This week's material closes in on one question from two directions: where the money goes in the AI era. On the hardware side, Indigo draws a physical choke-point map across memory, packaging, optical interconnects, and electricity. On the work side, he splits knowledge work into three parts and finds AI has eaten only the middle one. The two threads meet at pricing power.
2026.06.14 — 2026.06.21 · Once a week — identify signals, recalibrate.
This Week's Signal
Start with the facts. Indigo's most widely shared post on X this week drew 134 likes. It made a counterintuitive ranking: Micron, a memory maker, will end up worth more than Meta and its huge capex (capital expenditure). The same week, he pinned the next choke point, in 2028, on packaging. Industry numbers happen to pile up in the same direction: JPMorgan estimates the AI power semiconductor market will grow from $2.7bn in 2025 to $16bn in 2028; NFX counts five hyperscalers (the largest cloud providers) reaching $700B in US data center capex in 2026; Broadcom's in-package optical interfaces are already shipping, with bandwidth climbing 50→100→200 Tb.
Most readers will read this as a list of hardware beneficiaries. What's easy to miss is the framework behind the list. First layer: a bottleneck is a state that moves — memory is today's choke point, packaging and power delivery are 2028's, and electricity is the ultimate constraint. Judging which layer things are stuck at right now matters far more than memorizing a string of company names. The second layer is the week's hidden axis: which layer ultimately captures the value. Satya Nadella bets value will spread across the ecosystem; SemiAnalysis points to concentration in the model labs (gross margins going from 38% to 70%+); Matan Grinberg holds that value capture drifts over time, with no steady-state winner-take-all. Only when you put these two axes together do you get both sides of the question where does the money go in the AI era.
In the AI era, money does not flow to the smartest model. It flows to the physical choke point that is hardest to route around right now — and the choke point shifts layers. Tracking the shifts matters far more than arguing about the endgame.
Currents
#01 Whoever Holds the Bottleneck Takes the Pricing Power
The week's most widely shared post starts from a thought experiment: if you had to bind yourself to one company for ten years, which would it be? Indigo wrote: "If you had to hold shares for ten years — SpaceX, OpenAI, or Anthropic — which would you pick? … Aravind's answer is SpaceX, and mine is the same — because SpaceX is the only one of its kind … Micron will be worth more than Meta! Because 'whoever is the bottleneck holds the pricing power' … As long as it is still the bottleneck, the price hasn't peaked; meanwhile the market doesn't credit the marginal returns Meta gets for its massive capex — this is also the Rewired Index's investment thesis" (original post). The judgment traces to an interview with Perplexity founder Aravind Srinivas, and the evidence chain behind it is dense: memory has already risen 5x on a cost basis (COGS, cost of goods sold); DRAM (the main category of memory chips) is projected to compound at 25.5% a year from 2024→2031, with the data center portion reaching 34.1%; compute throughput tracks almost exactly the line of HBM (high-bandwidth memory stacked next to AI chips) capacity times bandwidth. And electricity is the ultimate constraint — 40 out of 100 data centers can't get built because of public resistance. A bottleneck was never a fixed title; it's a state that moves. The real question is which layer of the physical chain value is stuck at right now.
#02 In 2028, the Choke Point Moves from Transistors to Packaging
This week Indigo nailed down both the timing and the location of the next choke point: "After 2028, what limits AI compute hardware is no longer the transistor — it's packaging! How to seal multiple dies into one 240mm monster block, feed in 4kW of power, then get the heat and the light back out … CPO brings light directly into the package (the default route for 2028) · >4kW direct liquid cooling goes from on-chip to on-silicon · power delivery pushed inside the package · panel-level interposers · traditional substrates can no longer support >200mm mega packages" (original post). A die is a bare chip; CPO means sealing optical communication components directly into the chip package. Silicon photonics expert John Bowers's judgment: once copper interconnects hit a physical wall at high frequencies, light goes from an option to an architectural necessity. Supporting evidence on the power side comes from JPMorgan: the AI power semiconductor market compounds at roughly 82% a year, yet in the 2028 split silicon still takes 70.1%, SiC (silicon carbide) 19.3%, GaN (gallium nitride) 10.6% — fast growth does not mean more dollars. What deserves watching is not any one component maker's growth rate, but the content-inflation curve: semiconductor content per kW rising from $175 toward $285.
#03 AI Eats Execution; Judgment and Accountability Survive at Both Ends
Indigo splits knowledge work into three parts: deciding what to do, executing, and delivering with accountability. On X he said: AI has compressed execution, but judgment and accountability you have to do yourself! Once a decision can be handed to AI, it is no longer an advantage, and value migrates up one layer. Writing code was never the bottleneck — deciding what to write and being accountable for the outcome is … The Jevons paradox increases derived demand; demand for software engineers may rise in the future
(original post). The Jevons paradox: efficiency gains end up amplifying total usage. On the empirical side, Anthropic's study of 400,000 agent (AI that executes multi-step tasks autonomously) coding sessions shows: humans kept about 70% of planning decisions but only about 20% of execution decisions; what decided success was domain understanding, not programming background — success rates across the top ten occupations all fell within 7 percentage points of software engineers, and management occupations actually ranked highest. The reverse calibration also has hard data behind it: 59% of hiring managers admit using AI as a layoff excuse; in an HBR survey, 21% expectation matched 2% reality; of 25,000 laid-off workers in New York State, only 46 checked AI as the reason. What's scarce was never execution. It's judgment and accountability.
On the Ground
#04 An Immediate Signal of Token Commoditization
Indigo said on X: Claude -p is no longer billed separately — it stays included in the subscription plans! Is this compute being sufficient and owning up to a mistake? Of course — any longer without fixing it and the users would have all run off
(original post). Claude -p is Claude's programmatic invocation mode. Tokens (the usage-billed unit of model output) are commoditizing; the business of purely reselling tokens is running out of ground.
#05 The Second Term in the White-Collar Value Formula Goes to Zero
In a post quoting his own podcast, Indigo Talk EP49, he gave a formula: White-collar value = time × domain knowledge × relationships; large models knock the second term down to near zero, flattening junior and mid-level workers
(original post). This is the plain-language version of the three-part split in #03: once execution is leveled, value can only migrate toward taste, relationships, and accountability.
#06 All Companies Are Software Companies
He also quoted an assertion from @turingou: "In the end, all companies are 'software' companies" (original post). When AI pushes execution costs toward zero, what separates companies is no longer whether they use AI, but whether their core processes can be rewritten as software.
#07 Is AI Like Electricity or Like Social Media
Dwarkesh poses a top-level test: if AI is like electricity, value goes to downstream users; if it's like social media, the rents go to the platform. The deciding variable is how far open source trails the frontier — 6-9 months would make it look like electricity; the current measurement is about 4 months. Which layer captures the value is not settled; it's a variable that drifts over time.
#08 The Harness Balance Sheet
The harness (the software layer that plugs models into enterprise workflows) was stress-tested from both sides this week: the case for holds that the separation of generation and verification in the loop contributes most of the quality, and the only metric worth chasing is cost per accepted change; the case against warns that if the harness merely patches the holes of the previous model generation, it flips from asset to liability once the next generation trains those capabilities in. Moats shift layers too — this is the bottleneck logic mirrored in the software layer.
#09 The Same Law on the Battlefield
Schmidt's judgment in Noema offers an extreme sample: in March 2026, Ukrainian drones caused 96% of Russian casualties; Russia is targeting daily production of 1,000 Shaheds, while Lockheed produces only 600 interceptors a year. Victory never came down to how refined a single weapon is, but to loop speed and unit cost — bottleneck logic holds in war too.
Slow Thinking
Indigo did revise one judgment this week. The previous default narrative was AI speeding everything up — the whole chain, from R&D to deployment, being compressed. The current version is narrower: AI compresses only the middle segment, execution and design. The two ends — judgment and accountability — plus external physics, institutions, and clinical timelines are all still standing where they were. The current density of breakthroughs looks more like the lagged payoff of decades of past investment than AI's immediate output. The trigger was three independent pieces of evidence converging in the same week: a review in biology points out that this month's breakthroughs are really a lagging indicator of decades of funding, while the US is cutting roughly 80% of NCI/NHLBI budgets — a cost that will only show up years later as missing breakthroughs; on the semiconductor side, new wafer capacity won't take effect until 2030-32; drug pipelines still take 7-8 years. Fei-Fei Li's reminder is another calibration: she explicitly rejects the claim that the cost of intelligence goes to zero — language is only a lossy first-generation interface, while spatial intelligence took 500 million years of evolution to get here. Before pricing any AI-acceleration narrative, first work out which segment is actually being compressed: design and manufacturing, or clinical trials, capacity, and grid connection.
The other side: the strongest rebuttal to the bottleneck thesis is cyclicality. Memory is the most cyclical part of semiconductors, and a 5x cost rise is itself an invitation to expand capacity; once supply floods in, pricing power disappears just as fast. Most of the people calling bottleneck this week are also interested parties — Lip-Bu Tan is both Intel's CEO and a bettor on the outcome, and Bowers himself admits he was 25 years early on silicon photonics. The deepest hidden variable in hard-tech investing was never right versus wrong; it's timing. The falsification signals are also clear: memory prices falling back after capacity is released, the market starting to credit the marginal returns on Meta's capex, or the CPO 2028 timeline slipping again — if any one of these appears, this thesis should be downgraded.
Indigo on X
"If you had to hold shares for ten years — SpaceX, OpenAI, or Anthropic — which would you pick? … Aravind's answer is SpaceX, and mine is the same — because SpaceX is the only one of its kind … Micron will be worth more than Meta! Because 'whoever is the bottleneck holds the pricing power' … As long as it is still the bottleneck, the price hasn't peaked; meanwhile the market doesn't credit the marginal returns Meta gets for its massive capex — this is also the Rewired Index's investment thesis"
From @indigox, 134 likes
"The structural-vs-cyclical test: layoffs while earnings look great mean the job's historical mission has ended … White-collar value = time × domain knowledge × relationships; large models knock the second term down to near zero, flattening junior and mid-level workers … Two career paths, no third: burrow into the tip of the pyramid and work on the models themselves, or go horizontal as a Builder/Architect with 'program taste' — the CRUD door has closed … The window-period alpha: rebuilding everything for agents (UI/UX, payments, verification, workflow) is fleeting"
From @indigox, 76 likes
"After 2028, what limits AI compute hardware is no longer the transistor — it's packaging! How to seal multiple dies into one 240mm monster block, feed in 4kW of power, then get the heat and the light back out … CPO brings light directly into the package (the default route for 2028) · >4kW direct liquid cooling goes from on-chip to on-silicon · power delivery pushed inside the package · panel-level interposers · traditional substrates can no longer support >200mm mega packages"
From @indigox, 68 likes
Closing
One Thought
The 19th-century railroad era ran the same script: in the early mania, the bottleneck sat at locomotives and steel rails, and the steel segment held pricing power; once rail capacity caught up, the bottleneck moved to land, rights-of-way, and financing, and pricing power changed hands with it. In every round of technology mania, the big money usually goes not to the layer shouting the loudest narrative, but to the link hardest to route around at the time — and that link changes every few years. Today's memory, 2028's packaging, electricity further out: rather than arguing over who the eternal winner is, make the layer-shift itself the object of study. Bowers's 25 years applies equally to the railroad pioneers of that era: a correct thesis never guarantees a payoff; the payoff answers to the physical schedule.
One Exercise
Spend 30 minutes on a choke-point inventory: pick an AI company you follow, or your own team, and answer three questions on paper. First, which segment is its revenue anchored to? Split what it delivers into deciding what to do, executing, and delivering with accountability, and see whether what it occupies is execution or judgment. Second, if the underlying model were swapped for an open-source version tomorrow, what would it lose? The less it loses, the more real what it owns is — private evals and institutional memory both count. Third, which layer of physical bottleneck does it depend on right now — memory, packaging, or electricity — and when that layer shifts, is it a beneficiary or a casualty? Write the answers down and read them again against Indigo's updates next issue.