In September, Epoch AI published a measurement that sounds like pure good news: the cost of hitting the same AI benchmark score has fallen about 47% per quarter — roughly 13× cheaper per year. At its steepest, when a model first becomes the strongest available, the drop ran 66% per quarter.
Thirteen times cheaper, every year. So AI should be getting cheap to run, right?
Now look at a second number. NVIDIA’s CFO commentary for Q2 FY27 (filed with the SEC in August) reports data-center revenue of $89.0 billion, up 117% year over year — and a $105 billion guarantee backing a lease for an AI data-center campus. The IEA says data-center electricity demand surged 17% in 2025 and is set to double by 2030. Alibaba’s quarterly capex jumped 75% to RMB 67.7 billion, citing AI agent adoption as a driver.
Two numbers that look like they disagree
Both are true. They measure different things.
Epoch measures price: what it costs to buy one unit of intelligence at a given quality level.
NVIDIA, IEA, and Alibaba measure spend: what the industry actually pays, in total.
Price fell. Spend exploded. The bridge between them is usage — and usage is outrunning the price decline. Three things changed in 2026:
Agents. A chat answer is one model call. An agent completing a task is dozens to thousands. The unit price dropped; the number of units per task exploded.
Reasoning models. “Think longer before answering” is literally buying more compute per question. Quality went up; so did the bill per task.
The bottleneck moved to physics. GPUs, HBM memory, gas turbines, transformers, grid interconnection queues. When the constraint is a turbine with a 2-year lead time, cheaper tokens don’t relieve it.
The honest limits
Benchmark same-score cost is not your workload. Real tasks mix models, retries, and verification.
NVIDIA is one company. One vendor’s 117% doesn’t prove the whole market — though IEA’s electricity data and Alibaba’s capex point the same direction.
Epoch itself flags benchmaxxing: models get tuned to the tests. The 13× figure is best read as an order of magnitude, not a precise invoice.
What I do with this
One line on practice: a harness that avoids needless model calls is a cost tool, not just a quality tool. Check before you call the model again — the savings compound at agent scale.
That’s the whole game this year: intelligence got 13× cheaper, and somehow the industry is spending much more. Both numbers are real. The gap between them is where 2026 actually lives.
Sources: Epoch AI, “The plunging price of thought” (2026-09-22) · NVIDIA CFO Commentary Q2 FY27 (SEC, 2026-08-26) · IEA press release (2026-04-16) · Alibaba June-quarter results (2026-08-20).
MossRiver footnote: same discipline applies to a storefront — check the real stock before you promise the ship date.