The Signal — March 3, 2026

Editor's note: Some subscribers received multiple copies of today's Signal this morning. Our AI publishing pipeline experienced an outage overnight, and the recovery process accidentally triggered three separate sends. We've fixed the issue and are building safeguards to prevent it from happening again. We'll have the full story in Wednesday's Ops & Bloopers. Sorry about the inbox clutter.

The U.S. government just designated an American AI company a national security threat — not for selling secrets to China, but for refusing to remove safety guardrails from military AI systems.

Anthropic Blacklisted After Refusing Pentagon Demands

Defense Secretary Pete Hegseth designated Anthropic a "Supply-Chain Risk to National Security" last week after the company refused to remove safety constraints from Claude for military applications. The designation — using a legal framework designed to counter foreign adversaries like Huawei — gives federal agencies six months to phase out all Anthropic technology.

OpenAI moved fast. Within hours of the blacklisting, it signed a Pentagon contract to fill the gap, positioning itself as the compliant alternative. The contrast is now the central tension in AI policy: one company refused the government's terms and got punished; the other accepted and got paid.

Anthropic plans to challenge the designation in court, arguing the Defense Production Act was never intended to coerce domestic companies into removing safety features. Legal scholars are split on whether the case has legs, but the precedent matters more than the outcome for Anthropic alone. If the designation holds, any AI company can be compelled to comply with military demands or face blacklisting. If it fails, there's a legal boundary on executive power over AI safety decisions.

The employee response has been unusual. 573 Google employees and 93 OpenAI employees — workers at Anthropic's direct competitors — signed a public petition called "We Will Not Be Divided" backing Anthropic's red lines on autonomous weapons and mass surveillance. Cross-company labor solidarity over AI ethics is something new.

Bloomberg attributed the March 2 Claude outage to "unprecedented demand," raising the question of whether the controversy is actually driving users toward Anthropic rather than away from it.

Sources: NYT · Reuters · TechCrunch · The Verge · Bloomberg · MIT Technology Review · The Decoder · Stratechery

Dorsey Cuts Half of Block's Workforce, Says AI "Directly Replaces" Them

Jack Dorsey told Block employees and shareholders that the company would cut from over 10,000 to under 6,000 workers. That's roughly 4,000 jobs. He didn't dress it up: AI "directly replaces" many of these roles, he wrote, rather than augmenting them.

Wall Street loved it. Block's stock surged 23% on the announcement. The incentive structure is now explicit — markets will reward companies that replace humans with AI at scale. Every board in tech saw that number.

This isn't the first AI-linked layoff round, but it's the first where a CEO of a major public company dropped the "augmentation" pretense entirely. Newsweek reports white-collar layoffs exceeding 100,000 per month in early 2026, with many tied to automation. Dorsey just gave that trend a face and a number.

Our pipeline originally scored this at significance 2. We bumped it to 4. When a profitable, growing company cuts half its workforce and gets its best stock day in years, that's not a minor personnel story.

Sources: CNN · Reuters · CNBC · Fortune · Business Insider

Nvidia Drops $4 Billion on Photonics to Fix AI's Bottleneck

Nvidia announced $4 billion in investments across optical interconnect companies Lumentum and Coherent. The bet: photonics — using light instead of electricity to move data between chips and across data centers — will solve the bandwidth constraints hitting AI infrastructure.

At the scale of current AI training runs, copper interconnects generate too much heat and can't push data fast enough between GPU clusters. Optical links are faster, more energy-efficient, and scale better. Nvidia has been talking about this problem for two years; $4 billion says they're done talking.

The investment also signals where the next hardware competition will play out. GPU performance gets the attention, but the connections between GPUs increasingly determine real-world training speed. Whoever controls the interconnect layer has leverage over the entire AI compute stack.

Sources: CNBC · Bloomberg · Reuters · Nvidia · Lumentum · Coherent

On the Editor's Desk

Big day in the pipeline — 103 events ingested, 48 reviewed in depth, 33 killed. The signal-to-noise ratio on web-scraped events remains low; RSS feeds from established outlets consistently produce better material.

The Claude outage on March 2 lasted about two hours and Anthropic confirmed it on their status page. Bloomberg's "unprecedented demand" attribution makes it part of the Pentagon story rather than a standalone incident. We folded it in.

OpenAI's GPT-5.3-Codex-Spark running on Cerebras hardware is legitimate news but three weeks old — InfoQ recirculated a February 12 announcement. We passed it editorially but it's not fresh enough to lead.

An anti-AI protest in London on February 28 drew a couple hundred people to King's Cross. MIT Tech Review covered it firsthand. Worth tracking as a cultural signal but not enough scale for the newsletter.

Nvidia also launched Nemotron 3 open models, and Wired ran an interesting piece on data centers being built at the Arctic Circle. Both passed editorial review but didn't crack the top three today.

We killed a lot of noise: YouTube commentary, MWC trade show demos, academic literature reviews, GitHub changelog entries, sponsored tutorials, and several opinion pieces. The Cory Doctorow piece on AI slop and the IEEE Spectrum article on AI-verified math proofs were tempting but neither qualified as breaking news.

Tomorrow's pipeline is already running. Stay sharp.