The lazy question this morning is whether the AI boom is over. No press release answers that, and nobody guided capital over the weekend. The narrower question pays: what does Microsoft's document actually constrain, and what is actually at risk in the names that fell?
Microsoft published a draft Humanist AI Code of Conduct at 09:00 ET on 14 September, opening a six-week consultation. By 09:57 ET Micron was at $904.29, down 7.28% on Friday's close, SanDisk $1,528.02, down 6.45%, CoreWeave $84.46, down 5.09% and Nebius $213.62, down 4.87%. The selling had started in Asia hours before the document existed.
The document is 38 pages and never says "compute"
The PDF runs 15,203 words. Searched in full, it returns zero occurrences of compute, capex, capital expenditure, compute budget, gigawatt, datacenter, training FLOPs, spend or invest. The two appendices are a glossary and an evaluations list. The three hits for "pause" are a user's right to pause a task; the one for "billion" is "billions of people".
What it does contain is behaviour. Absolute constraints covering CBRNE, offensive cyber, loss of human control, manipulation at scale and child safety. A requirement that models comply with a request to "pause, redirect, cancel, or shut down" and never act "in a way that makes it harder to pause, change, direct or end the interaction". Its own preface says Microsoft is "not using it to train our models today".
It binds Microsoft's in-house MAI model family, not Azure, not OpenAI, and not Microsoft's cloud spending. The same announcement page advertises a compute roadmap "ramping quickly and extensively". Mustafa Suleyman told Reuters the document had been in preparation for five to six months and that "now's a good time for everybody to have this conversation and take a breath." No dollar figure anywhere.
What the assumption would actually be worth
Grant the narrative regardless, and price it against our own files. Every name sold this morning carries a published fair value that rests on capacity prices continuing to rise, so the test is mechanical: set those drifts to zero and re-run.
Our published base fair value, then with every price-drift line set to zero
| Company | Base, then flat | Cut |
|---|---|---|
| Micron | $706.33 → $568.67 | −19.5% |
| SanDisk | $2,439.91 → $1,696.77 | −30.5% |
| CoreWeave | $61.24 → $35.68 | −41.7% |
| Nebius | No price-drift line | n/a |
Drifts flattened: Micron cloud memory +9.0%, core data centre +10.5%, mobile and client +9.5%; SanDisk datacenter, edge and consumer all +14.0%; CoreWeave AI cloud +11.75%. Nebius carries a +57.5% quarterly growth assumption instead. Base cases are ours, as are the re-runs.
Micron's model assumes cloud memory prices drift +9.0% a quarter, core data centre +10.5% and mobile and client +9.5%. Flattening cloud memory alone takes the base from $706.33 to $646.22, an 8.5% cut. Flattening all three takes it to $568.67: $706.33 − $568.67 = $137.66, or 19.5% of the base. SanDisk is the cleanest case in the set because all three of its verticals carry the identical +14.0% drift, so no line inside the company survives a flat deck: $2,439.91 − $1,696.77 = $743.14 off its base, 30.5%.
CoreWeave carries the most leverage and two levers. Its single AI-cloud vertical assumes +11.75% quarterly price drift with utilisation gliding from 76.7% to 95%. Price flat alone costs 41.7%, to $35.68. Utilisation frozen at today's 76.7% costs 39.4%, to $37.09. Both together leave $16.72, a 72.7% cut — and the obligations do not flex, as $64.6bn of committed lease payments against $35.1bn of debt already showed.
Nebius does not belong in this frame
Nebius fell with the group and does not fit the mechanism. Its model carries no price-drift line at all: the AI-cloud vertical is a growth assumption of +57.5% a quarter. Drop that vertical and the base goes from $136.85 to −$2.83; set the growth to zero and it is −$26.10. That is a different and far more violent exposure, and it was sold on sentiment rather than on the mechanism measured here.
Two honest qualifications
These are changes to our own fair values, not forecasts of price. Three of the four already trade above our base: Micron 28.0% above, CoreWeave 37.9%, Nebius 56.1%. Only SanDisk trades below, 37.4% under its base. A reader who thinks our bases are too low should read the sensitivities as what the assumption is worth inside the model, not as a target.
And the moves are not single-cause. Micron and SanDisk were already inside a global chip selloff through the weekend, and the same discipline we applied to a published bear case in August applies here: a headline is evidence for the phrase in quotation marks, not for every assumption needed to reach the number.
What would overturn this: a Microsoft capital-expenditure guide cut, or any lab publishing a compute budget reduction. Neither has happened. Until one does, the tape moved 5–7% on a document that contains no price, while the assumption genuinely at risk is worth two to eight times that inside the models.
The Code of Conduct, its word count, the zero term hits, the appendix titles and every quoted passage are from the draft PDF at microsoft.ai, published and read 14 September 2026; the "ramping quickly and extensively" line is from its companion announcement page. Suleyman's remarks are Reuters', 14 September. Prices are Google Finance quotes taken at 09:57 ET on 14 September, intraday and still moving, against Friday's closes. Every fair value, drift assumption and utilisation glide is from our own published models, dated 18 August for Micron, 23 August for CoreWeave and 27 August for SanDisk and Nebius; the flat-price and frozen-utilisation runs are ours, produced by re-running each model with the named driver set to zero.
