• The companies that survive the next decade will not be those with the best algorithms or the most data. They will be those that have the courage to abandon how decisions get made.
• Consensus management was, in its time, a rational response to complexity.
• The first is speed. Decisions pass through gaggles of lawyers, marketing teams, PR, investor relations, risk managers, and compliance, each of whom is incentivized to mitigate personal and organizational risk.
• The second is that it distorts information. In our work across dozens of corporate turnarounds, the most dangerous threat to a company was often the systematic filtering of reality as information travels upward through layers of management.
• AI pushes us into rougher waters and turns both weaknesses into critical liabilities. On the first, because it rapidly accelerates the speed at which companies can—and must—operate

• The Architecture of Speed

   

  Leaders need to redesign their organizations around a different decision-making architecture, and to invest in advanced AI-enabled information systems.

• The “Autonomous Scrum”

   

  Most organizations give small teams permission to recommend. High-velocity organizations give them permission to act. This, of course, requires rebalancing priorities in addition to changing the decision-making architecture.
• The risks of this architecture producing a suboptimal decision pale against the paralysis of endless deliberation. AI agents equip the Autonomous Scrum with an army of savants, deep subject matter expertise, and almost instantaneous feedback on what actions are producing positive results.
• We’ve had success with a different approach: the OVIS framework. One person Owns the decision. Two or three people Veto or Influence it. Everyone else Supports the outcome.

• What Boards Must Demand

   

  While the CEO and her management team must execute these changes, it is unlikely to happen without a Board of true believers—devoid of the discomfort of radical mandates
• Boards can’t continue to operate as they have in the pre-AI era. If they’re working off filtered, committee-approved reports from the C-suite they are not exercising governance—they are perpetuating the distortion field.
• The executives who thrive in this environment will share a particular disposition. They will be comfortable— even excited—making consequential decisions on incomplete information. They will trust signals over instinct, speed over process, and small teams over consensus. They will demand a world of continuous feedback loops and recognize that the greatest sin is not making the wrong call but making no call at all.
• The Great AI Reorganization will not be won by companies with the most compute or the deepest pockets. It will be won by those with the clarity to see what the new environment demands, and the courage to build it.

11 more annotations...

• The only way out of that trap is a decision. A real one. Not a committee consensus, not a hedged bet, not a roadmap that tries to please every stakeholder in the room. A genuine, uncomfortable, this-is-who-we-are-and-this-is-what-we’re-building decision.
• But there is something specific that the statistical model underneath every LLM is not designed to do. It cannot generate exceptions. That’s not a criticism. It’s just math.

• It finds the most statistically significant answer to your question based on everything it was trained on. Everything that already exists. Everything that has already been built, written, described, and shipped by every company that came before you.

  Including all of those copied playbooks.
• The model learned from an industry that was already converging. It was trained on the teardowns, the frameworks, and the “what’s working right now” content that spread because something worked once, and then everyone grabbed it.
• That happened before anyone was prompting a model. It happened because the moment something worked, everyone noticed. Someone wrote a teardown. Someone posted the conversion numbers. Someone gave a conference talk about it. And then seventeen other teams rebuilt the same thing because it felt like the safe bet.
• When you ask the model what you should build, it reaches into that pile and hands you the consensus.
• The convergence doesn’t slow down with AI. It just stops requiring any effort at all.
• When your product decisions are coming out of a model, nobody has to own them. Nobody planted the flag. Nobody said “this is the bet we’re making and here’s why I believe in it.” The model said so. And the model can’t be held accountable.

• The LLM doesn’t know your customers the way you know your customers.

  • It doesn’t know the conversation you had with a client last Tuesday where they told you something that completely reframed how you understood their problem.

  • It doesn’t know the weird strength your team has that nobody else in your market has figured out yet.

  • It doesn’t know the contrarian bet you’ve been sitting on for a year because you haven’t had the nerve to make it.
• Saying out loud: we are going to be the best in the world at this specific thing for this specific kind of customer, and we are going to say no to everything that pulls us away from that.

• Humans generate exceptions. Always have. That’s not going away.

  ---
• When was the last time your product roadmap surprised someone? Not overwhelmed them. Not confused them. Surprised them in the way that makes a customer say “I didn’t know I needed that until you built it.”
• Because the market doesn’t need another version of what already exists.

11 more annotations...

• Anthropic’s Products Are Constantly Breaking Because It Doesn’t Have Enough Capacity, And Opus 4.7 Is Both Worse and Burns More Tokens
• This is a company that recently raised $30 billion dollars and gets talked about like somebody’s gifted child, yet Anthropic’s services seem to have constant uptime issues linked to a lack of capacity. 
• Since mid-February, outages for systems across Anthropic have become so common that some of its enterprise clients are switching to other AI model players. 
• In that post, Laurenzo wrote that Claude Code had regressed to the point that it could not be trusted for complex engineering work, then backed that claim with a sprawling analysis of 6,852 Claude Code session files, 17,871 thinking blocks and 234,760 tool calls.
• While Anthropic claims that it doesn’t degrade models to better serve demand, that doesn’t really square with the many, many users complaining about the problem
• I think it’s deeply bizarre that a huge company allegedly worth hundreds of billions of dollars A) can’t seem to keep its services online with any level of consistency, B) appears to be making its products worse, and C) refuses to actually address or discuss the problem
• Anthropic Has No Good Solutions To Its Capacity Issues And Shouldn’t Be Accepting New Customers — And More Capacity Will Only Lose It Money
• It is blatantly, alarmingly obvious that Anthropic cannot afford to provide a stable and reliable service to its customers, and its plans to expand capacity appear to be signing deals with Broadcom that will come online “starting in 2027,” near-theoretical capacity with Hut8, which does not appear to have ever built an AI data center, and also with CoreWeave, a company that is yet to build the full capacity for its 2025 deals with OpenAI and only has around 850MW of “active power capacity” — so around 653MW of actual compute capacity — as of the end of 2025, up from 360MW of power at end of 2024.   
• We’re likely hitting the absolute physical limits of available AI compute capacity, if we haven’t already done so,
• Yet it’s unclear whether “more capacity” means that things will be cheaper, or better, or just a way of Anthropic scaling an increasingly-shittier experience. 
• it doesn’t mean that they start turning away business or stop accepting subscribers, but that current (and new) subscribers will face randomized downtime and model issues, along with increasingly-punishing rate limits. 
• Anthropic is currently losing billions of dollars a year offering a service with amateurish availability and oscillating quality, and continues to accept new subscribers, meaning that capacity issues are not affecting its growth.
• As a result, adding more capacity simply makes the product work better for a much higher cost.
• Anthropic’s growth story is a sham built on selling subscriptions that let users burn anywhere from $8 to $13.50 for every dollar of subscription revenue and providing a brittle, inconsistent service, made possible only through a near-infinite stream of venture capital money and infrastructure providers footing the bill for data center construction.
• If Anthropic were forced to charge its actual costs — and no, I don’t believe its API is profitable no matter how many people misread Dario Amodei’s interview — its growth would quickly fall apart as customers faced the real costs of AI (which I’ll get to in a bit).

• Claude Mythos Was Held Back Due To Capacity Constraints, Not Fears Around Capabilities

• but my general feeling is that if Mythos was so powerful, how did Claude Code’s source code leak? 

  Did… Anthropic not bother to use its super-powerful Mythos model to check? Or did it not find anything? Either way, very embarrassing for all involved. 
• AI Compute Demand Is Being Inflated By Anthropic and OpenAI, With More Than 50% of AI Data Centers Under Construction Built For Two Companies, and Only 15.2GW of Capacity Under Construction Through The End of 2028
• As I’ve discussed in the past, only 5GW of AI compute capacity is currently under construction worldwide
• reached out to Sightline to get some clarity, and they told me that of the 114GW of capacity due to come online by the end of 2028, only 15.2GW is under construction, including the 5GW due in 2026. 
• So, to summarize, at least 4.6GW of the 15.2GW of data center capacity under construction is for OpenAI, with at least another 4GW of that reserved for Anthropic through partners like Microsoft, Google and Amazon. In truth, the number could be much higher. 
• The only way that these companies can continue to exist is by raising endless venture capital funding or, assuming they make it to IPO, endless debt offerings or at-the-market stock sales.

• NVIDIA Claims To Have $1 Trillion In Sales Visibility Through 2027, But Only $285 Billion GPUs Worth Of Data Centers Are Under Construction — NVIDIA Is Selling Years’ Worth of GPUs In Advance And Warehousing Them

• It’s also very concerning that only such a small percentage of announced compute capacity is being built, especially when you run the numbers against NVIDIA’s actual sales.
• NVIDIA claims it had a combined $500 billion in orders between 2025 and 2026, and $1 trillion of sales through 2027, and it’s unclear where any of those orders are meant to go other than a warehouse in Taiwan. 
• At this point, I think it’s fair to ask why anyone is buying more GPUs, as there’s nowhere to fucking put them. Every beat-and-raise earnings from NVIDIA is now deeply suspicious. 

• AI Is Really Expensive, With Companies Spending As Much As 10% Of Headcount Cost On LLM Tokens, And May Reach 100% of Headcount Cost In The Next Few Quarters
• Last week, a report from Goldman Sachs revealed that (and I quote) “...companies are overrunning their initial budgets for inference by orders of magnitude (we heard one industry datapoint on inference costs in engineering now approaching about 10% of headcount cost,
• To simplify, this means that some companies are spending as much as 10% of the cost of their employees on generative AI services, all without appearing to provide any stability, quality or efficiency gains, or (not that I want this) justification to lay people off. 
• Uber’s surging use of AI coding tools, particularly Anthropic’s Claude Code, has maxed out its full year AI budget just a few months into 2026, according to chief technology officer Praveen Neppalli Naga.
• Uber’s CTO also added that about “...11% of real, live updates to the code in its backend systems are being written by AI agents
• et whenever you ask somebody about it the answer is either “yeah, we’re writing all the code with it!” without any described benefits or “it costs so much fucking money, man.” 

• The Subprime AI Crisis Continues, With Microsoft Starting Token-Based Billing For GitHub Copilot Later This Year, And Anthropic Already Moving Enterprise Customers To API Rates
• As I reported yesterday, internal documents have revealed that Microsoft plans to temporarily suspend individual account signups to its GitHub Copilot coding product,
• This is a significant move, driven by (per my own reporting) Microsoft’s week-over-week costs of running GitHub Copilot nearly doubling since January.
• The move to token-based billing will see GitHub users charged based on their usage of the platform, and how many tokens their prompts consume —
• I’ll also say that the fact that Microsoft has stopped signing up new paid GitHub Copilot subscriptions entirely is one of the most shocking moves in the history of software.
• Microsoft is one of the few players capable of bankrolling AI in perpetuity, with over $20 billion a quarter in profits since the middle of 2023. 
• In any case, if Microsoft’s making this move, it means that CFO Amy Hood — the woman behind last year’s pullback on data center construction — has decided that the subsidy party is over. Though Microsoft is yet to formally announce the move to Token-Based Billing, I imagine it’ll be sometime this week that it rips off the bandage
• “AI agents make it so every other company on the planet starts to create software for bringing automation to their workflows in a way that would be either infeasible technically or unaffordable economically
• He is, however, increasingly-indicative of the hysterical brainrot of executive AI hysteria, where the only way to discuss the industry is in vaguely futuristic-sounding terms about “agents” and “inference” and “tokens as a commodity,”
• AI revenues are dogshit. They’re awful. They’re pathetic. The entire industry — including OpenAI and Anthropic’s theoretical revenues of $13.1 billion and $4.5 billion — hit around $65 billion last year, and that includes the revenues from providing compute generated by neoclouds like CoreWeave and hyperscalers like Microsoft.
• : I think the AI startups are misleading their investors and the general public about their revenues.
• Even if said numbers were real, the majority of OpenAI, Cursor and Anthropic’s revenues come from subsidized software subscriptions. Things have gotten so dire that even Deidre Bosa of CNBC agrees with me that AI demand is inflated by token-maxxing and subsidized services.
• overstating annualized revenues is extremely common, with AI startups booking “three-year-long” enterprise deals with the first year discounted and a twelve-month out:
• The setup: Company signs 3-year enterprise deals. Year 1 is discounted (say $1M), Year 2 steps up ($2M), Year 3 is full price ($3M). 
  
  They report $3M as “ARR” — even though they’re only collecting $1M right now.
• , Stevenson estimates that more than 50% of enterprise AI startups are using “contracted ARR” to pump their values.
• And at the top of the pile sits NVIDIA, the largest company on the stock market, which is selling more GPUs than can be possibly installed, and very few people seem to notice or care. 
• Where are you going to put them, Jensen? Where do the fucking GPUs go? There isn’t enough capacity under construction!
• is totally divorced from reality, discussing “agents” in terms that don’t match with reality and AI data centers in terms of “gigawatts” that are entirely fucking theoretical, all with a terrifying certainty that makes me wonder what it is I’m missing.

48 more annotations...