The flip side of AI tokens: most projects are preoccupied with financial interests rather than real-world impact

Original author: Gagra

Original source: gagra.vc

原文标题：Flipping the AI coin

编译:深潮TechFlow

Abstract of this article

• This is not yet another VC article that is optimistic about the "AI + Web3" space. We are optimistic about the merger of these two technologies, but this article is a call to action. Otherwise, this optimism will eventually lose its grounds.
• Because developing and running the best AI models requires huge capital expenditures on state-of-the-art and often hard-to-obtain hardware, as well as domain-specific R&D. Crowdsourcing through crypto incentives, as most Web3 AI projects do, isn't enough to offset the tens of billions of dollars invested by large companies that firmly control AI development. Given the hardware limitations, this may be the first big software paradigm that smart and creative engineers outside of the current organization don't have the resources to disrupt it.
• Software is "cannibalizing the world" faster and faster, and will soon grow exponentially with the acceleration of artificial intelligence. In the current situation, all this "cake" has gone to the tech giants, and the end users, including governments and big business, and of course, consumers, are more dependent on their power.

Incentive misalignment

None of this could have unfolded at a more inopportune time – at a time when 90% of decentralized network participants are busy chasing the big low-hanging gains that come with narrative-driven development. Yes, developers are following investors into our industry, not the other way around. It varies from public acknowledgment to more subtle subconscious motivations, but the narratives and markets that form around them drive a large portion of Web3 decision-making. The participants were too immersed in the reflexive bubble to notice the outside world, except for the narrative that helped drive the cycle further. And AI is clearly the biggest one, as it is experiencing a boom in its own right.

We've spoken to dozens of teams at the AI x Crypto intersection and can confirm that many of them are very capable, mission-driven, and passionate about building projects. But it's human nature that when faced with temptations, we tend to succumb to them and then rationalize those choices after the fact.

Easy access to liquidity has been a historical curse for the crypto industry, which has slowed its development and delayed useful adoption for several years. It will make even the most loyal crypto believers turn to "hype tokens". The rationalized explanation is that the more capital held the token, the better opportunities these builders may have.

The relatively low maturity of institutional and retail capital provides an opportunity for builders to make claims that are detached from reality while still benefiting from valuations as if those claims had been realized. The result of these processes is to actually lead to moral hazard and capital destruction, and few of these strategies are effective in the long term. Necessity is the mother of all inventions, and when needs disappear, inventions cease to exist.

The timing of this couldn't be worse. While all the brightest tech entrepreneurs, state leaders, and large and small businesses are racing to ensure that they benefit from the AI revolution, crypto founders and investors have opted for "rapid growth." In our view, this is the real opportunity cost.

Web3 AI Market Overview

With the above incentives in mind, the categorization of Web3 AI projects really boils down to:

• Legitimate (also divided into realists and idealists)
• Semi-legal
• counterfeiter

Basically, we think builders have a clear idea of what it takes to keep pace with their Web2 competitors, and in which verticals it's possible to compete and in which ones are more of a dream, but these are things that can be advertised to venture capitalists and the immature public.

The goal is to be able to compete at this moment. Otherwise, the speed of AI development may leave Web3 behind, and the world is moving towards Web4, a dystopia of Western corporate AI and Chinese national AI. Those who can't be competitive quickly and rely on distributed technology to catch up over a longer period of time are too optimistic to be taken seriously.

Obviously, this is a very rough generalization, and even the group of forgers contains at least a couple of serious teams (and perhaps more delusionals). But this article is a call-to-heart, so we don't intend to be objective, but rather to appeal to the reader to feel a sense of urgency.

legality

Middleware that "brings AI to the chain". The founders behind these solutions, though not many, have made it feasible or even impossible to understand the models that so far decentralized training or inference users actually want. Therefore, in order for the best centralized model to be connected to the on-chain environment so that it can benefit from sophisticated automation, this is a good enough first step for them. At the moment, hardware enclaves (TEEs, or "null isolation" processors) that provide API access points, bidirectional oracles (for bidirectional indexing of on-chain and off-chain data), and verifiable off-chain computing environments for proxies seem to be the best solution. There are also coprocessor architectures that use zero-knowledge proofs (ZKPs) for snapshot state changes, rather than verifying the full computation, which we also believe is feasible in the medium term.

A more idealistic approach to the same problem attempts to validate off-chain reasoning to align it with on-chain computation in terms of trust assumptions. In our view, the goal of this should be to allow AI to perform on-chain and off-chain tasks in a single, coherent runtime environment. However, most proponents of inferential verifiability talk about vague goals like "trust model weights" that will never even become important in the coming years. Recently, the founders of this camp began to explore alternative methods to verify reasoning, but initially all based on ZKP. Although a lot of smart teams are working on so-called ZKML, they are risking too much in anticipating cryptographic optimizations to outpace the complexity and computational needs of AI models. As a result, we believe that they are not suitable for competition at this time. However, some recent developments are interesting and should not be overlooked.

Semi-legal

Consumer applications that use wrappers for closed and open-source models (e.g., Stable Diffusion or Midjourney for image generation). Some of these teams are among the first in the market and have the attraction of actual users. Therefore, it would be unfair to call them false in generalizations, but only a few are thinking deeply about how to develop their underlying models in a decentralized way and innovate in incentive design. There are some interesting governance/ownership changes in this regard. But most projects in this category simply add a token on a centralized wrapper like the OpenAI API to capture a valuation premium or provide faster liquidity for the team.

The problem that neither of the above camps has solved is the training and inference of large models in a decentralized environment. Currently, there is no way to train a base model in a reasonable amount of time without relying on tightly connected hardware clusters. Considering the level of competition, "reasonable time" is the key factor.

Recently, there have been some promising research results where, in theory, methods such as differential data streaming could be extended to distributed computing networks to increase their capacity in the future (as network capabilities continue to match data flow requirements). However, competitive model training still requires communication between localized clusters (rather than a single distributed device), as well as cutting-edge computing power (retail GPUs are becoming less competitive).

Recently, there has also been progress in research on localizing inference (one of the two ways of decentralization) by reducing the size of the model, but there are no existing protocols that take advantage of it in Web3.

The problem of decentralized training and inference logically leads us to the last and most important of the three camps, and therefore the one that is most emotionally triggered for us.

counterfeiter

Infrastructure applications are mainly concentrated in the field of decentralized servers, providing bare hardware or decentralized model training/hosting environments. There are also software infrastructure projects that are pushing protocols such as federated learning (decentralized model training) or merging software and hardware components into a single platform where people can essentially train and deploy their decentralized models end-to-end. Most of them lack the sophistication needed to actually solve the problem in question, and the naïve idea of "token incentives + market winds" prevails here. Neither the solutions we see in the public or private markets can compete meaningfully at this moment. Some solutions may evolve into viable (but niche) products, but what we need now is fresh, competitive solutions. This can only be achieved through innovative design that solves the bottleneck of distributed computing. In training, not only speed is an issue, but also the verifiability of work done and the coordination of the training workload, which adds to the bandwidth bottleneck.

We need a competitive and truly decentralized set of foundational models that require decentralized training and inference to be effective. If computers become intelligent and AI is centralized, then there will be no world computers to talk about except some kind of dystopian version.

Training and inference are at the heart of AI innovation. While the rest of the AI world is moving towards a tighter architecture, Web3 needs some orthogonal solutions to compete with it, as head-to-head competition is becoming less viable.

The scale of the problem

It's all about computing power. Whether it's in training or inference, the more you put in, the better the results. Yes, there are some tweaks and optimizations here, and the computation itself is not homogeneous, there are now all sorts of new ways to overcome the bottlenecks of the traditional von Neumann architecture processing unit, but in the end, it all comes down to how many matrix multiplication operations you can do on how big a block of memory and how fast you can do it.

That's why we're seeing so-called "hyperscalers" building so strongly in data centers, all looking to create a full-stack with powerful processors for AI models as the top layer and the hardware that supports it as the bottom layer: OpenAI (model) + Microsoft (compute), Anthropic (model) + AWS (compute), Google (both) and Meta (increasingly getting involved in both by doubling down on data center expansion). There are many more nuances, interaction dynamics, and parties involved, but we won't discuss them here. The big picture is that hyperscalers are investing unprecedented billions of dollars in data center expansions and creating synergies between their computing and AI products, which are expected to pay off as AI becomes more widely used in the global economy.

Let's just look at the level of expansion that these 4 companies can expect this year:

• Meta expects capital spending in the range of $30 billion to $37 billion in 2024, which is likely to be heavily skewed towards data centers.
• Microsoft's capital expenditure in 2023 is around $11.5 billion, and it is rumored to invest another $400-50 billion in '24-25! This can be partly inferred from the huge data center investments announced in just a few countries: $3.2 billion in the UK, $3.5 billion in Australia, $2.1 billion in Spain, €3.2 billion in Germany, $1 billion in Georgia and $10 billion in Wisconsin. And these are just some of the regional investments in their network of 300 data centers in more than 60 locations. There are also rumors that Microsoft may spend another $100 billion to build a supercomputer for OpenAI!
• Amazon's leadership expects their capital spending to grow significantly in 2024, compared to $48 billion in 2023, largely due to the expansion of AWS infrastructure for artificial intelligence.
• Google spent $11 billion to expand its servers and data centers in the fourth quarter of 2023 alone. They acknowledge that these investments are made to meet the expected demand for AI and expect the speed and total amount of their infrastructure spending to increase significantly in 2024 due to AI.

Here's how much NVIDIA is already spending on AI hardware in 2023:

Nvidia's CEO, Jensen Huang, has been touting $1 trillion in AI acceleration over the next few years. He recently doubled that forecast to $2 trillion, allegedly due to the interest he sees from sovereign players. Altimeter's analysts expect global spending on AI-related data centers to be $160 billion in 2024 and more than $200 billion in 2025.

Now compare these numbers to those that Web 3 offers independent data center operators to incentivize them to scale up their capital expenditures on the latest AI hardware:

• The total market capitalization of all decentralized physical infrastructure (DePIn) projects is currently around $40 billion, and these tokens are relatively illiquid and mostly speculative. Essentially, the market cap of these networks is equal to the cap estimate of the total capital expenditure of their contributors, as they incentivize this construction with tokens. However, the current market cap is almost useless because it has already been issued.
• So, let's assume that there is another $80 billion (2x the current value) of private and public DePIn token market caps coming into the market as an incentive in the next 3-5 years, and assume that this is entirely for AI use cases.

Even if we divide this very rough estimate by 3 years and compare its dollar value to the cash spent by hyperscalers in 2024 alone, it becomes clear that applying token incentives to a range of "decentralized GPU network" projects is not enough.

Investors also need billions of dollars in demand to absorb these tokens, as the operators of these networks sell large quantities of these mined coins to cover costs such as capital expenditures. More billions of dollars are needed to drive up the value of these tokens and incentivize the growth of construction to outpace hyperscalers.

However, someone with an in-depth understanding of how most Web3 servers currently operate might expect that a significant portion of the "decentralized physical infrastructure" actually runs on the cloud services of these hyperscalers. Of course, the surge in demand for GPUs and other AI professional hardware is also driving more supply, which should eventually make it cheaper to rent or buy them in the cloud. At least that's what people expect.

But it's also important to consider: Nvidia now needs to prioritize the latest generation of GPUs for its customers. At the same time, Nvidia is also starting to compete with the largest cloud providers on its own turf, offering AI platform services to enterprise customers who are already locked into hyperscale servers. This will eventually prompt it to either build its own data centers over time (essentially eroding the lucrative profits they now enjoy, so it is unlikely) or significantly limit its AI hardware sales to the network cloud providers it works with.

In addition, NVIDIA's competitors have introduced additional AI-specific hardware, mostly using the same chips as NVIDIA produced by TSMC. As a result, basically all AI hardware companies are currently competing for TSMC's production capacity. TSMC also needs to prioritize certain customers. Samsung and potentially Intel (which is trying to return to state-of-the-art chip manufacturing soon) may be able to absorb the additional demand, but TSMC is currently producing most of the AI-related chips, and it will take years to scale and calibrate cutting-edge chip manufacturing (3 and 2nm).

On top of that, all of the current cutting-edge chip manufacturing is done by Taiwan's TSMC and South Korea's Samsung near the Taiwan Strait, and the risk of a military conflict could become a reality before the facilities currently being built in the United States to offset this (and no next-generation chips are expected to be produced in the next few years) could kick in.

Finally, China is largely cut off from the latest generation of AI hardware due to U.S. restrictions on Nvidia and TSMC, and China is competing for the remaining computing power, just like the Web3 DePIn network. Unlike Web3, Chinese companies actually have their own competing models, especially large language models (LLMs) from companies like Baidu and Alibaba, which require a large number of previous-generation devices to run.

As a result, due to a combination of one or more of the above reasons, there is a non-material risk that hyperscale cloud service providers will restrict external party access to their AI hardware as the AI-dominated war intensifies and takes precedence over cloud business. Basically, it's a scenario where they take up all of the cloud computing capacity related to AI for their own use and don't offer it to anyone else anymore, while also gobbling up all the latest hardware. When this happens, the remaining supply of computing will be more demanded by other large players, including sovereigns. And consumer GPUs are becoming less and less competitive.

Obviously, this is an extreme case, but for large players, the rewards are so great that they won't back down even if the hardware bottleneck remains. As a result, decentralized operators like secondary data centers and retail-grade hardware owners, which make up the majority of Web3 DePIn providers, are left out of the competition.

The other side of the coin

Before the crypto founders knew it, the AI giants were keeping a close eye on the cryptocurrency. Government pressure and competition may force them to adopt cryptocurrencies in order to avoid being shut down or heavily regulated.

The recent resignation of the founder of Stability AI to begin "decentralizing" his company is one of the earliest public hints. He had previously made no secret of his public appearances as he planned to launch the token after the company's successful IPO, which in some way revealed the authenticity of the intended motives.

Similarly, while Sam Altman is not involved in the operation of the crypto project Worldcoin, which he co-founded, its tokens do trade like an agent of OpenAI. Only time will tell if there will be a path to connect free internet money projects with AI R&D projects, but the Worldcoin team also seems to be aware that the market is testing this hypothesis.

It makes sense to us that the AI giants might explore different decentralization paths. The problem we're seeing here is that Web3 hasn't come up with a meaningful solution yet. "Governance tokens" are largely a meme, and only those tokens that explicitly avoid a direct link between asset holders and the development and operation of their networks, such as $BTC and $ETH, are truly decentralized at the moment.

The (in)incentives that slow down the development of the technology can likewise affect the development of different designs that govern cryptographic networks. Startup teams simply put a "governance token" label on their product, hoping to find a solution, and end up just getting caught up in the allocation of resources around the "governance theater".

conclusion

The AI race is underway, and everyone is taking it very seriously. We can't find a hole in the thinking of big tech, where more computing means better AI, and better AI means reducing costs, adding new revenues, and expanding market share. For us, this means that the bubble is justified, but all the crooks will still be purged out of the game in the inevitable jitter.

Centralized, large-scale enterprise AI is dominating the space, and legitimate startups are finding it difficult to keep up. The Web3 space is a late entrant, but it's also joining the race. The market rewards crypto AI projects too richly, and Web2 startups appear to have fewer rewards in this space in comparison, which has led to a shift in founders' interest from delivering products to driving token appreciation at critical moments, and the window to catch up is rapidly closing. So far, there hasn't been any orthogonal innovation here that can bypass scaling up computing to compete.

There is now a credible open source movement around consumer-facing models, initially driven by a few centralized players who chose to compete for market share with larger, closed-source competitors (e.g., Meta, Stability AI). But now the community is catching up and putting pressure on the leading AI companies. These pressures will continue to affect the closed-source development of AI products, but will not have a substantial impact in the case of open source catch-up. This is another big opportunity for the Web3 space, but only if it solves the problem of decentralized model training and inference.

So, while on the surface there is an opportunity for the "classic" disruptors, the reality is far from it. AI is primarily associated with computing, and this cannot be changed unless breakthrough innovations occur in the next 3-5 years, which is critical to determining who controls and directs AI development.

Even as demand drives supply-side efforts, the computing market itself cannot "blossom," and competition among manufacturers is constrained by structural factors such as chip manufacturing and economies of scale.

We are optimistic about human ingenuity and convinced that there are enough smart and noble people to try to solve the AI problem area in favor of the free world rather than top-down corporate or government control. But the odds look slim, and at best it's a speculative game, with Web3 founders preoccupied with financial gain rather than real-world influence.