• “The question is not whether AI will shape the world. It will,” Schmidt said, pausing for a moment as students booed. “The question is whether you will have shaped artificial intelligence.”
• The rise of artificial intelligence is the next industrial revolution.” One audience member jeered in response, “AI sucks.”
• The same cohort is also adopting the technology at a rapid clip, with 57% of U.S. college students reporting using the AI tools in their coursework weekly, and 20% using it daily,
• “The results were anomalous,” he told Fortune, noting 8% of his class getting a perfect score on the multiple choice section of the exam while many either struggled on the essay portion or gave written responses
• rinceton University faculty voted last week to rescind its 133-year-old honor code and proctor all in-person exams to mitigate cheating using AI
• a report she published last year found most college students use AI, but many don’t disclose it.
• She identified the phenomenon as a form of cognitive dissonance, a psychological pattern in which a set of behaviors may contradict a belief system, leaving individuals to alter either their attitude or actions toward a certain topic.
• “The job market already seems precarious to them, and so even the students that did acknowledge that,
• Indeed, a stagnant job market, along with tech leaders warning of mass AI job displacement, has instilled fear in many recent grads
• Concerns around AI taking certain jobs have already begun to materialize as anecdotal evidence, despite no widespread proof of AI markedly changing the labor market
• While students see AI as a threat, Das said, the proliferation of AI in the workplace, as well as in schools—where last year about 30% of teachers said they use AI at least weekly—has also created a justification for them to use the technology, even if it means cheating
• “They are thinking, ‘People rather than me are using AI. Why am I held to a different standard? Why can’t I use AI?’
• “Students feel that there’s a corporate mouthpiece narrative,” Das said. “They are facing this very real fear of not landing a job, and so especially the tech CEOs, when they come to these commencement stages and encourage and cheerlead AI, I think students feel a disconnect there.”
• Some of the blame, Shelley argued, lies with educational institutions themselves, which have advocated for students to use AI.
• Das noted that AI authorities, including higher education institutions, have done a poor job identifying what jobs will be created as a result of AI and subsequently encouraging the appropriate form of upskilling
• “The worst thing we could do is blame students here,” Shelley said. “It’s our job to teach them, to nurture them, to inspire them, to guide them. It’s our job to educate them, and it’s our responsibility as society to take a deep look and go, ‘Why has this happened?’”

14 more annotations...

• "I spend a lot of time scrolling through videos made by content-creators," says Lucie, also 16. "They're more interesting than the posts of people I know." She doesn't post except sometimes "stories" which disappear after 24 hours. 
• Whether it's TikTok, Snapchat, Facebook and Instagram, we are a long way from the "digital town square" of personal interaction that social media was even just a few years ago.
• In France, annual official Barometre du numerique 2026 shows 49% of social media users are "active only occasionally".
• "users have become more conscious that the traces you leave (on social media) stay
• Some don’t want the exposure to criticism that might be a risk when you post or the feeling that their post will seem poor alongside all the professional content".
• The goal is to make people laugh, not to tell people about their lives."
• That still happens, she says, but it’s moved from social media platforms like Instagram and Facebook to messaging sites like WhatsApp.
• What we're seeing is social media splitting in two," says social media consultant Matt Navarra, author of the Geekout Newsletter. "Big platforms like Instagram and TikTok are becoming more about entertainment and discovery. WhatsApp is becoming the place people go to actually be social. The catch is, those kinds of spaces are harder for companies to make money from."
• Now, says Matt Navarra, "Meta has built what it calls an AI system for unconnected content recommendations on Facebook and Instagram, which basically means, they're increasingly showing you stuff from people you don't follow because the machine thinks you're going to like it.
• The social platforms continue to be monetised predominantly by ad revenue. That is still the core business model. And ad revenue continues to grow – Matt Navarra
• Social media is evolving into something passive like television, albeit television that adapts as you zap.
• The transition from truly social media to entertainment platform does seem to be paying off. "The social platforms continue to be monetised predominantly by ad revenue. That is still the core business model. And ad revenue continues to grow,"
• AI powered digital ad targeting is becoming ever more effective and precise.

• the social platforms will have that information and will place the ads accordingly."
• . It costs more to place ads in the social media feeds of people who buy horses than people who buy ice-cream.

13 more annotations...

• the billionaire AI evangelist Marc Andreessen — who’s invested billions into AI development through his venture capital firm — struggled when asked to articulate AI’s benefits.
• “So, you’re saying that the people running AI have done a terrible job of selling AI,” Rogan said. “So sell it.”
• Andreessen goes on like this for the next minute, trying explain that Newton wanted to turn lead into gold, seemingly trying to draw a parallel to the tech industry’s drive to turn sand — silicon — into thought.
• At any rate, Andreessen immediately abandons the alchemy analogy to explain that AI is actually good because stuff like lawsuits, medical problems, and work is just too hard for humans to think about on their own.
• Nevermind that AI compute is incredibly expensive — it certainly won’t be free “for everybody” for much longer — it also isn’t even close to performing human-level thought. Keep in mind, these are basically massive predictive algorithms: a more accurate phrase would be “data at scale.”
• Of course, that’s a self-reinforcing argument, a cop-out which doesn’t explain why it’s important to have “thought at scale,” or why thought is currently insufficient at its current, human-sized scale.
• The fact that one of AI’s loudest and richest devotees can barely articulate why he thinks the technology is important should be a telling sign, especially as productive returns on the tech remain far out of reach.
• Commenters posting under clips of the interview were unimpressed. “Before AI, if I didn’t know something I would just google it,” one TikTok user wrote. “AI gets to the answer faster but it’s only correct half the time. How is that revolutionary?”

6 more annotations...

• Imas asks: if AI makes a wide range of cognitive work cheap, where does spending go next? His answer: spending flows toward goods and services where the human origin is part of what customers buy.
• When a sector becomes more productive, spending shifts away from it toward sectors with higher income elasticity
• René Girard argued that human desire is mimetic: we want what others want, and we want it more when they cannot have it
• So as AI cheapens commodity production, real incomes rise and spending shifts toward what Imas calls the relational sector: teaching, care, hospitality, craft and live performance, where the human element is part of what is being bought.
• Many jobs survive because firms do not buy isolated tasks. They buy bundles. And because organizations do more than process information. They allocate authority and settle conflicts. AI can help with all of this. It does not follow that it can replace the people who do it.
• But labour markets price jobs, not tasks.
• A job is a bundle of tasks. The real question is not whether AI can perform one component of the bundle. It is whether that component can be separated from the rest at low cost,
• or most of what agents used to do —searching flights, comparing hotel rates, issuing tickets— the bundle was weak. Separating the booking task from the human was cheap, and once it was cheap, the task was gone. But something else happened to the agents who stayed. They moved upmarket, charged planning fees, and joined luxury consortia that offer upgrades and personalized itineraries.
• In 2013, a study by Carl Frey and Michael Osborne put the probability that accountants and auditors would be automated at 94 percent. A decade later, the US Bureau of Labor Statistics counts 1.6 million accountants and auditors employed, median pay of $81,680, and projects the occupation to grow another 5 percent through 2034, faster than the average for all jobs.
• The accountant’s job is a strong bundle. She interprets tax law as it applies to a specific client, signs the audit that the bank and the SEC will rely on, and carries the legal exposure.
• Three traits make a bundle stronger.
• The first is unpredictable demand. If you could tell in advance which tasks a customer will need, you could assign each to the right worker, human or machine.
• The second is production spillovers. Some tasks belong together because doing one makes you better at the other
• The third is the measurement problem. Armen Alchian and Harold Demsetz pointed out that when several inputs jointly produce an output, their separate contributions are hard to disentangle

• Bundling understates the case. Organisations are not production functions. They are coalitions of people with legitimate but conflicting goals.
• Trust is an important lubricant of a social system: It is extremely efficient; it saves a lot of trouble to have a fair degree of reliance on other people’s word. Unfortunately this is not a commodity which can be bought very easily. If you have to buy it, you already have some doubts about what you’ve bought.

• Trust in this sense is not a belief about competence. It is the expectation that the other party will bear a reputational cost if it defects, and will be around tomorrow to pay it.
• A human manager carries a reputation with counterparties who will deal with her again. The firm can fire her if she fails. An AI agent has neither. It cannot be sued, cannot be replaced with fanfare to signal a reset.
• There is also the problem of the unforeseen. Every project involves thousands of situations nobody specified in advance.
• n the electrician and the plumber both need access to the same wall on the same day. The manager resolves this by exercising what Oliver Hart and John Moore called residual decision rights: the authority to decide matters that contracts and processes have not specified in advance.
• These residual decision rights are not a cognitive task, but an institutional one. The decider must hold tacit and relational knowledge the people around them will not share, because communicating it would give away their bargaining power.
• AI will not produce mass unemployment in rich economies that can fund the transition. Spending will shift toward work where humans add value. Labour share may fall in aggregate even as human-intensive sectors grow as a share of the economy.
• . I think much of the surviving employment will sit in strong-bundle, AI-augmented work and in the political-organizational core of firms.
• The disruption of junior career ladders is real, and we have written about it. But the argument that “half of entry-level white-collar jobs be gone in five years” confuses task automation with the extinction of jobs.
• The real world is messy. The mess is what happens when human beings with competing interests try to get things done together. The economy Imas describes is the economy of what customers want. The economy I am describing is the economy of what organizations need to do. The second is larger.

23 more annotations...

• Si nous devions résumer l’échange, on pourrait dire que nous ne faisons pas face à des problèmes isolés quant à l’impact de nos pratiques sur notre vécu du travail. De plus en plus d’écarts s’observent entre la promesse affichée de ces nouveaux outils de GenAI et la réalité vécue de leur utilisation..
• Nous vous partageons cinq paradoxes qui touchent quatre capacités humaines fondamentales : l’attention, la relation, la compréhension, le jugement
• Être hyperconnecté·e est censé nous aider à mieux suivre nos sujets : être plus informé·e, plus vite, plus réactif·ve. Sauf que, dans la vraie vie, l’hyperconnexion ralentit et dégrade le suivi.
• . Ce que nous appelons “multitâche” est en réalité du task switching : une alternance de contextes, avec un coût mesurable en temps, en erreurs et en charge mentale.
• au prix d’une hausse du stress, de la frustration et de la pression temporelle.
• Notre cerveau met entre 10 et 25 min pour se reconcentrer.
• La perte de productivité liée au manque d’attention est estimée à « 0,6 point de PIB par an pour la France »^[5].
• • Réduire la densité de bascules : des réunions en 50/55 minutes plutôt qu’1h, pour s’autoriser une respiration avant d’enchaîner.
  • Ritualiser un check-in cognitif au début d’une réunion (“tu arrives comment ? saturé·e, disponible, déjà sous l’eau ?”) pour arrêter de faire comme si tout le monde était en état de réception optimale.
• Des études menées en 1967 par Albert Mehrabian découpent la communication en présentiel en 7% (verbal), 38% (paraverbal), 55% (non verbal
• le produit, c’est l’équipe. OCTO le dit explicitement : la collaboration fluide d’une équipe intégrée est une condition de succès, pas un supplément.
• Donc oui, “être en contact” avec 5 000 personnes n’est pas “être en lien”. Et le coût, c’est une baisse de coopération réelle, donc un delivery plus fragile.
• • Réduire la FOMO en visio : moins d’onglets ouverts “au cas où”. Pas par morale, mais parce que la présence à l’autre est une condition de qualité relationnelle (et donc de coordination) même si on a accès au monde entier en un clic.
  • Désactiver sa miniature vidéo quand c’est pertinent : ne pas se regarder en miroir peut améliorer l’écoute et réduire l’auto-surveillance.
  • Maintenir des rétros comme espace protégé d’ouverture et d’écoute (le lieu et le moment où l’on peut dire “ça ne va pas”, où l’on peut prendre du recul).
  • Renforcer l’exposition au terrain utilisateur : on ne maintient pas le sens d’un produit sans contact régulier avec celles et ceux qui le “subissent”… ou en bénéficient.
• Avec la GenAI, une rupture supplémentaire s’installe : la production devient “très facile”.
• ces gains d’efficience peuvent produire un effet rebond. Autrement dit, gagner du temps ne garantit pas d’en libére
• Le travail se déplace de l’exécution vers le cadrage, la supervision et la prise de décision. Et ce déplacement peut être épuisant autrement : moins de production manuelle, plus d’arbitrages, plus de vigilance, plus de charge de décision.
• • Plan individuel : limiter le temps d’exposition, et surtout continuer à solliciter des humains (pair review, échanges, retours), parce que la qualité ce n’est pas juste des “prompts” de qualité.
  • Plan collectif : tester des espaces “sans IA” (pas pour interdire, mais pour protéger des temps de raisonnement et de débat).
  • Plan organisationnel : rappeler que la curiosité ne justifie pas l’hyperconnexion : aller “voir si l’agent a fini” à 21h, c’est un signal. Pas un exploit.
• ’accès à l’information ne signifie pas la compréhension.
• C’est un changement profond, et il est visible au quotidien : décisions prises sur des éléments insuffisamment compris,
• • rendre explicite l’incertitude (ce qu’on sait / ce qu’on suppose / ce qu’on doit vérifier)
  • collectiviser la compréhension (review, débat, confrontation)
  • Une astuce ? Une fois le prompt terminé, écrire la réponse ou ce qu’on a appris sur un papier. L’ancrage sera facilité

• Mais cette autonomie peut rester de surface. Plus on produit seul, plus on assume seul, le cadrage, les choix ou les arbitrages.

• Cela crée une nouvelle charge : non plus une charge d’exécution, mais une charge de décision qui peut faire émerger :

  • une attente implicite d’être “capable seul”,
  • une réduction des temps de confrontation collective,
  • une difficulté accrue à partager les doutes.
• uand le code est produit par un agent, l’explicabilité et la responsabilité sont diluées, rendant la correction plus difficile par l’humain si le collectif n’a pas maintenu un niveau de compréhension partagé.
• • Ritualiser la confrontation : garder des moments où l’on vient challenger des productions IA (pair prompting voir mob prompting et donc critiques collectives sur le prompt, sur le code généré, sur le contexte, les règles, les skills fournis)
  • Nommer explicitement l’incertitude : ce qui est sûr, ce qui est supposé, ce qui reste à valider
  • Partager les usages IA : expliciter comment on s’y prend, plutôt que laisser chacun bricoler dans son coin et construire un référentiel commun de skills, de règles.
• Ce qui devient critique n’est plus seulement ce que nous produisons ou les outils que nous utilisons mais dans quelles conditions nous pensons, décidons, agissons et collaborons.

• a recherche de ce “better way” se déplace :

  • Plus que de simplement exécuter, c’est l’attention au geste de travail qui importera de plus en plus
  • Au-delà d’avoir accès à de l’information, des données et des résultats, c’est la question de “comment je les comprends et comment je peux exercer un esprit critique dans la manière dont je les restitue et/ou les remets en cause”.
  • On pourrait penser que le “capable de faire seul”, le “besoin de personne” seraient la nouvelle norme. Ce serait accepter que regarder l’humain et soutenir le collectif ne règlent pas la grande majorité des problèmes.
• • Niveau 1 - Augmentation : je fais avec l’IA
  • Niveau 2 - Délégation : l’IA fait pour moi sur certaines tâches
  • Niveau 3 - Substitution : l’IA fait à ma place

• Nous avons longtemps pensé que les outils étaient un levier de performance. Aujourd’hui, ils deviennent un test. Un test de notre capacité à :

  • ne pas confondre vitesse et valeur,
  • ne pas déléguer notre jugement,
  • ne pas transformer l’autonomie en isolement.

25 more annotations...

• This means that for every dollar of revenue the company generated, it lost $1.22. 
• By that maths, that means that OpenAI lost $6.95 billion in the quarter, and because this is non-GAAP, it’s quite possible that losses are much higher, revenues are lower, and its margins are worse.
• The report also claims that OpenAI is “on track” to hit its goal of generating $30bn in revenue for 2026, although if it maintains these disastrous margins, it would end up losing $36.6bn. 
• Meanwhile, ChatGPT’s user growth has stalled. While weekly active users hit 920m in February, the average for the quarter sat at 905m, suggesting lower numbers in either (or both) January or March. OpenAI had expected to hit 1 billion weekly active users in 2025.
• As I’ve noted in the past, weekly active users are a fairly novel metric, with most companies using monthly active users to represent adoption. I’ve also speculated that the reason why OpenAI has favored this metric is because it’s easy to manipulate.
• OpenAI reportedly had 55m paying ChatGPT customers at the end of Q1 — up from 47m people at the end of the year. 
• Nevertheless, while this represents an improvement over the 2.583% rate in February of last year, it’s likely improved as a result of cheaper ad-supported ChatGPT “Go” subscribers at $5 or $8 a month, depending on geography.
• As I wrote up yesterday, Anthropic leaked (or had leaked) that it believed it would have a non-GAAP EBIT operating profit in Q2 2026 entirely as a result of Elon Musk discounting two months of compute costs for that specific quarter, and it makes me wonder why we’re suddenly, in the space of 24 hours, talking about operating margins or operating profits for two companies that have hidden behind annualized revenues and obfuscated financials for several years.
• If I had to guess, it’s likely that investors have begun to demand firmer, more “real company”-adjacent numbers,
• Anthropic clearly wants to position itself as the stable, reliable, economically viable alternative to OpenAI, but can only do so with a kind of financial engineering
• Nevertheless, it’s interesting that Anthropic appears to be leapfrogging OpenAI in revenue
• Nevertheless, if investors become convinced that OpenAI is falling behind, it’ll be much harder to raise another round at or above its current $852 billion valuation. 
• Perhaps that’s why OpenAI is rushing to go public - it realizes it might have tapped out private investors.

11 more annotations...

• The future of HR will not be decided by which vendor automates the most clicks; it will be decided by who can manage complexity across people, agents, policy, process and accountability.
• Firms do not buy isolated tasks, they buy bundles of tasks, embedded in roles, relationships and systems of accountability. That matters because labor markets price jobs, not individual moments of work.
• Much of the current conversation about AI in enterprise software treats HR as if it were mostly a collection of repetitive transactions waiting to be streamlined. But any HR executive knows the real work begins where the workflow stops being routine
• AI-native entrants will create real pressure on the layers of the HR stack that are admin-heavy, repetitive and experience-poor. There is no serious scenario in which the current architecture of HR technology stays untouched.
• But this is where the argument veers off course. A painful workflow is not the same thing as a replaceable enterprise core.
• In large enterprises, the most important HR decisions are difficult because the underlying work is genuinely complex, cross-functional and high consequence—not because the tech interfaces are bad.
• That distinction also helps explain why Workday still matters, even as the market changes around it. As the world changes, Workday is framing itself as an enterprise AI platform for managing people, money and agents, and its Agent System of Record is explicitly designed to give organizations visibility and control over digital labor—including third-party agents—in what Workday calls a “blended workforce.”
• The Agent System of Record will be a crowded space. Some players will focus on orchestration, some on workflow, some on permissions or employee experience.
• The most likely outcome is not a clean replacement story, but a re-layering of the stack.
• “If companies use AI simply to compress the bottom of the pyramid, they may solve a productivity problem and create a capability problem.
• Simply put, while the future of HR is certainly an IT modernization story, it is also a training story, and an operating model story, and a leadership story.
• Which parts of our HR operating model are truly ripe for automation? Where do we still need human judgment and decision rights? What should remain anchored in the system of record, and what should move into agent-enabled workflows? How do we govern a workforce that now includes both people and software agents without creating new operational, ethical, or regulatory exposure?
• The future of enterprise HR will absolutely be more agentic—more open, more conversational, and more dynamic. But it will also remain governed, political, interdependent, and deeply human.
• The task is not the job. And the workflow is not the system.

12 more annotations...

• Remember that SAP, unlike Workday, is a true ERP. In other words, the system manages all forms of business resources, including financial, human capital, parts, inventories, manufacturing in-process, and all the procurement, contracts, suppliers, vendors, and outsourced staffing and contingent labor we use.

   
• Without AI that kind of question sets off a lot of analysts looking at data. Now, with SAP’s “autonomous enterprise AI” stack you ask Joule the question, it may ask you to clarify a few things, and it runs an analysis to find the answer.

• This is powerful stuff if it works.

   
• In our area of HR and human capital, there are equally interesting and difficult questions to answer
• SAP seems to believe that your SAP system will “run without your help” at some point,
• Are these “automations” designed to simplify something that took a long time to do by hand? Or are they agents that redesign how HR really operates. It’s not clear quite yet.
• Our research shows that there are “types” of agents, not just “flavors.” Some are monitoring agents, some are action agents, some are rules agents, and some are data and information agents.
• What I noticed, however, is that a lot of the value is focused on the automation part, not the “process redesign”
• This is fine, by the way, because most SAP systems are so complex that these automation workflows add massive value. But ultimately the potential for AI is much greater as we redesign, which SAP may have to do over time.
• As you can see, there’s a big blue AI layer that includes a data fabric and many models, including a large context window that lets SAP put lots of business data into the blue AI model.
• In the middle is this tiny bubble called SAP Knowledge Graph. This is a big thing.
• this is the “brains” of the system that maps all the thousands of business entities, structures, and rules you have in SAP into a semantic layer you can talk to.
• On the top, in purple is Joule, the copilot interface for users and also the development tool for agents
• The new Joule Studio, shown below, looks like an enterprise class development tool (not a vibe coding system) that lets you design, build, test, integrate, and manage any complex or simple agent you want.
• So what SAP has done is not “oversimplify” the tools (which I’m afraid many LLMs are doing, listen to this podcast for more) but rather give IT teams and SAP developers a very robust system to build all types of agents.
• If people gravitate to Joule to build agents, they may just “redesign SAP” from Joule
• And as with ServiceNow and Workday, you can access other systems through SAP’s data layer and interoperate with non-SAP agents through their agent management system.

• I would expect that Workday’s new Sana Studio will be similar, but so far Joule is the most advanced I’ve seen.

   
• Another casual mention discussed building massive context windows to store an entire “company memory.”
• Imagine the AI model(s) could store an entire data set of operational knowledge (customers, products, processes, rules, and even documents and emails) and use that “company corpus” to analyze, model, and continuously improve. Kind of a big big idea.
• SAP now offers AI Agent Hub, a management system similar to tools available from Workday and ServiceNow.

• In other words, the SaaS Apocalypse is baloney.

   

  SAP CEO Christian Klein, in a parallel fashion to Aneel Bhusri of Workday, simply says this is not going to happen any time soon.
• Remember the goal here is not to start over but rather leverage AI to bring cross-domain process innovation, customer and employee experiences, and smarter systems into our companies.

• This is why Christian Klein believes SAP is an AI company, and it’s starting to make sense.

   
• SAP is not abandoning software: it’s making software less visible by embedding execution into agents that work across the suite

• And best of all, SAP’s financial model will shift to consumption and outcomes, not just seat licenses.

   

24 more annotations...

• McKinsey has described the rise of the agentic organization as a new operating model where humans work alongside virtual and physical AI agents to create value
• Before companies scale AI agents, they need a clear map of the people who understand the work deeply enough to direct it, challenge it and carry accountability for the result.
• A human capability map is different from a skills inventory.
• A capability map goes further. It shows where judgment sits,  who carries institutional knowledge, who can connect work across functions and who can supervise digital output without losing sight of risk, quality or context.
• They show who occupies a role, but they do not always show how someone thinks when the answer is unclear, how they respond to weak information or how they make decisions when speed and responsibility collide.
• The real value moves to the person who can define the task correctly, review the result, notice what is missing and decide whether the output is strong enough to act on. This is not just AI literacy, it’s supervision.
• T-shaped employees have deep expertise in one area and enough understanding of adjacent areas to work across boundaries. M-shaped employees have several areas of deeper expertise and can connect different domains with greater independence.
• These people are valuable because AI-enabled work rarely stays inside one clean function.
• Deloitte’s 2026 Global Human Capital Trends report makes a similar point. It argues that advantage is shifting from allocating talent in static structures to orchestrating people, skills, data and technology in real time.
• It may remove roles that look administrative while losing people who quietly held the context that made the process work. It may reward employees for using AI often, while failing to see whether they are using it responsibly.
• Gartner has predicted that more than 40% of agentic AI projects will be cancelled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls

• Where must human judgment remain close to the work?

   

  Who understands the process beyond the task list?

   

  Who has the institutional knowledge that should be protected before a workflow is redesigned?

   

  Who can review AI output with enough context to challenge it?

   

  Who has the cross-functional thinking needed to coordinate people, systems and agents?
• The organizations that handle this well will not treat human capability as an afterthought. They will identify their future AI orchestrators early.
• A person’s value inside an organization is often larger than their job description
• For CHROs, this is the strategic opportunity. AI agents may change how work moves through the business, but HR can help decide whether that change strengthens or weakens the organization.

13 more annotations...

• all of these numbers are mysteriously leaking exactly while it raises its funding round!
• .it is unclear what accounting methods Anthropic has used to book revenue and costs, as the company isn’t yet required to follow the financial-reporting requirements of a public company.
• The company might not remain profitable for the full year as it plans spending increases due to its vast computing needs. 
• So Anthropic may be profitable very specifically in Q2 2026, but might not be afterward.
• Remember that deal Anthropic signed with SpaceX to take over Colossus-1? Well it’s also taking over some or all of Colossus-2, paying SpaceX $1.25 billion a month starting in May and June… when it’ll have a reduced fee as it ramps up!
• That’s $15 billion a year in compute costs, but reduced to an indeterminately-discounted level for the precise months that Anthropic is using to tell investors and the media that it has an operating profit
• If Anthropic paid full-rate for its compute in those two months, its economics would shift back to what they’ve always been
• Per The Information in January, Anthropic missed on its gross margin projections, saying that its inference costs were 23% higher than the company had anticipated.
• For large enterprises, Anthropic is taking prepayment of tokens —-- say, $50 million intended to be spread over 12 months that it takes in as revenue.
• Anthropic is already offering discounted tokens for Claude users through the “buy extra credits” page on their accounts,
• Anthropic could be front-loading annual commitments of any kind
• Anthropic could have ratcheted down training to ease the burden on its infrastructure to provide inference. 
• Anthropic has said it brought in $4.8 billion in revenue in Q1 2026, and projects to hit $10.9 billion in Q2 2026.
• As a reminder, ARR is an accounting tool largely used primarily by startups, where a snapshot of a single month’s income is taken and multiplied by twelve.
• Dario Amodei would claim Anthropic had reached $19bn in ARR — which works out to $1.58bn per month. Two days later, on March 9, Krishna Rao — Chief Financial Officer at Anthropic — would declare under oath in a court filing that Anthropic had brought in revenues “exceeding $5 billion to date.” 
• I find it doubtful that the CFO of Anthropic would, under oath, lead the court to believe its business was 30% to 40% smaller than it was,

• this means that Anthropic:

  1. Made over 90% of its lifetime revenues in the first quarter of 2026., 
  2. Made virtually no revenue in its previous years, and… 
  3. Leaked completely imaginary run rates to the media for years.
• Dario Amodei and Elon Musk worked out a sweetheart deal, which they - framed as a “ramp-up,” - that allowed Anthropic to artificially depress its costs
• I also question how much of a ramp-up there really was, or what Anthropic’s actual compute constraints were, because it immediately loosened rate limits for Claude subscribers on announcing the deal,
• In July, Anthropic will start paying SpaceX $1.25 billion a month,  - or $15 billion a year, - on top of all of its other compute deals with Google, Amazon and Microsoft. 
• The company might not remain profitable for the full year as it plans spending increases due to its vast computing needs. 
• nobody does financial engineering and a press-led information war better than Anthropic.
• We’ve got its CFO saying numbers that don’t match up with these leaks or Anthropic’s own marketing materials, and the aggressive and deluded way in which many people ignore them is equal parts frustrating and depressing. 
• for example, WeWork claimed to be profitable since the second month of its operations, and repeated claims of profitability throughout its existence, and it turned out that it was only “profitable” if you removed things like “some of the costs of doing business.”
• Anthropic timed this leak to focus on a specific quarter where it artificially suppressed costs, and gave you the flimsiest proof imaginable, specifically-crafted for you to share it as a triumph and spread the idea that “AI labs are actually profitable,” when their core economics haven’t changed. Costs increase linearly with revenue, and will continue to do so in perpetuity. 

23 more annotations...