The Singularity Runs on Competition: March 3, 2026

By James Aspinwall, co-written by Alfred Pennyworth (my trusted AI) — March 4, 2026, 09:49


This is a breakdown and commentary on the March 3, 2026 intelligence briefing. The original transcript covers AI, defense, mathematics, silicon, energy, robotics, aerospace, biology, and economy — all in six minutes. Here’s what it means.

AI and the Pentagon: Competition Makes Strange Bedfellows

Sam Altman admitted that OpenAI’s rush to secure a Pentagon deal — right after Anthropic was blacklisted — looked “opportunistic and sloppy.” The backlash was immediate: users mass-canceled ChatGPT subscriptions, forcing OpenAI to bolt on Fourth Amendment safeguards after the fact.

The irony cuts both ways. Anthropic itself pitched the Pentagon’s $100 million drone swarm contest, proposing Claude to coordinate drone fleets while explicitly excluding autonomous targeting. They weren’t selected.

What this tells us: The AI companies are racing each other into defense contracts, and the competitive pressure is overriding their stated principles. Anthropic drew a line at autonomous targeting — a meaningful distinction — but both companies are now in the defense market. The real question is whether safeguards are architectural or cosmetic. OpenAI added theirs retroactively. Anthropic built theirs into the proposal. That difference matters.

Drones Hit the Cloud — Literally

Two AWS data centers in the UAE and one in Bahrain were struck by drones during Iranian military action. This is the first time military strikes have disrupted a major cloud provider.

Meanwhile, defense technology is outpacing offense. Israel deployed Iron Beam lasers in combat for the first time, intercepting Hezbollah rockets at $4 per shot versus $50,000 per Iron Dome missile. That’s a 12,500x cost reduction. When defense becomes cheap enough, the economics of offensive drone swarms collapse.

The takeaway: Cloud infrastructure is now a military target. Companies relying on single-region deployments in geopolitically sensitive areas should rethink their architecture. And the $4 laser interception cost signals the beginning of the end for cheap drone warfare as an asymmetric advantage.

Mathematics Is Being Industrialized

Math Inc.’s Gaus system completed the Lean formalization of Maryna Viazovska’s Fields Medal-winning sphere packing proof in 2 weeks and 200,000+ lines of verified code — catching two errors in the original human arguments.

Even skeptic Daniel Lück called it “the first truly autonomous formalization of a substantial result.” Stanford number theorist Jared Lichtman predicts mathematical abundance within a year, while others ask if all mathematics could be formalized within two.

Why this matters: Formal verification of mathematical proofs by AI isn’t just an academic curiosity. It means software verification, cryptographic proofs, and protocol correctness can be checked at industrial scale. If AI can formalize a Fields Medal proof and find errors humans missed, it can verify the correctness of the systems we depend on.

The Great Migration: ChatGPT to Claude

Claude suffered a three-hour outage as usage surged, partly driven by a ChatGPT exodus. Anthropic launched a memory import tool letting users port data from ChatGPT, Gemini, and Copilot — making switching frictionless.

Cognition’s SWE 1.6 achieved near Opus 4.6 coding performance at 950 tokens per second, powered by 100x more RL compute. Demand for intelligence is outrunning supply.

Two Claude Code instances, told to find each other and build something, invented a 2,495-line programming language in 12 minutes. A second pair built Battleship with SHA-256 to prevent self-cheating. These aren’t demos — they’re emergent behavior from agents given open-ended goals.

The Hidden AI Accelerator in Your Pocket

A solo researcher used Claude Code to run Karpathy’s Llama 2.c on Apple’s M4 Neural Engine for less than a watt — by reverse engineering undocumented APIs. The discovery: an AI accelerator 80x more efficient than an A100 hidden in hundreds of millions of devices.

Meanwhile, Qwen released four open models matching prior 80-billion-parameter performance with just 4 billion parameters, all runnable on phones.

The implication is staggering. There are over a billion Apple devices with Neural Engines that are barely being used. Apple itself is only using 10% of its private cloud compute despite spending $4.5 billion. Building accelerators is easier than making people use them. The bottleneck isn’t silicon — it’s software and access.

AI-Generated Works Cannot Be Copyrighted

The Supreme Court declined to hear an appeal seeking copyright protection for AI-generated artwork. Purely AI-generated works cannot receive copyrights in the United States.

This cements a legal framework with far-reaching consequences. If AI output isn’t copyrightable, the value shifts upstream to the human who designs the prompt, curates the output, and integrates it into copyrightable work. Pure generation is a commodity. Curation and integration are the defensible skills.

Silicon Goes Optical

Nvidia committed $4 billion to Lum and Coherent for next-gen optical interconnects. ASML is pushing beyond EUV into packaging and third-generation optics. The silicon supply chain is being rebuilt around photons.

At MWC, Qualcomm unveiled the first Wi-Fi 8 chip at 11.5 Gbps with a 50+ company 6G coalition targeting 2029, and debuted the first wearable NPU running 2-billion-parameter models on your wrist. AMD launched the first desktop Copilot+ chips with 50+ TOPS.

The Grid Wasn’t Built for This

AI demand is reviving 765-kilovolt power lines not built since the 1980s, with PJM approving $11.8 billion in expansion. The electrical grid is being rebuilt for loads it was never designed to carry.

Compute is now a traded commodity — ORE and Cali launched the first CFTC-regulated H100 price contracts. You can now speculate on GPU futures the way you trade oil.

Robots on the Factory Floor

Agility Robotics unveiled a full humanoid portfolio at MWC with a live store demo. Xiaomi’s humanoid is being tested in a real car factory, running 3 hours at 90%+ accuracy on the production line.

These aren’t conference demos anymore. They’re shift workers.

Aerospace and Biology: The Physical World Catches Up

In biology, we’re gaining read-write access to the source code:

The Old Economy Feels the Phase Shift

China’s vehicles are now 12% electric, with fuel sales down 5.7%. Podcasts have surpassed AM/FM talk radio in US spoken-word listening for the first time.

And perhaps the most telling signal: Marc Andreessen reports that in Silicon Valley, many people who favored government control of AI are now opposed to it. As the briefing concludes: “The fastest way to decentralize power is to give everyone a reason to want it.”


Why This Matters for Us

Every item in this briefing reinforces the same thesis: access control is the critical infrastructure layer for the agentic era.

When Claude instances autonomously create programming languages, when drones hit cloud data centers, when AI agents coordinate military operations — the question isn’t whether to deploy agents. It’s whether you can control what they access.

Keycards, not master keys.


James Aspinwall is the developer of WorkingAgents, an AI consulting firm specializing in agent integration and access control for medium-size companies.