Week in AI: Anthropic Blacklisted, Memory Chip Crisis Hits Smartphones, and Perplexity Launches Digital Workers
The week's biggest AI stories — Pentagon blacklists Anthropic over safety guardrails, AI-driven memory shortage pushes smartphone prices to $523, Perplexity Computer orchestrates 19 models, Burger King deploys surveillance headsets, and Standard Intelligence trains on 11M hours of video.
This was the week artificial intelligence stopped being an abstract policy debate and started producing consequences that real people can feel. The Pentagon blacklisted one of America's leading AI companies. A chip shortage driven by data center demand pushed the average smartphone past $500 for the first time. And two new product launches — one from Perplexity, one from Burger King — demonstrated that AI agents are arriving in places nobody expected.
Here's everything that mattered in AI this week.
Week in Review
1. Pentagon Blacklists Anthropic After Safety Guardrail Standoff
The biggest story of the week — and arguably the most consequential AI policy event of 2026 so far — was the escalating conflict between the Pentagon and Anthropic that culminated in Defense Secretary Pete Hegseth formally designating Anthropic a "supply chain risk" to national security.
The dispute centered on Anthropic's refusal to remove two specific restrictions from its Claude AI models: prohibitions on mass surveillance of American citizens and on powering fully autonomous weapon systems. The Pentagon's position, articulated through an ultimatum with a 5:01 PM ET Friday deadline, was that these restrictions impeded military operations and violated the terms of a contract worth up to $200 million.
Contract at Stake
$200 Million
Pentagon AI contract that triggered the Anthropic blacklisting
Anthropic CEO Dario Amodei released a public statement refusing to bend, arguing that the restrictions represent foundational safety principles — not negotiable contract terms. Hours after the blacklisting, OpenAI announced it had struck a deal with the Defense Department to provide AI technology for classified networks, with CEO Sam Altman claiming he shares Anthropic's "red lines" while simultaneously accepting the contract.
The paradox was immediate and obvious: OpenAI claimed to share the same safety principles as Anthropic while accepting the exact contract Anthropic was blacklisted for refusing. Whether OpenAI's deal includes the same guardrail restrictions remains unclear — the company hasn't published its terms.
I covered the full timeline, the Defense Production Act threat, and the OpenAI deal in my deep analysis of the Anthropic blacklisting. The earlier reporting on the Pentagon ultimatum from Tuesday covers the Defense Production Act implications — a Korean War-era law that has never been invoked against an AI company.
| event | severity | precedent |
|---|---|---|
| Ultimatum Issued | 8 | 9 |
| DPA Threat | 9 | 10 |
| Blacklisting | 10 | 10 |
| OpenAI Deal | 7 | 8 |
Why it matters: This sets a precedent that AI companies can be blacklisted from all government and military contracting for maintaining safety restrictions. Every AI company now faces a choice between government revenue and safety principles — and the first company to face that choice just lost $200 million.
2. AI-Driven Memory Chip Shortage Pushes Smartphones Past $500
The second major story this week quantified something the semiconductor industry has been warning about for months: AI's insatiable demand for memory chips is creating a consumer electronics crisis.
A CNN Business report, backed by IDC data, revealed that the average smartphone selling price will hit $523 in 2026 — a 14% increase — while total smartphone shipments are expected to fall to 1.12 billion units, the lowest annual volume since 2013. The culprit is straightforward: AI data centers are consuming memory chips faster than manufacturers can produce them, and consumer electronics are getting what's left.
AI's Memory Chip Impact on Consumer Electronics
Data Center Demand
Consumer Impact
The mid-range segment is getting squeezed hardest. Budget phones can cut features to hit price points. Flagships have margin to absorb cost increases. But mid-range devices — the $200-$400 category that represents the majority of global smartphone sales — are caught between rising component costs and price-sensitive consumers.
I covered the broader semiconductor supply chain dynamics in my analysis of Meta's TPU diversification and ASML's EUV breakthrough, which explains why companies like Meta are spending billions to reduce NVIDIA dependence. This week's memory chip data adds a consumer dimension to that infrastructure story — the $650 billion flowing into AI data centers isn't just an enterprise spending story. It's directly raising the price of the phone in your pocket.
| Name | Value |
|---|---|
| AI Data Centers | 62 |
| Smartphones | 18 |
| PCs and Laptops | 11 |
| Gaming Consoles | 5 |
| Other Electronics | 4 |
Why it matters: AI infrastructure spending is now producing visible consumer-level inflation in electronics. This is the first time AI costs have directly impacted the average consumer's purchasing decisions at scale.
3. Perplexity Launches "Computer" — 19 AI Models, One Digital Worker
Perplexity AI launched Perplexity Computer on Tuesday, introducing a multi-model orchestration platform that deploys 19 specialized AI systems to handle complete projects autonomously. Users describe a desired outcome, and the system breaks it into subtasks handled by specialized sub-agents that research, draft, code, deploy, and manage workflows — potentially running for hours or months.
Describe Desired Outcome
Natural language description of the project goal. No task decomposition required from the user.
19-Model Task Routing
System breaks the goal into subtasks and routes each to the optimal specialized model.
Sub-Agent Workflows
Specialized agents handle research, writing, coding, image generation, and external API calls.
Continuous Operation
Workflows can run for hours or months, with progress tracking and intermediate deliverables.
The platform is currently available only to Perplexity Max subscribers ($200/month) with 10,000 monthly credits and a 20,000-credit launch bonus. Pro and Enterprise tiers will follow after load testing.
What makes Perplexity Computer notable isn't any single capability — it's the multi-model architecture. Rather than relying on one foundation model for everything, it routes tasks to whichever of its 19 models is best suited. This mirrors the multi-model consensus approach that Perplexity pioneered with its Model Council feature earlier this year.
Perplexity Computer
19 Models
Specialized AI systems orchestrated as a single digital worker
Why it matters: This is the clearest signal yet that the future of AI isn't one model doing everything — it's orchestration layers routing tasks to specialized models. The $200/month price point positions it as a professional tool, not a consumer product, and directly competes with OpenAI's Frontier platform for enterprise workflows.
4. Burger King Deploys AI Surveillance Headsets in 500 Locations
In the week's most dystopian product launch, Burger King announced that 500 locations are now equipping employees with OpenAI-powered headsets called "Patty" that monitor drive-through conversations for "friendliness" — specifically tracking whether workers say phrases like "Welcome to Burger King," "please," and "thank you."
The headsets are part of a broader platform called BK Assistant that also monitors inventory levels, removes sold-out items from digital menus, and alerts managers to operational issues. Burger King's chief digital officer Thibault Roux described it as a "coaching tool" while acknowledging the company is "iterating" on capturing conversational tone.
| feature | surveillance | operational |
|---|---|---|
| Friendliness Scoring | 95 | 20 |
| Inventory Tracking | 30 | 90 |
| Menu Updates | 10 | 85 |
| Manager Alerts | 60 | 75 |
| Tone Analysis | 90 | 15 |
The reaction was swift and polarized. Labor advocates called it "next-level micromanagement." Privacy experts noted that continuous audio monitoring of minimum-wage workers raises significant consent and dignity concerns. Burger King framed it as operational efficiency.
Why it matters: This is the first major fast-food chain to deploy real-time AI monitoring of employee speech patterns at scale. Whether framed as "coaching" or "surveillance," it establishes a precedent for AI-powered behavioral monitoring in service industries — and the OpenAI partnership means the technology will only get more sophisticated.
5. Standard Intelligence Trains FDM-1 on 11 Million Hours of Video
Standard Intelligence released FDM-1, billed as the world's first general-purpose computer action model. Unlike existing computer-use models that learn from screenshots, FDM-1 trains directly on video — 11 million hours of screen recordings labeled using an inverse dynamics model.
The efficiency gains are striking. FDM-1's video encoder compresses nearly 2 hours of 30 FPS video into 1 million tokens. For comparison, Gemini handles 775 frames and Claude handles 162 frames with the same 200,000 tokens.
Video Processing Efficiency
FDM-1
Current Models (200K tokens)
Demonstrations showed FDM-1 constructing a gear in Blender, finding software bugs, and driving a real car through San Francisco using arrow keys. The generality is the breakthrough — one model trained on video that can operate across CAD software, web browsers, and physical vehicles.
Why it matters: If FDM-1's approach scales, it could make computer-use AI dramatically more capable by learning from the same medium humans use — continuous video rather than discrete screenshots. The 11M-hour training dataset is orders of magnitude larger than anything previously used for computer action models.
6. Google Gemini 3.1 Pro Reshuffles the Benchmark Race
Google launched Gemini 3.1 Pro on February 19, claiming 13 of 16 benchmark wins. The headline numbers are impressive — 77.1% on ARC-AGI-2 (more than double Gemini 3 Pro), 94.3% on GPQA Diamond, and aggressive pricing at $2/$12 per million tokens.
But as I detailed in today's deep analysis of the Gemini 3.1 Pro benchmarks, the "13 of 16 wins" claim carries significant asterisks. GPT-5.3-Codex appears in only 2 of 16 benchmarks. Gemini trails Anthropic by 300 Elo points on enterprise tasks (GDPval-AA). And on Chatbot Arena — where humans evaluate outputs blindly — Gemini and Opus are statistically tied.
Why it matters: No single frontier model dominates all categories. The three-way race between Google, Anthropic, and OpenAI is fragmenting by use case, making multi-model strategies essential for enterprises.
Week Ahead
What to watch for March 2-6, 2026:
- Anthropic's response: How does the company navigate post-blacklisting? Employee retention, investor confidence, and commercial customer reactions will all unfold this week.
- OpenAI Pentagon details: The terms of OpenAI's classified network deal remain undisclosed. Expect pressure from journalists and Congress for specifics on guardrail terms.
- Memory chip pricing data: Samsung and SK Hynix earnings guidance will quantify Q1 2026 HBM demand. Watch for forward allocation percentages between AI and consumer segments.
- Perplexity Computer early reviews: The first wave of Max subscriber reviews will establish whether 19-model orchestration delivers on the promise or creates coordination overhead.
- MWC 2026 aftermath: Mobile World Congress wrapped this week. Expect delayed announcements about AI-powered device features and carrier partnerships through the first half of March.