|
| 1 | +<script> |
| 2 | + import BlogHeader from "$lib/BlogHeader.svelte"; |
| 3 | + import bg from "./bg.avif"; |
| 4 | +</script> |
| 5 | + |
| 6 | +<BlogHeader {bg} title="CrofAI updated my worldview" /> |
| 7 | +<p> |
| 8 | + <a href="https://ai.nahcrof.com/home">CrofAI</a> is an AI inference provider I learned about a few |
| 9 | + days ago. I used to think inference was fast OR cheap - all providers were either mid, cheap but slow |
| 10 | + or unreliable, or fast but expensive. CrofAI is both fast and cheap. |
| 11 | +</p> |
| 12 | +<p> |
| 13 | + Numbers show this best. Let's say you're using AI, specifically Kimi K2, to clone a set of |
| 14 | + calculators. Across 100 calls, you used 10K input tokens and 100K output tokens. How slow and |
| 15 | + expensive would this be on your favorite provider? |
| 16 | +</p> |
| 17 | +<table> |
| 18 | + <tbody> |
| 19 | + <tr> |
| 20 | + <th>Please consult the table</th> |
| 21 | + <th>Avg time</th> |
| 22 | + <th>Total cost (~)</th> |
| 23 | + </tr> |
| 24 | + <tr> |
| 25 | + <td>Groq (OR's fastest)</td> |
| 26 | + <td>6.32s</td> |
| 27 | + <td>$0.31</td> |
| 28 | + </tr> |
| 29 | + <tr> |
| 30 | + <td>CrofAI Turbo</td> |
| 31 | + <td>4.67s</td> |
| 32 | + <td>$0.21</td> |
| 33 | + </tr> |
| 34 | + <tr> |
| 35 | + <td>DeepInfra (OR's cheapest)</td> |
| 36 | + <td>14.99s</td> |
| 37 | + <td>$0.23</td> |
| 38 | + </tr> |
| 39 | + <tr> |
| 40 | + <td>Chutes (OR's cheapest)</td> |
| 41 | + <td>15.73s</td> |
| 42 | + <td>$0.05</td> |
| 43 | + </tr> |
| 44 | + <tr> |
| 45 | + <td>CrofAI non-turbo</td> |
| 46 | + <td>29.5s</td> |
| 47 | + <td>$0.04</td> |
| 48 | + </tr> |
| 49 | + </tbody> |
| 50 | +</table> |
| 51 | +<p>Almost certainly, your favorite provider is beat by CrofAI.</p> |
| 52 | +<p> |
| 53 | + Gemma 3n is also on CrofAI, which raises an interesting point: <em>if</em> |
| 54 | + <code>input_audio</code> |
| 55 | + blocks worked, we could transcribe at the price of |
| 56 | + <abbr title="(1/(160ms)) * ($0.03/million) + (150/minute) * ($0.05/million) in $/hour" |
| 57 | + >$0.001/hour</abbr |
| 58 | + >, dethroning <a href="/blog/humanresearch/phi-4/">Phi 4 Multimodal</a> |
| 59 | + and transcribing a whole year of audio (including night) for just $9.86. This is what "the price of |
| 60 | + intelligence [going] to zero" looks like. |
| 61 | +</p> |
| 62 | +<p> |
| 63 | + CrofAI is a story about markets in a sense. They're performing arbitrage: there are a lot of cheap |
| 64 | + 3090s and 4090s out there, but nobody went to the effort of finding the cheapest ones, manually |
| 65 | + optimizing inference, and selling it as a service with a little markup. Doing this is how they |
| 66 | + become better than any other choice and prove themselves to costumers, and eventually get into |
| 67 | + OpenRouter (OR please add CrofAI thanks). |
| 68 | +</p> |
| 69 | +<p> |
| 70 | + Who's the closest to catching up to CrofAI? Arguably Chutes or another entity on Bittensor. Their |
| 71 | + prices are already decent, and if there's an incentive for making kernels faster, they could go |
| 72 | + down even more and kill CrofAI in the process. However, I doubt this is soon: Chutes moves slow, |
| 73 | + and CrofAI has already acknowledged Chutes as competition, planning even more price cuts to K2 and |
| 74 | + other models within the next week. |
| 75 | +</p> |
0 commit comments