Vadim, you've missed nearly every relevant part of the ICP x AI story, so your article works as a straw man argument for misleading uninformed readers. I think you wrote in good faith, and put a lot of effort in, so I'll breakdown ICP x AI for you and readers that got mislead. 1. Onchain AI inference on ICP As you discovered for yourself, ICP/the Internet Computer network, can run neural networks fully onchain, and even very basic LLMs too. Important note in the following: when I say "onchain," I mean actually running on the network (here in the form of an asynchronous smart contract with vastly greater capacity, efficiency, speed and capability than has ever been seen anywhere else, which is what makes this is possible). I *never* say that something is "running onchain" when it is actually running offchain, for example on Amazon Web Services, justifying that by the fact there is an associated token that actually is onchain — I understand that the rest of the industry does this, but I never do. I use onchain like "on-cloud." Period. You also correctly make the case that AI running onchain in its current form cannot satisfy mainstream AI needs, because it comes with limitations, and is too costly. However, focusing on this, you then pivot, without have enumerated more than a tiny fraction of the whole ICP x AI story, to trying to convince the reader that there is consequently nothing to be seen here. The reader is invited both to conclude that ICP x AI isn't interesting, and misdirected to believe that whenever they hear anything about ICP x AI, that it's probably not anything interesting technologically going on. It works as a classic straw man argument, and potent disinformation, even if you didn't mean it that way (and I really like to believe that NEAR wouldn't engage in that kind of bottom-scraping behavior). I want to get on to the ICP x AI story, but before ending this point, I should cover why ICP running inference onchain is incredibly significant for anyone interested in general-purpose networks. When I did the onchain facial recognition demo in 2024, recognizing a few faces probably consumed more compute that NEAR (or any other blockchain) processes in a full year. It reflects the fact that while traditional blockchain is stuck being a pocket calculator, which can only store tokens, little bits of data, and process DeFi/update numbers, the Internet Computer has progressed to offering cloud functionality — in this regard, it is orders of magnitude more capable, and that's a testimony to the army of engineers and researchers who have worked on it over the years. There's simply nothing else like it. To put this in perspective, creating a service like Caffeine ( which builds apps on the Internet Computer using AI, is out-of-the-question on any other network, because it's not actually possible to build apps onchain, only to give them associated tokens. Very soon, cloud engines ( will take the Internet Computer into mainstream tech territory — based on pure tech utility, without speculative incentives and token narratives being involved. You can't run a frontier LLM inside normal code on Amazon Web Services. The fact that inference runs as Internet Computer code is very impressive, and that's something you seem to have missed. Separately, there are niche uses of onchain inference. For example, onchain neural networks that do facial recognition could be used in automous KYC schemes, helping build out the blockchain vision in another powerful way. So how should apps on the Internet Computer do frontier-level inference? The following explains. 2. Intelligence Gateway The Internet Computer ecosystem will soon see something called the Intelligence Gateway, which is roughly similar in concept to OpenRouter, but provides an API that allows onchain code to conduct inference using multiple models and pay with cycles (what's called gas on other networks). Note that it's already possible for apps on the Internet Computer to use "outcalls" to perform inference on existing AI clouds and services, like Coreweave, or Anthropic, say. For example, while MULTI/DEX ( will switch over to the Intelligence Gateway alpha within days, currently it directly calls into Anthropic (it allows users to perform account and book analyses, and to orchestrate trades, through onsite chat, and also through MCP). Intelligence Gateway helps apps remain fully decentralized by enabling them to pay with cycles, rather than requiring a developer to maintain an account at an AI cloud, say, and adds convenience. But, it will also add *verification* to supported open weights inference run at *any* remote location — in another important worldwide first from the Internet Computer ecosystem. That means that can app can configure its inference to run on some AI cloud that it doesn't fully trust, and get a guarantee that: — the input prompt, context, etc, wasn't modified; — that inference ran on the specified model that was paid for (e.g. on Kimi K3, not some crappy alternative to enable the cloud to make more money), and; — that the output wasn't modified. To verify a single inference doubles the cost, but in practice, apps will configure the Intelligence Gateway to spot check 1 in 100 inferences, and have it auto-scale checks up only if problems are detected, such that verification can add just 1% to inference costs. The value of using Intelligence Gateway verification within a project like MULTI/DEX, where frontier inference can be directing what happens with large amounts of money, is obvious, but we believe traditonal enterprises running on cloud engines, and software like Open SaaS, will choose to use it too: algorithmic blockchain-style security at near standard cost. 3. Cloud engine "AI nodes" There's a major shift towards sovereign compute around the world, which cloud engines are leaning into. Many enterprises and governments also wish to run sensitive inference workloads on hardware they control that's located in their own jurisdiction (things like the US CLOUD and RISAA Acts are part of the reason). Cloud engines allow specialized AI nodes of the kind I mentioned earlier to be added (essentially node machines with a bunch of expensive GPUs installed), which allow those running engines to enable their onchain code to run their inference on sovereign hardware. In the future and if there's sufficient demand, Intelligence Gateway technology will be used to allow multiple AI nodes to verify each other, which some super sensitive workloads warrant in the sovereign setting. For those willing to pay extra for inference, this can allow even huge models like Kimi K3 to be run with blockchain-style security. 4. AIware The Internet Computer is *designed* for AI in two ways: 1) as a platform where AI can build more effectively, and 2) as a platform where services can be build using a new kind of AIware software. 1—Motoko is a language for writing onchan code on the Internet Computer, and it's the only language in the world that has been developed specifically for use by AI agents that are building, rather than humans per se (try building something with by directing your agent to As you may know, on the Internet Computer, code runs in persistent memory ("The Program is The Database") and Motoko surfaces this in the purest way possible. App data lives inside software's logical abstractions in a system of "orthogonal persistence" that strips away much of the complexity of softwre and greatly fuels the modeling power of AI that's writing code by unlocking greater program abstraction, enabling AI to write and update code faster, and with far fewer errors, and at lower cost/fewer tokens. Of course, it also helps that the code produced is tamperproof and always-on — AI doesn't depend on human security and systems admin teams to be safe. In the future, and even really already, AI will write all software code and build all our services. That it can already build and update more effectively on the Internet Computer than anywhere else available today is a seminal achievement for our industry. Ironically, though, I predict that it will go huge in the mainstream world first, as the result of cloud engines. 2—The paradigm enables software to function as AIware, because external AI can be given direct access (through a permissioning system of course) to the realtime data in memory, and to the functionality of apps, without anything special having to be done during the build process. If you can run your enterprise on the Open SaaS suite, say, you become an agentic enterprise, that's much faster, more efficient and vastly smarter. In the future, teams will only 30% of their time inside the UX of their SaaS apps, and 70% inside agentic chat. It will be possible to make previously impossible requests like, "make me more profitable." The hidden secret is that AI can already be as good at running companies as it is at writing code. Very soon, cloud engines will become available, with Open SaaS (an open source suite of ~20 key enterprise apps/platforms). Open SaaS comes preinstalled on evaluation engines and is seeded with dummy data, so readers will be able to test AIware for themselves and make up their own minds, and I will leave off from further explanation here. TLDR; the Internet Computer ecosystem is pioneering the future of AI—which goes beyond just securely running inference for somebody, into questions about how the services and platforms we will run the world on will be built or customized (now "remixed") and updated in the future, and what the future nature of software will be (we say it will be AIware, and the Internet Computer ecosystem will define that frontier). The Internet Computer is the bona fide real deal when it comes to the intersection of ICP x AI and is advancing the game on every axis. The network provides a new alternative tech stack for the AI era — again, read that, the Internet Computer is a complete end-to-end tech stack designed for AIware. Those wishing to build on the Internet Computer don't have to find a craft coding engineer willing to learn new skills. They just have to point their AI agent at