This week, Indigo said the same thing at three scales. On the industry map, value is fleeing the model layer toward scarce resources at the bottom and customers at the top. For individuals, only rare data, taste, and trust remain. In method, once generation approaches free, verification becomes the new threshold. This issue puts the three scales back into one picture.
2026.08.02 — 2026.08.09 · Once a week: spot the signals, recalibrate.
This Week's Signals
In the first week of August, Indigo posted three judgments on X in quick succession. The model layer is being commoditized (products lose differentiation, leaving only price competition), while opportunities for the Harness (the software layer that plugs models into enterprise workflows and fits them with context and constraints) are already appearing in volume on the enterprise side. An individual's core assets in the intelligent economy come down to rare data, taste, and trust — this one got his widest reach of the week. And all three of Google's original Gemini leads left; he read this as role optimization rather than crisis — this one drew 67 replies and 18.3k views, the most contested of the week. One hard fact landed in the same window: Jeff Dean left to found Discovery Loop, and Alphabet turned around and took a stake as a founding investor, bundling in cloud services while it was at it.
Most readers will scroll past these as three unrelated news items, but they are projections of the same sentence at three scales. On Chamath's map splitting the AI industry into six layers, the moat is called self-generated data and enforcement points (the spots that actually execute decisions and block actions, not the ones that only produce answers). At the individual level, it is called rare data, taste, and trust. At the level of method, it is called verification and the exoskeleton (writing your personal judgment process into a system that AI can amplify). These things share exactly one trait — they cannot be rented. The ability to generate is visibly turning into a commodity, and commodities never earn a premium.
When the cost of generation approaches zero, the moat is not what you can make; it is what you hold that others cannot rent.
Currents
#01 The Harness Is the New Enterprise Application Layer
This week Indigo gave his most systematic public judgment in a while: My judgment: at the current pace of model evolution, Application has no opportunity yet, but Harness opportunities are already appearing in large numbers, especially on the enterprise side. Knowledge internalized by an enterprise can only land through Context and constraints, so the Harness is the new enterprise application layer
(original post). As for why this layer in particular is worth money, Coatue supplied first-hand data: Together AI's processing volume rose from 30 billion tokens (the unit models use to measure text) per month to 400 trillion tokens — but more tokens does not mean better results; routing, evals, and monitoring sit in between. Baseten's data is even more blunt: ordinary users' open-source model utilization is only 3-5%, while the very top users sit at 40-50%. The difference is not in the models at all; it is in whether a harness has been built around them. The model layer's problem is not that it is too weak; it is turning into a commodity. So value flees toward both ends at once: scarce resources at the bottom, customer relationships at the top.
#02 Individuals Have Only Three Assets Left, in Decreasing Order of Copyability
In his most widely shared original post of the week, Indigo wrote: The core assets an individual can bring into the intelligent economy really come down to three: rare data, taste, and trust
(original post). These are not three parallel options; they are a ladder of decreasing copyability — data can be bought away, taste has to grow slowly over time, and trust can only be given, never self-declared. Applied to investing, this has a name: the exoskeleton. LLMs (large language models) score only 60-70% accuracy on financial benchmarks, and 60% accuracy on Wall Street equals being fired 100% of the time. The only fix is systematic cross-validation across multiple models, plus human checks tailored to each thesis — a 97% match rate is not a passing grade, it is an alert waiting to be investigated. Indigo draws the boundary hard: "The exoskeleton must be personal! Because investment conviction is close to religion, 'you cannot use someone else's agents'; the bottleneck in building the system is not technology, but writing down tacit knowledge" (original post). Anyone can generate an answer; only a judgment that can be verified carries a price.
#03 Google May Not Be Fighting the Frontier Race at All
The facts first: all three original Gemini leads left this week. Q2 revenue grew 24% year over year to $120B, and cloud grew 82%, but record capital spending pushed free cash flow (cash from operations minus capital spending) negative for the first time since the 2004 IPO, and Google's stock fell about 4-5% intraday that day. Indigo's reading runs opposite to the pessimist frame: This is a role optimization for Google, not a crisis
(original post). The next day he added: "Gemini is about 6 months behind frontier models in coding ability… Google is still the most likely to win the consumer AI race (in-house TPU compute + an existing consumer base + control of the world's most popular mobile OS); the reason it has not made a 'super AI App' yet is that the world isn't changing that fast" (original post). The TPU is Google's in-house AI chip, and even competitor Anthropic is now a customer — from this, O'Reilly reads Google's path as a diffusion race, not a frontier race. The narrative has swapped catch-up premium for infrastructure stability. That is not a pure positive; it looks more like a deliberate downgrade that trades ceiling for certainty.
On the Ground
#04 Electricity Is the Real Open Question for Next-Generation Compute
The true constraint on next-generation compute is not chips; it is electricity. Lam Research founder 林杰屏 says the power problem remains unsolved to this day. People in three positions — equipment, chips, and system architecture — converged independently on the same judgment. The only dissent so far comes from supply-chain analysts, whose evidence is one power plant that expanded from about 495MW to 1.7GW in five months. Electricity happens to be exactly the kind of asset that is hardest to rent.
#05 The New Lithography Battle Is Only Over the Light Source
The FEL (free-electron laser) camp is attacking the light source, not the whole machine. One light source can feed 8-20 scanners, with 4-16x better wall-plug efficiency; xLight aims for a prototype by 2028. But today, EUV (extreme ultraviolet lithography, the key step in leading-edge chipmaking) production runs 100% on ASML alone, and the claimed 30-40% cost reduction comes entirely from one side's own statements. Whole-machine integration is exactly the kind of position that cannot be copied away.
#06 One 6:5 Blind Test Said Two Things
Copying generation is already near zero cost; replacing execution is still far off. A scanner that an undergraduate vibe-coded (had AI write quickly from a rough description) in one week went up against the $4,790/year Nessus Professional in 15 paired blind tests, and Nessus barely won 6:5. That score proves two things at once, but most coverage only told the first half. Balaji estimates the verification economy will grow 10 to 100x as a result.
#07 Reinforcement Learning Taught No New Moves
RLVR teaches when to use which move, not new moves. RLVR (a training method that does reinforcement learning with automatically verifiable rewards) may take 20% to over 50% of training compute yet contribute only about 0.01% of the information. In domains without a verifier, every capability claim deserves a discount. This is the mechanical side of the moat's retreat toward verification.
#08 Use-It-or-Lose-It Now Has Longitudinal Evidence
When AI reads and calculates for people, it touches the very variable that decides whether you decline in old age. German longitudinal tracking shows literacy peaks on average at 46 and numeracy at 41. People above the median in usage show little decline on average through 65; those below the median start declining at 35. Taste, as an asset, only keeps if it is used continuously.
#09 Platforms Are Starting to Invest in Their Own Departures
When a platform cannot keep its top talent, it invests in their departure instead. Alphabet took a stake in Jeff Dean's new Discovery Loop as a founding investor, with cloud services bundled in. This supported spin-out (a platform taking equity and compute stakes in a departing team) is becoming the standard capital structure of the compute era — yet almost no one on X is talking about it.
#10 Microsoft's Arithmetic
For Microsoft, moving the structure pays better than building compute. After renegotiating away the 20% revenue share, running OpenAI models directly on its own compute has revenue potential of about $100M/MW/year; renting that compute to OpenAI yields only about $14M/MW/year. The top layer is about customers, and value ultimately goes to whoever holds the layer between customers and models.
Slow Thinking
The biggest change this week happened in Indigo himself. Back in May, his call on the harness was still to stay away — this layer would end up thin, brutally competitive, and eventually eaten by rising model capability. By August 2, he publicly wrote down the opposite conclusion: the Harness is the new enterprise application layer, and Application has no opportunity yet. The trigger was not one great essay, but three independent pieces of evidence arriving at the same time. Chamath's six-layer map listed the harness as its own layer. The Baseten data cited by Coatue gave quantitative evidence — top users' open-source model utilization is 40-50%, ordinary users' only 3-5%, and the difference is whether there is a harness. Trask added the mechanism path (integrations will occupy the Pareto frontier, and the harness layer will follow up with real-time routing). This is a directional reversal, not a parameter tweak. In the same week, his frame on Google also moved from can Gemini catch up
to it may not be fighting that race at all.
The test point is clear: if within 12 months there is a case where a model vendor's native capability leaves third-party harnesses with no reason to be paid for, this reversal must be walked back.
On the other side, the counterargument is simply May's Indigo himself — when the slope of model capability is steep enough, everything in the middle gets eaten. The 306 prompts (prompt-writing techniques) Karpathy taught a year ago are all obsolete today, which shows how fast this layer can depreciate. And the diffusion-race frame carries a crack of its own: electricity has no quality tiers, models do — if frontier capability can diffuse smoothly through APIs, the frontier winner and the diffusion winner will be the same company, and the analogy collapses. To prove this week's theme wrong, watching two things is enough: cases of frontier models directly swallowing enterprise harness use cases, and the pace of Gemini 4's productization along with the new leadership's iteration rhythm.
Indigo on X
"The core assets an individual can bring into the intelligent economy really come down to three: rare data, taste, and trust… People will not only buy your 'data', they will also buy your 'taste' judgments; and when even your taste does not cover something, they will buy your 'trust network'… These three map almost one-to-one onto the Indigo-Mind I built: scarce data = the digested material in libs/ + notes/; taste = those 'my standards' in refs/; trust = my readers on X and the Rewired community"
From @indigox, 56 likes
"If you feel your taste is sharp, your perspective unique, and you are good at choosing, then you should 'productize' yourself… AI can always give people the most consensus-driven, most in-depth answers; but we can offer the most anti-consensus, most unique answers. Be an 'irreducible' human"
From @indigox, 34 likes
"Gemini is about 6 months behind frontier models in coding ability… Google is still the most likely to win the consumer AI race (in-house TPU compute + an existing consumer base + control of the world's most popular mobile OS); the reason it has not made a 'super AI App' yet is that the world isn't changing that fast"
From @indigox, 31 likes
Closing
One Thought
In history, the winners of technological revolutions are often not the frontier inventors but the side that completes diffusion. Britain started the Industrial Revolution, but America, which completed electrification, took the baton. Japan led in chips in the 80s, yet lost the information revolution. O'Reilly calls Google's choice this week a Westinghouse-style bet — a bet on diffusion, not the frontier. The moat's retreat from can generate
to cannot be copied
is the same historical law replaying on AI: capability will diffuse in the end, and once diffusion is complete, what remains are only the positions that could never be moved in the first place.
One Exercise
Indigo said this week: "If you feel your taste is sharp, your perspective unique, and you are good at choosing, then you should 'productize' yourself… AI can always give people the most consensus-driven, most in-depth answers; but we can offer the most anti-consensus, most unique answers. Be an 'irreducible' human" (original post). Here is an exercise that takes under 30 minutes. Take out a sheet of paper or a blank document. Write down one judgment in your field that runs against mainstream consensus and that you would publish under your own name, in no more than 200 words. Then write two more lines — what evidence would prove it wrong, and whether it relies on your rare data, your taste, or your trust. If you cannot write it, that confirms this week's line: the bottleneck is not technology, it is writing down tacit knowledge.