Scaling Laws Explained: Why Bigger Models Keep Winning (For Now)
Scaling laws are the empirical rule that predicts how a language model's performance improves as you add compute, parameters, and training data, and they've been the single best predictor of AI progress since 2020. DeepMind's Chinchilla paper corrected the original formula in 2022, showing labs had been building models too large for the data they fed them. As of mid-2026, the live debate isn't whether the law holds — it's whether returns are bending at the high end and whether training data and inference cost, not GPU count, are now the binding constraint. For PMs, that shift changes whether the smarter bet is renting a frontier model, fine-tuning a smaller one, or architecting around inference cost rather than waiting for the next parameter jump.
Here's the uncomfortable question: are you still assuming next quarter's model upgrade will be free, that the vendor eats the cost of getting smarter and you just inherit the gains? That assumption held for years. It's getting shakier now that labs are spending the marginal dollar on inference-time reasoning instead of pretraining, which means "smarter" increasingly comes with a per-request bill your product absorbs directly, not a training cost buried in someone else's cap-ex. This week, pull your actual API logs and tag every request by whether it truly needed frontier-model reasoning or could have been served by a smaller, cheaper model with a well-written prompt or a lightweight router in front of it. Most teams find a large share of traffic doesn't need the expensive model at all; they just never built the routing layer because the frontier model was "good enough" not to bother questioning. To be fair, sometimes the frontier model genuinely is the only one that clears your quality bar, and a router adds real engineering cost and a new failure mode to maintain. Don't rip out a working architecture over a blog post. But if you haven't measured which fraction of your traffic actually needs the expensive tier, you're making a budget decision on vibes in exactly the moment the industry stopped doing that. Model access is about to get cheaper and more differentiated at the same time, and the teams that measured first will capture the savings before the pricing catches up.
DeepMind's 2022 Chinchilla paper showed a 70-billion-parameter model beat a 280-billion-parameter Gopher on identical compute, proving prior practice had over-sized models relative to their training data.
In March 2022, DeepMind trained a 70-billion-parameter model called Chinchilla with less than a quarter of the parameters of its own 280-billion-parameter Gopher, on the same compute budget. Chinchilla won, clearly, on benchmark after benchmark — because it saw four times more training data, not because it was bigger.
That result should have killed the phrase "bigger is better" outright. It didn't. The industry absorbed it, redefined "bigger" to mean more total compute rather than more parameters, and kept building.
Here's why it matters whether your team is renting a frontier model's API, fine-tuning something smaller, or training your own weights: scaling laws are genuinely real and predictable, but "just make it bigger" was always a sloppy paraphrase of a stricter rule, and the constraints on that rule (data supply, inference cost, the shape of the loss curve itself) are now the whole ballgame.
My verdict: the scaling law isn't broken. The blank check that funded it is.
Why It Kept Working — To Be Fair to the True Believers
Ignore the slogan for a second. The empirical finding, first documented by OpenAI researchers in 2020 and then corrected by DeepMind's Chinchilla paper in 2022, is that model loss falls in a smooth, predictable power-law curve as compute increases, provided you split that compute correctly between parameters and training tokens. Chinchilla's headline number: roughly 20 training tokens for every parameter, split evenly between "make it bigger" and "show it more."
Before that correction, labs were doing something closer to the opposite. OpenAI's original 2020 scaling paper implied parameters mattered more than data, and GPT-3, Gopher, and Megatron-Turing NLG were all trained on that assumption. Chinchilla's real message was narrower and less exciting than "bigger is better": the field had been misallocating compute, attributed to a specific, falsifiable resource-allocation formula, not a vibe.
Here's the part skeptics undersell: this formula predicted capability jumps before labs spent the money to build them. That's not normal in engineering, let alone in AI, where progress usually gets discovered by accident. A team could plot a loss curve, extrapolate two orders of magnitude, and describe roughly what a model that size would be able to do, before it existed.
That predictability is what actually justified the capital. Billions of dollars in GPU purchases aren't rational bets on vibes; they're rational bets on a curve that, as of mid-2026, has held up across multiple model generations and several architecture families.
To be fair to the scaling maximalists: they weren't wrong that bigger, fed correctly, reliably got smarter. They made a bet with unusually good odds, and for several years running, it paid out on schedule.
Where the Counter-Evidence Actually Bites
The genuine debate in 2026 isn't whether the power law exists; researchers broadly agree it does. It's whether the curve has started bending, and whether the inputs to it are running out. Both are separately documented, and they're different problems.
On the curve itself: research examining large numbers of trained models has found that returns decelerate faster than the standard formula predicts once training compute passes roughly 10^24 FLOPs. Each further order-of-magnitude jump buys a noticeably thinner slice of benchmark improvement than the last one did. That's a measured deceleration, not just a feeling among skeptics, though its exact scope and permanence are still being argued over.
On the inputs: the "data wall" isn't hypothetical anymore. High-quality, non-redundant human text is finite, and labs now routinely train past the Chinchilla-optimal ratio just to squeeze a cheaper-to-run model out of a fixed data pool.
Meta's Llama 3 405B model, for instance, was trained on roughly 15 trillion tokens against 405 billion parameters, a ratio well north of Chinchilla's 20-to-1, specifically because a smaller, over-trained model is dramatically cheaper to serve at inference. That's the tell: labs are already trading training-time optimality for inference-time economics, which only makes sense if parameter count stopped being the only lever worth pulling.
Scaling a frontier model, at this point, is a lot like carb-loading before a marathon. More fuel reliably helps, right up to the point where you stop converting it into speed and start just carrying dead weight past mile eighteen.
Nobody disputes that carb-loading works. The argument is entirely about where mile eighteen is, and whether the field is past it.
The Calculus Has Moved to Inference
The most consequential shift isn't in pretraining at all. Labs found a second scaling axis: OpenAI's o-series models showed that spending more compute at inference time, letting a model reason longer before answering, could match gains that used to require a bigger pretrained base.
Reports on o3's highest reasoning setting describe tens of millions of tokens and several minutes of runtime for a single hard question. That's not a training cost.
It's a cost your product pays every single time a user hits "send," and some infrastructure analysts now project inference will account for the large majority of total AI compute spend by the end of the decade, a real inversion from the training-dominated years.
Who benefits from that shift? Labs with enough scale to run mixture-of-experts routing, sending easy queries to cheap models and reserving the reasoning tax for queries that actually need it.
Who bears the cost? Anyone who bolted an "always call the frontier model" architecture onto their product back when a periodic training-driven capability jump was the only variable that mattered.
The harder question for anyone building right now isn't whether the next model will be smarter. It probably will be, on the current trajectory; nobody credible is calling an end to progress. It's whether your product's unit economics survive on whichever tier of model your quality bar actually requires, for every request, forever, because that number, not parameter count, is what's actually being fought over inside every frontier lab's roadmap.
Get this in your inbox. AI Rundown Daily delivers original briefings every morning — free. Subscribe →
Frequently Asked Questions
Because parameter count stopped being the number that mattered, and disclosing it invited crude comparisons that don't reflect the real design tradeoffs labs are making. Modern frontier systems mix pretraining size, data ratio, mixture-of-experts routing, and inference-time reasoning budget, and a single parameter figure doesn't capture that combination. It's also plausible some labs prefer not to invite a "diminishing returns" conversation about their next model, though that's interpretation, not a confirmed motive. Treat round-number rumors about model sizes as unverified until a lab states them directly.
No; model choice still matters, especially for tasks near the edge of a smaller model's capability, but cost is no longer a rounding error you can ignore. Start by benchmarking your actual task against two or three model tiers rather than assuming the newest, largest model is required by default. Build or buy a router that sends easy queries to cheap models and reserves expensive reasoning for queries that measurably need it. Cost gaps between tiers are often large enough that even modest routing accuracy pays for itself within a billing cycle.
The documented risk is that returns per unit of compute keep thinning at the frontier, so the next model generation may cost dramatically more to train while delivering a smaller capability jump than the last one did; that's a bet with worsening odds, not a broken bet. A second, less discussed risk is data quality: heavier reliance on synthetic training data to work around the data wall carries a real, published risk of model collapse if labs don't keep mixing in fresh, real-world data. Neither risk means progress stops; both mean you shouldn't plan your roadmap assuming the next jump arrives on the old schedule or at the old price.
