Anthropic's Super Bowl Ad: Who Won & Lost? | Sierra Hits $150M ARR: Is Customer Support Too Crowded?
Summary
Anthropic’s 2029 forecast only reconciles if AI expands the economic pie rather than reallocating today’s software budget. Its optimistic $149 billion ARR target, alongside OpenAI’s cited $180 billion, approaches half of the roughly $700 billion global software market before counting Microsoft or the rest of the application stack. Because one customer dollar can appear as Atlassian revenue, AWS revenue, and then Anthropic revenue, the panel cautions against adding every layer at face value—but even after adjusting, “there better be more money coming,” potentially from the roughly $1 trillion consulting-services pool and faster overall technology-budget growth.
The strongest rebuttal to “software is dead” is Atlassian’s operating evidence, but incumbents still need to prove that AI reconnects infrastructure spending to revenue. Mike Cannon-Brookes called category-wide extinction “ludicrous”: Atlassian was cited as growing roughly 23%, cloud revenue 26%, and RPO 44%, with customers making three-year commitments while gross margins improved amid AI deployment. Jason’s pushback was temporal—most public SaaS companies are still decelerating, and after 12–18 months of capable models, “you’ve got to show me the money in 2026”; Mike’s simpler rule was, “You have to be good.”
AI creates a consequential split between output-unbounded product and engineering work and input-constrained functions such as legal, support, HR, and systems integration. Engineering road maps are never finished, so greater productivity can produce more software without reducing teams; a company does not buy a second NetSuite implementation merely because the first became cheaper. Jason therefore sees almost every non-product category at “existential risk of shrinking seats,” although Jevons-style demand expansion could still turn previously uneconomic leads, legal questions, or support interactions into new work.
Harvey’s $200 million round at an $11 billion valuation proves extraordinary demand, not yet an extraordinary risk-adjusted return. At roughly $190–$200 million ARR and a cited path toward $600 million, Harvey could triple and then double to make the entry price roughly 10 times two-year-forward revenue—but today’s price is about 50 times run-rate revenue. With approximately 100,000 active users, current revenue implies only about $2,000 per user; to penetrate the cited $200 billion pool of associate and paralegal labor, Harvey must move from “next-generation tool revenue” toward $5,000–$10,000 per lawyer or $50,000–$100,000 agents that genuinely absorb work.
Customer support is simultaneously one of AI’s clearest ROI markets and one of venture’s most dangerous funding clusters. Sierra reportedly exceeded $50 million in a quarter and $150 million in ARR, while Mike estimated service and support at 3%–7% of virtually every business and said Atlassian’s service collection could stand alone as a public company. Yet Harry counted 14 recent entrants that had each raised more than $100 million, alongside Atlassian, ServiceNow, Salesforce, Zendesk, Intercom, in-house systems, and fast-moving specialists; the years-two-and-three test is whether vendors keep adding value after the first labor savings are “baked into the business.”
Venture capital is responding to category ambiguity by paying almost any price for perceived winners. The panel cited a striking pattern: 40% of newly minted Q1 unicorns had completed at least one further up-round by Q4, reflecting a “kingmaker, double kingmaker, treble kingmaker” market. Harvey fits the venture equivalent of “you can’t get fired for buying IBM”—a consensus leader bought with a 1× preference can feel like the safer bet, but that is also “classic top-of-the-bubble stuff.”
Anthropic’s Super Bowl attack on OpenAI’s advertising looked more like an expensive signal to talent and its rival than a consumer-acquisition strategy. Rory argued that Claude’s weakest position is precisely consumer usage, making “we will not advertise inside the app” an odd message for the broad audience; perhaps the real recipients were roughly 10,000 engineers Anthropic might hire. Mike resisted calling the ads a definitive market top: when model companies can burn enormous sums and keep raising, an $8 million placement is immaterial—a “sign of the times” and of unconstrained capital allocation.
The investable management signal is whether incumbents can accept disruption, fund the future, and still execute without exhausting their leaders. Mike said public-company discipline made Atlassian better at forecasting, planning, and delivery, provided those practices do not replace strategy; his answer to the selloff is, “We’re going to create our way out of this problem,” because “part of creation is destruction.” He now starts around 5 a.m., but argued that founders should ask whether they would choose the job again and preserve enough family, exercise, and friendship to avoid becoming irrational: “It’s tough out there. It’s going to be hard. Welcome to the technology industry.”
Deep dive
1. Anthropic’s forecast requires more than budget redistribution
The opening arithmetic was deliberately stark: Anthropic’s optimistic 2029 target is $149 billion ARR, while OpenAI’s cited estimate is $180 billion. The panel rounded their combined appetite to roughly $350–$380 billion against a global software industry of about $700 billion—before assuming Microsoft, already around $200 billion, simply “rolls over and dies.”
Mike’s first qualification was revenue stacking. In his illustrative example, a customer spends $1 million with Atlassian, Atlassian incurs $200,000 of AI expense through AWS, and AWS passes perhaps $150,000 to Anthropic; those proportions were explicitly hypothetical, but the point was that one end-market dollar can appear as revenue at several layers.
Harry offered Cursor and Anthropic as the cleaner abstraction: each might report $1 billion, yet economically it can be “pretty much the same billion dollars.” If model providers, clouds, and application vendors all grow, the end market must expand faster than its 20-year trend line.
Mike said the bet has to be TAM expansion rather than a purely zero-sum software budget. Harry argued that businesses buying productivity, speed, quality, and new products should redirect and increase spending, as technology budgets have done for 30 years; he also pointed to the roughly $1 trillion consulting-services pool as one plausible source if software can consume some of that work.
2. Model companies are suppliers, partners, and competitors at once
Atlassian uses Anthropic heavily but also Gemini, OpenAI, Llama, Mistral, and other models. Its gateway continually selects for “the best model for the best cost, with the best quality outcome and the best speed,” making model portability part of the application vendor’s job.
Mike accepted that partnership and competition coexist: Atlassian can work deeply with Google on Gemini while Confluence competes in some respects with Google Docs. He doubted model companies could reserve their best models exclusively for themselves when three or four other model competitors—and eventually Chinese models—remain available.
The durable defense is not category entitlement. Mike’s answer was simply, “We have to be good,” because Atlassian must compete with anyone delivering the same customer value, even while buying infrastructure from them.
3. Consulting demand rises first, then automation attacks the rote work
Harry argued that Accenture-like firms could generate more revenue than foundation-model companies over the next two or three years because large enterprises need implementation and integration help early in the cycle. Rory agreed for AI deployment, but separated that from repetitive SAP, Oracle, and NetSuite integration work that agents may automate.
Harry’s concern was talent rather than demand: B2B companies are trying to convert customer-success teams into forward-deployed engineers, yet he doubted even 10% could operate like product engineers. The panel’s joke captured the bottleneck—consulting might die because the industry “runs out of consultants” smart enough to do it.
Harry also made the vendor-incentive rebuttal. If only “wizards” can deploy a product, vendors exhaust the wizard supply and then revenue; mass adoption therefore forces “mass simplification,” guardrails, and tools ordinary teams can control. Mike joked that this could mean “the death of consulting because we’re going to run out of consultants.”
Atlassian customers typically begin with small automations that improve one process. The immediate business gain may be minor, Mike said, but the organizational learning is major: companies gradually discover what agents can do rather than release millions of agents across their stack on day one.
4. Software survives, but an architectural shift raises the casualty rate
Mike’s historical baseline was unsentimental: many Atlassian competitors documented in 2000, 2005, 2010, and 2015 were later acquired, merged, or extinguished. New challengers replaced them, which is “the history of the technology industry,” not evidence that software itself has ended.
Prepackaged technology remains efficient; businesses do not write everything in assembly language and are unlikely to start generating every system from scratch. Mike expects many SaaS companies to disappear over five to ten years and many others to prosper—“software is alive and kicking.”
Harry sharpened the bear case without accepting the slogan. In ordinary periods capitalism might produce a 5%–10% annual death rate, but a once-in-10-or-15-year architectural transition could hypothetically produce something closer to a 50% two-year casualty rate.
Asked whether LLM software is a quantum change or an organic progression, Mike answered, “It’s both.” The production method is changing quickly, but incumbents can use the same tools; the relevant question is which organizations adapt fast enough.
5. Atlassian is rebuilding its substrate, not bolting on features
Atlassian has roughly 10,000 people in R&D, including pockets where teams “Claude Code all day every day.” Mike’s reaction to coding acceleration was enthusiasm: the company can build “a lot more stuff a lot better,” while laggards that preserve old workflows may not survive.
The investment extends below visible AI features. Mike described large teams building context, search, and chat infrastructure that is necessary even if it is not sold separately, plus an infrastructure organization devoted to managing model selection and inference expense.
Contrary to predictions that AI COGS would crush application margins, Atlassian’s inference costs are falling. Some features now run approximately 1,000 times more cheaply than at launch because older models remain adequate after their prices collapse.
Mike said gross margin had risen over the previous six or seven quarters despite widespread AI deployment. The operating claim was not that inference is free, but that application companies can optimize it like any other infrastructure layer.
6. Public SaaS must now connect AI spending to reacceleration
Mike described the software selloff as a fear-driven sectoral rotation with a real underlying cause. Investors are funding semiconductors and data centers first, while the application layer’s eventual share of the revenue and profit stack remains unresolved.
Atlassian offered contrary operating evidence: overall growth was cited around 23%, cloud revenue grew 26% at well north of $5.6 billion, and RPO increased 44%, accelerating for a third consecutive quarter. Mike emphasized that RPO reflects customers signing multi-year, multimillion-dollar agreements rather than casual purchases from an online ad.
Jason nevertheless drew a hard line around persistent deceleration. Claude 3.5 and 3.7 had already been capable for months; incumbents might get 12 or 18 months to monetize AI, but by 2026 “all that infrastructure spend” must flow through inference into applications.
Product and engineering may be “above the fold,” Jason argued, because companies are making radically more software and still need issue tracking and coordination. Mike added that service is hardly collapsing at Atlassian: its service collection is the company’s largest at-scale business and is growing faster than the whole company.
7. Public SaaS averages exclude too many young winners
Harry recalled that median public SaaS growth stayed near 30% for roughly 15 years partly because companies falling below 10% received the “red card,” while new issuers arrived growing 60%–80%. With few strong IPOs entering the cohort, today’s average naturally ages downward.
Jason added two selection effects: private equity and large technology companies have acquired many listed vendors, while ample private capital lets companies such as Stripe remain private longer. Figma was offered as a private-company example that appeared well positioned in the current environment.
Of approximately 16 SaaS CEOs who formed a weekly support group in 2020—10 or 11 then running public companies—Jason said only Atlassian, Zoom, Box, and Shopify remained public. Most peers had been acquired, producing a strange survivor set between private-equity size and big-tech size.
Mike added that AI-driven changes in marketing channels further muddy comparisons. Multi-decade companies must repeatedly rebuild distribution as audiences migrate, while some weaker SaaS businesses face both product disruption and the loss of previously reliable acquisition channels.
8. Harvey defeats the “wrapper” insult by having customers
Harvey raised $200 million at an $11 billion valuation while producing roughly $190–$200 million ARR, with a cited expectation of approaching $600 million by year-end. The panel treated the growth as exceptional and legal work—text-heavy and inefficient—as unusually ripe for transformation.
The timing exposed a contradiction: Claude Code produced 13 plugins, including a legal plugin that helped trigger claims of a SaaS apocalypse, yet investors simultaneously valued a legal application built on foundation models at $11 billion. Harry’s point was not to dismiss Harvey as a “GPT wrapper,” but to show that both narratives cannot comfortably be true.
Jason reframed every software business as a wrapper around a process: Atlassian “orchestrates humans,” while Harvey packages legal work into a better, repeatable interface. Customers like the product, distribution works, and law firms capable of building their own Harvey generally have not done so.
The open moat question is how far Harvey extends beyond its current interface. Customer relationships and stickiness are advantages, but durable value requires absorbing more of the legal workflow rather than assuming the underlying models remain inaccessible.
9. Harvey’s product proof and investment math are different tests
Jason’s bullish rule was to “let the revenue show us the path to TAM.” If Harvey moves from $200 million to $600 million and then doubles to $1.2 billion, an $11 billion entry price becomes roughly 10 times two-year-forward revenue—scarce, but comprehensible.
Harry’s objection was opportunity cost: investors can buy a proven public company such as Atlassian at depressed levels, whereas Harvey must triple and then double merely to reach that forward multiple. At today’s approximately $190 million run rate, the valuation is close to 50 times revenue.
Mike’s Accel history illustrated the asymmetric upside and the base-rate danger. Accel invested $60 million of secondary capital in Atlassian in 2010, when it had roughly $50 million of revenue and $20 million of profit; a forecast of 30%, then 20%, then 15% growth was beaten by approximately 50%, 60%, and 70%, turning a hoped-for 2–3× into roughly 100×.
Jason’s conclusion was unforgiving: Atlassian only needed to outperform a moderate model, whereas Harvey must sustain growth “well north of 300%–400%” for several years to beat a 50-times entry price. “If you do, then it works.”
10. Legal AI must graduate from tool pricing into labor economics
Harry cited approximately $200 billion spent annually in the US on non-partner lawyers, associates, and paralegals. His changed view was explicit: he began skeptical of sweeping job-replacement claims, but growth this fast suggests previously inaccessible labor TAM might now be capturable.
Rory’s countercalculation started with 100,000 active Harvey users and $190 million ARR—about $2,000 each, reportedly below what Westlaw gets from a lawyer. With perhaps 400,000–500,000 relevant large-law-firm lawyers, that supports only a roughly $1–$1.2 billion tool market.
To escape that ceiling, Harvey must raise value per lawyer to $5,000 or $10,000, or sell $50,000–$100,000 agents that eliminate meaningful associate work. Jason argued customers will readily pay when agents “don’t quit,” demand raises, or object to weekends.
Mike supplied the industry-specific complication: law firms bill time, so completing work faster may undermine their existing revenue model rather than simply expand margins. Value-based pricing could resolve that tension, but he presented it as an open question, not a forecast.
11. Input constraints predict where efficiency removes seats
Mike’s most useful schema separated input-constrained functions from creation functions. In-house legal teams receive a finite flow of problems, while customer-service questions scale with customers; making either team twice as efficient does not automatically double the incoming work.
Engineering is different because “the road map is never finished.” Better tools can increase output rather than reduce headcount, preserving the humans and coordination systems needed to create more software.
Harry applied the framework to systems integration: automating a Salesforce, SAP, or NetSuite deployment does not cause a customer to purchase “two helpings of NetSuite.” Efficiency primarily lowers spending, making such markets less expansive than generative product creation.
Mike supplied the Jevons-style case. Atlassian used Agentforce on event and sponsorship leads too low-value for humans and reached five or six times more inputs; similarly, cheap legal review or excellent support could unleash questions customers previously abandoned.
12. Customer support offers visible ROI—and a difficult second act
Sierra’s cited $50 million-plus quarter took it beyond $150 million ARR, prompting Harry to ask whether new agents replace Zendesk, serve previously unsupported users, or perform entirely new jobs. Mike’s answer began with scale: service and support represent perhaps 3%–7% of every business.
Atlassian’s service collection, focused heavily on internal IT, HR, finance, and employee workflows, is large enough that Mike said it “could go public by itself.” Twitter/X alone reportedly runs 300–400 service desks on it, illustrating how many distinct teams support one organization.
Support sells cleanly because customers can compare existing expenditure with measurable savings and the product price. The harder challenge arrives in years two and three, when the original labor savings are embedded in the cost base and vendors must produce continual incremental value.
Mike attributed current acceleration to collapsing support hiring, dramatic “aha” moments that mature SaaS had lost, and the possibility that the principal support agent becomes a Trojan horse for sales, research, and marketing. Legacy businesses can stagnate while their agentic revenue explodes.
13. Support winners will be determined below the category label
Mike pushed back on treating support as one market. High-volume B2C password or ticket questions differ from a B2B P1 incident involving six or seven people across departments; internal help desks, external service, company size, and chat, voice, or video entry points create distinct workflows.
Mike highlighted a hidden complement: an agent answers only as well as the knowledge it can access. Some adopters are writing far more detailed, stepwise documentation, potentially replacing question-answering staff with more writers who structure knowledge for agents.
The largest leap is from answering to acting. An HR agent that explains parental leave is useful; one that files the application, completes forms, submits an expense, or resets a password “gets your job done” and makes the entire business move faster.
Harry’s investment objection remained: 14 young companies had reportedly raised more than $100 million each, while ServiceNow, Atlassian, Salesforce, Zendesk, Intercom, technology-first in-house systems, Sierra, and Decagon compete. Being third in a narrow subsegment is precarious when an adjacent winner can bundle it.
14. Venture consensus is crowning winners earlier and repeatedly
Rory connected Harvey’s financing to category uncertainty: when submarkets may disappear or merge, the seemingly safest decision is to “pay the winner at any price.” A 1× preference makes the bet feel safer, and nobody gets fired for backing the consensus champion.
The supporting statistic was extreme: of the unicorns newly minted in Q1 of the prior year, 40% had already completed one or more up-rounds by Q4. Capital was not merely selecting kings; it was “double kingmaking” and “treble kingmaking” them.
Rory’s historical warning was that every technology revolution attracts enormous capital, creates a handful of genuine winners, and torches money across the next 10 or 15 imitators before the market sorts itself out over five or six years. AI changes the world without repealing that cycle.
Portfolio signaling also matters. Harry noted that buying a stake in OpenAI, Anthropic, or Harvey at a huge valuation may not resemble classic early-stage venture returns, but it puts a coveted logo on the firm’s wall and “early investor” in its marketing.
15. The Super Bowl fight was a capital-abundance signal
Harry reconstructed the sequence: Anthropic’s ads mocked OpenAI for having ads, prompting angry public responses from OpenAI before OpenAI aired a broad “you can build things” commercial. A cited analysis framed the strategies cleanly: fewer than 5% of consumers subscribe, so a mass consumer product needs ads; an enterprise product may not.
Rory found Anthropic’s choice strategically odd because consumer usage is Claude’s weak point: it advertised “the worst product” for that audience instead of its transformation of coding and product creation. His more charitable reading was a $5 million message to perhaps 10,000 desirable engineers and to OpenAI itself.
Rory cited Wix’s experience: its CMO had purchased six prior Super Bowl spots and bought two that year, for Wix and Base44, because the impressions could justify the price. Harry replied that truly superior channels would be used consistently across capital cycles, not chiefly when money is abundant.
Mike split effectiveness from ego. Doritos can rationally pay roughly $8 million because it sells more chips; some CEOs buy the same slot because unconstrained capital permits it. For model companies burning enormous sums, the ad is cheap—“a sign of the times,” though not necessarily proof of a market top.
16. Public-market discipline helps only when strategy survives it
Rory’s deeper question was whether public companies, constrained by EPS and current growth, can compete with private challengers that barely price marginal spending or stock compensation. Mike’s answer was that strong leaders must manage the short term and invest for five-year relevance simultaneously.
Atlassian spends heavily on R&D and AI while delivering public-company commitments. Mike argued that any listed software company failing to fund fundamental AI work is “probably in trouble,” although financial statements reveal little about which R&D dollars are genuinely repositioning the product.
Public status made Atlassian better at forecasting, planning, and execution. The danger is allowing those capabilities to replace strategy; a company still has to enter new areas, hire for the new era, and explain why current investment creates durable customer value.
Mike acknowledged that strong quarterly delivery had not translated into a supportive share price. His response was neither denial nor capitulation: “You’ve got to accept reality,” then build, because “part of creation is destruction” and “we’re going to create our way out of this problem.”
17. The founder test is whether disruption remains worth choosing
Mike described a wave of “SaaS therapy” among executives, while Rory rejected moral judgment for those who step aside. A leader who has built hundreds of millions in revenue but no longer wants another reinvention may make the braver choice by handing control to a capable successor.
Mike’s own pace has increased: after his co-founder retired roughly a year and a half earlier following 23 years together, Mike said he often starts work around 5 a.m. Disruption, speed, and competitive change demand more effort, but only if the leader still enjoys the underlying work.
Their diagnostic was deliberately slow: would you choose this job again today? Over 90 days or a year, do customers, creation, and the challenge still provide energy? If the answer is no, “no harm, no foul”—technology is hard, and ego should not trap someone in the arena.
Balance is operational rather than ornamental. Mike argued that founders who abandon children, exercise, friendship, and time outdoors eventually make worse decisions; the lasting reward is also tribal—teams that endure “trenches and scuffles and fights” can share tea ten years later and know they built something together.