AI in March 2026: The OpenClaw Race, the Layoff Wave, and the One-Person Company

March moved on every front.

Anthropic had its best month on record after Claude hit #1 in the App Store and shipped Computer Use. OpenAI shipped four models, killed Sora, and committed to a fully automated researcher by 2028. OpenClaw went viral and every major lab raced to ship a counter.

Oracle, Meta, Atlassian, and Dell cut tens of thousands of jobs, and AI became the #1 cited layoff reason on record. Shenzhen logged 17,000 one-person company applications in ten days.

Palantir's Maven became an official Pentagon program, quietly running on Claude. Jensen Huang pitched tokens as compensation. Apple started opening Siri to outside models.

Here is what happened.

Anthropic's March

Anthropic's Pentagon standoff from February spilled into March in a way nobody predicted.

Claude hits #1 in the App Store. Claude overtook ChatGPT as the most downloaded free iPhone app in the US heading into March. The trigger: public backlash against OpenAI's Pentagon deal, which came hours after Anthropic was designated a "supply chain risk" for refusing to loosen guardrails on autonomous weapons and mass surveillance. Free Claude users are up over 60% since January, and business subscriptions have quadrupled since the start of the year.

Claude for Excel and PowerPoint get shared context. On March 11, Anthropic shipped shared context between Excel and PowerPoint. Analyze sales data in Excel, switch to PowerPoint, and Claude keeps the full conversation and data in memory while it builds the deck. The update also introduced one-click Skills, letting users save a multi-step workflow as a reusable action.

Computer Use ships. On March 23, Claude Computer Use entered research preview for Pro and Max subscribers. Claude can see the screen, click buttons, open apps, and complete multi-step workflows without a human in the loop. macOS only for now. Windows support is planned but Anthropic has not shared a specific timeline.

The Pentagon loss was supposed to be the story. Instead, Anthropic had its best month on record.

OpenAI's March

OpenAI shipped four models, two acquisitions, and its ads business, all while quietly killing its highest-profile consumer product.

GPT-5.4. Released March 5 in Thinking and Pro variants, with Mini and Nano following on March 17. GPT-5.4 is OpenAI's first general-purpose model with native computer use built in. On OSWorld-Verified, it scored 75%, up from GPT-5.2's 47.3%, and just above the human average of 72.4%. Context window: 1M tokens. Factual errors down 33% versus GPT-5.2.

Codex Security. Launched March 6 in research preview. An app-sec agent that identifies vulnerabilities, verifies them, and ships fixes as pull requests. In a 30-day window, it flagged 792 critical and 10,561 high-severity issues across 1.2 million commits, with a false-positive rate below 6%.

Promptfoo and Astral. Two acquisitions in two weeks. OpenAI announced the acquisition of Promptfoo on March 9, folding its red-teaming and evals stack into the Frontier platform. Ten days later, it acquired Astral, the team behind Python tools uv and Ruff. Codex now has 2 million weekly active users.

The automated researcher. On March 20, OpenAI made its "North Star" public. The company plans to ship an autonomous AI research intern by September 2026 and a fully automated researcher by 2028. Jakub Pachocki now sets the long-term research agenda.

Ads tip $100M ARR. The ads pilot launched in February crossed $100 million in annualized revenue less than two months in. Only Free and Go tiers see ads. Paid tiers do not.

Sora shuts down. On March 24, OpenAI killed Sora. The app peaked at one million users, then collapsed to under 500,000 while burning roughly $1 million a day. The web and app experiences end April 26, and the API ends September 24. The Disney partnership died with it.

Catching Up to OpenClaw

OpenClaw's breakout forced everyone else to catch up.

Perplexity Personal Computer. Announced March 11. A local agent that runs on a dedicated Mac mini, touches your files, and orchestrates up to 20 models (Claude, Gemini, Grok) behind the scenes. Perplexity pitches it as a more secure OpenClaw alternative. Requires a $200/month Perplexity Max subscription plus the hardware.

Manus My Computer. Shipped March 16. A desktop app for Mac and Windows that reads, analyzes, and edits local files, and controls local apps through your terminal. Manus was acquired by Meta in late 2025. Paid tier starts at $20/month, or $17/month with annual billing.

Anthropic ships Claude Code Channels. Anthropic shipped Claude Code Channels, which VentureBeat called an "OpenClaw killer." The feature lets developers drive Claude Code from Telegram, Discord, and iMessage, turning the CLI into a persistent, message-driven agent.

Google's Gemini App Actions. The March Pixel Drop added Gemini App Actions, letting Gemini execute real-world tasks across installed Android apps at the OS level. Google also banned paid accounts using OpenClaw.

Tencent, Alibaba, Baidu all ship their own. Tencent plugged ClawBot into WeChat on March 23, giving 1+ billion users OpenClaw access as a chat contact. Alibaba launched Wukong, a multi-agent business platform. Baidu shipped OpenClaw-based agents for desktop, cloud, mobile, and smart home. Zhipu released GLM-5-Turbo, a model built exclusively for OpenClaw workflows.

The race to catch OpenClaw now spans every major lab on both sides of the Pacific.

China's One-Person Company Push

China is treating the AI solo founder as industrial policy, not a meme.

Shenzhen's OPC rush. Since opening a dedicated one-person company registration channel in March, Shenzhen logged 17,000 applications in ten days. Plug and Play China formally stood up an OPC community in Shenzhen's Luohu district by late March.

Subsidies across cities. Shanghai's Pudong district covers compute costs up to 300,000 yuan (around $44,000). Wuhan offers special loans for solo AI founders. Hangzhou runs a free coworking incubator through state-owned facilities. Zhejiang data centers are giving away office space and compute. Suzhou has committed to 30 OPC communities and 1,000 enterprises by 2028.

US "tiny teams," same bet. Silicon Valley is running the same experiment through capital. Cursor hit $500 million ARR with 245 employees in December 2025 and raised at a $29.3 billion valuation. Midjourney is operating at $500 million ARR with 163 employees, profitable six months after launch, and has never taken a dollar of VC. Last year, Dario Amodei put the odds of the first billion-dollar one-person company by 2026 at 70 to 80 percent. Sam Altman said there is a tech CEO group chat betting on the exact timing.

Same bet, different playbook. China is industrializing the solo founder through policy: compute subsidies, free office space, state-backed loans, and fast-lane registrations. The US is doing it through capital and tools: Cursor, Claude, Copilot, and venture checks written straight into one-person LLCs. Both countries now treat "one" as a serious answer to "how many people does it take?"

Payments for Agents

Agents are starting to hold credit cards.

Stripe LLM Token Billing. Launched March 2 in private preview. Developers meter tokens, model calls, and agent tasks through Stripe. Markup is configurable. Works with the OpenAI Agents SDK, Vercel AI SDK, LangChain, and CrewAI.

Virtual cards for agents. Crossmint gives agents their own virtual Visa and Mastercard cards. Visa shipped tokenized "AI-Ready Cards" that swap card numbers for credentials scoped to a specific agent or merchant. Santander and Mastercard completed Europe's first live end-to-end agent-executed payment inside a regulated banking framework.

Put together: agents hold cards, buy things, and get metered through Stripe. The economic plumbing for autonomous commerce is being laid in real time.

Consolidation and Acquisitions

March was a heavy month for deals, especially at Meta.

Meta buys the viral "social network for AI bots." On March 10, Meta acquired Moltbook, the AI-agent-only social network where bots post, upvote, and reply to each other. Founders Matt Schlicht and Ben Parr joined Meta Superintelligence Labs.

Meta hires the Gizmo team. On March 5, Meta acqui-hired the ex-Snapchat engineers at Atma Sciences who built Gizmo, a vibe-coding app for mini-games. Also into Superintelligence Labs.

Meta takes Dreamer's founding team. On March 23, Meta absorbed Dreamer's co-founders, including former Meta and Google executive Hugo Barra. Dreamer had raised $56 million at a $500 million valuation in 2024 and had launched its public beta five weeks earlier. Meta gets a non-exclusive license. Dreamer stays a standalone entity.

Amazon buys Fauna. On March 24, Amazon acquired Fauna Robotics, maker of Sprout, a $50,000 bipedal robot designed for homes and schools. Amazon's second robotics acquisition this month after Rivr.

OpenAI's acquisitions. Promptfoo (March 9) and Astral (March 19), covered above.

Hardware and Compute

Arm ships its own chip. On March 24, Arm launched the Arm AGI CPU, the first in-house silicon in its 35-year history. Up to 136 cores per chip, 272 cores per blade, 8,160 cores per 36kW rack. Meta is the first customer. OpenAI, Cloudflare, and SAP signed on too. Arm stock jumped 16% on the news.

Project Prometheus. On March 19, reports emerged that Bezos is raising up to $100 billion for Project Prometheus, an AI-first manufacturing vehicle. The fund plans to buy industrial companies and retool them with AI, focused on pre-production (prototyping, materials, machinery) rather than assembly robots. The company already raised $6.2 billion in November.

Amazon's AI phone. Reuters reported Amazon is building an Alexa-centric smartphone codenamed "Transformer", led by ex-Microsoft Xbox exec J Allard. Two variants: a conventional phone, and a stripped-down device aimed at screen addiction. No confirmed timeline. Amazon's first phone attempt (Fire Phone, 2014) flopped.

Defense and Policy

Palantir's Maven becomes a program of record. In March, the Deputy Secretary of Defense ordered Maven classified as an official Pentagon program of record, guaranteeing long-term funding and forcing adoption across every service branch. Palantir CTO Shyam Sankar said the Iran war will be remembered as the first major AI-driven conflict. Maven generated roughly 1,000 prioritized targets in the first 24 hours of Operation Epic Fury. Maven runs on Claude.

The irony is sharp. The same Claude the Pentagon flagged as a supply-chain risk in February is the brain inside the targeting system it just made official.

Platform and Partnerships

Apple opens Siri to outside models. Apple plans to let rival AI assistants power Siri in iOS 27, beyond the existing ChatGPT partnership. Gemini is already the backend for the March 2026 Siri overhaul in iOS 26.4. Apple is reportedly paying Google around $1 billion a year.

Nvidia's enterprise Agent Toolkit. At GTC, Nvidia launched an enterprise agent platform with 17 launch partners, including Adobe, Salesforce, SAP, ServiceNow, Cisco, CrowdStrike, and Siemens.

Snowflake's Project SnowWork. Snowflake previewed Project SnowWork, an autonomous enterprise AI platform aimed at non-developer business users.

Mistral Forge. Also at GTC, Mistral announced Forge, an enterprise platform that lets companies train proprietary models on their own data using Mistral's internal training recipes.

Tokens as Compensation

Companies are starting to pay engineers in compute, not just cash.

Nvidia leads. At GTC, Jensen Huang pitched tokens as a new kind of compensation. "If that $500,000 engineer did not consume at least $250,000 worth of tokens, I am going to be deeply alarmed," he said on the All-In podcast. Nvidia already spends around $2 billion a year on tokens for its engineering team. Huang said tokens are becoming "one of the recruiting tools in Silicon Valley," sitting next to equity.

Candidates start negotiating tokens. Microsoft EVP Charles Lamanna told GeekWire that job candidates are asking for AI token budgets during hiring, ranging from "$100 to hundreds of dollars of token cost per day." TechCrunch is calling it the fourth pillar of tech compensation, alongside salary, equity, and bonus.

The AI Layoff Wave

Oracle cuts up to 30,000. On March 31, Oracle began terminating between 20,000 and 30,000 employees globally, roughly 18% of its 162,000-person workforce. TD Cowen estimates the cuts free up $8 to $10 billion in cash flow to help fund Oracle's $156 billion AI infrastructure commitment. Affected staff got termination emails from "Oracle Leadership" at 6 a.m. local time, with no prior HR warning.

Meta cuts hundreds more. On March 25, Meta laid off staff across Reality Labs, recruiting, and sales, affecting fewer than 1,000 employees globally. Reality Labs has absorbed the heaviest cuts, after a January round that took out 1,000 from the same division.

Atlassian cuts 1,600 jobs. On March 11, Atlassian cut 10% of its global workforce. Over 900 of the cuts were in software R&D. CTO Rajeev Rajan stepped down and was replaced by two AI-focused co-CTOs. The company expects $225 to $236 million in charges. CEO Mike Cannon-Brookes: "It would be disingenuous to pretend AI doesn't change the mix of skills we need or the number of roles required."

Dell reveals 11,000 cuts in its filing. A March 18 annual filing disclosed that Dell's fiscal 2026 headcount dropped by roughly 11,000 to 97,000, the third straight year of 10% cuts. Severance ran $569 million. Dell also booked a $43 billion backlog of AI-related product orders in the same year.

AI tops the Challenger report for the first time. March saw 60,620 US layoff announcements, up 25% from February. 15,341 were attributed specifically to AI, making it the #1 cited reason for job cuts in the history of the Challenger report.

Block in February. Atlassian, Dell, Meta, and Oracle in March. The pattern is now routine: cut the workforce, rename AI as a "teammate," keep hiring ML engineers.

Elon's March

Macrohard. On March 11, Musk unveiled a joint Tesla and xAI project called Macrohard, a "jab at Microsoft." The pitch: Grok acts as the navigator, a Tesla-built agent processes screen video and keyboard and mouse actions, and together they emulate entire software companies. Runs on Tesla's $650 AI4 chip paired with xAI's Nvidia servers.

xAI pays bankers $100/hour to train Grok. xAI is recruiting Wall Street analysts, traders, and credit experts as data annotators. Pay: up to $100 an hour. Target skills: leveraged loan syndication, distressed investing, CLOs, MBS. xAI is lagging OpenAI and Anthropic in enterprise finance. This is the catch-up plan.

New Benchmark

ARC-AGI-3 launches. On March 25, the ARC Prize Foundation released ARC-AGI-3. Versions 1 and 2 measured static reasoning. Version 3 measures interactive reasoning through hundreds of game-style environments. At launch, humans scored 100%. Every frontier AI scored below 1%. François Chollet and Sam Altman did a public fireside at YC the next day. Over $2 million in prizes are on the line, with results announced in December.

This is the cleanest measurement yet of the gap between models that look smart on text and agents that can act in an interactive environment.

New Models

March saw another heavy month of releases. The wave of computer-use native, agent-native, and specialized retrieval models is accelerating.

Model	Provider	Type	Highlight
GPT-5.4 Thinking / Pro	OpenAI	LLM	Native computer use. 75% on OSWorld-Verified. 1M context.
GPT-5.4 Mini / Nano	OpenAI	LLM	Smaller variants, released March 17.
Gemini 3.1 Flash-Lite	Google	LLM	$0.25 / $1.50 per million tokens. 2.5x faster TTFT than 2.5 Flash.
Gemini 3.1 Flash Live	Google	Voice	Real-time audio-to-audio for voice agents.
Gemini Embedding 2	Google	Embedding	First natively multimodal embedding. Text, image, video, audio, PDF in one 3,072-dim space.
Lyria 3 Pro	Google	Music	3-minute tracks (up from 30s). Understands intros, verses, choruses.
Mistral Small 4	Mistral	LLM	119B MoE. Merges Magistral, Pixtral, Devstral into one model. Apache 2.0. 256K context.
Voxtral TTS	Mistral	TTS	~4B params. Zero-shot voice cloning. Beats ElevenLabs Flash v2.5 in 68% of blind tests.
Leanstral	Mistral	Code	Lean 4 formal proof agent. $290 vs Opus 4.6's $1,650 at pass@16. Apache 2.0.
Qwen 3.5 small series	Alibaba	LLM	0.8B to 9B. Apache 2.0. 201 languages. Runs on laptops. 9B beats GPT-OSS-120B on GPQA Diamond.
GLM-5-Turbo	Zhipu	LLM	Built exclusively for OpenClaw agent workflows.
MiniMax M2.7	MiniMax	LLM	230B total, 10B active. Self-evolution training loop ran 100+ rounds. $0.30 / $1.20 per million tokens.
MiMo V2 Pro / Omni / TTS	Xiaomi	LLM / Multimodal / TTS	1T+ params, 1M context. Ranked 2nd in Chinese LLMs.
MolmoPoint	Ai2	Vision	Lets models point by selecting input features, not by generating coordinates. SOTA on PointBench (70.7%) and ScreenSpotPro (61.1%).
Qianfan-OCR	Baidu	OCR	4B params. End-to-end image-to-Markdown. SOTA on OmniDocBench v1.5.
SAM 3	Meta	Vision	Segment by noun phrase. 2x accuracy gain on Promptable Concept Segmentation. Companion SAM 3D powers Facebook Marketplace's "View in Room."
Chroma Context-1	Chroma	Retrieval	20B agentic search model. Decomposes queries, parallel tool calls, self-prunes context.
Luma Uni-1	Luma	Image	Autoregressive (not diffusion) image generation. Beats Nano Banana 2 and GPT Image 1.5 on reasoning benchmarks at 10 to 30% lower cost.
Composer 2	Cursor	Coding	Built on Kimi K2.5. $0.50 / $2.50 per million tokens. 61.3 on CursorBench.

The themes are clear. Computer use is table stakes. Chinese open source (MiniMax, MiMo, Qwen, GLM, Qianfan) keeps closing the gap at a fraction of the cost. Retrieval and grounding models (Chroma Context-1, MolmoPoint, Gemini Embedding 2) are becoming a category of their own. And architectures keep diverging: diffusion, autoregressive image, grounding tokens, self-evolving MoE.

Looking Ahead

The story starts in the boardroom. Every executive spending on AI is weighing the same trade-off. Do I run my current business with fewer people, or do I run a much bigger one with the same team I have today? One is cost control. The other is growth. Both depend on whether agents can actually do the work.

In March, they got a lot closer. Agents crossed a threshold. They got capable enough to handle complex, multi-step work end to end, with less and less human supervision. Claude, Perplexity, Manus, and Gemini all shipped products that run a computer on your behalf. Claude Code Channels turns the CLI into a background worker you message over Discord. One person can now pilot many agents at once.

The "same with less" side showed up in the layoff reports. Oracle began cutting up to 30,000. Dell quietly shed 11,000. Atlassian cut 1,600 and said AI changes the number of roles required. AI topped the Challenger report as the #1 cited reason for US layoffs for the first time ever.

The "more with the same" side showed up in Shenzhen. 17,000 one-person company applications in ten days. Pudong pays compute bills up to 300,000 yuan. Midjourney runs $500M ARR with 163 employees and no VC. Dario Amodei's 70 to 80 percent odds on the first billion-dollar solo founder stopped sounding like a flex.

More agents, fewer people, smaller companies. That was March.