Alphabet stock rises 0.6% the same day judge rejects forced sale of its ad exchange

A divestiture order is a forced sale, and in a forced sale the discount is the whole of the argument, because a seller working to a court's calendar negotiates from a worse position than one that can walk away. That is the risk the judge took off the table on Wednesday when he declined the Justice Department's request that Google sell its ad exchange ^[1], and the market's answer was 0.6%, which did not cover the more than 1% Alphabet had shed the session before ^[4], leaving the two days net negative ^[15].

Two readings are available. Either the overhang had already been discounted, which is plausible given that this is not the first ruling to go Google's way against the government's antitrust efforts ^[2], or Alphabet spent its summer being repriced on model quality, departures and a DeepMind restructuring ^[7], in which case a legal win does not touch the variable under dispute. I lean to the second, which makes the price sheet the more interesting document.

Gemini 3.8 Flash arrived at 75 cents per million input tokens and $3.75 per million output, the same introductory rate as the Flash model before it, with Google claiming better coding and multi-step performance ^[8][3]. Output priced at five times input ^[21] is the geometry that matters for agents, since an agent that plans and revises generates tokens rather than merely reading them, and Google left that rate alone. The concession went into packaging instead: token discounts of up to 20%, monthly caps on agent spending, and a base subscription priced at zero ^[10], with Google's own materials naming Microsoft's and Anthropic's recurring seat fees as the thing being undercut ^[13].

Three Flash models in six weeks is one about every fortnight ^[3][14], a cadence the smaller models make affordable because they are cheaper to run and faster to iterate than the frontier systems ^[18]. The strategic version of that bet is Demis Hassabis's, who told the G20 Innovation meeting that Gemini could act as a general-purpose layer coordinating cheaper specialised models and agents, so that breadth counts for as much as the single best model ^[19]. The financial version is Thomas Kurian's: nearly three-quarters of Google Cloud customers already use the AI products, and those customers are spending about 50% above their original commitments ^[17]. Overage is the cleanest figure here, because it is revenue nobody contracted for and it is what has to carry the infrastructure bill ^[12]. The cybersecurity release, gated to vetted government and enterprise defenders through the Fairwind Program ^[9], is not yet carrying anything.

The ruling's practical effect is on what Alphabet no longer has to do: no sale process for the exchange, and no carve-out of shared ad infrastructure ^[1]. This is probably wrong, but I read the remaining risk as commercial rather than legal, and what would prove it wrong is measurable: attach and overage flattening while capex holds ^[17][12], or that introductory 75 cents turning out to be a rate Google cannot hold once the packaging discounts are spent ^[8][10]. A zero-dollar base subscription is an acquisition cost, and the bill arrives in tokens.