Week 46, 2026•

Nvidia Without Jensen, 145B Models on Laptops

Jensen Huang exits Nvidia after 33 years while Edge0 proves you can run frontier-class AI on a single laptop with 29GB of RAM.

The man who bet Nvidia on CUDA in 2006 walked out the door this week. Meanwhile, two teams independently proved that frontier-class inference no longer needs his GPUs — or anyone's data center. The centralization thesis for AI compute is cracking from both ends.

The Big Story

Jensen Huang resigned as President, CEO, and board member of Nvidia on November 12, effective immediately. No reason was given. The press release thanked him for "immeasurable contributions" over 33 years — co-founder to the architect of the GPU computing revolution — and named CFO Colette Kress as interim CEO while the board searches for a permanent successor. Harvey C. Jones, former Synopsys CEO, was appointed Nvidia's first-ever independent Chair of the Board.

The timeline matters. Nvidia's Q3 fiscal 2027 earnings report drops November 17 — five days after the announcement. The board published a separate statement from its independent members two days before the resignation, which strongly suggests this was negotiated, not impulsive. Reuters has reported on a relationship with a second Nvidia employee, though the company has not connected the two events.

For the AI infrastructure community, the operational question is straightforward: does Nvidia's product roadmap continue? The board says yes, explicitly stating that customer commitments and the roadmap proceed "without interruption." Kress, who has been CFO since 2013, knows the business inside out. The near-term execution risk is low.

The strategic risk is harder to size. Jensen's singular contribution wasn't managing Nvidia — it was placing bets that the semiconductor industry considered insane at the time. CUDA in 2006 when GPUs were for games. Tensor Cores when nobody knew what to do with them. The DGX platform priced so aggressively it locked out competitors before they shipped alternatives. These were founder-conviction decisions, not MBA-consensus decisions. The board is searching for a permanent CEO who will consider both internal and external candidates. The market will tell you whether it thinks that search will produce another founder-type or a caretaker through how NVDA trades in the next two weeks.

The bigger structural story: this exit happens as the inference market is fragmenting in real time. AMD's MI series, Google's TPUs, Amazon's Trainium, Apple silicon for local inference, and — as this week's Edge0-145B demonstrates — entirely new architectures designed from the ground up for single-machine deployment. Nvidia's training monopoly is intact. Its inference moat is eroding. A committee-run Nvidia can defend the first. It takes a founder to build the second.

The Hacker News thread hit 3,896 points and 3,339 comments — the single largest AI-adjacent story on the platform this year. The industry clearly understands the stakes.


This Week in 60 Seconds


Deep Dive: How Edge0 Fits 145B Parameters on a Laptop

Edge0-145B is the most architecturally interesting model release in months — not because of its benchmarks (competitive but not dominant), but because of how it runs. This model was trained natively as an on-device architecture from day one, not distilled or quantized from a cloud-scale checkpoint after the fact. That design decision changes everything downstream.

The core trick: shared expert matrices. Standard Mixture-of-Experts gives you N independent expert networks with a router selecting K of them per token. Edge0 does something different. Each group of 64 experts shares two large weight matrices, and each expert retains only a small private parameter set — essentially, N lightweight deltas on top of shared compute. Total parameter count: 145 billion. Active parameters per inference step: approximately 10 billion. The ratio is what makes local inference possible.

Three-tier memory hierarchy. This is where it gets genuinely interesting for systems engineers. Edge0 treats the machine as a unified memory hierarchy rather than treating VRAM as the only tier that matters:

The scheduler runs GPU compute on cached experts while the CPU prepares incoming experts in parallel — a producer-consumer pipeline where the bottleneck moves between tiers depending on the expert access pattern. Speculative decoding adds another layer: the model verifies multiple candidate tokens simultaneously, amortizing the cost of each expert load.

Measured performance on compatible hardware (the blog lists RTX 4090, RTX 5090, Apple M-series chips, and the new RTX Spark): 1,100+ tokens/sec prefill, 80+ tokens/sec decode, 29GB peak memory, ~120W typical power draw. The target machine is a 48GB laptop. Context window holds at 128K–256K tokens with reportedly stable throughput — though that claim needs independent testing before you trust it for production agent loops.

The benchmark picture. Edge0 reports a six-benchmark mean of 72.8 against Qwen3.8-Flash-Next's 71.7 and DeepSeek-V4-Flash-0731's 70.1. It trades blows rather than dominating: it loses on HLE (35.4 vs. Claude Opus 4.6's 40.0) and GPQA-Diamond (90.4 vs. Qwen's 91.7), wins on DeepSWE (64.6 vs. 58.7) and Terminal-Bench 2.1 (86.52 vs. 83.3). All of these are Edge0's internal evaluations — the technical report and independent reproductions are still pending.

But the headline comparison isn't quality. It's the resource envelope. Here's the estimated memory for running each model at 4-bit precision:

That's the difference between "buy a workstation" and "use the MacBook you have." If Edge0's quality claims hold up under independent testing, the local-inference versus cloud-API tradeoff just shifted for a large class of coding-agent workloads. The weights are on Hugging Face (NVFP4 variant), but full license terms haven't been published yet. Run the benchmarks yourself before making architecture decisions.


Open Source Radar

Mika — An on-device AI agent that operates Android apps by reading the screen and simulating touch, scroll, and type — the same way a human would, with no per-app APIs. It runs on a hidden background screen so your phone stays usable, supports any LLM provider (OpenAI, Anthropic, Gemini, local models via Ollama/LM Studio), and stores everything encrypted on-device with zero server component. Tasks can be scheduled, interrupted mid-flight, and refined conversationally. 8.5K stars, #1 trending Kotlin repo this week.

Cloudflare security-audit-skill — Turns any coding agent into a structured security auditor. Six phases: reconnaissance maps trust boundaries, coverage-led hunting assigns isolated sub-agents to each surface, then — crucially — a separate verifier agent tries to disprove each candidate finding. Ship with attack-class playbooks covering memory safety, AI/LLM-specific targets, web auth, supply chain, cloud deployment, and more. The adversarial verification pattern (finder ≠ verifier) is the right architecture for agent-driven security work, and 10.7K new stars in a single week suggests the market agrees. 71.7K stars total.

smolmaxxing — The opposite end of the spectrum from Edge0: squeeze the biggest model possible onto the smallest hardware you own. Custom 1-bit lookup-table kernels, MoE expert paging from flash storage, and a headless mode that stops background services to reclaim RAM. Published results across three test devices: Qwen3.5-2B at 7.7 tok/s on a Xiaomi TV Box with 1GB usable RAM; Qwen3-Next-80B at 2.6–3.1 tok/s on a OnePlus Nord CE 2 with 3GB free. A 2013 iPad Air (jailbroken, 1GB RAM) runs Bonsai-1.7B at 24 tok/s. All patches against llama.cpp, 165 experiments documented in a public lab notebook.


The Numbers

  • 33 years: Jensen Huang's tenure at Nvidia, from co-founding in a Denny's booth in 1993 to resignation this week — one of the longest-serving CEO departures in the semiconductor industry's history
  • 10.7K stars in one week: Cloudflare's security-audit-skill is now the fastest-growing agent skill in the ecosystem, suggesting structured security is the next must-have capability layer
  • 7.7 tok/s on 1GB RAM: smolmaxxing running Qwen3.5-2B on a $50 TV streaming box — the floor for useful local inference keeps dropping, and it's approaching appliance territory

Aaron's Take

We're watching the AI compute stack decentralize in real time, and this week crystallized it from three directions simultaneously. Nvidia's founder-CEO exits while the inference market fragments beneath him. A 145B-parameter model runs on a single laptop at 80 tokens per second with an architecture that treats your machine's memory hierarchy as a first-class design constraint. And someone got an 80B model running on a mid-range phone by patching llama.cpp.

The training side still needs data centers and billion-dollar GPU clusters — that hasn't changed and won't change soon. But the inference side, where all the user-facing value actually lives, is racing toward the edge faster than most infrastructure teams have planned for. The Claude outage this week underscored the point: when your agent's brain lives in one data center, a bad deploy takes out every tool in your chain. Local inference isn't just a cost optimization. It's a reliability architecture. If your system assumes cloud-only inference in 2027, this is the week to revisit that assumption.


Sources


— Aaron

You Might Also Like

Braintrust vs Langfuse vs Promptfoo vs DeepEval: LLM Evals

Braintrust, Langfuse, Promptfoo, and DeepEval compared on eval workflow, CI gates, red teaming, self-hosting, and pricing. Pick the right LLM eval stack.

AI Engineering

WebMCP Security: Can a Website Hijack Your AI Agent?

WebMCP lets websites hand tools to AI agents. A 9-of-9 attack result shows why tool descriptions are untrusted input, plus the guards that stop it.

AI Engineering

LlamaIndex vs LangChain: Agent and RAG Orchestration

LlamaIndex vs LangChain for RAG and agents: indexing depth, create_agent vs AgentWorkflow, checkpoints, human approval and when to combine both.

AI Engineering