• The move signals that AI is evolving from a market defined by venture funding, breakthrough models and rapid experimentation into one increasingly shaped by public markets, infrastructure investments and demands for sustainable growth.
• Anthropic's filing suggests the industry may be entering a new phase in which investors, customers and regulators begin asking different questions about revenue, profitability, infrastructure requirements and long-term business viability.
• This week, Alphabet revealed plans to raise as much as $80 billion to expand its AI infrastructure,
• Meanwhile, SoftBank committed to investing $87.3 billion for AI infrastructure in France,
• Taken together, these developments suggest the AI industry is becoming less defined by individual model releases and more by the economic realities required to support them. 
• AI is no longer behaving like an emerging technology sector. Increasingly, it looks like an industry entering its next stage of maturity, one in which financial performance, infrastructure investment and operational execution matter as much as technological breakthroughs.

4 more annotations...

• AI is a tool. That is not a criticism. Tools matter. But tools do not compound. Tools do not govern themselves. Tools do not produce structural advantage simply by being deployed.
• The organisations that will define the competitive landscape through the rest of this decade are not asking whether they have AI. They are asking whether they have an Intelligence Architecture (IA).
• AI is the capability. Intelligence Architecture is the system that makes the capability count.
• AI asks: what can this capability do? IA asks: how is intelligence structured, governed, and compounded across the organisation over time? One question produces a tool. The other produces a structural asset.
• The capability was never architecturally embedded. The intelligence was never captured, preserved, or owned. There is no memory, no governance layer, no compounding. That is not an AI problem. That is an architecture problem.
• That gap – between acquired capability and architectural foundation — is where the competitive advantage of the next decade will be won or lost.
• Most organisations are buying intelligence while leaving memory behind
• Because organisations are discoverin that deploying intelligence and governing intelligence are fundamentally different challenges -and that the gap between them is where value is lost.
• The next question is not whether organisations will build Intelligence Architectures. The next question is what those architectures are built from — and who owns them.

7 more annotations...

• Après avoir bâti tout Copilot sur le moteur d'OpenAI, Microsoft dévoile cette semaine à Build 2026 ses propres modèles
• C'est ce qui a fait que certains ont exigé des plateformes entreprises dont l'étanchéité était garantie par contrat. D'autres ont préféré une étanchéité physique avec une plateforme séparée. Microsoft, avait le choix de sa stratégie Copilot, du moteur de recherche au tableur, en passant par Teams et GitHub, tout reposait intégralement sur les modèles GPT d'OpenAI. Un partenaire dont il était à la fois le premier investisseur, le premier client et, déjà, le concurrent silencieux.
• le géant de Redmond avait, a minima sous-traité son cerveau.
• Sur la scène de Build, le patron de Microsoft AI, Mustafa Suleyman, a dévoilé une famille de sept modèles maison. Ils sont regroupés sous le préfixe MAI et, c'est le point important, entraînés from scratch, avec les données et la propriété intellectuelle de Microsoft. C'est une rupture avec ChatGPT, entraîné par OpenAI, parfois avec des données dont le droit d'usage était en "zone grise"...
• Mustafa Suleyman a lui-même livré la clé : "Le moment décisif a été de renégocier notre contrat avec OpenAI", ce qui a autorisé Microsoft à entraîner des modèles à plus grande échelle, avec sa propre IP, sans distillation.
• Et oui chaque token de GitHub Copilot transitant par l'infrastructure d'inférence d'OpenAI, lui coûtait. Après, comme OpenAI achetait de l'Azure à Microsoft, on ne sait pas qui payait qui. Mais à la fin, les modèles maison sont d'abord une reconquête de marge. Et OpenAI a déjà repris sa liberté de confier ses inférences à d'autres hyperscaleurs.
• C'est une stratégie de plateforme, pas de produit. Et c'est exactement ce qui la rend redoutable.
• Satya Nadella a articulé l'ensemble de Build autour d'une seule formule : "agent-first", y compris pour Windows. Donc pas que Azure.
• Windows doit devenir un runtime pour des agents IA qui travaillent à votre place. Ce n'est pas une fonctionnalité de plus, mais une ré-architecture de la vision du poste de travail.
• Scout, d'abord. Un agent "toujours actif" qui opère à travers Teams, la messagerie, les calendriers, SharePoint et OneDrive.
  Microsoft le décrit moins comme un chatbot que comme une nouvelle identité de bureau : un logiciel qui agit entre vos réunions et vous représente en votre absence.
• Le runtime d'agents, ensuite, avec les Microsoft Execution Containers.
  C'est une couche d'exécution pilotée par une politique, qui déclare ce qu'un agent a le droit de faire, au sens de l'accès aux fichiers, aux réseaux…
• Agent 365, enfin, le tableau de bord censé gouverner tout cela à l'échelle. Et c'est là que Microsoft lâche un aveu intéressant : Agent 365 doit aussi reprendre en main les agents que vos équipes utilisent déjà localement sans l'autorisation de la DSI
• Le dispositif de sécurité présenté à Build tranche avec les débuts en mode "test & learn" de l'IA générative. Fini les boutons Copilot qui se multiplient partout,
• Chaque agent reçoit une identité Entra ID et hérite des mêmes contrôles d'accès basés sur les rôles qu'un utilisateur humain.
• Pour la première fois, l'agent est traité comme ce qu'il est réellement : une identité non-humaine qu'il faut authentifier, autoriser, tracer et révoquer. Microsoft a livré là un modèle de gouvernance que beaucoup de DSI n'avaient pas formalisé pour leur propre RPA.
• Elle est opérée par un seul fournisseur soumis au Cloud Act. Donc, là où ce dernier n'opérait que sur le Cloud, il va falloir maintenant regarder ce qui se passe sur le poste de travail. Et si on a déjà sa propre pile de gouvernance, comment superviser ces agents sans ajouter une couche de plus et se faire de nouveaux cheveux blancs.
• La vraie question est : suis-je prêt à ce que la couche de gouvernance de mes agents soit mono-fournisseur?
• Il s'agit de confier son référentiel d'identités, ses journaux d'audit, sa prévention de fuite de données et la cartographie complète de ses agents autonomes à un acteur unique.

• Le DSI français va être pris en tenaille :

  • d'un côté une pile agentique prête à l'emploi et richement outillée mais un renoncement plus fort à la souveraineté;
  • de l'autre une ambition souveraine sincère mais immature à l'instant, et un régulateur Européen qui hésite.

• Mais avant de déployer Scout ou Agent 365, trois réflexes sont peut être à inscrire à l'agenda du prochain Comité de Direction :

  • Savez-vous déjà cartographier les identités non-humaines comme on cartographie les humaines, avant qu'elles ne dépassent en nombre les comptes utilisateurs. Cela va être le sujet de votre prochaine gouvernance et d'une décision forte pour adopter Microsoft IA.
  • Posez, à votre commercial Microsoft favori, la question de la portabilité des logs d'audit vers un SIEM tiers, pour que la traçabilité exigée par l'AI Act et le RGPD ne soit pas captive.
  • Préparez-vous à débattre sur le fait de sanctuariser, ou non, le mode "suggestion uniquement". C'est une décision de gouvernance de votre Comité de Direction, qui peut réduire des risques mais aussi perdre les gains de productivité que Microsoft fera certainement miroiter à ses membres.

18 more annotations...

• The Dow was down, but only 1.35%. Chip companies (NVidia, Broadcom, Micron) GPU leasers (CoreWeave and Nebius) and some other major AI related companies (Oracle, Microsoft, Meta, etc) all took larger hits:

• Bailing out OpenAI is crony socialism and corruption (remember those $25M donations from Greg Brockman and his wife?), preventing capitalism from taking its natural course.
• As I noted on X, nobody is going to trust an American AI company that is partly owned by the US Government. Just the way the US doesn’t trust Huawei, Europe and Asia are not going to trust companies like OpenAI and maybe even Google

• But back to economics. Who can forget the hapless Meta, which may ultimately win the prize for most money burned vs least results?

• Meanwhile, it also came out yesterday that Musk is leasing another 110,000 GPUs to Google, on top of the 220,000 it is leasing to Anthropic.

   \"On June 5, 2026, we entered into a Cloud Service Agreement with Google with respect to access to compute capacity. The ","username":"SawyerMerritt","name":"Sawyer Merritt","profile_image_url":"https://pbs.substack.com/profile_images/1837744842715082752/xH9vYixL_normal.jpg","date":"2026-06-05T18:58:57.000Z","photos":[{"img_url":"https://pbs.substack.com/media/HKEmAM9WsAAsAQd.jpg","link_url":"https://t.co/hLFkJofdzi"}],"quoted_tweet":{},"reply_count":377,"retweet_count":870,"like_count":7735,"impression_count":1709678,"expanded_url":null,"video_url":null,"belowTheFold":true}">

  
• This year, SpaceX’s AI division is leasing GPUs left, right, and center, because they can’t figure out what to do with them.
• Whether SpaceX is making money on the deals with Google and Anthropic or losing money, they are waving the towel on winning the frontier model race— by arming their competitors rather than themselves.
• Elon clearly bought a lot of hardware he did not have demand for. He is probably not the only one that bought substantial capacity long before there was demand or a use case for it.”
• “Then of course the fact that even Google and Meta, who are [historically] very cash flow positive companies, need to sell equity to continue to fund their AI investments. Just shows you this whole thing is a black hole.”

7 more annotations...

• The report, released this week, found that the environmental cost of AI is being “systematically mismeasured” because current assessments focus on the carbon emissions from training large language models while overlooking the tech’s broader water and land footprint.
• The water footprint comes from cooling and powering the data centers, and the land footprint comes from the energy infrastructure and supply chains that go into building and running them.
• raining OpenAI’s GPT-4 model consumed up to 70 gigawatt-hours of electricity, for example. But running ChatGPT, it’s estimated, uses a monstrous 383 GWh from answering billions of prompts per day.
• Factoring in inference costs, data centers powering AI will use 945 terawatt-hours of electricity by 2030, the report found, which is triple the combined electricity use of Pakistan, Bangladesh, and Nigeria — which, altogether, are home to over 650 million people.
• By that same year, AI’s thirst will see it consume 9.3 trillion liters water — which is equal the basic annual water needs of all 1.3 billion people in Sub-Saharan Africa.
• “If we keep judging AI sustainability by carbon alone, we might think that renewables make AI infrastructure clean but that is solving one problem while creating other problems, often in places that didn’t ask for it,” she added.
• “A lot of people think that the environmental footprint of AI reduces, as technology improves and processes become more efficient. But that is only a partial picture of the overall problem,”

5 more annotations...

• Then we had The Times of London weighing in with its piece “Sack HR? Maybe not, but definitely cut its mission creep,” Harry Wallop noting that the top comment under the paper’s own article read: “HR are nothing but a retardant on the day-to-day activities of a company.”
• With so many jumping on the same bandwagon, something is clearly in the air. What we have to decide is whether it’s fresh air, hot air or smoke for mirrors.
• Many people don’t understand that the HR function only assumes the role a company’s leadership allocates it, or allows it to assume.
• When a company is doing well, you never hear anyone say “great HR,” but leadership teams receive accolades and massive bonuses.
• This is not to say that all criticism is invalid. Even Breslow Bolt’s CEO, suggested, albeit delivered with the finesse of a flamethrower, that HR adds most value in stable times and at scale, but that survival mode requires something different.
• It’s a context-specific argument, not a universal one. But his People Ops function of two, is still generous for a 100 person company, and hardly start-up lean.
• HR in the UK has nearly doubled in size in fifteen years, to 502,000 people. As the headcount grew, so did the mandate. “Head of people and culture” replaced “HR director.” Wellness, belonging, inclusion, culture-shaping, all of it landed in HR’s portfolio
• Wallop’s assertion is that too many in HR came to believe profits were secondary to a wider mission of fairness and inclusion.
• Ryan Breslow, for example, championed eliminating HR while omitting key context: Bolt had laid off 30% of its workforce, faced reports of disputes involving unpaid contractors, and became synonymous with the “Coldplay kiss-cam” scandal that exposed leadership indiscretions amid allegations of a toxic culture.
• Ironically, the HR function he eliminated was the very team responsible for managing many of these risks.
• Similarly, Uber’s cuts represented 23% of its People and Places division, encompassing HR, recruiting, workplace operations, and culture. This is the same company that became a case study in workplace toxicity,
• What is known is that reducing back-office functions is often applauded by markets, even when the longer-term organisational consequences remain hidden.
• In my long experience many managers find the “people” side of business tiresome. They frequently lack skills managing people and many organisations promote technical experts into leadership roles, without developing their interpersonal capabilities.
• In the US, 73.3% of HR managers are women, according to the Bureau of Labor Statistics. Globally, women hold approximately 72% of human resource positions. It is, by any measure, one of the most female-dominated professions in the corporate world.
• The language is telling: “infantilising,” “nannying,” “policing,” “mission creep,” a “productivity-killing machine.” These are not neutral words and carry the weight of cultural anxiety about who gets to set the rules, and what kind of authority is considered legitimate.
• This doesn’t mean the criticism is unwarranted, it can be at times. Any function that doesn’t serve an organisation’s mission is a problem, regardless of who runs it.

• We are living in times when trust in leadership is fragile and 58% of emlpoyees would trust a stranger before their boss. If that doesn’t give pause for thought, nothing will.
• The sad fact is workplaces need legislation to ensure that people are treated even half-way decently, and even with legal obligations to be met, workplace bullying remains widespread and sexual harassment continues to be underreported
• These situations are the lived reality in many organisations, and notably, they are exactly the kinds of problems that originally led to the creation of HR functions in the first place.
• : the biggest determinant of employee experience is the quality of immediate management, not the quality of HR policy. “People leave managers, not companies” is a cliché because it is true. If trust is declining, engagement is falling, and bullying persists, those are fundamentally leadership failures, not HR failures.
• Rather than asking “do we need HR?”, the better question is: what kind of HR do organisations need in a world where trust is low, engagement is falling, and workplace harm remains common?
• If sexism and harassment racism, bullying, burnout, and toxic leadership continue to affect workplaces, then expertise in people, culture, conflict, ethics, and inclusion becomes more important, not less
• What we are watching right now is not a reckoning with HR overreach, but a permission structure for founders to remove accountability and for managers to avoid scrutiny.
• Would this result in the turning back of clocks to a time where employee rights and benefits didn’t exist and an era when “move fast and break things” was the standaard operating procedur
• hat gets broken is trust, credibility and reltionships, all of which are harder to fix and have significant impact on the very things leaders want to protect - productivity and profit.

23 more annotations...

• The more organizations invest in AI adoption, the slower their decision-making becomes.
• This dynamic can be described as the innovation–bureaucracy paradox.
• The paradox is simple. The more uncertainty leaders face, the more control mechanisms they introduce. And the more control mechanisms they introduce, the less capable their organizations become of adapting to change.

• Each step feels responsible. Each additional voice feels like protection.

    

     
• Studies describe what is known as the threat-rigidity effect: When leaders face uncertainty or perceived threat, they instinctively centralize decisions, narrow information flows and rely more heavily on established procedures.

• they try to manage the new reality using the control systems designed for the old one.

   

• And once such bureaucratic processes are introduced, they often become the cultural norm: “This is how things are done around here.”

   

• Experienced skiers know the counterintuitive truth: The safest and most effective way to ski down a steep slope is to lean forward.

   
• this reaction often produces exactly what leaders fear most: more confusion, slower decisions, less ownership and ultimately, more bureaucracy.
• Instead of falling back into control, they must lean forward by encouraging ownership amidst uncertainty: clearer decision rights, stronger ownership and cultures that support responsible experimentation.
• They simplify decision processes, reduce approval layers and push ownership to the teams closest to the technology and the customer.

9 more annotations...

• Residents are packing themselves into local county meetings in incredible numbers and calling on their representatives to oppose gargantuan data center projects, developments that could cause electricity prices to spike, drain water supplies, and generate copious amounts of noise.
• Farmers are being hailed as heroes for rejecting millions of dollars to turn their land into data centers,
• at least seven in ten Americans would oppose a data center being built near their home
• 51 percent saying they were against having a data center project near their home,
• “The public has swung 49 points against data centers in just nine months,
• The widespread opposition has turned into a rare bipartisan issue, with conservatives and liberals joining sides to an astonishing degree.
• Younger voters were particularly strong in their opposition, with an overwhelming 83 percent
• Meanwhile, many rural Americans continue their fight while are struggling to have their voices heard

• Worse yet, the region is already facing a water crisis, a dire environmental predicament that could be made far worse by the development.

  Advertisement

7 more annotations...

• At the start of the gen AI boom, the scarcest asset seemed obvious: access to the frontier model.
• Now, beneath all of that, a new constraint is emerging: electricity. The new scarcity is not intelligence but the energy-intensive infrastructure required to produce and deliver it.
• Our research on competitive positioning and optimal distinctiveness and platform globalization and renewable-energy finance suggests that what’s happening today with AI fits a recurring pattern in how firms capture value during technological transitions
• . A new technology creates a scarce control point, and value then pools at that layer, because customers can’t get enough of it.
• The managerial mistake is to keep investing as if yesterday’s scarce layer will endure
• you have to ask what is becoming abundant, what is becoming standardized, and what bottleneck is forming underneath.
• The first era involved infrastructure: the key constraint was connectivity, and firms such as Cisco and AT&T harvested value by controlling the pipes
• The second era involved attention: the key constraint was discovery, and firms such as Google and Meta harvested value by organizing access to information, products, and people
• The third era involved intelligence: the key constraint was compute, and firms such as OpenAI, Anthropic, and Nvidia harvested value from frontier models, chips, and AI infrastructure
• Now the loop is moving into its fourth era: energy & physics. The key constraint in this new era is power, and the bottleneck is shifting away from digital intelligence and back into the physical world: electricity, cooling, land, and grid connections.
• This new era began when AI demand stopped being measured only in parameters, tokens, or cloud budgets and started being measured in megawatts.
• The clearest signal is how aggressively hyperscalers are now moving upstream into generation itself—signing 20-year nuclear-power purchase agreements, acquiring data-center sites adjacent to reactors, issuing RFPs for gigawatts of new power
• Intelligence is getting cheaper but expanding the number of economically viable uses, driving total demand up rather than down.
• Energy is not an input whose price can simply be renegotiated annually. It is local, permitted, slow to build, and politically contested.

• 1) Make energy intensity visible.

   

  Most firms track cloud spend, model accuracy, and AI adoption. Far fewer can answer a more basic question: How much electricity does a given AI workflow require?
• he goal is not perfect measurement. It is to make “intelligence per watt” a management metric, not an invisible engineering variable.

•  2) Reduce demand before buying supply.

   

  Many companies are overpaying for AI because they send too many tasks to models that are larger, faster, and more energy-intensive than the job requires.

• 3) Contract for optionality, not ownership.

   

  Microsoft has signed a 20-year power-purchase agreement with Constellation Energy that is intended to enable the restart of Three Mile Island Unit 1 as the Crane Clean Energy Center, adding roughly 835 megawatts of carbon-free electricity to the grid if the restart is completed.

• 4) Redesign where compute runs.

   

  Cloud-region selection used to be mainly about latency, compliance, and vendor architecture. It’s now also an energy decision. AWS’s acquisition of Talen’s data-center campus adjacent to the Susquehanna nuclear station, for example, and Google’s agreement with Kairos Power for advanced nuclear capacity show that hyperscalers are already organizing AI infrastructure around energy access.

• 5) Make someone accountable.

   

  Create a standing Compute and Energy Council, chaired jointly by the CIO, CFO, procurement leader, and sustainability or operations leader. Some companies are already moving in this direction: Salesforce has embedded sustainability metrics into its AI development process, and Microsoft provides cloud-emissions visibility and sustainability-governance tools that connect technology, finance, procurement, and sustainability decisions.
• As electricity becomes the new constraint, leaders thinking about AI strategy need to stop treating energy as a background input.
• . And they’ll need to recognize that in the next phase of competition, AI strategy and energy strategy will become inseparable.

20 more annotations...