Fractile's Mark Rose 6.5x on a $250M Order for Silicon That Doesn't Ship Until 2027 — In the Same Week OpenAI Cut $150 Billion Off Its Own

These three marks are not comparable in size, but they are comparable in what they price. Fractile's is a forward price on silicon that will not reach a data centre until 2027. OpenAI's and Anthropic's are prices on equity claims on the cash those same data centres will eventually produce. A model lab's mark can fall because the equity is worth less; a chip company's mark can rise because the buyer has committed. Only one of those two mechanisms requires anyone to pay cash today.

Fractile's re-rating has no revenue behind it, because there is no revenue. Share allotment returns filed with Companies House on 15 and 28 October show £577.4 million of Series C preferred shares allotted between 9 September and 23 October — roughly $774 million at the period's exchange rate — in a company that has not shipped a production chip and discloses no sales. What moved in the same window was Anthropic's initial agreement to take roughly $250 million of Fractile's inference silicon for deployment in data centres in 2027, an order worth about 4% of the resulting valuation. Fractile re-rated 6.5x on a promise, not a shipment.

Anthropic's own registration statement, filed 26 October, is the cleanest statement of what changed. First-half 2026 revenue was $16.286 billion, operating loss was $693 million, adjusted operating income was $1.62 billion against a $1.85 billion loss a year earlier, and the net loss was $80.748 billion on 2,024,972,990 weighted-average shares. The net loss came from convertible-bond accounting, not from the model layer's operations.

The commitment schedule, not the demand forecast, is what turns a model's equity mark into a supplier's revenue. NVIDIA has gone further: its 10-Q for the quarter ended 26 July 2026 discloses $105.0 billion of guarantees supporting leases for roughly 4.25 gigawatts of IT load at an OpenAI affiliate's Ohio campus, plus $3.5 billion of other AI-cloud land, power and shell guarantees, and $366 billion of future purchase commitments of which $279 billion is supply and capacity. The chip vendor is now underwriting its own customer's credit.

Nothing in the operating results of the listed supply chain supports the October selloff. NVIDIA grew quarterly revenue to $96.2 billion with data centre revenue of $89.0 billion, 92.5% of the total, on a 66.2% operating margin. Broadcom posted $16.7 billion of AI semiconductor revenue, up 221% year on year. TSMC lifted quarterly operating profit 75% to NT$875.5 billion. SK hynix ran an 80% operating margin. Micron earned an 86.7% gross margin. Samsung Electronics posted KRW 107.4 trillion of operating profit. Every one of those prints was a record. Every one of those shares fell. The market was not disputing demand. It was disputing who pays for it and at what cost of capital.

The single most consequential number in the AI buildout this quarter is Anthropic's first-day price, because it sets the equity value against which at least $518 billion of non-cancellable obligations will be carried. At $1.5 trillion, committed compute equals about 35% of market capitalisation; at $2 trillion it was about 26%. That gap decides whether Anthropic's commitments look like investment or like leverage.

The split is not compute versus models. It is private capital versus public capital — and inside private capital, sovereign money versus everyone else. NVIDIA put $108.5 billion of guarantees behind an OpenAI affiliate's leases; AMD agreed to buy up to $5 billion of Anthropic stock; Anthropic committed $518 billion to counterparties that are simultaneously its suppliers and its competitors. The money is going round in a circle, and the circle now has a credit market that reprices faster than the equity market. Meanwhile the UK's Sovereign AI Fund paid up for a chip company whose only named customer is the company whose valuation fell 25% that same week. The equity in the model layer is worth less while the silicon it has committed to buy is worth more.