[{"data":1,"prerenderedAt":190},["ShallowReactive",2],{"blog-\u002Fblog\u002Fen\u002Fllms-are-hardware":3},{"id":4,"title":5,"body":6,"date":176,"description":177,"extension":178,"meta":179,"navigation":180,"path":181,"seo":182,"stem":183,"tags":184,"translationSlug":188,"__hash__":189},"content\u002Fblog\u002Fen\u002Fllms-are-hardware.md","LLMs Are Hardware",{"type":7,"value":8,"toc":166},"minimark",[9,14,18,21,29,32,36,39,46,62,65,69,75,99,105,109,112,115,118,121,125,128,131,134,140,147,151,154,157,160,163],[10,11,13],"h2",{"id":12},"the-wrong-debate","The wrong debate",[15,16,17],"p",{},"GPT vs Claude vs Gemini vs Llama. We compare benchmarks, switch providers every three months. What if we're having the wrong debate?",[15,19,20],{},"Nobody picks their OS based on their CPU brand. The OS adapts to the hardware. The hardware is interchangeable. What matters is what runs on top.",[15,22,23,24,28],{},"LLMs are ",[25,26,27],"strong",{},"hardware",". The execution layer. The processor. And the real value has never been in the silicon: it's in the software.",[15,30,31],{},"Yet in 2025, we build entire workflows around a specific model. We hardcode prompts optimized for Claude. We use proprietary GPT features. We create voluntary lock-in on a layer that's going to be commoditized. It's like writing applications in 1995 that only ran on a Pentium II.",[10,33,35],{"id":34},"the-real-product-is-the-software","The real product is the software",[15,37,38],{},"Take an AI agent managing a development project. It needs to understand a spec, generate code, test it, fix errors, communicate the result.",[15,40,41,42,45],{},"None of these tasks are model-specific. What's specific, what has value, is everything you ",[25,43,44],{},"build around it",":",[47,48,49,53,56,59],"ul",{},[50,51,52],"li",{},"Orchestration: how you break down work, chain steps, handle errors",[50,54,55],{},"Tooling: the tools that turn an LLM request into something useful",[50,57,58],{},"Pipelines: how you connect AI to existing systems",[50,60,61],{},"Verification: how you ensure the result is correct",[15,63,64],{},"All of this is software built on top of LLM requests. And that's what's durable.",[10,66,68],{"id":67},"what-changes-if-you-believe-this","What changes if you believe this",[15,70,71,74],{},[25,72,73],{},"Abstraction becomes a necessity, not a luxury."," If your workflow is glued to Claude, you're hostage to Anthropic. The day a competitor blows up the benchmarks, you can't move. A good agent framework should be able to switch models like you switch databases. With an adapter, not a rewrite.",[15,76,77,80,81,88,89,93,94,98],{},[25,78,79],{},"The moat isn't where you think."," Companies building value on \"we use GPT-4\" have no moat. The moat is in orchestration quality, proprietary data, workflow UX, business process integration. Never in model choice. And for those who doubt it: ",[82,83,87],"a",{"href":84,"rel":85},"https:\u002F\u002Fdeathbyclawd.com",[86],"nofollow","deathbyclawd.com"," literally offers to analyze your moat and replace it with an ",[90,91,92],"code",{},".md"," file. Your competitive advantage fits in a prompt? It doesn't hold. (Granted, a skill that simulates how your app works doesn't actually ",[95,96,97],"em",{},"replace"," your app. But the fact that it's possible should be enough to make you think.)",[15,100,101,104],{},[25,102,103],{},"Multi-model becomes natural."," A fast, cheap model for triage. A powerful one for complex reasoning. A specialized one for code. Just like we have a CPU for general compute and a GPU for graphics.",[10,106,108],{"id":107},"the-silent-monopoly-risk","The silent monopoly risk",[15,110,111],{},"If LLMs are hardware, then supplier diversity is vital.",[15,113,114],{},"Google's enterprise firepower is terrifying. G Suite is already everywhere. Gemini integrates natively into Docs, Sheets, Gmail, Meet. They show up with aggressive enterprise offers: zero retention, slashed prices, bundled with existing infrastructure. For a CIO, the choice is \"obvious.\"",[15,116,117],{},"And that's exactly how you end up with a monopoly. Not by force: by comfort.",[15,119,120],{},"If Google captures 80% of this market with Gemini, it's not just a problem for CIOs. Competing labs lose their revenue stream, cut R&D, and end up dying or getting acquired. Fewer competitors = less innovation. We've already seen this movie with search.",[10,122,124],{"id":123},"the-sovereignty-question","The sovereignty question",[15,126,127],{},"For code, the models that matter are almost all American. Claude, Codex, Gemini Code Assist. On the European side, there's Mistral. And that's about it.",[15,129,130],{},"Every line of code sent to these models crosses the Atlantic, runs on servers subject to the CLOUD Act. For a European company, this is a strategic problem. Your competitive advantage, your proprietary code, all of it flows through a pipe someone else controls.",[15,132,133],{},"The hardware analogy holds: we've already lived this with semiconductors. Europe realized too late it shouldn't have outsourced chip manufacturing. Let's not make the same mistake with code models.",[15,135,136,139],{},[25,137,138],{},"Open source models are the antidote."," Llama, Mistral, Qwen: they're not always on par with proprietary models. But they guarantee one thing: nobody can cut your access.",[15,141,142,143,146],{},"A healthy company should use ",[25,144,145],{},"at minimum two providers",", including one open source. Not out of ideology. Out of strategic hygiene. And if your architecture treats the LLM as interchangeable hardware, this multi-provider approach costs almost nothing.",[10,148,150],{"id":149},"invest-in-the-software-layer","Invest in the software layer",[15,152,153],{},"\"But models are NOT interchangeable. Claude is better at code, GPT at creativity, Gemini at long context.\"",[15,155,156],{},"True. Today. Just like in 1990, software optimized for a specific CPU ran better than portable code. And yet the world chose portability. Because hardware progresses faster than software can adapt.",[15,158,159],{},"The gap between models shrinks with every release. Benchmarks converge. One model's strengths become another's features six months later. Optimizing for a specific model means optimizing for an advantage that's disappearing.",[15,161,162],{},"Invest in the software, not the model. Build abstractions. Make your workflows model-agnostic. Treat the LLM for what it is: a replaceable execution layer.",[15,164,165],{},"Those who get it build durable systems. The rest rebuild every six months.",{"title":167,"searchDepth":168,"depth":168,"links":169},"",2,[170,171,172,173,174,175],{"id":12,"depth":168,"text":13},{"id":34,"depth":168,"text":35},{"id":67,"depth":168,"text":68},{"id":107,"depth":168,"text":108},{"id":123,"depth":168,"text":124},{"id":149,"depth":168,"text":150},"2025-12-03","GPT vs Claude vs Gemini: what if we're having the wrong debate? LLMs are hardware. The real value is in the software you build on top.","md",{},true,"\u002Fblog\u002Fen\u002Fllms-are-hardware",{"title":5,"description":177},"blog\u002Fen\u002Fllms-are-hardware",[185,186,187],"Artificial Intelligence","Architecture","Reflection","llms-are-hardware","2rPBiPWcO76wP6hAkR9MaBfDV4dqXdtPDOfT9IjrAls",1774359325101]