Some Thoughts on the DeepSeek Bloodbath

1) Equities Selloff

The equities selloff should be a reminder that so much of the capital activity in AI across the last couple of years has been a story of compute buildout (ie: the supply side) subsidized by the cashflows of the most profitable companies in history - which is why things are hypersensitive to demand shocks In other words, the capital flows in AI to date are not indicative of AI being here today at the scale implied by those flows.

We’re still deep in the experimentation phase – very likely technologically, but absolutely economically

2) Overblown Fears

The sell pressure on public equities feels somewhat overblown

- In part because the LT business case has always centered on inference (ie: using this stuff) – and falling input costs should only accelerate experimentation and adoption

- Also, the most obvious near term beneficiaries of AI at scale are still big tech platforms with plug + play distribution - which leads to a pretty circular compute setup (ie: deploying / channeling value back to hyperscaler infra anyway)

^ IMO even if open source models proliferate, a big portion of the wrappers and applications built on top of them will likely still run through conventional compute sources

For the biggest AI labs - they've barely figured out a sustainable business model. But that isn't news - and neither is the phenomenon of smaller teams using creative training techniques + distillation to build performant models

The selloff DOES prove how effective the strategy has been of screaming AGI to raise tens of billions and distract investors from the awkward topic of profitability

3) Geopolitics

DeepSeek's progress is indicative of just how misguided the US's ongoing approach to protectionism really is

Brilliant builders will always find a way - and trying to choke them off by limiting key resources will at best simply delay their success and divert it to different parts of the world

4) Crypto x AI Agents

AI Agents in crypto had a massive few months – but this selloff is a reminder that we didn’t find PMF at the AI Application layer, despite the many billions in volume and market caps

Onchain agents are largely speculative proxies atop the broader AI meta that offer people a more interesting, granular way to get exposure

^ To be clear, novel speculation mechanisms have their worth, but the main problem is that to date we’ve obfuscated what’s really playing out

Inference costs have already been plummeting for some time, and OSS innovations like those from DeepSeek will continue to drive input costs down while enabling materially better performance

But the major question is a familiar one: is the infra finally good enough to enable compelling, breakout consumer use cases?

5) Infra -> Apps

IMO – much of the investment and building activity in Crypto x AI has concentrated on infrastructure (compute, data, model development / deployment)

But too much of this has been predicated on transitory theses, temporary constraints or misunderstood technical and economic assumptions subsidized by incentives or ideology

I think there’s still room for world class Web3 builders to make an impact at the AI infra layers, but we need to get serious about where we have unfair advantages and can build with differentiation

Tokenized API wrappers that live on a Web2 social platform aren’t driving the future forward - but my sincere hope is that continued technical and economic breakthroughs in OSS AI empower developers with a renewed drive to unlock the future of AI applications onchain