🦍 State of the Market..

The overall Crypto x AI market cap remained relatively flat, rising a meagre 1.3% to $22.7B this week

All eyes were on NVIDIA’s Q2 earnings release. Despite beating estimates (EPS of 68 cents vs. 65 cents expected), the stock took a nearly 10% hit, reflecting the market’s sky-high expectations for the chipmaker.

In typical crypto fashion, NVDA’s post-earnings drop sparked a selloff in several crypto AI tokens. TAO, the AI sector’s bellwether, fell 15% this week.

Meanwhile, a few small-cap AI tokens saw a nice pump. NOS, a GPU marketplace on Solana, surged over 50% without any clear catalyst or news—maybe some insider action?

 📊 Chart of the Week

Akash CPU Usage Returns to Highs

For the past six weeks, CPUs on Akash have been in high demand, with the number of CPUs leased in August nearing 8,000—almost back to the yearly highs. Suppliers are earning ~$0.8 and $1.6 per CPU thread per month.

Some of the standout builders on Akash include Venice.ai and Nous Research.

Nous Research recently made a breakthrough in distributed training using Akash. Similarly, NVIDIA acquired Brev, which was using Akash for compute. All in all, this year has been a whirlwind for Akash.

You can hear founder Greg Osuri talk about Akash’s growth here

🏆 Caught Our Eyes..

Project Updates

• Space and Time, an AI ecosystem project, closes a $20M Series A bringing total funding to $50M. Arkstream Capital explains why they invested.

• Skyfire uses USDC as a stablecoin network to enable purchasing power for AI agents

• Atoma Network releases its technical whitepaper

• Mizu AI, a group building large open-source datasets, releases the Mizu bot, a telegram bot for mining rewards through data processing.

• Nous Research released a new open dataset, Hermes Function Calling V1, which gave Hermes 2 Pro its tool use and output capabilities

• RISC Zero is now fully open-source, releasing its source code and compiler tech to the public

Bittensor

• Bittensor starts 7-day competitive testing of dynamic TAO

• Brody summarizes the past week of Bittensor updates

• Masa, a contribute data-to-earn project, launched its subnet 42 mainnet on August 27th

Decentralised Compute

• Hyperbolic makes the Llama 3.1-405B Base model available on their platform

• Prime Intellect added full H100 nodes to their marketplace

• Mira and Hyperbolic partner, with Mira leveraging Hyperbolic’s GPU marketplace

• Aethir announces their 6-month roadmap, doubling down on Enterprise AI

Incentives and Launches

• Sharpe AI had its TGE on August 27th. It’s currently trading 20% above its IDO price.

• Arweave makes DAI available for AO token minting; rewards begin on September 4th at 1100 EDT

• Nillion Network announces the Nillion Verifier Program, with rewards for early participants

• Open Gradient announces its mission statement

🧠 Here’s How We Get to GPT-8

Scaling laws in AI have sparked much debate.

In simple terms, scaling laws state that as you increase the model size, data, and compute power used to train an AI, its performance gets better. So far, this has held true.

But here’s the catch: there are concerns that we’re hitting the ceiling on available data, compute power, and even electricity to fuel the next generation of massive AI models.

Epoch AI’s analysis suggests otherwise: AI scaling can continue its current trend through 2030. This means that AI training runs could be 10,000 larger than GPT-4.

So we’re likely to see GPT-8 hit the market eventually. It’s starting to feel like the iPhone cycle—new version, better features, same hype.

Who knows? Maybe soon, we’ll be lining up for the latest GPT like we do for the newest iPhone.

🔥 On X..

Can your X account handle a Claude roast?

AI hack: Use this prompt

How decentralised systems can reduce pre-training costs

Nous Research’s report on Distributed Training, gets 3K+ likes

Spectral walks through their AI agent technology BROS

Wei Dai on why prediction markets need zkTLS and zkLLM

Shayon discusses the path forward for DePIN