This week compressed all the tension in the 2026 AI narrative. The largest IPO in history landed in the industrial layer. The harshest controls fell on the model layer. And in Shanghai, Indigo gave his first complete public walkthrough of how he reads this board. This issue twists the three threads into one judgment and leaves it for time to test.
2026.06.07 — 2026.06.14 · Once a week: identify the signals, recalibrate.
This Week's Signals
Three things crowded together within two days. On June 12, SpaceX completed the largest IPO in human history, with a first-day market cap settling at about $2.107 trillion. On June 13, Indigo gave his most complete public demonstration of his methodology to date, in Shanghai, using SpaceX — listed just the day before — as a live example. The same week, Anthropic announced export controls: frontier models Fable 5 / Mythos are barred to non-US persons. And only recently, Fable 5 had shown the first signs of separate pricing for frontier capability.
Most people will scroll past these as three unrelated news items. They actually share one structure. The model layer is being eaten by two forces at once. One is commoditization — distillation (training follower models cheaply on a frontier model's outputs), open source, and three to six companies doing the same thing with the same chips, making it hard for anyone to hold a premium. The other is politicization — export controls treat frontier models as military items; Karp has spent six months telling anyone who will listen that Palantir will eventually be nationalized; Vancouver saw an anti-AI march; a New Jersey data center drew protests; Maine banned data center construction outright; and only 20% of people remain optimistic about AI. Indigo's summary at the Workshop stings: AI evolves by the day, society digests by the year. SpaceX stands on the exact opposite side of this structure — it vertically integrates launch capacity, compute, and applications into one whole, selling shovels while mining its own ore, which naturally sidesteps the trap of value being captured by someone else.
Indigo's core judgment this week: the model layer is not where value ends up; it is the pipe value flows through — the real moats grow in vertical integration and the harness layer.
Trends
#01 One ticker, one vertical stack: SpaceX rewrote the grammar of valuation
On June 12, SpaceX listed at $135 per share, raising about $75 billion — a historical record. It closed the first day at $160.95, up 19.22%, for a market cap of about $2.107 trillion. Indigo wrote on X that day: Congratulations to SpaceX on its IPO 🚀 We have witnessed the largest IPO in human history. To the sea of stars — Infinite Scale
(original post). His breakdown nests three layers: Starlink's cash flow ($11.4 billion in 2025 revenue, 61% of the total) feeds the AI infrastructure (a $45 billion, three-year compute contract signed with Anthropic), with a long-dated option thrown in on top — orbital compute (compute in orbit, i.e. data centers built in space). The odds on that option are not conjured from nothing: Nebius founder Chernin disclosed that because of public backlash, roughly 40 out of every 100 ground data center applications fail to clear approval. SpaceX is not a rocket company. It is a vertically integrated stack running from launch capacity through compute to applications. But the same judgment carries a warning: the first-day $2.1 trillion is clearly above the $1.25 trillion combined valuation in the listing documents, which means the market has already priced in that long-dated option.
#02 The model layer's double erosion: beyond commoditization, politicization is now a variable of equal weight
The same week, Anthropic announced export controls: frontier models Fable 5 / Mythos are barred to non-US persons, including foreigners inside the United States. Indigo's reaction was blunt: Non-US persons, including those inside the US, banned from accessing Fable 5 / Mythos! This federal government has lost its mind 🫠
(original post). A few days earlier he was chasing another question: "What does this 'Included until June 22' for Fable 5 mean? Do subscribers only get to try it for free before June 22? Will it be charged separately after that?" (original post) — that post drew 44.5K views and 61 replies, his widest-reaching post of the week. The two events point at the same structure: separate pricing is a struggle to find pricing power under commoditization pressure; export controls are political power treating frontier capability as strategic material. Analyst Benedict Evans's telecom-carrier analogy footnotes the commoditization side: mobile data traffic grew 1,500-2,000x over twenty years, carriers built $1 trillion in revenue, and their stocks barely moved for two decades. Indigo's conclusion needs two entries: export controls are not simply bad news; they are a negative on the revenue side plus an endorsement on the moat side — the controls themselves acknowledge frontier models as strategic assets.
#03 Chatbots are a transitional form; the value is in harness and ontology
Two of Indigo's product observations this week point to the same judgment. The first is about Apple: Siri AI now harnesses the entire OS ✨ But the Siri app is still an outdated chatbot 😜
(original post). The harness (the software layer that plugs a model into real workflows and gives it context and tool-calling) is the real entry point; the chat box is just a shell. The second is self-evidence from the consumer side: NotebookLM is now powered by Gemini 3.5. I really do use this product more than the Gemini app now 😃
(original post) — the winner is the harness-type product, not the bare chat box. The enterprise counterpart is called ontology (encoding institutional knowledge and business processes into machine-readable structure): Palantir's deployment at Hertz covers 500,000 vehicles, 15,000 frontline employees, and 4 million customer touchpoints per month; each percentage point of recovered fleet utilization equals roughly tens of millions of dollars in annual revenue. Karp offers the strongest theoretical version: everything that can scale will eventually be commoditized; moats only grow where things cannot scale. The first question when evaluating any AI product is not how strong the model is, but whether it will be a chatbot headed for obsolescence or a harness with switching costs.
On the Ground
#04 Karp's counterweight
Massive token consumption is not productivity; it is digital masturbation — checking the weather, tagging every email, building yet another dashboard. Once companies actually run ROI (return on investment) audits, these uses will be the first to get cut. This was the only dissenting note in this week's not-enough-compute chorus, and it is worth recording precisely because it grates.
#05 The other side of penetration
Enterprise AI penetration is under 1%; the infrastructure layer is around 10% and sold out worldwide. Chernin added a line: the only use case that has truly worked is coding, and its history is only a few months old. To Sacerdote, AI adoption is not an S-curve heading for saturation but an L-curve with no visible top — which is exactly why, even with the model layer under pressure, he stays bullish on compute demand.
#06 The Jevons paradox, evidenced
The week DeepSeek launched, compute cloud provider Nebius fell 40% — yet that same week was the best sales week in the company's history. The market read cheaper
as not needed
; the truth is precisely the Jevons paradox (falling unit cost expands total consumption): the cheaper a unit of intelligence gets, the more people consume.
#07 Where earnings and stock prices diverge
Micron grew quarterly earnings 80%, far above expectations, yet the stock turned down — Loeb pins this on forced liquidations by quant and trend-following funds. Where earnings and prices diverge is exactly where human judgment still pays; it is really the other side of Karp's line that taste cannot scale.
#08 Capability is a function of inference budget
A Noam Brown piece drew 1.02 million reads on X: benchmark scores increasingly depend on inference budget (test-time compute — the compute a model spends while answering). The UK's AISI burned through 100 million tokens and the models were still not saturated. Capability has no natural ceiling, so compute demand has no natural ceiling either.
#09 Hardware is de-commoditizing
AI workloads grow 10x per year, pushing every link toward physical limits: AI server PCBs now reach 40 layers (ordinary servers have only 10), and DRAM/NAND/PCB shortages have hit 30%. But Sacerdote offers a ruler: benefiting is not the same as holding pricing power — watch the rate of change in AI sales share and category market share.
#10 When erosion becomes a punchline
Distillation squeezing the frontier premium has long been an open joke inside the industry. Reposting a meme, Indigo added only one line: From 1 to 100, driving the price down 😜
(original post). The moment a risk becomes a punchline, it stops being a hypothesis and becomes background noise.
Slow Thinking
Indigo made two clear reversals this week. The first concerns valuation method. In his L×T framework (a scoring method that places a ticker by industry layer L1–L5 and time bracket T1–T4), his habit was to assign each ticker a single cell. He now believes a species exists where one ticker equals one vertical stack: SpaceX holds Starlink's near-term cash flow, AI infrastructure's mid-term growth engine, and orbital compute's long-dated option all at once — the valuation should be split into parts and priced by different time brackets, not forced into one cell. The trigger was the June 12 IPO and the June 13 Shanghai Workshop live demonstration landing back to back — which also explains why the market was willing to pay $2.1 trillion on day one: the option part was priced separately. The second concerns the unit of scarcity. Since 2023 his default answer has been the GPU — get the cards and you win. This week at the Workshop he publicly changed it: the scarce thing is not GPUs, it is Gigawatts (GW, a unit of power capacity). By 2030 the industry will add 150 GW, a $7.5 trillion heavy-asset cycle, with annual investment at roughly 5% of US GDP. The trigger was the real build sequence Chernin described: lock in power and land first, then build the data center, and only then fill in the GPUs — NeoClouds (new cloud vendors specializing in AI compute rental) look more like developers, and model companies look more like tenants.
The other side: the strongest counterarguments this week came from Sacerdote and Loeb. Sacerdote lists five moats for Anthropic — critical intellectual property; enterprise brand (ask any CIO, a company's chief information officer, and the first answer is Claude); a harness ecosystem analogous to AWS in 2013; escape velocity; and recursive self-improvement — with the ARR (annual recurring revenue) trajectory from $100 million to $1 billion and on to $9 billion as evidence. Loeb argues this industry has barely dug beneath the surface, and that the difference from the dot-com bubble is that these companies invest off their own balance sheets and generate enormous amounts of real cash. If three things show up next — Fable 5 starts charging separately after June 22 and users pay without flinching; the capability gap between frontier models and distilled followers widens rather than narrows; enterprise procurement concentrates further toward single vendors — then the model-layer-erosion judgment should be overturned: pricing power lived in the model layer after all.
Indigo on X
Congratulations to SpaceX on its IPO 🚀 We have witnessed the largest IPO in human history. To the sea of stars — Infinite Scale
From @indigox, 38 likes
Non-US persons, including those inside the US, banned from accessing Fable 5 / Mythos! This federal government has lost its mind 🫠
From @indigox, 29 likes
"What does this 'Included until June 22' for Fable 5 mean? Do subscribers only get to try it for free before June 22? Will it be charged separately after that?"
From @indigox, 20 likes
Closing
One Thought
In the last infrastructure mania, telecom carriers grew mobile data traffic 1,500-2,000x and built $1 trillion in revenue, genuinely changing everyone's lives — yet their stocks barely moved for twenty years, because all the value migrated to the layers above. This AI round's annual investment runs at roughly 5% of US GDP, landing right between the railroad era's 6% and the dot-com bubble's 1.2%: railroad-scale investment, internet-speed rollout. History does not guarantee the ending repeats, but it hands us the right way to ask: the question is not whether the infrastructure gets built, but which layer the value stops at once it is. Who plays the carrier this round, and who plays the winner growing on top of the infrastructure — that is worth tracking every week, more than any index level.
One Exercise
Spend 30 minutes on a harness audit. Open the three AI products you use most, and for each one write down answers to three questions. One: has it accumulated context that belongs only to you — history, preferences, data? Two: can it call other tools on your behalf and act beyond the chat box? Three: if you switched it for a competitor, would the cost be five minutes or five weeks? Three no's: a chatbot headed for obsolescence. Three yes's: a harness with switching costs. When you are done, apply the same three questions to the products of any AI company you follow — it makes clear, faster than ten research reports, why this week's judgment is: the value sits in the layer above the chat box.