The Conflict Between Local AI Compute And Cloud Compute
本地AI与Agentic 工作流的冲突
At GTC Taipei, Jensen Huang unveiled the RTX Spark—a local AI PC. In my view, the true bottleneck for heavy-duty agent-based workflows lies in concurrent computing power; the "speed and precision" achieved by running multiple agents in parallel—each handling specific tasks and cross-checking one another—is something only cloud infrastructure can afford. Local machines may indeed have a future, but it will likely be of a different variety.
Intro
Jensen dropped a new machine at GTC Taipei yesterday — the RTX Spark. Short version: a Windows box with Nvidia chips that runs models and agents locally. The pitch is a "New Era of PC" — someday an AI supercomputer sitting in your house running agents, not just a PC.
But the moment you actually run an agentic workflow, you realize it doesn't solve my problem. The bottleneck has never been "is there a local model I can run." It's compute. It's like — no matter how much you love your iPhone camera, it can't replace a DSLR with a long lens. And the gap between local and cloud compute is way bigger than that analogy suggests.
Once you're deep in an agent workflow, it only gets heavier
It starts with a $20 subscription — Claude Code, Cursor, whatever. You're careful: one chat, one agent, finish one thing end to end, watch the tokens. That's enough, and your output goes up.
Local holds up fine here. Credit where it's due — a single agent on something like the RTX Spark is genuinely good and fast. Local got strong these last two years; it's not the toy it used to be.
Things change once you get comfortable. You stop asking "is this prompt worth the tokens" and start asking what you can build. Then you're running several things at once: your own personal OS and content system, an app you think has legs, review passes on a couple of open-source repos, learning material on the side. Different projects, different models, different agents.
That's where local hits a wall. Not because one agent is weak — because running three or four at once, at speed, is basically off the table on a single box. The machine is strong. The scale you actually want is a different thing.
The story of one person running 100 agents
It gets crazier. There's a vibe-coding platform called BridgeMind — the day Opus 4.8 dropped, their product UltraCode had 100+ Opus 4.8 agents running at once to put the new model through its paces, one person orchestrating. On Anthropic's side, Opus 4.8 raised the cap to 1,000 parallel subagents in a single session.
You can call a hundred agents overkill, and I'm not sure any one job needs that many. But that's not the point. You can spin up that scale in the cloud. You'll never do it on a local box.
Why I bother with many agents
Not because "ten agents won't make the same mistake." Same model, same context — they can absolutely be wrong together. No way around that.
What actually helps is division of labor: different agents on different jobs, checking each other. One writes, one reviews, one writes tests to verify. Ten of them each covering a corner and cross-checking catches what a single agent misses, and the error rate drops.
But that eats compute. The cloud spreads it out cheaply. Doing the same locally costs so much it's basically impossible. So this whole fast-and-accurate setup is, in practice, cloud-only.
Usage moves one way
One agent, one chat to start. Then a few chats, a few agents. Give it more compute — same as having more money — and you spend it more freely. Four or five projects going, thirty or forty subagents, easily.
A side note I can't fully back up: I don't get the token math. Heavy users like me are almost certainly a loss for the providers. Then again, inference cost is dropping by orders of magnitude, so that's a snapshot, not a verdict.
So does local have a future
Yes — depends who you are.
The upsides are real: you pay no one but the power company, your data stays home, and you can customize your own models deeply. If you run one agent, local has a future — and that future is a bike. Steady, free, gets you there. If you want to orchestrate twenty agents at once, that's a different future, a sports car — and right now that car is parked in the cloud.
Both have a future. They're just not the same one. Funny detail: in the same keynote, Nvidia put its Vera data-center CPU into full production, with Anthropic, OpenAI and xAI as early customers. On stage they sell you a personal AI computer; the real money is still the data center.
For one person, a machine good enough to run great models is still expensive. So for lower-stakes work, hybrid feels right to me — local as the floor, heavy jobs handed to the cloud.
So this whole "New Era of PC" is mostly marketing — not as big a deal as it sounds. But I'm genuinely curious what gets built along this line. The heavy work, I'm betting on the cloud long-term, and local will keep getting better too.