Intelligence Is Abundant. Authority Is Not.
The structural read behind today's Morning 10 — why the same forty-eight hours bought an application layer and sold a component layer, and what that says about where the scarcity now sits.
Ten contests, not one
The public version of the AI argument is still “open models versus closed”, with “local versus API” as the sophisticated follow-up. That is two of about ten. Laid out properly the enterprise market is decomposing along model ownership, deployment, architecture, context, integration, agent design, execution, distribution, development and economics — and most of the large players sit on both sides of at least one line. Microsoft and Google straddle nearly all of them. The reason this matters for a portfolio rather than a panel discussion is that the ten lines do not carry equal weight. They resolve into a single question: not whose model is smartest, but who controls the context, the permissions, the execution and the customer relationship.
The scarcity inverted
Rank the layers by how hard each is to obtain and the ordering has quietly reversed. Models are becoming plentiful — open weights, falling prices, a new frontier release every few weeks. Connectivity is being standardised, with a tool-calling protocol the labs themselves now converge on. Enterprise context is still difficult. Permissions and governance are scarcer. Operational execution — the right to stop a production line, approve a claim, move money, write the change back into the system of record — is scarcest of all. Owning the model does not confer that right. This is the six-layer framing arriving at its uncomfortable conclusion: the durable control point may not be the foundation model at all, but the layer deciding which model receives a task, what context it may see, which actions it may take, how the result is verified, and who remains accountable when it is wrong.
The abundance is not a forecast
The top of the stack is already behaving like a commodity. Moonshot released Kimi K3 open-weight in late July at 2.8 trillion parameters. On OpenRouter, coding has gone from roughly a tenth of routed token traffic in early 2025 to more than half by the middle of this year, with total routed volume up around fourfold — from about five trillion to over twenty trillion tokens a week. That is what abundance looks like in the data: more capable weights, given away, consumed in vastly greater quantity for a narrowing set of high-value tasks. Meanwhile the layer above it is being built in public — Microsoft brought its IQ family to general availability on 2 June, an ontology-grounded context layer sitting across work, data and web. Only 18.7% of enterprises had adopted outcome-based software pricing by the first half of this year. The pricing model for the scarce layer has not been settled yet, which is usually a sign that the layer itself is still being claimed.
What the tape paid for last night
Palantir printed $1.94bn against roughly $1.80bn expected — revenue up 93% year over year, the fastest the company has posted, at a scale where growth normally decays. US commercial revenue rose 149% to $764m, accelerating rather than fading, and US government rose 90% to $809m. Adjusted operating margin reached 62%, for a Rule of 40 score of 155%; net income was $1.07bn against about $329m a year ago; adjusted free cash flow was $1.22bn. The company closed 220 deals worth $1m or more, seventy of them above $10m, and guided the full year to roughly $8.15bn — about 82% growth. The stock closed at 125.65 and traded near 145.
The chief executive spent the call arguing that enterprises are engaged in what he calls tokenmaxxing — paying heavily for model consumption, receiving little measurable operational value, and handing their proprietary process knowledge to the model providers along the way. He has a direct commercial interest in that framing, and it should be discounted accordingly. It is also, structurally, the same argument the map makes: that spending on intelligence and capturing value from it have come apart.
The price chart accompanying this piece shows the name before the print, and it is the part worth sitting with. Going into the quarter the stock was down 29.3% on the year, closing at 125.65 — 39.5% below the high of 207.52 printed on 3 November, and only 18% above the low of 106.37 set on 25 June. Draw the line off that November high and it runs down across the whole of 2026, sitting in the mid-130s by the start of August: nine months in which every attempt to reclaim it failed, while the business compounded underneath and the last three earnings beats were graded by being sold. A print near 145 does not merely bounce inside that structure; it opens above the line that defined it. Even so, paying up fifteen percent leaves the name roughly thirty percent below where it traded nine months ago, on a quarter that nearly doubled revenue at a 62% margin. That is not a re-rating. It is a market beginning to correct a mispricing it maintained for three quarters — the same shape we documented at length in Crushed While Correct.
What it sold in the same forty-eight hours
The other half of the statement is the part that got sold. On Monday, while the Dow closed at a record and the large platforms rose four to six percent each, the component layer went the other way: analog and storage were marked down, with MPWR off about five percent and WDC down 4.6%. Overnight the same trade repeated in Seoul with the volume turned up. The Kospi opened more than a percent higher, reversed the whole move inside half an hour and traded down near 6,200 with Samsung Electronics and SK Hynix each off around three percent — while the Kosdaq, the small and mid-cap board, ran more than seven percent. A market that splits its own exchange in half in a single session is not de-risking. It is choosing, and what it chose against was the layer that supplies capacity.
Read the two halves together and it stops looking like rotation from hardware into software. It looks like the stack being repriced by scarcity. Compute is on its way to being abundant, which is what a sold semi complex into a record tape actually means. Intelligence is on its way to being abundant. Authority — context, permission, the right to act — is not, and that is where the bid went.
Our own instruments, saying it in index form
This is not a read we need to import. The Rubin 100, our build-out index, is flat on the week and deeply negative on the month, and all four of its layers — Architects & IP, Manufacturing, Memory & Packaging, Substrates & Power — finished the week red while the broad tape rallied. HALO and Euro-AI, which sit closer to the monetisation layer, lead the family on the week. The instrument measuring who builds the capacity is being sold; the instruments measuring who books the revenue are leading. Our own tape has been making the argument for a month, in the one form that cannot be talked out of it.
What this does not settle
Three things stay genuinely open. The first is the counter-case, which is serious: if a universal interface becomes capable enough, it may absorb the application layer rather than supply it, and the labs are explicit that this is the plan. Historically the application owner captures more of the customer-specific value than the component supplier — unless the component becomes an unavoidable platform, and nobody yet knows which of those two sentences describes a frontier model. The second is that one quarter is not a regime. Three sold beats became one bought beat overnight; the third reading arrives with AMD tonight, and a thesis that took nine months to be ignored will not be settled by a single gap. The third is price. A structural argument about where scarcity sits says nothing about what any given company is worth on a given morning, and a name that moves fifteen percent on a print has already been paid for some of it. Probability, not prophecy — the map tells you which contests are being fought, not which morning they are resolved.
C · members block
Into today
What the book is actually watching. The map is a multi-year claim; the tape gives it one test at a time. Today’s is the second reading, not the first. Palantir broke the sold-beat pattern as a class — the question tonight is whether AMD, carded at nine earnings beats against a 3-1-6 reaction record, confirms that the tape has stopped refusing to pay for delivery, or whether Palantir was a single-name exception with a nine-month spring underneath it. A compute name is the harder test precisely because it sits on the layer this piece argues is becoming abundant.
The tell to watch is not the print, it is the pairing. If capacity really is being repriced as plentiful while orchestration is repriced as scarce, then the pattern should keep showing up as a split rather than a direction: the AI-revenue layer bid and the component layer sold on the same tape, the same session, sometimes the same exchange. Seoul gave the cleanest version of it overnight. The build-out layers inside Rubin are where it would show up in our own instruments — a floor there, while the monetisation indices keep leading, would argue the repricing is finishing rather than starting.
And the discipline stays what it was. Nothing here is a level or an instruction. The position is unchanged; a confirmation that runs the book’s own way is not a reason to add to it, and a framework that explains yesterday convincingly is exactly the kind of thing that talks people into paying up for tomorrow. The map earns its keep by telling us which questions are load-bearing — who owns context, who owns permission, who owns the right to act — not by telling us what to do before the open.