adobe leadership story: no follow-up. Adobe appeared only as a rival's punchline.
• Ramp spend data and "AI demand cracks": no new monthly figure. The theme carried on in softer form (SAP token limits, Everyday AI spend caps, Anthropic "flatlining" chatter).
• Still dark: DDOG, ASAN, MNDY (category mentions, an ad and one anecdote). No AI-feature gross margin from any of the seven (about the 15th straight week). No reported NRR from any of the seven; Atlassian's 2x ARR growth is the closest proxy.

Airtime shift: TEAM up sharply (one full CEO interview). CRM and ADBE down. HUBS steady at anecdote level. Microsoft displaced Salesforce as the main source on pricing models.

Numbers revised vs. last week

• Anthropic run rate: last week, ">$100B annualized by year-end" expectations. This week, a ~$60B claimed run rate per the Enterprise AI Show, ~$11.5B Q2 revenue, and "flatline" talk (More or Less). Expectations are moving down.
• Anthropic IPO: last week, "slipped to November." This week, "at risk" / "may be delayed even more" (All-In).
• Token prices: last week, GPT-6 half price and Opus 5.5 40% cheaper. This week, "50% less" restated (All-In), but the top model, Fable 5, is priced at $10/$50 per million tokens (Everyday AI), which is above the $4/$20 for Opus 5.5.
• Company guidance: no changes from any of the seven.