🦍 State of the Market..

The total Crypto x AI market cap is down a whopping 23% from last week ($28.6B). Some of the biggest losers this week are FET, OCEAN, and AGIX, all are down about 20% since the launch of Artificial Superintelligence Alliance (ASI) on July 1st.

This drawdown is not specific to AI tokens. BTC & ETH are down >10% this week, with several investors calling for a bear market.

On a lighter note, Grayscale released some research on how AI tokens have outperformed this year:

“Year to date, an equally weighted basket of AI-adjacent Crypto Sectors tokens has increased 80%, compared to the small decline for the crypto market”

There may be a brighter future for Crypto x AI in the coming months.

 📊 Chart of the Week

Ora Protocol

Since June 24th, we have noticed that the number of inference calls on Ora Protocol has jumped significantly.

This is likely due to their release of a points program for using their Onchain AI Oracle (OAO). On June 23rd, users could earn 3-6 points repeatedly per each inference call they made through the OAO. Additionally, ORA announced a $20M fundraise last week.

Looks like something is brewing over at Ora Protocol.

🏆 Caught Our Eyes..

• Sentient closes an $85M seed round led by Founders Fund, Pantera, and Framework

• Bittensor network gets hacked for ~$8M due to a malicious package manager, publishes a post-mortem on the validator hack

• Apus Network launching its benchmark Proof-of-Concept on July 15th

• Cloud computing firm Prodia raised a $15M seed round led by Dragonfly Capital

• Kaito.ai reaches profitability in two years, upper percentile amongst GenAI startups

• Allora introduces the Allora Research Forum - a center for research on inference synthesis, parameter optimization, and model creation

• Phala partners with Allora to integrate data privacy and security on Allora using Phala’s TEE network

• Prime Intellect announces its Compute Exchange, a decentralised compute marketplace. They also partnered with Akash to provide NVIDIA H100s on their network

• Grass.io open sources their Reddit dataset with over 600 million posts and comments

• Compute Labs raises $3M in a pre-seed funding round to tokenise GPUs on Solana

• Netmind AI releases their Q3 roadmap

🐰 Research Highlight — Prime Intellect

Every week, we present early, interesting crypto AI projects that caught our eye during our independent research.

Prime Intellect is building critical infrastructure for decentralised AI training at scale.

But they’re more than just a DePIN project, and here’s why.

Prime Intellect’s grand plan has four parts to it:

1. Aggregate global compute

2. Develop distributed training frameworks for collaborative model development

3. Collaboratively train open-source AI models

4. Enable collective ownership of AI models

A GPU Marketplace Aggregator

This week, they kicked off Part 1 by launching their GPU marketplace.

This aggregates compute resources from major centralised and decentralised GPU providers, including Akash, Io.net, Vast.ai, Lambda Cloud and more.

The goal is to provide the cheapest rental rates for the user by aggregating provider supply and providing the tooling to do this easily. So, you can use the Prime Intellect platform rather than going directly to Akash or Io.net and comparing prices.

Their live beta platform is intuitive and easy to use. You can get your cluster up and running within minutes, with no KYC needed.

• You select which location you’d like to lease GPUs from and the security level of the network (i.e. secure or community cloud).

• They also have a “cheapest option” available.

They offer a range of GPUs from the top-end H100s to the RTX3000 and 4000 series. However, in terms of cluster size, they’re capped at 8 GPUs at a time. Prime Intellect is working on increasing this to 16-128 GPU clusters.

Decentralised Training at Scale

Part 2 of their master plan—developing distributed AI training frameworks—excites us the most.

Here’s the situation today:

Building your own data centres is often necessary to train large foundational AI models.

It involves high-speed networking, customised data storage, privacy considerations, and optimizing efficiency—capabilities that cannot be fully realized by simply renting multiple GPUs.

No wonder big tech companies like Microsoft, Google & OpenAI dominate the space. Smaller players don’t have the resources to do this.

Prime Intellect, on the other hand, will enable the training of models across multiple distributed GPU clusters.

Several challenges need to be solved in decentralised training:

• Communication between nodes in different parts of the world, optimizing latency and bandwidth for data transfer

• Accommodating different types of GPUs used in these networks

• Fault tolerance: The training process must be resilient to changes in GPU cluster availability since they can drop in and out

This involves scaling some of the cutting-edge research today into actual production systems:

• Distributed Low-Communication Training (DiLoCo): A method for data-parallel training on poorly connected devices, synchronizing gradients every 500 steps instead of at each step.

• Distributed Path Composition (DiPaCo): A routing mechanism where each worker processes data specific to one path only.

• Google Gemini’s Ultra Training: Using approximately 18 superpods across different data centres, utilizing proprietary software to transmit large amounts of data efficiently.

• SWARM Parallelism: A parallel training algorithm that optimizes based on the speed and success rate of each worker in the cluster.

If Prime Intellect can solve this, it will have a huge impact on how models are trained and the utilization of resources to train them.

The last feature that Prime Intellect is building is a protocol that rewards participants for contributing compute, code, and capital and enables collective governance of AI models.

This fits in with the decentralised AI ethos and encourages users and participants to be part of decentralised AI. We expect they’ll likely utilise crypto as a medium of exchange and ownership here.

Our Thoughts

• GPU marketplaces are a dime a dozen today and not particularly compelling. Although some have managed to aggregate supply using token incentives, the demand side remains weak due to the challenges of decentralised training we mentioned.

• The global decentralised GPU market is very competitive. Stacy Muur compares the rates of several GPU providers:

  

• If Prime Intellect can make decentralised training of AI more effective and efficient, it will open up the floodgates for GPU demand.

• Prime Intellect has some notable investors like Clem Delangue (Hugging Face), Erik Voorhees (Shapeshift), and Andrew Kang (Mechanism Capital).

🧠 Open Source Corner

• Stable Diffusion 3’s disastrous launch could change the AI landscape forever

• The Open Model Initiative has launched. It’s a community-driven effort to promote open AI models for image, video and audio generation. This is in response to community frustration with Stable Diffusion

• Meta’s LLama-3-405B model is likely coming soon, and it could be our first open-source GPT-4 class model.

🔥 On X..

Our friends Liang & Sven wrote a great article on coordinating intelligence using crypto

YouTube: Crypto x AI's effect on global markets

Stacy Muur on why Aethir Cloud is the flagship GPU provider

Erik Vorhees shares about Venice.ai - a generative AI app for non-technical people

X Spaces: The future of decentralised AI from venture capitalists’ Point of View

Michael Hanono on AI Agents

The Atlantic: The Perils of AI Agents

Const on the Bittensor hack