# Epoch AI’s new index says FLOP/s badly undersells modern AI chips

> Source: <https://cryptobriefing.com/epoch-ai-inference-performance-index-chips/>
> Published: 2026-11-18 18:07:33+00:00

**Recommendation:** Evaluate inference hardware with simulated concurrent-request capacity at 50-400 tokens per second per request rather than FLOP/s alone, since memory bandwidth and interconnect can shift rankings by up to 10x.
More in the playbook "Inference and self-hosting": <https://wpnews.pro/playbooks/inference-serving>

# Epoch AI’s new index says FLOP/s badly undersells modern AI chips

The Inference Performance Index finds agentic inference capacity grew 4.6 times in a year, well ahead of raw compute growth

For years, the AI hardware world has kept score with one number: FLOP/s, or how many floating-point calculations a chip can perform each second. Epoch AI thinks that number has been telling only part of the story.

The AI research organization launched its **Inference Performance Index** (IPI) Version 1 on November 9, 2026. Its first major finding is that [Nvidia](https://cryptobriefing.com/markets/nvidia/)’s GB300 delivers approximately 25 times the inference capacity of the H100, despite offering only about 2.5 times the eight-bit FLOP/s.

That is a tenfold gap between what the spec sheet suggests and what the index measures.

## What the index actually measures

The IPI focuses on inference, the work a model does when it answers a request, rather than training. More specifically, it targets agentic workloads, where AI systems carry out multistep tasks.

In technical terms, Epoch AI simulates how many concurrent requests a system can handle across a range of output speeds. Its capacity curves run from 50 to 400 tokens per second per request.

The index also accounts for factors that FLOP/s ignores entirely. Those include memory bandwidth, communication between chips, and compute capability. Memory bandwidth governs how quickly a chip can feed data to its processors, while inter-chip communication determines how well many chips work together as one system.

Those details matter more as models handle long context inputs, which the index treats as a key consideration.

## The scoreboard

Epoch AI normalizes every result to Nvidia’s GB300 NVL72, which receives a score of 1.0 in a unit the organization calls GB300e. Everything else is measured relative to that baseline.

### AI, tech, and the markets they move—in one daily briefing.

Daily. Free. Join 34,000+ readers across crypto, finance, and policy.

The scores look like this:

- H200 HGX: 0.12
- B300 HGX: 0.54
- MI355X: 0.55
- GB300 NVL72: 1.0
- VR200: 2.18
- MI455X: 2.89

The H200 HGX lands at roughly an eighth of the GB300 NVL72’s capacity. The MI355X and B300 HGX sit almost neck and neck, separated by a hundredth of a point. Meanwhile, the MI455X posts the highest score on the list at 2.89, ahead of the VR200 at 2.18.

## Capacity is growing faster than compute

A follow-up announcement on November 18 applied the index to the broader installed base. Epoch AI found that agentic inference capacity grew 4.6 times from Q2 2025 to Q2 2026.

Over the same window, compute grew 2.6 times. So the ability to actually serve agentic AI requests expanded considerably faster than the raw calculation power behind it.

Nvidia’s Blackwell GPU series drove 74% of that increase in inference capacity.

Epoch AI built the methodology on Nvidia’s AISimulate projections and published its code on GitHub.

## Why a better yardstick matters

If inference capacity diverges from FLOP/s by as much as the GB300 and H100 comparison suggests, then compute-only comparisons could steer buyers toward the wrong conclusions. The IPI gives them a figure closer to the job they actually need done: serving lots of users at usable speeds.

The MI355X matching the B300 HGX, and the MI455X topping the list outright, show that rivals can compete when the measurement captures memory and interconnect rather than raw arithmetic alone.

The index relies on simulation rather than live production traffic, and its projections come from Nvidia’s own AISimulate tool while results are normalized to an Nvidia system. The open GitHub release gives competitors and independent researchers a way to test whether the curves hold up.

It is also labeled Version 1, which signals the methodology may evolve.

**Disclosure:** This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our [Editorial Policy](https://cryptobriefing.com/editorial-policy/).
