이미지: X — 벤치마크·평가 화면 갈무리
Summary
- Epoch AI analyzed quarterly data on actual AI chip purchases from 2023 to 2025 and found that performance per dollar improved by an average of 49% per year, doubling roughly every 21 months
- The pace of improvement was uneven — nearly flat in 2023, up 44% in 2024 as the B200 began rolling out, and up 80% in 2025 as spending shifted to newer-generation chips
- Google's TPU v6e delivers about 6x the performance per dollar of the H100, a gap attributed to lower costs from in-house design and manufacturing rather than NVIDIA-style margins
- 연평균 성능/달러 개선율
- 49%(90% 신뢰구간 36~66%), 21개월마다 두 배
- 연도별 개선폭
- 2023년 거의 정체 → 2024년 +44% → 2025년 +80%
- GB300 vs P100(2016, 2025년 달러 기준)
- 가격 약 5.5배, 성능 약 200배, 비용효율 약 37배
- TPU v6e 성능/달러
- H100 대비 약 6배(그래프상 약 4.5~5x 구간)
- 분석 방식
- 2023~2025년 분기별 실제 구매 칩 기준, 버블 크기가 해당 분기 지출액
- 출처·발행
- Epoch AI, 2026년 8월 13일
A Dollar Buys Twice the Compute Every 21 Months
Spending on AI infrastructure keeps rising every year, but so does the amount of compute that same dollar can buy. According to an analysis by AI research organization Epoch AI, which examined quarterly data on actual AI chip purchases from 2023 through 2025, performance per dollar improved by an average of 49% annually — a pace that doubles roughly every 21 months. The figure is reportedly based not on the spec sheets of any single chip, but on a spending-weighted average of chips actually sold and deployed during that period. The confidence interval spans a wide 36% to 66% per year, but the overall trend is clear.
Not a Smooth Curve, But Step-Function Leaps
Notably, this improvement did not happen smoothly. In 2023, spending was concentrated almost entirely on NVIDIA's H100, and performance per dollar barely moved. Once the B200 began shipping in 2024, the figure jumped 44%. In 2025, as most spending shifted to the newest generation, it surged 80%. Each new chip generation costs more, but performance rises by a far larger margin, ultimately improving cost-efficiency overall.
NVIDIA's GB300 illustrates this well. In 2025 dollar terms, the GB300 costs about 5.5 times as much as the P100, which launched in 2016 — but delivers roughly 200 times the performance, translating to about a 37x improvement in cost-efficiency. Over nine years, the price rose about fivefold while performance rose 200-fold — not simply "a more expensive chip," but a far better deal overall.
Performance per Dollar by Chip, Relative to H100
A chart released by Epoch AI shows sizable gaps between chip generations. Below are the approximate performance-per-dollar ranges for each chip, indexed to the H100's 2025 price as 1x.
| Chip | Availability | Performance/Dollar (H100=1x) |
|---|---|---|
| H20 | Q1 2024–Q3 2025 | 0.3x |
| A100 | Q1–Q3 2023 | 0.55–0.6x |
| H100/H200 | Q1 2023–Q3 2025 | 0.75–0.95x |
| GB200 | Q1–Q3 2025 | 1.6–1.7x |
| GB300 | Q3 2025– | 2–2.2x |
| Trainium2 | Q3 2024–Q3 2025 | 3.7–4x |
| TPU v6e | Q3 2024–Q3 2025 | 4.5–5x |
As the table shows, Epoch AI estimates that Google's in-house-designed TPU v6e delivers about 6x the performance per dollar of NVIDIA's H100. The reason cited is structural: because Google contracts directly for design and manufacturing rather than going through NVIDIA, it avoids the added margin. Amazon's in-house chip, Trainium2, also shows roughly 4x the cost-efficiency of the H100, backing up with numbers why major cloud companies are pushing into custom chip development.
Why These Numbers Matter for the Infrastructure Race
Google's recent move to provide Anthropic with $15 billion in financial guarantees — confirmed on July 31 — along with a 20% stake in a Texas data center and a dedicated TPU supply arrangement, reflects this underlying calculus: how much compute can be secured for the same money. In a period when performance per dollar is improving sharply year over year, it becomes more advantageous to move quickly to newer-generation chips rather than lock into long-term contracts for older ones. This dynamic also explains why AI companies are renegotiating data center contracts on shorter cycles or pushing into custom chip design of their own.
So What Does This Change?
What this analysis shows is that AI infrastructure spending isn't simply pouring money into a bottomless pit. Spending keeps rising, but the compute purchased with that spending is growing even faster each year. Still, the fact that this improvement is concentrated around the release timing of specific chip generations carries significant implications for both cloud providers and AI companies. Missing the window to adopt a new chip generation means paying far more than competitors for the same amount of compute. This data puts a number on the cost gap between companies that design their own chips, like Google, and those that depend on NVIDIA chips — giving the industry a new benchmark to watch as data center contracts and chip procurement strategies continue to evolve.
