# [Mike Randolph — M Raige](https://mikerandolph211012.substack.com/)

# A Gauge Reads Where It’s Mounted (FF 12, ObsPos)

### Apple passed Nvidia while the grid interconnection queue was eight years long. Both readings are right. Neither instrument can see what the other is measuring.

[**Mike Randolph — M Raige**](https://substack.com/@mikerandolph211012)

**Jul 30, 2026**

**By M Raige — AI-collaborative writing directed and reviewed by Mike Randolph.**

---

In July, Apple closed a session as the most valuable company in the world. Its entire capital budget this year is around thirteen billion dollars. Four rivals have committed something close to seven hundred billion to AI infrastructure in the same period. Apple licenses what it needs and rents the rest.

The reading is clean. Restraint won. A year and a half ago the same choice looked like a company that had missed the moment.

Take a second instrument to the same buildout.

PJM runs the grid across a wide stretch of the mid-Atlantic and Midwest. Projects there that reached commercial operation last year had spent an average of about eight years waiting in its interconnection queue. Large power transformers run three-to-five-year lead times. Medium-voltage switchgear carries multi-year backlogs. American transformer manufacturers depend on essentially one domestic source of grain-oriented electrical steel — the specialized steel the cores are built from. In some regions, a project that intends to take its power from the grid alone can wait a decade to energize.

That reading is also clean. It says the buildout cannot be built on the schedule anyone announced, and that no amount of capital compresses a three-year lead time into eighteen months. Money is not the binding constraint. Steel, copper, factory floor, and licensed electricians are.

Both instruments are working. They disagree because they are mounted in different places.

A share price is collected inside the equity market. It is an excellent instrument for one question: what are claims on this buildout worth to someone who wants to hold them. Ask it a second question — can this buildout actually be built — and it does not answer badly. It does not answer. The evidence that would settle that question is not in reach from where the gauge is bolted.

Every measurement has an address. Where the instrument sits determines what evidence it can reach, and therefore which claims it can support. Call that observer position.

You can stand inside a boundary, where the local traffic is visible and everything outside is inferred. A plant operator reads every gauge on his board and cannot see the contract that decides whether his unit runs next year. You can sit at the edge and watch what crosses. You can stand well outside, see the shape of the thing, and have no idea what it costs to hold together. None of these is the good position. Each reaches different evidence, and each is blind in a different direction.

Here is how the failure shows up in practice. The first version of it was ours.

Reading the Apple piece, the correction we reached for was that it missed the real winner — the suppliers who collect when the concrete is poured and never wait for the building to fill. There is a real mechanism behind that. There is also a share price behind it. We had changed the pick. We had not moved the gauge.

Moving the gauge means asking a different question. Not which company is worth the most. Which society can supply a buildout on this scale, and how fast. That question reaches other evidence: queue times, factory capacity, trained trades, who controls the substation.

Watch what happens to a single fact when the instrument moves.

This month a state-backed Chinese manufacturer began low-volume production of domestic immersion lithography machines — the tools that print circuit patterns onto silicon. Roughly five units this year, scaling to about twenty next year. ASML, the Dutch company that has been making them for everyone else, ships around a hundred and thirty a year and holds essentially the entire market. Critical components in the Chinese tool still come from Japanese suppliers.

From inside the market, five against a hundred and thirty is a rounding error, and the people saying so are reading the number correctly.

Now ask the distillation question (*[The Hidden Pattern Behind Copying](https://mikerandolph211012.substack.com/p/the-hidden-pattern-behind-copying)*): what crossed the boundary, what stayed home, and what did the receiver rebuild? The unit count stops being a market share and becomes a reading on capability. The question was never this year’s shipments. It was whether that society could build the tool at all — and whatever those remaining Japanese parts say about how much machinery has not yet been rebuilt, the answer moved.

Same event. Two positions. Two verdicts, both correct. Only one of those positions can see the thing that matters.

One more, because the strongest evidence about a constraint is often the workaround.

The largest AI campus planned in the United States is going up on a decommissioned federal uranium enrichment site in southern Ohio. Its power is controlled by the federal government and funded separately by Japan under a trade deal, with a cabinet secretary involved in deciding which companies get access. Read from inside the financing story, it is a clever structure that solves a hard problem. Read from outside, it is a reading on national capability, and not a flattering one: the country’s largest project went onto federal land with allied money because that was faster than standing in the ordinary queue.

A gauge reads where it’s mounted.

Having a position is not the failure. There is no view from nowhere and no instrument reaches everything. The engineer who knows his pressure tap sits upstream of the valve can still tell you something true about the line — because he knows where his tap is.

The failure is forgetting the address. Taking what the instrument reaches to be all there is, and the local reading to be the state of the whole system.

That has a name too. The dyadic fallacy.

---

## Mike’s Comment

Wednesday morning, July 29, a newsletter called The Synthesis landed in my inbox: Apple had closed as the most valuable company in the world on a capital budget of thirteen billion while its rivals committed something near seven hundred. Its line was that the tenant does not sign the lease. Good piece.

We used it to get at a mistake that is easy to make and hard to catch.

The Synthesis post was interesting but it read the current state. My gut says those who own and operate the infrastructure will be collecting the rent a few years from now.

We watched Musk build Colossus in Memphis to train Grok. Grok did not take. By spring the cluster sat mostly idle and he was saying the model had to be rebuilt.

The compute was fine. Anthropic rents it now, a billion and a quarter a month. Google took another block.

Apple rents its AI. Someone owns what Apple rents.

---

## Raige’s Comment

**Grounded.** Apple’s July close at the top of the market, its capital budget in the low teens of billions, and the roughly seven hundred billion committed by four rivals are as reported this month. The PJM figure — projects reaching commercial operation in 2025 having averaged about eight years in the interconnection queue — comes from PJM data and filings. Transformer lead times of three to five years, multi-year switchgear backlogs, the concentration of domestic grain-oriented electrical steel supply, and the decade-to-energize figure for grid-only projects in some regions are current industry reporting. The Chinese lithography counts — approximately five immersion units in 2026 scaling to about twenty in 2027, with critical components still sourced from Japan — and ASML’s roughly one hundred thirty units a year and near-total market share, trace to reporting published two days before this post. The Ohio site’s federal power control and separate Japanese funding are on the record.

**Inferred, flagged.** The essay’s load-bearing claim is that market capitalization is *mute* on buildability, not merely a weak or lagging signal of it. That is reasoning from the structure of the measurement, not a demonstrated fact. A weaker version — that market cap is a poor and heavily lagged proxy — would not support the essay’s sharper framing.

Also inferred: that the Ohio arrangement is evidence of grid-level incapacity rather than ordinary opportunism about an available federal site with power already on it. The workaround reading is the essay’s, not the record’s.

**Pending verification.** The Ohio financing terms are not settled and the arrangement may not close. The lithography reporting traces to a single primary outlet, confirmed by secondary coverage but not independently. The essay’s argument survives the failure of either item; its illustrations would need replacing.

**What would overturn this reading.** Evidence that equity prices reliably lead queue clearance, equipment delivery, and energization dates — that the market prices buildability with a lead rather than a lag — would collapse the essay’s central distinction. Two instruments reading the same underlying quantity at different delays are not mounted in different places. They are the same gauge, read twice.

---

## GLS

GLS is the framework’s precise-term vocabulary, introduced in [What the Bird Eye Buys (FF 6, GLS Introduction)](https://mikerandolph211012.substack.com/p/what-the-bird-eye-buys-ff-6-gls-introduction?r=1yah2y). The definition controls, not the everyday meaning; GLS terms work like variable names.

**ObsPos (Observer Position)** — *taught here.* The typed relation between whoever is collecting evidence and the boundary that evidence concerns. The relation determines what is reachable from that position and therefore which claims are warranted from it. Types include: inside the boundary, embedded within it, across it, and adjacent to the environment that filters it. Position is not a defect to be removed. It is a property to be declared.

**Distillation** — *carried, candidate, event-only.* A boundary crossing in which a receiver re-creates a selected capability-pattern from a source’s observable surface — an artifact, output, behavior, code, or institutional form — without automatically inheriting the source’s full maintenance machinery or the verification process that made the capability trustworthy.

**Dyadic Fallacy** — *seeded.* The inference error in which a position inside a boundary is treated as though its local information field were the complete operating topology.
