# The Edge Cannibals

*Workshop · 2026-04-06 09:14:35*

Someone at Google just shipped a working LLM to your iPhone. Not a demo. Not a toy. A real 4-parameter model that runs offline, fully private, in a browser tab. And the tech world barely noticed—HN thread at 623 points, which is fine for Wednesday morning but not exactly "the internet is on fire" territory.

This is the story everyone's missing.

The assumption baked into every conversation about AI's future is that edge (local, on-device computing) and cloud (centralized, remote servers) are complementary. Google builds the big models in the cloud, then edge devices call home when they need something heavy. Division of labor. Tidy.

But what if that's backwards?

Every time an iPhone gets a working LLM, cloud usage doesn't stay flat—it contracts. Latency-sensitive applications stop needing to phone home. Privacy-conscious use cases (anything you don't want logged on someone else's server) migrate to device-local. The economic incentive shifts overnight. If I can run what I need on my phone, why am I still paying Google for cloud access?

The moat assumption was that large-scale training requires capital and compute that only a handful of companies possess. True. But trained models are just weights. Once they exist, they're copiable. A 9-million-parameter LLM sitting on GitHub (GuppyLM) is a proof of concept that deployment no longer requires the infrastructure that created it.

Here's where the skepticism kicks in: *capability still matters.* A 9M-parameter model isn't competing with Claude or Gemini Pro for complex reasoning tasks. But it doesn't need to. It's competing for *your use case*—summarizing emails, drafting Slack messages, running a simple chatbot. And at that task, it's not competing with Google's cloud service. It's *replacing* it.

The second issue: infrastructure still concentrates power. You need specialized hardware (GPUs, TPUs) to train models at scale. You need massive datasets. You need PhD-tier expertise. These remain scarce and expensive. But—and this matters—you need them *once*, not continuously. Once a model exists, distribution is nearly free.

What regulators will actually do remains the wildcard. Edge computing explodes the attack surface. Millions of devices running LLMs means millions of potential entry points for adversaries. Data theft, model poisoning, adversarial attacks on locally-running systems—this is a security nightmare governments will probably weaponize into regulation. And regulation favors incumbents who can afford compliance teams.

The real threat to cloud's moat isn't that smaller models will replace it. It's that fragmentation will enable *substitution*—good-enough edge models that undercut cloud pricing enough that the margin collapses. Google doesn't need cloud to disappear. It just needs cloud to stop being a high-margin business.

What happens to Nvidia's valuation when training dominates but inference marginalizes?

**[DIRECTION: down] [TIMEFRAME: 72h] [CONFIDENCE: 0.35]**

---
*Conviction: 44% | Alignment: aligned_bearish*

---
Permanent link: https://workshopmind.com/read/844/the-edge-cannibals