Spicy Takeaways From Token 2049

Gm! Did you miss us? I was on the ground at Token 2049, and it was incredible to meet so many Chain of Thought readers in person. I was blown away by the cool things some of you are working on and the stories you all have to share.

Now that I’m back at the desk and my brain has recovered from the mush over a whirlwind week, I’m penning down my thoughts while they’re still fresh.

You can expect much more Crypto AI research from us—there’s a ton in the pipeline.

By the way, even more important to me than intellectual learning is the people I catch up with.

Big shoutout to our friends at Onchain Outpost and TAO Times, who consistently cover the top developments in Crypto AI. Their updates, especially on the Bittensor ecosystem, are second to none.

General Conference Views

• Crypto AI was a hot topic. Every day, there was an “AI summit” somewhere, but with many repetitions—decentralized AI, training, and VC investing were common themes. Often, the same speakers rotated across multiple panels throughout the day, making it feel a bit redundant.

• SG is buzzing. This year's scale is massive. Last year, it was primarily people from Asia (10K+); this year, it feels like everyone flew in—20-hour flights included (20K+). I recognised crypto folks everywhere, not just around Marina Bay Sands but all over the city—Bugis, Raffles Place etc.

• Cautious optimism: The 2021 insane bullish vibes aren’t quite back, but there’s a steady, cautiously optimistic feel. There’s still some excess (note to self: maybe short anyone spending millions of investor money to book out a club for a few nights)

Crypto AI Funding 

• VC skepticism lingers: Several VCs remain wary of funding Crypto AI startups. Common tropes:

  • Valuations are still generally high

  • Hard to pinpoint precisely where crypto adds direct value to AI or why blockchain needs to be part of the product.

  • AI is capital-intensive, and decentralization tends to drive up costs.

  • To deploy capital, there needs to be a real conviction that the market opportunity is big enough—more crucial than the valuation itself.

• Liquid funds: This is not specific to AI, but the fund managers I spoke to are optimistic and patient, waiting on the November elections and future rate cuts before deploying more capital. This means a lot of money is sitting on the sidelines. If you believe we’ve hit an inflection point with the 50bps Fed rate cuts, we could see more capital flow into risk assets like crypto over the next six months.

Personal Takeaways on Crypto AI

• Compute is king: It’s clear to me that every Crypto AI startup needs access to compute, whether by building their own decentralized networks/marketplace, or by leveraging existing ones. This raises the question—will we see one player aggregate all the compute supply, or will it stay fragmented? Decentralized AI relies on compute being accessed permissionlessly.

• Decentralized training: A lot of the foundational research on distributed training has been done in the academic space, but now it’s about testing it in production and seeing how far we can push the boundaries. Network latency remains an issue (slow + expensive), but a distributed setup could make certain parts of training costs cheaper. Edge devices like mobile phones aren’t yet capable of training AI at scale.

• There’s a growing agreement that Inference and computing power could shift shift towards decentralized networks, especially in latency-tolerant applications. OpenAI’s o1/strawberry model is a sign of this (slower but better outputs)

• Verifiable inference: The market for verifiable inference (zkMP, opML, etc.) is still uncertain, with few convincing use cases beyond the moral need for trustless verification. That said, privacy is crucial, and its role in AI is becoming more important.

• Private data = better models: Private data, such as in healthcare, can lead to more valuable AI models. Personalization is key to the next AI unlock; stripping out personally identifiable information diminishes the data’s utility.

• Fully Homomorphic Encryption (FHE) allows AI models to train on personal data while keeping it private, but the computation cost is very steep. FHE is still in the research phase, probably 3-5 years away from production-ready applications.

• AI Agents are a hot topic. Some interesting perspectives floating around:

  • Could prediction markets help AI agents make better decisions?

  • Are we over-anthropomorphizing AI agents? Maybe we shouldn’t view them as doing “human” work.

  • On-chain data is a major PITA (pain in the ass) to work with

  • Use cases for on-chain agents are still up for debate. No one’s quite sure yet. Ideas like intent-based MEV and Uniswap hooks are being tossed around, but nothing’s concrete.

• Several Crypto AI founders I spoke with are seriously considering relocating to San Francisco/Bay Area. There’s real alpha in being on the ground in SF—you get wind of the latest AI developments weeks ahead of the rest of the world.

Bittensor

Opinions are sharply divided about the most intriguing Crypto AI protocol. Everyone I spoke to had something to say.

For those hesitating, here are some concerns that came up in my conversations:

• With a $10B fully diluted valuation, how much upside is left realistically? Getting to a 5x will be a tough climb outside of market mania.

• dTAO is a significant upgrade with major implications for incentive allocation and user experience. It could lead to unintended consequences if not rolled out carefully (slowly and iteratively).

• There are concerns about the OpenTensor Foundation (OTF). OTF still holds centralized power, and there’s been tension between subnets and OTF over reward distributions. Some worry that OTF might have checked out.

On the flip side, there’s plenty to be excited about. xponentcrisis did an excellent recap of the Bittensor event this week.

It’s obvious to me that Bittensor is the Schelling point for Crypto AI today. Everyone in the space—builders, investors, researchers—is paying close attention. Funds are indexing on TAO as their bet on Crypto AI, avoiding the hassle of diving into individual startups. New AI projects are seriously considering launching as subnets, where they can instantly gain access to rewards and users. The community is growing, strong, and full of smart, passionate individuals—groups like this are hard to fade.

If the market runs & AI continues on its exponential trajectory, those who believe in the way of the TAO could do very well.

Our analyst, Josh, also attended several Crypto AI events over the week. Here are some of his key takeaways.

NEAR Keynote

Alex Skidanov(Co-founder, NEAR) discussed the rapid expansion of AI applications within the NEAR ecosystem and how these apps seamlessly interact across the AI stack, from data to applications.

NEAR is strongly focusing on user-owned AI, where users have complete control over their data and ownership in AI systems.

He also shared a demo of Web Sim, an AI-powered tool for website generation. Web Sim offers a glimpse into the future of web interfaces, where users can build their own UI instead of relying on traditional front-ends.

An AI model can generate the website by the user simply specifying the features they want through a prompt. As long as there’s an API or smart contract to interact with, users will enjoy full personalization of their web experience.

Panel: Onchain Intelligence on Robust Decentralised AI Infrastructure

Kartin from Ora highlighted that zkML is currently too expensive for on-chain AI, while OpML cost few cents per inference, making it a much more viable solution for verifiable inference. It’s the method of choice for Ora Protocol.

Michael Heinrich from 0G Labs addressed the challenge of competing with closed-source models, pointing out that there simply aren’t enough H100s in Crypto AI to match their scale.

However, distributed training could be the game-changer that helps Crypto AI catch up. He also suggested a shift in approach—using smaller, specialized models trained as experts in specific topics and combining them rather than relying solely on massive foundational models that demand enormous resources.

Morpheus’s deAI summit

Data labelling could become the largest gig economy in the world.

Sapien AI emphasized that expert knowledge is essential for developing AI models. Frontier data—the highest-quality, most insightful data—will be scarce and incredibly valuable. Take, for example, a doctor’s process for diagnosing a patient. In the end, human experts remain a critical part of AI's future growth, as their expertise fuels the training of these models.

Final Thoughts

Most people I spoke with at Token 2049 are genuinely excited about Crypto AI and are paying closer attention.

I’ve never seen anything like this before: crypto tied to a rapidly accelerating tech trend (gen AI → AGI) that’s guaranteed to transform the world in the coming years.

It definitely feels like the next major crypto narrative after DeFi and NFTs, and it will take its own unique shape. Many new products and tokens will be launching over the next 6–12 months.

The real debate is around timeframes. How soon will this play out—one year or five? And how big a role will crypto play?

Personally, I think it’ll be huge. And we’re just getting started.

Cheers,

Teng Yan

Note: These represent our personal opinions and not those of any other organisation or individual.

This newsletter is intended solely for educational purposes and does not constitute financial advice. It is not an endorsement to buy or sell assets or make financial decisions. Always conduct your own research and exercise caution when making investments.