Two exponentials are pulling the cost of frontier AI in opposite directions. Epoch AI estimates the cost to train frontier language models has risen about 3.5 times per year since 2020, while pre-training compute efficiency improves roughly 3 times per year. One curve prices labs out, the other keeps the door open, and the narrow gap between them sets the pace of the field.
The arithmetic is unforgiving at the top. The largest known training run, Grok 4, used around 5e26 FLOP, and power use per run roughly doubles every year. Efficiency is the only force working in the buyer's favour, letting the same capability be reached for less compute as time passes.
The race between cost and efficiency
The two rates are what matter, and they nearly cancel. Training cost climbs about 3.5 times a year while efficiency improves about 3 times a year, so the real cost of holding a fixed capability bar rises only around 1.2 times annually. After three years the headline training bill is up more than 40-fold, yet the price of matching last year's frontier has barely moved. Efficiency is doing almost all the work of keeping AI affordable.
Who this favours
- Labs at the frontier: absorb the 3.5x and chase the largest runs.
- Everyone else: ride the 3x efficiency gain and arrive a year later for far less, the same logic behind last year's model on a laptop.
- Investors: watch the gap. If efficiency stalls, the cost curve wins.
The frontier is expensive on purpose. The rest of the market runs on the efficiency dividend. Both rates are documented by Epoch AI.
The gap decides market access
The difference between 3.5x cost growth and 3x efficiency growth looks small on paper. Over time, it decides who can afford to stay near the frontier. If costs grow faster than efficiency, the leading edge becomes more exclusive even while older capability gets cheaper. If efficiency catches up or pulls ahead, more organizations can compete with less capital.
That is why the efficiency rate deserves as much attention as training budgets. A new architecture, better data mixture, improved optimizer, or more efficient serving method can change the economics for everyone downstream. It may not make the largest run cheap, but it can make last year's largest capability available to many more users.
The effect is strongest outside the frontier. A company that does not need the absolute best model can wait for efficiency gains to lower the cost of a previously expensive capability. This is the economic path from lab demo to ordinary software feature.
Why headline training bills keep rising
Frontier labs spend more because the prize is positional. The first lab to reach a new capability can win customers, investors, talent, and strategic leverage. That race encourages spending even when efficiency improves. Savings are often reinvested into larger runs rather than returned as lower budgets.
This creates a strange public picture. The technology gets more efficient, but the largest checks get bigger. That is not a contradiction. It means labs are using efficiency to climb the curve faster. The same force that lowers the cost of a fixed capability can increase the ambition of the next training run.
Power use follows the same pattern. Better efficiency reduces the compute needed for a given target, but frontier targets keep moving upward. The grid still feels pressure because the field spends the efficiency dividend on scale.
What investors should monitor
Investors should watch whether efficiency gains remain broad or become harder to find. If the 3x improvement slows materially while training ambitions keep rising, frontier economics deteriorate quickly. More capital would be needed for smaller gains, and the number of credible frontier players would shrink.
They should also monitor how quickly efficiency diffuses. Some improvements are published, copied, and absorbed into open tooling. Others remain proprietary to the largest labs. The more private the efficiency gain, the more concentrated the market becomes.
The cleanest signal is cost to reach a fixed capability level. If that cost keeps falling, the broader AI market can thrive even as frontier spending grows. If it stops falling, the field becomes more dependent on a few companies willing to fund massive runs.
The practical takeaway
For most builders, the smart strategy is to ride the efficiency curve rather than chase the frontier. Use the newest models to learn what will become possible, then move stable workloads onto cheaper models as soon as quality is good enough. That approach captures capability without inheriting the largest capital burden.
For labs, the calculation is harsher. They need efficiency research to keep the frontier affordable, and they need scale to stay ahead. The tension between those two needs is the central economics of AI development in 2026.
The cost curve is therefore not only a warning about expensive models. It is a map of how capability spreads: first through capital-heavy frontier runs, then through efficiency gains that make the same work cheaper for everyone else.
The policy angle
Efficiency also matters for national and regional strategy. A country that cannot match the largest frontier budgets may still benefit if efficiency gains make strong models trainable and deployable on smaller clusters. Public funding for data quality, evaluation, open tooling, and energy-efficient infrastructure can therefore widen access even without funding the largest runs.
The opposite is also true. If efficiency gains concentrate inside a few private labs, the market becomes more dependent on those labs for both capability and cost reductions. Watching the efficiency curve is a way to watch the openness of the AI economy, not just its technical progress.
