This week's material comes from the close of Indigo's three-week Silicon Valley trip. For the first time, his front-line interviews, public capital data and his public statements on X all lock together: the middle layer is collapsing, physical AI has entered its delivery phase, and judgment is becoming the last scarce resource. Three threads that look separate are in fact one.
2026.04.12 — 2026.04.19 · Once a week: spot the signals, recalibrate your thinking.
This Week's Signals
In the week of April 12 to 19, Indigo wrapped up more than a week of intensive interviews in SF and Silicon Valley. When he got back, he did something he had not done before — he publicly stated the strongest consensus from the trip: AI is eliminating the middle layer. As it happens, three sets of public facts collided in the same week. The compute specs of Tesla's next-generation in-car chip, AI5, were confirmed against physical hardware. Tesla's Optimus v3 hand patent and Rocket Lab's Gauss electric propulsion system (an engine that uses electricity to accelerate propellant so satellites can change orbit) surfaced almost simultaneously. And a Euclid VC report showed vertical AI already accounts for 53% of early-stage deal volume, rising to 55% in Q4.
Most people would file these three sets of facts into three separate drawers: jobs, hardware, venture capital. Indigo reads them differently — he sees them as one thing: the value distribution function is being rewritten. In public markets, software stock valuations have fallen back to 2022 levels. In private markets, liquidity is concentrating in leaders like OpenAI, SpaceX and Anthropic. At the application layer, capital is pouring into vertical AI. Three markets are drawing the same shape at the same time — what he calls the U-shaped structure (an industry pattern where value concentrates at the two ends, the very best talent and pure-automation companies, while the middle layer collapses). Semiconductors eating software's profits is the physical support underneath that shape.
This week's takeaway: value is no longer spread evenly along the industry chain. Judgment and automation are the only two high grounds left.
The Winds
#01 The U-Shaped Structure: The Disappearance of the Middle Layer Said Out Loud
After more than a week of Silicon Valley interviews, Indigo took a judgment he had previously only discussed in private and put it on the table for the first time: AI 正在消灭中间层，头部通吃一切！在 SF 与硅谷呆了一周多……这是整趟行程中最强的共识信号
(AI is eliminating the middle layer — the top takes everything! After more than a week in SF and Silicon Valley... this is the strongest consensus signal of the whole trip
) (original post). Behind that statement is a timeline specific enough to be frightening: junior programmer jobs are already gone, senior engineers have maybe 1 year left, researchers about 3 years. The numbers on the capital side confirm it: the Euclid VC report shows vertical AI at 53% of early-stage deal volume, rising to 55% in Q4, with the early-stage share of investment climbing from 53% to 60%. The most extreme sample is the engineering company CREAO — 99% of its production code is written by AI, and the organization has nothing left but architects and operators at the two ends, with nothing in between. The disappearance of the middle layer is not a layoff headline. It is the value distribution function being redrawn. For investors, the screening criterion changes with it: from is this the right sector
to does this company stand at one of the two ends.
#02 Physical AI Moves from Slide Decks to the Shop Floor: AI5, Optimus v3 and Gauss Appear in the Same Week
This week, at a friend's Tesla repair shop, Indigo pulled an old HW3 in-car computer out with his own hands, compared it against the real hardware, and reached a verdict: AI5's real compute is five times a dual AI4, on par with an H100, at 2,000-2,500 TOPS (trillions of operations per second, the unit for measuring chip compute) — data-center-class compute being packed onto wheels. The same week, Tesla's Optimus v3 hand tendon patent (WO 2026/080687) disclosed its details: 2 linear actuators controlling 22 degrees of freedom, a key engineering milestone in the robot's move from prototype to factory worker. Rocket Lab also launched its Gauss electric propulsion system, with planned annual capacity of 200+ units, and acquired the laser communications company Mynaric the same day. Silicon Valley interviewees diverge widely on the timeline — from investable now
to 3-5 years to 5-10 years. One view Indigo relayed: robots still need 3-10 years, because embodied intelligence is harder than large models — the cerebellum took 300 million years to evolve and has 7 times the neurons. Physical AI is no longer a narrative. It is a set of engineering milestones you can verify one by one.
#03 Judgment Is a Precondition, Not a Product
On the human side, Indigo delivered a symmetrical judgment this week: "判断力不是分析能力的产物，而是分析能力的前提！人类对'好'的判断是前智识的 (
Judgment is not a product of analytical ability — it is its precondition! The human sense of what is 'good' is pre-intellectual") (original post). Borrowing the philosopher Pirsig's concept of Quality (that direct, pre-analytical feel for what is good
), he derived a workable division of labor: humans handle dynamic Quality — feeling, direction, judgment; AI handles static Quality — execution, consistency, scale; then humans circle back to calibrate AI's output, over and over, climbing level by level. He also reposted and endorsed a claim that AI-native cognitive systems must go through three mutations — first learn to befriend uncertainty, then learn to hand over the hard drive and keep only the index, and finally let creativity move up into curation. The training path is almost classically simple: look at good things, experience good things, stay close to good things. In an era when AI can do everything, the one ability humans must train is the ability to feel what is good and what is not.
On the Ground
#04 The Information Business Is Industrializing
SemiAnalysis started from semiconductor research. By 2024 it had reached 50,000 Substack subscribers at $500 each — 25 million in subscription revenue, with estimated annual revenue over 100 million. Now that information edge is being packaged into an early-stage fund of about $400 million. Information is not content — it is an asset. The information gap itself is becoming the infrastructure of the AI era.
#05 The Clip Economy: Livestreams Are Factories
TBPN, a tech livestream show, averages 7,000 live viewers, but the Clips (short videos) cut from the stream average 257,000 — a 37x gap. Revenue rose from 5 million in 2025 to a projected 30 million in 2026, and OpenAI acquired it for about $200 million. A livestream is not a product — it is a Clip factory. Media value is flowing to the two ends as well: the scarcity of the live moment, and the scale of distribution.
#06 Models Are Getting Passports
Anthropic is testing user identity verification (KYC, i.e. real-name verification). Indigo's extrapolation: once access to a model is tied to where you were born and which jurisdiction you sit in, digital intelligence gets a passport of its own. KYC is not a compliance detail — it is the national border of digital intelligence. After the middle layer collapses, even the right to use AI is being tiered.
#07 The Next Shape of Compute Clouds
Several Silicon Valley interviewees pointed to the same gap: NeoCloud (a new type of cloud provider specializing in renting out GPUs and other AI compute). As Agents (AI programs that carry out multi-step tasks on their own) stretch tasks to weeks or months, companies need long-term, stable compute. Some have even started Tokenmaxxing — converting as much of the operating budget as possible into model calls. Compute is not a cost line — it is a new major heading on the budget.
#08 The Moon Base Is an Engineering Checklist
NASA published a three-phase blueprint for its moon base — from 4000kg to 60000kg to 150000kg — and laid its capability gaps directly in front of industry. The same week, Rocket Lab's Gauss filled one bottleneck: propulsion systems have never been reliably supplied at any scale.
Space is not a distant story — it is an engineering checklist with the gaps marked.
#09 The Other End of the Blockade
In an interview, Jensen Huang warned that China holds 60% of mainstream chip capacity and 50% of AI researchers, and the real risk of a blockade is that DeepSeek could run at its best on Huawei chips from day one. A blockade is not a moat — it is a catalyst. This echoes the Silicon Valley consensus that China's optimal AI strategy is to fast-follow.
Slow Thinking
Indigo's biggest change of position this week happened on the claim that the middle layer is disappearing.
Before, it was just a phrase he used in private conversations with Silicon Valley friends — he agreed with it, but had never laid it out systematically in public. This time, he put seven interlocking judgments on the table at once: from talent structure, to how profits split between semiconductors and software, to the replacement timeline for programmer jobs, to the pace of robotics, to the demotion of the VC's role. What triggered the shift was not any single conversation but the density of corroboration — founders, investors, enterprise buyers and public capital data, standing in entirely different positions, described the same shape: the investment side sees semiconductors eating software's profits, the enterprise side sees budgets restructured around model calls, the capital side sees vertical AI taking 53% of deal volume. When six or seven independent sources all point to the same conclusion, he judged that this is no longer an insider hypothesis but a structural call he can put his name behind. The same week, his argument about humans also hardened — from a Renaissance
-style metaphor into a teachable, trainable Quality loop model. One piece placed at each end.
The opposing case carries weight too. The strongest version: the middle layer is disappearing
may just be a self-reinforcing narrative inside Silicon Valley's mood cycle — the interviewees know each other, read the same reports, show up at the same dinners, so the supposedly independent corroboration may not be independent at all. Vertical AI taking 53% of early-stage deal volume shows precisely that capital is still betting heavily on the application middle layer, not abandoning it. And the physical-AI timeline splits wildly among interviewees — from investable now
to 5-10 years
— which shows delivery is far from consensus. To prove this week's call wrong, watch for the following over the next few quarters: if mid-sized software companies' margins stabilize, junior engineer hiring recovers, and independent vertical AI companies keep growing rather than being absorbed by the leaders, the U-shaped structure should be withdrawn and re-examined.
Indigo on X
AI 正在消灭中间层，头部通吃一切！在 SF 与硅谷呆了一周多……这是整趟行程中最强的共识信号
(AI is eliminating the middle layer — the top takes everything! After more than a week in SF and Silicon Valley... this is the strongest consensus signal of the whole trip
)
From @indigox, 201 likes
"判断力不是分析能力的产物，而是分析能力的前提！人类对'好'的判断是前智识的 (
Judgment is not a product of analytical ability — it is its precondition! The human sense of what is 'good' is pre-intellectual")
From @indigox, 104 likes
Closing
One Thought
After Gutenberg's printing press appeared, the huge middle layer of scribes vanished within a generation or two. But the value did not vanish — it flowed to the two ends: publishers who could judge what was worth printing, and authors who could write things worth printing. The Renaissance was not delivered automatically by the press. It was the product of judgment meeting the power to copy. Today, AI is the new printing press: execution is copied without limit, and scarcity returns to two things — the judgment that sets direction, and the original work worth scaling. The two main threads Indigo laid out this week are the same old play in a new staging.
One Exercise
Spend 30 minutes on a two-ends test.
Open the role chart of your company (or your own team) and tag every role into one of three categories: judgment (deciding what to do, judging what is good), execution (turning decisions into output), and relay (moving information and progress between the two). Then ask just one question: which of the relay roles could be replaced within a year by an Agent plugged into company data? Write the answer in three lines and post it where you can see it. Look back next quarter — you will know exactly how far the U-shaped structure is from your own desk.