The Meltdown — March 1-7
The week AI companies learned that saying no to the government is great marketing, until it isn't.
01 — Anthropic said no to the Pentagon, then said maybe
Anthropic refused to strip Claude's guardrails on autonomous weapons and mass surveillance. The White House responded by ordering federal agencies to drop Anthropic products. Defense Secretary Hegseth tagged the company a "supply chain risk to national security." Their $200M Pentagon contract evaporated overnight.
Then something wild happened. Claude shot from 42nd to #1 on both app stores by March 3. Free signups surged 60%+. For about 48 hours, principled refusal looked like the best growth hack in tech history.
By March 4, Anthropic quietly reopened talks with the Pentagon. The Wall Street Journal reported that Claude was already being used in the Iran conflict. I don't know what the final arrangement looks like, but the pattern is familiar. Take a stand, absorb the brand boost, then negotiate from a stronger position. Maybe that's cynical. Maybe it's just how you survive when your biggest potential customer can also designate you a national security threat.
02 — OpenAI raised $110B and immediately went shopping for NATO
OpenAI closed the largest private fundraise in history. $730B pre-money, $840B post-money. Amazon put in $50B, NVIDIA $30B, SoftBank $30B. They're projecting $280B in 2030 revenue and targeting $600B in total compute spend by then. Those numbers are so large they've stopped feeling like numbers.
Days later, OpenAI began pursuing a NATO contract for unclassified networks. Sam Altman called the Pentagon deal "definitely rushed" with optics that "don't look good." So the pitch to NATO comes with red lines: no mass surveillance, no autonomous weapons, no automated high-consequence decisions. Sound familiar? Those are nearly identical to the guardrails Anthropic got blacklisted for keeping. OpenAI just drew them after signing the check.
03 — GPT-5.3 killed the moralizing preamble
GPT-5.3 Instant dropped March 3 with 26.8% fewer hallucinations on web-search queries and the removal of those painful "As an AI, I should note that..." preambles. People are calling it the "anti-cringe" update. I love this. The model still declines requests it should decline, but it stops lecturing you about it. Available via API as gpt-5.3-chat-latest. If you build anything consumer-facing with GPT, this update matters more than benchmark gains.
04 — Gemini 3.1 Flash-Lite is absurdly cheap
Google released Gemini 3.1 Flash-Lite on March 3 at $0.25 per million input tokens. One-eighth the cost of Pro. 2.5x faster on time-to-first-token, configurable "thinking levels," and it topped six benchmarks, beating GPT-5 mini and Claude 4.5 Haiku.
I think this is the most consequential release of the week and nobody's talking about it. The Anthropic drama and OpenAI's mega-round sucked all the oxygen out of the room. But cheap, fast, good-enough models are what actually change how people build. When inference costs drop by 8x, applications that were impossible last month become viable today. That shift is quiet and it's enormous.
05 — Apple confirmed Gemini-powered Siri in iOS 26.4
Apple is rebuilding Siri on Google's Gemini, white-labeled and running on Apple's Private Cloud Compute servers. On-screen awareness, multi-step task chains up to 10 sequential actions, data isolated from Google. Expected late March.
Apple spent years trying to build this in-house and couldn't. So they did the pragmatic thing. Took the best available model, wrapped it in their privacy infrastructure, and shipped. What gets me is the admission baked into this decision. Apple decided that falling further behind was scarier than depending on Google for their most personal user interaction.
06 — NVIDIA bet $4B that copper is done
NVIDIA invested $2B each in Lumentum and Coherent for optical interconnects. Jensen Huang is talking about "gigawatt-scale AI factories." Copper between chips is becoming the bottleneck in large training clusters, and NVIDIA isn't waiting for someone else to fix it. Infrastructure spending that won't show results for years. But it tells you where the biggest AI hardware company thinks the constraint is moving. Not chips. Wires.