Newsletter · · Ashutosh Agarwal

Enterprises Throttle Their AI Token Bills as Consumption Pricing Cuts Both Ways - Is SaaS Broken? - Week of July 11–17, 2026

Software newsletter for the week of July 11–17, 2026. Enterprises flipped from token maxing to throttling their AI usage bills, turning consumption pricing from SaaS's rescue into a fresh risk, while Salesforce and Datadog pushed back with operators and product substance.

Is SaaS Broken?

Week of July 11–17, 2026: Enterprises Throttle Their AI Token Bills as Consumption Pricing Cuts Both Ways

For most of this year the story was simple and scary: AI agents would collapse seat counts, and the software companies that survived would do so by switching to "consumption" pricing, charging you for what you use, like a utility. This week the consumption story grew a dark side. Multiple podcasts described the same whiplash: companies that spent the winter shoveling money at AI ("token maxing") spent this summer discovering the bill and yanking it back. If usage-based pricing was supposed to save software, then a wave of customers cutting their usage is not a rescue, it's a new way to miss the number.

Underneath that, three things happened at once. Salesforce sent its own executives onto podcasts to answer last week's downgrade with an actual adoption number. The chairman of OpenAI (who used to run Salesforce) said the future isn't seats or tokens, it's paying for outcomes, which should terrify anyone selling a subscription. And a respected software investor gave the cleanest "is software dead?" framework of the year. Meanwhile a new, nearly-free Chinese model landed, semiconductors fell into a bear market, and the phrase of the week on trading desks was "AI reckoning."

TL;DR

Consumption pricing cut both ways this week. Enterprises flipped from "use all the tokens you want" (December 2025–March 2026) to "rein it in" (May–July), as bills blew past budgets. One group of AI chiefs cut token costs from $6,000 a month to $600; a company told Chamath Palihapitiya its token costs were doubling every 45 days for maybe 5% more productivity. The pricing model that was meant to re-rate SaaS is also the one customers optimize against.

The pricing endgame got named, and it's not seats or tokens. Bret Taylor, OpenAI's chairman and CEO of AI startup Sierra, said flatly: "you're not paying for tokens. You're paying for business outcomes delivered." A CNBC anchor said out loud what every software holder was thinking: that "would instill fear in a lot of software companies... used to the subscription model."

Salesforce and Datadog fought their corners; the two numbers that would settle everything are still missing. Salesforce put two operators on the mic with a real Agentforce stat (one team now runs "a hundred individuals and 400 agents," generating "over a hundred million in pipe"), and Datadog's chief scientist detailed its observability AI. But for the sixth straight week, not one of the seven names disclosed an AI-feature gross margin or a fresh net-retention number. Adobe and HubSpot got marketing-only interviews; Atlassian, Asana and Monday.com were essentially silent.

What's new

Ranked by what actually matters for positioning a book this week.

1. "Token maxing" is over, and the hangover is the real story. [Analyst]

This is the week's most important shift, because it hits all seven names at once. On Everyday AI, Ep 820 (July 16), host Jordan Wilson described a clean, dated arc: "from December 2025 to... March 2026,... enterprise companies were like, yes,... go use as many tokens as you can. And then it's like, wait, these token costs are getting higher and higher... So now we've seen, you know, in the May, June, July, this whiplash of companies being like, wait, we have to start reining spend in." He was blunt about why nobody cared before: at "20 to $50 a month per seat, most employees couldn't get through that... we were playing with monopoly money until about four months ago."

The numbers behind the retrenchment are striking. On The Exchange (July 17), analyst Ray Wang said he'd just been "at a conference with a bunch of AI chief officers and they were able to get their token costs down from $6,000 a month to $600. And they know they're going to get down to $60 with more efficient coding." On All-In (July 11), Chamath Palihapitiya relayed a founder telling him "our token costs are doubling every 45 days" while "downstream productivity... maybe 5% max," because "you need to use a lot more tokens to get to this next iteration of improvement because we've effectively already asymptoted." And The Information's coverage of the UBS software conference (July 15) named the trend directly: "token optimization... becoming a dominant subject in tech circles," with enterprises "finding their compute token costs far exceeding what they budgeted for... And they're starting to throttle it back."

Why it moves numbers: "consumption pricing" only re-rates a software business upward if customers happily consume more every quarter. This week showed the opposite reflex, buyers treating AI usage like any utility bill and hunting for cuts. That is directly bearish for the usage-priced names (Datadog, and Monday.com's AI-credit model), and it undercuts the whole "seats are dying but consumption will save us" thesis. The AI Daily Brief (July 16) summed up the mood shift precisely: companies "have now actively shifted from how do we get our people to use tokens to being uncomfortable with their token cost ballooning without a clear line to revenue."

2. The pricing endgame has a name now: outcomes, not seats, not even tokens. [Operator]

On Squawk on the Street (July 16), Bret Taylor gave the single most quotable line of the week. Taylor is worth listening to twice here: he is chairman of OpenAI, CEO of the AI agent startup Sierra, and the former co-CEO of Salesforce, so he sits on top of the model layer, the application layer, and the incumbent he used to run. His pitch: "with a Sierra agent, you're not paying for tokens. You're paying for business outcomes delivered. You're paying for that appointment being booked... it's our job to worry about the tokenomics." He predicted the whole market moves this way: "I think we're going to move to a world of paying for outcomes."

The interviewer immediately drew the obvious conclusion, and it's the thesis of this newsletter in one sentence: that shift "would instill fear in a lot of software companies. If they are used to the subscription model, if we are starting to pay for outcomes, that might be a longer conversation." Taylor's other important point was defensive, and it's the incumbents' escape hatch: as models commoditize, "what do you have? You have your relationships with your customers and you have the memory of all those interactions... this compounding asset that's your moat."

Why it moves numbers: an outcome-priced world is the version of "SaaS is broken" that even a bull has to respect. If buyers get comfortable paying per booked appointment or per originated loan, the per-seat subscription, predictable, high-margin, the thing the multiple is built on, becomes the old way. But note Taylor's escape hatch is exactly what Salesforce (below) is leaning on: proprietary customer data as the moat that AI can't commoditize.

3. Salesforce answered last week's downgrade, with operators and a real number. [Operator]

Last week Salesforce took its first hard sell-side hit of the AI era (a KeyBank downgrade on "slowing adoption in Agentforce"). This week the company pushed back through its own executives. On Venture with Grace (July 11), Kris Billmaier, EVP & GM of Sales Cloud, gave the concrete adoption anecdote the bulls have been waiting for: the leader of Salesforce's own SDR team (the reps who chase sales leads) "runs roughly a hundred SDR individuals. We assigned an agent to pick up all those 75% of leads that we never work and generated... over a hundred million in pipe that would have been unseen otherwise. So now she's managing a team of a hundred individuals and 400 agents." His framing is the direct rebuttal to the seat-erosion thesis: "I don't think it's a replace. I think it's really... a partnership effectively." He also cited customer Batteries Plus spinning up "10 agents... handling tens of thousands of conversations a day in email."

On Marketing Beyond (July 15), Kevin Siminski, Salesforce's Chief Customer Officer for Agentforce Marketing, hammered "day zero" adoption, designing incentives so customers use the product "from the very first moment", and confirmed two strategic moves: Salesforce is "bringing MCP support to marketing cloud engagement... to enable essentially a headless experience," and has agreed to buy headless-CMS vendor Contentful, "scheduled to close in... Q3." He also confirmed the Anthropic investment and an "open" agent-to-agent stance ("super agents").

Why it moves numbers: this is Salesforce trying to convert "Agentforce is stalling" into "Agentforce is a partnership that expands what sellers can do." The $100M-in-pipe stat is exactly the kind of proof point that could settle the Cramer/KeyBank-vs-Benioff fight, if it shows up as reported revenue rather than an internal anecdote. Caution: it's a first-party number from a Salesforce executive on a founder-interview podcast, not an audited disclosure. Treat it as directionally encouraging, not verified. And note what's still missing: no seat-vs-consumption pricing detail, no net-retention figure.

4. "Software isn't dead. It's gotten harder." The cleanest structural frame of the year. [Analyst]

On The Official SaaStr Podcast, Ep 868 (July 15), Rory O'Driscoll of Scale Venture Partners laid out the whole debate in numbers. The macro backdrop: hyperscalers are spending "$688 billion this year" on AI, while "what's coming out the other side... about $110 billion in actual revenues, most of it, $89 billion... from the two foundation models." In plain terms, "we're spending half a trillion dollars more than we're taking in," and on his math it "takes until 2031 or 2032 before the revenue from those two companies surpasses the total CapEx", five or six years still in "invest mode," with plenty of room for a mid-journey "hiccup."

His key reframe: when people ask "is software dead?" they don't mean OpenAI and Anthropic (which are software, and very much not dead). "Sometimes they mean is old school... plain vanilla SaaS dead," and sometimes they mean "how does economic value get allocated between the foundation models and modern software companies." That, he said, is "the biggest question of them all. How does enterprise want to consume $1 trillion worth of AI over the next 5 or 6 years?", buy it all directly from Anthropic (in which case, he joked, the software industry can "go home"), or through companies that "add value on top." He also gave a vivid sense of the money at stake: to reach $1 trillion, foundation models need "15%, 17% of all knowledge worker dollars," and "well north of 25%" of software-developer pay, "for every $200,000 software [developer], they're spending $50,000 on tokens."

Why it moves numbers: this is the framework a PM should carry into every one of these seven names. The question is not "will SaaS die", it's "who captures the value: the model makers or the app companies sitting on top." The bull case for all seven rests on the same bet O'Driscoll poses: that enterprises want value-add software, not a raw model.

5. Datadog got product substance for the "infrastructure winner" call. [Operator]

Last week Datadog was crowned the archetypal AI winner on a general framework (up 88% year to date, the "MRI scan" for AI systems). This week it got real substance. On The Data Exchange (July 16), Datadog chief scientist Amit Talwalkar walked through the company's "time series foundation models", Toto, which started at "100 million parameters" with newer models "up to like 2.5 billion", built on the mountain of monitoring data Datadog already collects. A telling detail on the economics: the AI approach is actually cheaper to run than the old statistical method, because "it's actually cheaper for us to use [a] 100 million parameter zero shot model... faster at inference time than actually tuning an ARIMA model on the fly." The ambition is a "world model" that can "simulate the behavior of a distributed software system" and predict incidents by fusing "metrics with logs, traces, topology, code, and alerts." Separately, on Motley Fool's Hidden Gems (July 12), a veteran software executive listed "Snowflake, Databricks, Datadoc" [Datadog] among the data-infrastructure names "showing massive benefit... market caps are up."

Why it moves numbers: this is the picks-and-shovels bull case getting a factual backbone rather than a metaphor. Datadog isn't just monitoring AI, it's building AI models on its own data, and crucially claims its inference is cheap. The bear counter is unchanged and now sharper given item #1: Datadog is a consumption-priced business, so the token-throttling reflex sweeping enterprises is the exact risk to watch on its next print.

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

The bear case (it's breaking). The unit SaaS was built on, a human paying monthly to sit inside an app, is being pried loose from the work, and the "consumption" model meant to replace it is turning out to be something buyers cut, not grow. This week gave that its clearest evidence: the token-maxing-to-throttling whiplash, the $6,000-to-$600 cuts, costs "doubling every 45 days" for 5% more output (All-In, July 11). The model economics underneath stay ugly and got fresh confirmation: on Tom Bilyeu's Impact Theory (July 14), reacting to bear analyst Ed Zitron, the host cited OpenAI's audited financials, "OpenAI burned $20.9 billion in 2025... their costs increase linearly with their revenues. There is no proof that they can improve their margins", while Motley Fool (July 12) pegged the burn at "$3 billion a month" and noted "as many as 40 percent of enterprise AI projects could ultimately be abandoned." Bilyeu also relayed Palantir CEO Alex Karp's jab that enterprises feel they "chillax and waste... time with tokens... get no value and they [the labs] get my IP," and the sharp point that the labs "encourage waste... they don't charge on outcomes... because you can't with large language models. They're inherently hallucination prone." Add Bret Taylor's own prediction that pricing moves to outcomes (bad for subscriptions), a new nearly-free Chinese model (Kimi K3, "almost 3 trillion parameters," which Databricks CEO Ali Ghodsi warned on The Exchange "means more margin compression"), and semiconductors falling into a bear market, and you have an industry visibly re-pricing risk. Chamath's ROI probe is the punchline: strip out NVIDIA chip sales and the S&P 493's AI-driven earnings-per-share lift is "somewhere between 0 and 2%."

The bull case (it's re-rating, not dying). The steel-man is that this is ordinary disruption, and the well-run incumbents adapt. O'Driscoll's "software isn't dead, it's gotten harder" frame (SaaStr, July 15) is the anchor: someone captures that trillion dollars of value on top of the models, and it won't all be the model makers. The most direct rebuttal to seat-erosion came from Glean co-founder Arvind Jain on 20VC (July 11), who insisted "teams will get bigger not smaller." His logic: "per person productivity is going to shoot up, but so will the demands. To make the same amount of revenue, you have to produce a 10X better product." Jain's own company is going from "over a thousand people now" to "hopefully 5,000." He also punctured the commoditization panic, "90% or greater of [enterprise] use cases cannot be fully handled by... open source models", and called Anthropic's move into verticals "quite shallow... it's net new always... expanding the market." Wix founder Avishai Abrahami on 20VC (July 13) backed the stickiness point from the trenches: his own team of "professional developers" spent a week, then a stronger team two more weeks, and still couldn't rebuild a hairdresser's business logic in a vibe-coding tool, "those things are complicated"; "you're not going to vibe code Shopify... the business stack is too hard." (He even confessed to owning "a lot of Atlassian" on the same logic, "all the developers will replace it with their own thing, which obviously doesn't happen.") And Salesforce's operators showed agents adding pipeline to seat-based teams rather than replacing them. The unifying bull thread, from Taylor to Billmaier to Abrahami, is that proprietary customer data and codified business logic are the moat AI can't cheaply copy.

Where the swing sits. Both sides now agree on the destination: some blend of seats, consumption, outcomes, and services, with proprietary data as the defensible core. They disagree on speed and on who adapts. But the single most important fact is the one this newsletter has flagged for six straight weeks and still can't see: no in-scope SaaS company has disclosed an explicit gross margin on its AI features, and none has printed a fresh net-revenue-retention number. The only margin data points this week were adjacent, AI-native services firms moving from "35 to 45 percent gross margin... to a 50 and 60 percent" (AI to ROI, July 16), and frontier labs reportedly running "90% plus inference margins" at the model layer (AI Daily Brief, July 13). Neither tells you what an AI feature does to Salesforce's or Adobe's gross margin. Until those land, the bull and bear are arguing over a number neither can see.

Stocks in play

This week's direct, substantive coverage went to CRM (two operator interviews) and DDOG (a chief-scientist interview). ADBE and HUBS got marketing-only interviews, brand and product color, no financials. TEAM appeared only as a passing mention; ASAN and MNDY were dark. Flagged honestly below.

Salesforce (CRM), directly covered; operators on offense.

  • Bull: Two senior executives made the "partnership, not replacement" case with a concrete stat, an SDR team now running "a hundred individuals and 400 agents" and "over a hundred million in pipe that would have been unseen otherwise" (Venture with Grace, July 11). Plus a "day zero" adoption push and the Contentful acquisition to close in Q3 (Marketing Beyond, July 15). Its data moat is exactly the defensible asset Bret Taylor and others say survives commoditization.

  • Bear: Everything Salesforce cited is a first-party anecdote, not an audited metric, and the "pay for outcomes, not seats" world its own former co-CEO now predicts is a direct threat to Salesforce's per-seat model. Still no seat-vs-consumption pricing detail and no net-retention figure.

  • Next catalyst: Any reported (not anecdotal) Agentforce revenue/consumption metric on the next earnings call, and how the Contentful deal is framed for monetization.

Datadog (DDOG), directly covered; product substance added.

  • Bull: Chief scientist detailed real, in-production AI (Toto foundation models, 100M–2.5B parameters, a "world model" for incident prediction) built on Datadog's proprietary observability data, and claimed the AI is cheaper to run than the old method (The Data Exchange, July 16). Named again among data-infrastructure winners (Motley Fool, July 12).

  • Bear: Datadog is consumption-priced, so this week's dominant theme, enterprises throttling token/usage spend (The Information, July 15), cuts straight at it. Its own scientist's point that AI observability can be cheap is double-edged: cheap for Datadog to serve, but also cheaper (less billable) for customers.

  • Next catalyst: Usage/consumption trends and any customer-spend commentary on the next print, the read on whether the throttling reflex is reaching Datadog.

Adobe (ADBE), marketing-only interview; no financial update.

  • Bull: CMO Lara Balazs talked up Firefly as an "all in one AI studio" and a newly launched "agentic layer... Firefly AI assistant across... InDesign, Illustrator, Photoshop," plus the SEMrush acquisition creating "a unified platform to do both SEO and GEO search" (Uncensored CMO, July 15).

  • Bear: This was a brand-and-creativity conversation with zero financials, no update on the empty C-suite, the freemium pivot, or the subscription-revenue pressure flagged last week. Silence on the numbers is not reassurance.

  • Next catalyst: A permanent CEO/CFO appointment, and any hard figure on how the freemium shift and Firefly monetization affect subscription revenue.

Atlassian (TEAM), passing mention only.

  • Bull: The stickiness argument got an unexpected endorsement: Wix's founder said he owns "a lot of Atlassian" precisely because the fear that "all the developers will replace it with their own thing... obviously doesn't happen" (20VC, July 13).

  • Bear: Still a per-seat, teamwork-tool vendor squarely in the seat-erosion zone, and no Rovo adoption or consumption metric surfaced this week.

  • Next catalyst: A Rovo adoption or consumption number, and whether it converts seat pricing toward a hybrid or consumption blend.

HubSpot (HUBS), marketing-only interview.

  • Bull: CMO Kipp Bodnar discussed Breeze in the context of AI-era brand discovery and customer data (The Agile Brand, July 15), the customer-data angle is the same moat theme the bulls lean on.

  • Bear: No Breeze monetization, attach-rate, seat, or net-retention data, read-through only. SMBs remain the customers most able to "hack around" with cheap agentic tools, and HubSpot's per-seat core sits in the erosion zone.

  • Next catalyst: A first Breeze attach or monetization figure; any net-retention update.

Asana (ASAN), dark this week (no coverage).

  • Bull: Management has already conceded the low end is exposed and is moving upmarket, arguably the right defensive move.

  • Bear: Work-management seats are the textbook "one agent replaces five seats" target, and the consumption-throttling theme is a headwind for any usage-based upside. Silence is absence of information, not comfort.

  • Next catalyst: Any AI Studio usage data or enterprise-cohort retention number.

Monday.com (MNDY), dark again (multiple weeks).

  • Bull: An AI-credit consumption model is, in theory, well-suited to the consumption era.

  • Bear: That theory is exactly what this week undermines, if enterprises are actively cutting token/credit consumption, an AI-credit model turns from upside into a headwind. And we still have zero fresh data on whether it's working.

  • Next catalyst: A first disclosed AI-credit consumption or attach metric.

Read-throughs

Seat-heavy SaaS (TEAM, HUBS, ASAN, MNDY). Little to no direct airtime, but the week's central risk lands squarely on them, and it evolved: it's no longer just "agents collapse seats," it's "the consumption model that was supposed to rescue you is the one buyers are now cutting." For Monday.com specifically, an AI-credit model is a headwind, not a hedge, in a token-throttling environment. The offsetting hope is the moat thread running through every bull this week, proprietary customer data and codified business logic (Salesforce, Wix, Bret Taylor), plus Arvind Jain's contrarian "teams get bigger" argument. Watch which of these four is first to disclose an actual AI attach or retention number; the silence is itself a data point.

Model and inference vendors (OpenAI, Anthropic, Bedrock, Azure, DeepMind). The economics stayed brutal and better-documented: OpenAI "burned $20.9 billion in 2025" and roughly "$3 billion a month," with Oracle building "7.1 gigawatts... just for one customer" and, per the bear case, needing "$75 billion of revenue annually" to cover Stargate (Impact Theory, July 14). Two structural cross-currents matter for SaaS. First, the token subsidy is enormous but finite: Semi-Analysis figures cited on AI Daily Brief (July 13) put the $20/month tier at "$400 of usage for Anthropic, or $700 from OpenAI," and the $200/month tier at "$8,000... from Anthropic, or a staggering $14,000... from OpenAI." Second, AWS "accelerated their growth to 28%" (The Information, July 15), and the same analysts warned that as customers "down tier to cheaper, smaller models," the frontier labs will "vertically integrate up into the software space", meaning more head-to-head competition between the model makers and the seven names over the next several years. On the cheap-model front, Bridgewater reportedly fine-tuned a Chinese open-source model to beat "the best tested frontier model while costing 13.8 times less" (Everyday AI, July 16).

Multiple de-rating risk. The whole software/AI complex re-priced this week. Gavin Baker's framework (via AI Daily Brief, July 13) is the one to hold: if share shifts from "frontier labs with 90% plus inference margins towards cheaper models," margin dollars "get redistributed from the frontier labs to AI infrastructure providers." That's constructive for the infrastructure/observability layer (Datadog) and negative for anyone whose value is a thin wrapper on an expensive model. And the tape confirmed the anxiety: on The Exchange (July 17), semiconductors entered a bear market, the Nasdaq had its worst week in nearly a month, and money rotated into healthcare and financials, an "AI reckoning." IPO watch: SpaceX's IPO ("$75 billion at $1.75 trillion... trading at $2 trillion on roughly $35 billion of forward revenue") is the template, with Anthropic (rumored ">$100 billion in revenue") likely to precede OpenAI (rumored "~$70 billion" run rate) to market, both potentially above $1 trillion (All-In, July 11). Bret Taylor gave "no update on IPO plans" (Squawk on the Street, July 16).

What changed vs. last week

New: the consumption thesis flipped from savior to risk. Last week's frame was that incumbents could re-rate by moving from seats to consumption pricing. This week multiple podcasts documented enterprises cutting consumption, the "token maxing" to "token efficiency" whiplash, $6,000-to-$600 cuts, spend caps. That's a genuine escalation: it means the escape route from seat erosion has its own trapdoor. Most directly bearish for the usage-priced names (DDOG, MNDY).

New: Salesforce answered the KeyBank downgrade with operators and a number. Last week CRM was the lead bear story (a Buy-to-Hold cut on stalling Agentforce). This week two Salesforce executives went on the record with a concrete "partnership, not replacement" adoption stat ("100 individuals and 400 agents," "$100 million in pipe") plus the Contentful acquisition. The debate is now live and two-sided rather than one-directional, though still unsettled without audited numbers.

New: "pay for outcomes" got a marquee operator voice. Bret Taylor (OpenAI chair / Sierra CEO / ex-Salesforce) explicitly said pricing moves beyond tokens to business outcomes, a cleaner, scarier articulation than last week's "hybrid seat + consumption" synthesis, and pointed straight at subscription software.

Corroborated: the OpenAI burn story got its audited headline number. Last week: leaked figures ($1.60 spent per dollar, a ~$21B loss). This week: "$20.9 billion" burned in 2025 from the audited financials, plus "$3 billion a month," plus Oracle's 7.1GW single-customer build. The margin-floor thesis is better documented, not resolved.

Evolved: Datadog moved from metaphor to mechanism. Last week DDOG was the "MRI scan" winner on a framework. This week its chief scientist detailed the actual AI products (Toto foundation models, world models) and claimed cheap inference, substance behind the story, though the consumption-throttling risk now looms larger.

Still missing (six weeks running): the two numbers that would end the debate. No explicit AI-feature gross margin from any of the seven, and no fresh net-retention print. The nearest proxies (AI-native services margins of 35–45% rising toward 50–60%; frontier-lab inference margins of "90% plus") describe other layers of the stack, not in-scope SaaS.

Quiet: Adobe and HubSpot went qualitative, and three names went dark. ADBE (CMO on Firefly/agentic layer/SEMrush) and HUBS (CMO on AI search) got marketing-only airtime with no financials. TEAM surfaced only as a passing mention; ASAN and MNDY had no substantive coverage again.