Newsletter · · Ashutosh Agarwal

Anthropic Doubles Inference Margin to 70 Percent as the AI Token Bill Hits SaaS Vendors - Is SaaS Broken? - Week of June 26, 2026

Enterprise software newsletter for the week of June 26, 2026. Anthropic reportedly lifted inference gross margins from 38% to over 70%, signaling AI-layer COGS can re-rate, even as Uber, Meta and Microsoft capping token use and a SaaS CEO putting numbers on a multiple de-rating showed the bill for using that intelligence came due.

Is SaaS Broken?

Week of June 26, 2026: Anthropic Doubles Inference Margin to 70 Percent as the AI Token Bill Hits SaaS Vendors

Two Fridays ago the bear case got its number: OpenAI's leaked 39% gross margin, "proof" the AI layer is half as profitable as software. This week, the rebuttal, from the same place. Anthropic reportedly took inference margins from 38% to over 70%. Inference can re-rate; so can the cost line under every SaaS vendor's AI feature. The catch: the bill for everyone using that intelligence came due the same week, and it is ugly.

TL;DR

  • Anthropic's inference gross margin reportedly jumped from 38% to over 70%, the cleanest sign yet that AI-layer COGS can climb toward SaaS levels. The "52% vs 75-85%" floor just got less scary.
  • The AI bill came due, loudly. Uber blew its entire 2026 AI budget in Q1; Meta and even Microsoft are capping engineers' token use, token COGS is now an earnings-call complaint, not a theory.
  • A SaaS CEO put numbers on the de-rating: public SaaS multiples more than halved, now below the S&P 500. Workday -54%, Monday and Klaviyo -50%+. "AI features won't save you."

What's new

1. Anthropic flipped the margin script. On Elon Musk Podcast, 6/21 [Analyst, citing the Ramp AI Index], Anthropic "dropped their compute spending from 71 cents per revenue dollar down to 56 cents... inference gross margins surged from 38% to over 70%," on a "$47 billion" run rate, via prompt caching and a shift to its priciest tier. Translation: the AI-COGS floor under SaaS isn't fixed at 52%, it's an engineering problem.

2. The token bill came due, with names attached. On RiskReversal, 6/22, CNBC's Deirdre Bosa [Analyst] said Wall Street is "saying on earnings calls token costs are eating into our margins." Danny Moses [Analyst] had receipts: Uber "blew through their entire 2026 budget" in Q1, and Microsoft, "a 25% stake in OpenAI," was among the first "putting limits on their own engineers using tokens." Tech Brew, 6/24 [Analyst, citing The Information] added Ensemble Health saving "$700,000 per year" on a model "1/20th as expensive." When usage is the unit, customers fight the bill.

3. A SaaS CEO quantified the carnage. On NC Tweener Talks, 6/25, Levitate founder/CEO Jesse Lipson [Operator] said "public SaaS multiples have been more than cut in half... trading below the overall S&P 500," naming "Workday still down 54 percent... Salesforce, HubSpot... Klaviyo and Monday... down 50 plus percent." His warning hits our seven: "you should not be proud that you are building a lot of AI features... they're getting crushed." His "removal test": pull the AI out, does the product still work? Bolt-on AI doesn't.

4. Open source is now eating the lucrative customer. On Everyday AI #804, 6/23 [Analyst], one engineer reported "spending $300 a day on Claude, switched to GLM, spent $3.82 today"; China's GLM-5.2 now ranks top-three, ahead of Google. On Everyday AI #803, 6/22, Microsoft is eyeing DeepSeek V4 at "87 cents per million tokens compared to Anthropic's $50" and moving Copilot CoWork "to a usage based pricing model"; four of five top OpenRouter models in early June were Chinese (Patrick Boyle, 6/21) [Analyst]. Per 20VC, 6/25 [Analyst], open source targets the labs' "lucrative" 40-70%-margin enterprise customer.

5. Operators say consumption pricing isn't optional. On Run the Numbers, 6/22, Confluent's CFO [Operator]: "every AI-native business has some form of usage-based pricing... Otherwise you're probably going to lose money," a "one-way door." Monday.com's CPTO Daniel Leria [Operator] echoed it on AWS, 6/22: hybrid seat + AI-credit pricing, "we want to be the ones doing the work for them with AI."

The debate

Bear, per-seat SaaS is structurally broken. The de-rating already happened (Lipson: multiples below the S&P), AI features buy nothing, and intelligence routes to the cheapest model. Ed Zitron [Analyst] went further on Monetary Matters, 6/21: even the headline AI margins are fiction, "to consider their gross margin without sales and marketing is just kind of laughable." And on TECHtonic, 6/26 [Analyst], the kicker: some firms find "it's cheaper to put a junior coder on" a task than the AI. So much for one agent replacing five seats.

Bull, incumbents own the data and the plumbing. Anthropic's 38% to 70% says COGS is solvable; if intelligence commoditizes, value flows to whoever owns the workflow and the data. Even Lipson concedes "CRMs are somewhat insulated... you're the source of truth." Pegasystems CEO Alan Treffler [Operator] (Tech Talks, 6/19) showed the escape hatch, amortize reasoning into reusable "recipes" so "we don't have to charge for tokens." Swing factor unchanged: does consumption revenue outrun seat bleed? This week the bear had the data; the bull had the trajectory.

Stocks in play

  • ADBE, Bull: deep switching costs, cheap and buyback-heavy near multi-year lows (Investor's Podcast, 6/21) [Analyst]. Bear: AI-disruption fear plus CEO-and-CFO turnover mid-pivot. Catalyst: next print under new-CFO margin framing.
  • CRM, Bull: the "source of truth," ~80% share; bought AI-support vendor FIN for $3.6B, partly to cut its own lab bills (Tech Brew, 6/24). Bear: named in the down-50%+ cohort (Lipson) [Operator]. Catalyst: Agentforce consumption-ARR.
  • DDOG, Bull: already consumption-priced, so volume is a tailwind. Bear: if customers cap AI usage, metered revenue slows too. Catalyst: next NRR. (No coverage.)
  • TEAM, Bull: Rovo rides agents on a sticky cloud base. Bear: still seat-priced into seat erosion. Catalyst: Rovo consumption metrics. (No coverage.)
  • HUBS, Bull: SMB "source of truth" gets Lipson's CRM-insulation logic. Bear: most seat-exposed, named in the de-rating cohort (Lipson) [Operator]. Catalyst: Breeze attach and NRR.
  • ASAN, Bull: AI Studio agents as teammates. Bear: smallest and most seat-dependent. Catalyst: AI Studio monetization. (No coverage.)
  • MNDY, Bull: furthest along, 250,000 customers, ~$1.2B revenue, hybrid seat + AI-credit pricing built to "do the work for you" (CPTO Daniel Leria, AWS, 6/22) [Operator]. Bear: SMB seat base, named down 50%+ (Lipson). Catalyst: first AI-credit consumption disclosure.

Read-throughs

  • Seat-heavy SMB SaaS (HUBS, ASAN, MNDY) already took the multiple hit; the question is NRR now. Lipson's "removal test" is the screen, bolt-on AI gets 3-6x revenue at best.
  • Model/inference vendors got a bull surprise (Anthropic's 70%) and a bear one (open source targeting their best customer); a declared OpenAI price war caps lab pricing power, good for the application layer that resells tokens.
  • Multiple de-rating splits by proof: AI-native and consumption-durable names get the benefit of the doubt; bolt-on seat-counters trade below the S&P until they print otherwise.

What changed vs last week

Last week's hard number (OpenAI's 39%) got its mirror, Anthropic's 38% to 70%; the inference-margin debate now cuts both ways. The "SaaSpocalypse" stopped being a pundit's word: a SaaS CEO put named comps on it, and the token bill turned concrete with Uber, Meta and Microsoft capping usage. Monday.com got its first direct operator airtime in three weeks; Adobe and Salesforce drew real discussion (CRM's $3.6B FIN deal is new). Still dark: Datadog, Atlassian, Asana. Still missing two weeks running: an AI-feature gross margin from an in-scope vendor, and a hard NRR print from any of the seven.