AI Chip Performance Per Dollar Is Growing 49% Per Year, Doubling Every 1.7 Years

On August 13, 2026, Epoch AI published a data insight that crystallizes one of the most consequential questions in the AI industry: is the hardware foundation of the AI boom getting cheaper fast enough to sustain it? The answer, according to their rigorous analysis of 24 AI accelerators sold between Q1 2023 and Q4 2025, is a resounding yes — but the rate of improvement is accelerating in ways that carry profound implications for who wins, who loses, and whether the trillions being poured into AI infrastructure will pay off.

The headline finding: since 2023, the average dollar spent on AI chips each quarter has yielded approximately 49% more performance each year, measured in constant 2025 dollars. That translates to a doubling time of just 1.7 years. For an industry where training a frontier model can cost hundreds of millions of dollars in compute alone, a 49% annual improvement in price-performance is not a marginal optimization. It is the difference between AI economics that compound and AI economics that collapse.

What the Data Actually Measures

Epoch AI’s analysis is methodologically precise. For each quarter, they calculate a dollar-weighted average of performance per dollar across all AI chips sold. The formula multiplies each chip’s Total Processing Performance (TPP) — a standardized metric that accounts for compute, memory bandwidth, and interconnect capacity — by the number of units sold, then divides by total spending. The result captures not just how fast individual chips are getting, but how much compute the industry is actually buying per dollar in aggregate.

This is a critical distinction. A pure benchmark comparison might show a new chip generation delivering 2× the raw FLOP/s of its predecessor. But if that chip costs 1.8× more, the real-world price-performance gain is much smaller. Epoch’s approach weights by actual sales volume, meaning it reflects what buyers are truly experiencing — not what spec sheets promise.

The dataset covers 24 accelerators spanning the entire competitive landscape: Nvidia’s A100, A800, H100/H200, H800, H20, GB200, and GB300; AMD’s Instinct MI250X through MI355X; Google’s TPU v4 through v7; Huawei’s Ascend 910B and 910C; Amazon’s Trainium2; and Cambricon’s Siyuan 590. This breadth is essential — it captures not just Nvidia’s pricing power but the full ecosystem of alternatives that constrain or fail to constrain it.

Spending is adjusted for inflation using the Consumer Price Index, converting all figures to Q4 2025 dollars. A log-linear regression over 12 quarters yields the 49% annual growth rate, with a 90% confidence interval ranging from 36% to 66% per year (equivalent to doubling times between 1.4 and 2.3 years). In nominal dollars, the growth rate is 45% — still remarkable.

The Acceleration Story

Perhaps the most striking finding is not the headline number but the trajectory behind it. The growth is not smooth. It comes in violent spurts tied directly to chip generation transitions.

From 2023 through mid-2024, price-performance was nearly flat — growing at only about 6% per year. This was the Hopper era: the H100 dominated spending, but its price had not fallen significantly, and no fundamentally new architecture had displaced it. Buyers were paying premium prices for incrementally improving hardware.

Then Blackwell arrived. As Nvidia’s GB200 and GB300 chips began shipping in volume and capturing an increasing share of spending, price-performance growth surged to roughly 2× per year — far exceeding the long-term trend. The numbers tell the story starkly: chips purchased in 2024 averaged 23% better performance per dollar than 2023 purchases. Chips purchased in 2025 averaged 91% better than 2024’s. That is not a gentle curve. It is a step function.

The acceleration has a simple explanation. Each new chip generation delivers a larger performance jump relative to its price increase than the generation before. The H100 offered roughly 3× the performance of the A100 at about 2× the price — a meaningful improvement. The GB200 offers roughly 2.5× the performance of the H100 at a modestly higher price point per unit of compute. Because performance has risen faster than price across recent transitions, and because the industry rapidly shifts its spending toward each new generation, the weighted average of what buyers actually get per dollar is climbing at an accelerating rate.

The Memory Bottleneck

This price-performance revolution has a shadow: the cost structure of AI chips has shifted dramatically toward memory. In a separate but closely related data insight published in May 2026, Epoch AI revealed that high-bandwidth memory (HBM) now accounts for 63% of total AI chip component costs — up from 52% in Q1 2024. Total component spending on AI chips grew from approximately $22 billion in 2024 to $52 billion in 2025, with HBM alone accounting for the lion’s share of that increase.

This matters because the 49% annual improvement in performance per dollar is a measure of what has shipped — not of what the underlying economics can sustain. If memory costs continue to climb as a share of total chip cost, the next generation of improvements will depend increasingly on memory innovation rather than logic die advances. HBM supply is concentrated among just three manufacturers — SK Hynix, Samsung, and Micron — and the four largest AI chip designers consumed approximately 90% of global CoWoS advanced packaging capacity and HBM supply in 2025.

The implication is that the 49% growth rate, while extraordinary, may be difficult to sustain at the same pace if memory supply chains cannot scale at the same velocity as logic improvements. The next doubling in performance per dollar may require not just better chips but a fundamental restructuring of how HBM is manufactured and priced.

The Hyperscaler Question

Epoch AI’s data also intersects with another finding from their research: the financial sustainability of the buyers. In a June 2026 data insight, Epoch reported that aggregate cash capital expenditure across Microsoft, Amazon, Alphabet, Meta, and Oracle is on track to overtake their combined operating cash flow by Q3 2026. These five hyperscalers are projected to spend roughly $770 billion in 2026 — a figure that quadrupled since GPT-4’s release and is growing at approximately 70% year-over-year, against operating cash flow growth of only 23%.

This creates a tension that the 49% price-performance improvement partially resolves. If chips are getting 49% more cost-effective each year, then the same dollar of capex buys dramatically more compute over time. A hyperscaler spending $150 billion on AI infrastructure in 2026 is getting roughly 3.5× the compute-per-dollar it would have received for the same spend in 2023. This compounding effect is what makes the aggressive spending levels defensible — each year’s investment buys substantially more capability than the last.

But the math cuts both ways. If the 49% improvement rate does not hold — if it reverts to the 30-37% range that Epoch measured in earlier analyses, or if memory costs continue their upward trajectory — then the gap between capex and cash flow becomes a structural problem rather than a transitional one.

Context: This Is Faster Than Historical Trends

Epoch AI’s own earlier analyses underscore how dramatic the acceleration has been. Their February 2026 trends page reported AI chip performance per dollar improving at 37% per year, doubling every 2.2 years. A separate October 2024 data insight measured improvement at roughly 30% per year. And in their comprehensive chips topic overview published in July 2026, they noted that across chip generations, performance per dollar had roughly doubled every 2.5 years since the early 2010s.

The new 49% figure represents a sharp upward revision. The acceleration is driven almost entirely by the Blackwell generation’s impact on spending patterns — when buyers shift from H100-class hardware to GB200-class hardware, the price-performance ratio jumps dramatically, pulling the weighted average upward. Whether this rate persists depends on whether subsequent generations — Nvidia’s Rubin, AMD’s MI400 series, Google’s next TPU — can deliver similar or larger performance-per-price jumps.

Raw machine learning performance per chip has grown at roughly 1.6× per year since 2015, which is faster than the rate at which prices have risen. The gap between performance growth and price growth is what produces the price-performance improvement. In the Blackwell era, that gap widened substantially.

Limitations and Caveats

Epoch AI is refreshingly transparent about what their data does and does not capture. The analysis measures chips sold, not deployed — newly purchased hardware takes time to become operational, and older chips depreciate. The performance metric uses peak theoretical specifications from spec sheets rather than real-world workload performance, which depends heavily on software optimization, model architecture, networking topology, and memory bandwidth utilization.

Pricing methodology introduces a subtle bias. Nvidia, AMD, and Huawei chips are priced at external sale prices, which include the designer’s profit margin. Google’s TPUs and Amazon’s Trainium chips are priced at procurement cost — what it costs to manufacture them through partners, without the cloud rental margin — which flatters their price-performance relative to what an external buyer would actually pay. This means the aggregate number somewhat understates the cost that non-hyperscaler buyers face.

The dataset also excludes several notable accelerators: Intel’s Gaudi, Groq’s LPUs, Cerebras’s wafer-scale engines, Microsoft’s Maia, Meta’s MTIA, and Tesla’s Dojo. These represent a growing but still relatively small share of total AI compute, so their exclusion does not materially affect the headline finding — but it means the analysis captures the established landscape rather than potential disruptors.

What This Means for 2026 and Beyond

For the AI industry, the 49% figure carries several concrete implications. First, it validates the massive infrastructure investment thesis. If compute is getting nearly 50% more cost-effective each year, then the return on AI capex compounds aggressively — today’s billion-dollar data center will deliver dramatically more value than one built two years ago.

Second, it underscores Nvidia’s strategic position. Because Nvidia accounted for the majority of spending from 2023 through 2025, the weighted average of price-performance is heavily influenced by Nvidia’s product cadence. The Blackwell generation’s favorable price-performance ratio is what drove the acceleration. Nvidia’s ability to maintain this cadence with Rubin and beyond will determine whether the 49% rate holds.

Third, it raises the stakes for memory. With HBM now consuming 63% of chip component costs, the next phase of price-performance improvement may depend less on who designs the best logic die and more on who can secure the cheapest, highest-bandwidth memory. This is why every major chip designer — Nvidia, AMD, Google, Amazon, and now Anthropic with its custom silicon ambitions — has been aggressively locking in HBM supply contracts.

Finally, it provides a quantitative anchor for the broader debate about AI sustainability. The 49% annual improvement in price-performance is a powerful counterargument to those who fear the AI buildout is a speculative bubble. But it is also a warning: the rate is driven by a specific set of conditions — Blackwell’s favorable economics, massive spending concentration, and a supply chain that has bent but not broken. If any of those conditions change, the trajectory could shift rapidly.

Epoch AI’s contribution here is not just the number. It is the rigor. By grounding the analysis in actual sales volumes, real transaction prices, and inflation-adjusted dollars across 24 chips from seven designers, they have produced the most credible measurement to date of how quickly the hardware economics of AI are actually improving. The answer — 49% per year, doubling every 1.7 years — is faster than almost anyone expected. Whether it can last is the multi-trillion-dollar question.