The most consequential line in Meta's latest GEM disclosure is not a conversion lift. It is the claim, in the research paper behind GEM's new Kunlun architecture, that Meta's ad models now improve along a predictable curve as compute is added, the way large language models do. If that holds, Meta's infrastructure spending stops being a wager on whether AI pays off in advertising. It becomes a function the company can budget against: more training compute, lower prediction error, higher advertiser return. For two years the market has argued over whether Meta's capex is defensive or speculative. A scaling law reframes it as an allocation problem. The variable that decides the outcome is one the earnings headlines never report: how efficiently a model that never serves an ad passes what it learns to the models that do.
The Teacher That Never Serves an Ad
The architecture explains why that transfer step matters. When an Instagram user opens the app, Meta runs an auction she never sees, and it finishes in under a second. Take a user who has spent months watching trail-running Reels, saved a half-marathon training plan and clicked a running-shoe ad last week without buying. Her request first reaches Andromeda, the retrieval system. Andromeda reduces tens of millions of eligible ads to a few thousand plausible candidates. Its job is not to pick the winner. Its job is to make sure the winner is somewhere in the pile.
Ranking narrows those few thousand candidates to a handful. The Adaptive Ranking Model decides how much computation each request deserves. A user scrolling idly gets a standard model. A user who clicked a shoe ad last week may be routed to a much larger, LLM-scale model that still answers in about 100 milliseconds. The survivors enter the auction, where each bid is weighted by the predicted probability that she acts.
Behind this live path sit two systems that never touch the request directly. Lattice has spent the years since 2023 merging hundreds of narrow ranking models into fewer, larger ones that share learned patterns across surfaces, so behavior observed on Reels can inform predictions on Stories. GEM sits above everything as the teacher. It is trained on behavior across all of Meta's ad surfaces, and its knowledge is distilled into the smaller models that do the serving. CFO Susan Li said on the 3Q'25 call that running GEM for live inference would be "too cost prohibitive."
That one sentence holds the economic logic of the whole stack. GEM's gains are second-order by construction. Its quality reaches revenue only through the student models it trains. A scaling law on GEM is therefore a scaling law on the teacher. What investors need is a scaling law on the transfer.
Reading the Lift Numbers Without Adding Them Up
Meta has given the market a steady sequence of attributed improvements, and the habit of treating them as a running total is the most common misreading of the company's AI disclosures.
When Meta first quantified GEM's impact in 2Q'25, it cited a 5% lift in conversions on Instagram and 3% on Facebook Feed. The Adaptive Ranking Model, launched on Instagram in Q4 2025, was credited with a 3% increase in conversions and a 5% increase in click-through rate, but only for the users it targets. Lattice-driven consolidation, which folded Facebook Stories and other surfaces into the main Facebook model alongside other back-end changes, was tied to a 12% increase in ad quality in Q4 2025.
These figures use three different denominators. One is a surface-level conversion lift. One applies to a routed subset of users. One is an "ad quality" measure that is not a conversion rate and may combine several changes. They cannot be summed, and none of them is revenue. A 5% conversion lift on Instagram does not mean 5% more Instagram ad revenue. Better conversion raises the return advertisers get per dollar. How much of that improvement Meta keeps depends on how the auction clears and whether advertisers raise their budgets in response.
The new GEM table Meta published this month continues the series. The source analyst's reading of it is that the improvements are not plateauing. We have not reproduced the table's figures here. Even taken at face value, it measures the model, not the business. The trend is encouraging. It also sits one step further from the income statement than the headline suggests.
Why a Predictable Curve Changes the Capital Question
A scaling law in the LLM sense makes a specific claim: model loss falls along a smooth power law as compute, data and parameters rise. Smooth does not mean cheap. A power law delivers diminishing returns per unit of compute, and each equal step down in error costs a multiplicatively larger step up in compute. Its value is predictability. A board can underwrite a curve. It cannot underwrite a hunch.
Meta's ad stack is unusually well suited to exploit that predictability, and the reason is the teacher-student split. GEM's cost is overwhelmingly training compute. Training compute behaves like a fixed cost: it is spent once and amortized across every ad request Meta serves, which across Facebook, Instagram, Messenger and WhatsApp is an enormous base. Inference compute, the cost that grows with usage, is controlled separately by Adaptive Ranking. Its routing logic spends heavy computation only on requests where the expected value justifies it.
So the costs scale in two different ways. Spending on the teacher can rise along the curve without a matching rise in serving costs, because the expensive model never serves. Serving costs can be held down by pointing heavy models only at high-intent users. Most consumer AI products lack this separation. A chat assistant's inference bill grows with every query, and its monetization is still largely unproven. An ads teacher model converts training compute into better predictions inside a revenue engine that already exists and already prices every impression in real time.
This is the basis for the source analyst's view that Meta's ad engine may remain the best primary home for its compute, and the logic is sound as far as it goes. The ad system has a measurable objective, a fast feedback loop through clicks and conversions, and an auction that turns better predictions into higher willingness to pay within days. Few other internal uses of a GPU offer all three. The link we cannot verify is the size of the payoff at the margin, meaning the next block of compute rather than the last one.
The Countercase: Offline Curves and Online Dollars
The strongest objection is that the scaling law describes the wrong layer. Ad-model scaling results are typically measured on offline prediction metrics, which track how well the model forecasts clicks and conversions on held-out data. Between that metric and Meta's revenue sit at least three lossy steps.
The first is distillation. If the student models absorb a shrinking share of each improvement in GEM, then a smooth curve at the teacher becomes a flattening curve at the point of sale. Meta's disclosures do not quantify transfer efficiency, and that gap is the most important one in the story.
The second is the auction. Better predictions raise advertiser ROAS, and the gain is split between advertisers and Meta. Advertisers often raise spend when returns improve, but budgets are finite and partly set outside the platform. Some of the surplus stays with the advertiser, which supports retention but not pricing.
The third is the range of the fit. A power law fitted across the compute levels Meta has tested says nothing reliable about levels it has not tested. The LLM field has repeatedly debated whether its curves bend as high-quality data runs out. Meta's ad data has a structural advantage here: user behavior regenerates every day, so the training set renews itself. The constraint on that advantage is signal loss from privacy changes, which limits the conversion data that feeds the loop.
There is also opportunity cost. The same accelerators that train GEM are wanted for Meta's frontier-model ambitions. A predictable ad curve can make the ad engine the default home for compute. It can also make the ad engine the place compute gets taken from when the frontier effort needs more, because the ad return is the one management can forecast.
None of this breaks the thesis. It moves the burden of proof from the engineering blog to the income statement.
Where the Curve Has to Show Up
The scaling law needs confirmation in the quarterly filings, specifically in the two drivers Meta breaks out for advertising: growth in ad impressions and growth in average price per ad. Impression growth comes mostly from users and engagement, so it reflects how much ad inventory exists. Price per ad is where better predictions should show up. If GEM and its students raise conversion rates, advertisers should pay more for each impression, and price per ad should hold up even as impression growth slows with maturing engagement.
The second test is the gap between revenue growth and depreciation growth. Capex arrives on the income statement as depreciation, with a lag. A working ad scaling law shows up as revenue keeping pace with the depreciation that the GPU fleet generates.
The read invalidates under a specific combination: price per ad decelerating in quarters when Meta is still citing larger model gains and still raising capital spending. That pattern would mean the curve is improving the model faster than it improves the auction, with the transfer step or the advertiser's share absorbing the gains. Until the filings show that, the cleaner reading is that Meta has found a way to turn compute into ad pricing along a curve it can forecast, and that the market is still valuing its capex as if the payoff were unknown.
