This week's material spans labor economics, capital structures, and semiconductor physics. On the surface, unrelated. Underneath, one question: at which layer is the value AI creates being captured? Indigo's call this week: the answer is not in the models. It is in the least glamorous places — how organizations are built, how companies are structured, and how packaging is done.
2026.05.03 — 2026.05.10 · Once a week: spot the signals, recalibrate.
This Week's Signals
Start with the facts. Stanford's WORKBank study (a task-level survey of worker preferences covering 1,500 workers and 844 tasks) found that 41% of the money in the YC (startup accelerator Y Combinator) portfolio is bet in the wrong direction — what workers want does not match where that money is placed. Microsoft's 2026 Work Trend Index measured an even sharper ratio: among the factors that decide how fast AI spreads, organizations account for 67% and individuals only 32%. The same week, Anthropic joined Blackstone, Hellman & Friedman and Goldman Sachs to form a $1.5B AI services joint venture, and OpenAI launched the $10B The Development Company on the same day. Anthropic also signed a deal with SpaceX to add over 300 megawatts of compute capacity within a month. And TSMC, at its technology forum, drew its advanced packaging roadmap all the way out to 2029.
Most people would treat these five items as five unrelated news stories: one labor study, two deals, one roadmap. But Indigo spent the week hammering the same nail — the value AI creates is leaving the most visible layer (models, jobs, wafers) and sinking into the most structural layer (how organizations are built, how companies are structured, how packaging and power get supplied). The line between winners and losers was never about how strong the model is. It is about whether the organization can absorb it: the model's capability is already sitting there, and the side that cannot absorb it gets repriced first.
Every thread this week lands in the same place: the first domino of the AI revolution falls on software valuations and organizational structure, not on the employment statistics.
Winds
#01 What's being replaced isn't the job — it's the software the job runs on
Stanford's WORKBank used a survey of 1,500 workers across 844 tasks to draw a four-quadrant map: 45.2% of occupations want equal collaboration with AI, and only 1 occupation wants full automation — editors. Indigo's verdict on the study is blunt: "Stanford used 1,500 workers and 844 tasks to tell YC: 41% of your money went in the wrong direction … which means 'programmers replaced by AI' may be a false premise (original post). A harder signal comes from the US Census Bureau: among companies using AI, 16% used it to replace existing software and only 2% used it to replace employees — the former is 8 times the latter. The first thing getting eaten is not the job. It is the SaaS (subscription enterprise software) that job is currently using. The application-layer question changes direction too: no longer
which occupation disappears," but which company can turn the signals users leave behind into training fuel for its own model.
#02 The two-tier company: frontier model labs reshape themselves for IPO
This week, Anthropic announced a $1.5B AI services joint venture (JV) with Blackstone, Hellman & Friedman and Goldman Sachs, and OpenAI launched the $10B The Development Company on the same day. Indigo's judgment: "Anthropic, Blackstone, Hellman & Friedman and Goldman Sachs announced a new AI services (JV) company … The frontier AI companies of 2026 are no longer just model companies — they are using PE capital and the Palantir model to remake themselves into a two-tier structure of 'high-margin software parent + independent services subsidiary,' tailor-made for the period around an IPO roadshow" (original post). The logic is simple: hand the FDE business (forward-deployed engineers stationed at customer sites doing implementation) to PE (private equity funds) — these services run margins of only 30-50%, while software can run 80%+ — and the parent's financials get to tell a clean, pure-software story. The background numbers are just as striking: Anthropic raised $50B in 48 hours at a $900B valuation, and inference margins are climbing from 38% toward 70%+. This is not just a funding headline. It looks like a dress rehearsal for an IPO roadshow.
#03 The limit on compute isn't the wafer — it's packaging, memory, and power
TSMC's technology forum laid out this roadmap: the interposer (the silicon base that carries multiple chips) grows from 3.3 reticle areas to more than 14, HBM (high-bandwidth memory, the fast memory stacked next to the GPU) per package goes from 8 stacks to 24, and the whole semiconductor industry passes $1.5T by 2030. Indigo's read: the packaging envelope stretches to 2029 … this is the physical feasibility boundary for NVIDIA Rubin → Rubin Ultra → Feynman … the gap between 48× compute growth vs 34× memory bandwidth growth can only be filled by HBM generational leaps + CPO adoption
(original post). Here envelope
means the physical size limit of a single chip package; Rubin through Feynman are the codenames of NVIDIA's next chip generations; CPO (co-packaged optics) means building optical communication parts directly into the chip package. Meanwhile, on the ground, the power grid has already hit the wall first: the US PJM grid ran a 6.6GW shortfall in December 2025, and transformer lead times stretch to 128 weeks. Value is migrating from the wafer stage of chipmaking to the scarcer stages: packaging, memory, and power.
On the Ground
#04 AI adoption is stuck at management, not at employees
The bottleneck in AI adoption is not whether employees are willing. It is whether management acts. Indigo wrote on X: Of what affects AI penetration, organizational factors account for 67%, individual factors only 32%! … That 10% of blocked agency is the key … the cheapest intervention is for managers to start using AI themselves — it improves results by 17-30%
(original post). Blocked agency
means employees who want to use AI but are stopped by organizational process — the lever for change sits on the organization's side, not the individual's.
#05 The valuable half of Palantir is called the ontology
What Palantir sells is not software. It is the ability to turn a company's experience and know-how into a structure machines can read. Indigo's comment: "Palantir's AIP platform is a great example of end-to-end Agent architecture! The Ontology is the core of enterprise know-how — without it they really are just a large systems-deployment company" (original post). An Agent is an AI agent that executes tasks on its own; the ontology is exactly the half that stays with the software parent in the two-tier company structure.
#06 The list of valuable skills is being rewritten
When AI takes over information processing, the system companies use to price skills has to be rewritten too. He said on X: "Core human skills are shifting from 'information processing' to 'interpersonal relationships' … as AI erodes information-processing ability, AI-native HCM will redefine 'what is valuable' inside a company" (original post). HCM is human capital management software — and this points the same way as the Stanford study's finding that interpersonal and organizational skills are gaining value.
#07 Between efficiency and output sits the old organization
Efficiency went up, output did not follow — the layer in between is the thickness of the old organizational structure. Indigo's summary: Driving skills can atrophy; thinking skills cannot … 10x efficiency has not brought 10x output — the losses are all inside old organizational structures; the middle class is collapsing
(original post). That sentence compresses the week's three threads — labor, organization, capital — into a single warning.
#08 Compute competition enters the alliance era
Compute competition has entered a geopolitics-style alliance phase. Indigo wrote: The enemy of my enemy is my friend … Anthropic signed an agreement with SpaceX … gaining over 300 megawatts of new capacity within this month (more than 220,000 NVIDIA GPUs)
(original post). When power and transformers become scarce, the compute map gets redrawn along alliance lines.
#09 Enterprise AI coding crosses the threshold
Enterprise AI coding is no longer a pilot project. It is standard issue for everyone. All 23,000 engineers at Mercado Libre are using Claude Code; Stripe migrated 50K lines of Scala code to Java in just 4 days, 25 times faster than planned. When adoption is measured in units of the whole staff,
the main tension shifts from a technical problem to an organizational one.
#10 The courts split the US and China into two tempos
AI commercialization in the US and China is being cut into two entirely different tempos by their legal systems. Court rulings in both Hangzhou and Beijing established the same principle: deploying AI is a strategic choice the company makes on its own, and does not qualify as the major change in objective circumstances
described in Article 40 of the Labor Contract Law. At-will employment in the US lets companies replace labor aggressively; judicial constraints in China force a slower pace — the same technology, two entirely different curves of organizational impact.
Slow Thinking
Indigo made one clear revision this week. Earlier, he had read the Anthropic–Blackstone joint venture as a distribution story aimed at the mid-market. Now he has upgraded it into a full paradigm: pre-IPO margin protection, mid-market expansion, and a compliance moat, three legs stacked together — the low-margin services business spun off the books, the software parent traveling light with a clean story. What triggered the revision was a run of consecutive signals: the same day, OpenAI also formed the $10B Development Company; right after, Anthropic struck the compute deal with SpaceX; the person representing Anthropic was CFO Krishna Rao, not Dario; and Goldman Sachs appears on Anthropic's list while none of OpenAI's 19 investors is an investment bank — that contrast turned tailor-made for IPO
from a guess into a working hypothesis. One smaller update: he previously had only a vague intuition that large language models lack a sense of time and long-term memory; this week he turned it into an explicit thesis — and Claude's release of Dreams the same week (a mechanism for organizing and improving memory offline) supplied the engineering counterpart.
The other side: the strongest argument against this week's theme is that software taking the hit before jobs may just be an illusion created by a time lag. The Census Bureau's 16% versus 2% ratio may only show that replacing software is faster than replacing people, not that the latter will not happen. In the cases Galloway relays — investment analyst teams cut from 5 to 1, administrative assistants from 10 to 3, junior lawyers cut by a third — the employment shock is not theoretical. And Stanford measured what workers want,
not what employers choose
— the two can diverge. The falsification path is clear as well: if over the next few quarters the revenue and net retention (the rate at which existing customers renew and expand) of enterprise software companies do not systematically deteriorate — Sierra's 105x revenue multiple is a living counterexample — while white-collar job data deteriorates first, then this week's call should be written the other way around.
Indigo on X
"Stanford used 1,500 workers and 844 tasks to tell YC: 41% of your money went in the wrong direction … which means 'programmers replaced by AI' may be a false premise"
From @indigox, 195 likes
"Anthropic, Blackstone, Hellman & Friedman and Goldman Sachs announced a new AI services (JV) company … The frontier AI companies of 2026 are no longer just model companies — they are using PE capital and the Palantir model to remake themselves into a two-tier structure of 'high-margin software parent + independent services subsidiary,' tailor-made for the period around an IPO roadshow"
From @indigox, 170 likes
"Palantir's AIP platform is a great example of end-to-end Agent architecture! The Ontology is the core of enterprise know-how — without it they really are just a large systems-deployment company"
From @indigox, 164 likes
Closing
One Thought
The last general-purpose technology revolution was electrification. For the first few decades after factories installed electric motors, the productivity dividend failed to arrive — because factory floors were still laid out around the central drive-shaft logic of the steam era. Only when factories were rebuilt around the logic of the electric motor were the gains released. Today's counterpart is plain to see: Microsoft measured organizational factors at 67% of what drives AI penetration; and when Chesky flattened Airbnb's management layers, his reference point was the Catholic Church — across two thousand years of history it has kept just 4 layers. Once the technology is in place, the bottleneck is never the technology itself. It is the organization that grew up around the old technology.
One Thing to Try
Spend 30 minutes building your own WORKBank: break your job into 10 to 15 concrete tasks, give each task two scores — how deeply you want AI involved, and how well AI can actually do it today — then plot them on four quadrants. Count the tasks that land in the "you want to hand it to AI but it can't do it well yet" quadrant. That square is where new tools are most likely to appear in your industry. Then add one more observation: does your direct manager use AI openly? By the data Indigo cites, that single intervention alone raises a team's AI adoption results by 17-30%.