AI in February 2026: Swarms, Selloffs, and Consolidation

February was the month AI stopped being just a tech story and became a market story.

Here's what happened.

The SpaceX-xAI Merger

On February 2, Elon Musk merged SpaceX and xAI in a $1.25 trillion all-stock deal, making it the largest merger in history. SpaceX was valued at $1 trillion, xAI at $250 billion.

Musk's pitch: "orbital data centers" powered by the Sun. He claims Earth cannot sustain AI's energy demands long term. SpaceX applied to the FCC for authorization to launch up to one million satellites to serve as orbital compute infrastructure. The FCC's Space Bureau accepted the filing five days later.

The financial picture underneath tells its own story. xAI was burning roughly $1 billion per month with minimal revenue relative to its costs. The merger positions the combined entity for a planned SpaceX IPO targeting a $1.5 trillion valuation in mid-2026. Critics say the deal gives xAI investors an exit via the IPO.

Meanwhile, xAI's founding team is hemorrhaging. Half of xAI's twelve co-founders have departed, including Tony Wu and Jimmy Ba in February alone.

Whether you believe the orbital data center vision or not, the financial engineering is clear. This is as much about the IPO as it is about infrastructure.

The SaaS Selloff Continues

Anthropic triggered three separate selloffs in February, hitting SaaS, cybersecurity, and legacy tech stocks. Each time, a single product or blog post moved billions in market cap.

1. The Cowork Selloff

Anthropic’s industry-specific plugins for Claude Cowork triggered a broad selloff starting February 3. RELX fell 14%, Wolters Kluwer fell 13%, and London Stock Exchange Group dropped over 8%.

The selloff spread globally by the next day. Thomson Reuters dropped 16%. Salesforce, Adobe, and Intuit each fell over 2%. Alternative asset managers Blue Owl Capital, Ares Management, and KKR lost roughly 10% each. The S&P 500 software and services index declined nearly 13% over five sessions, down 26% from its October peak. By the end of the week, cumulative losses approached nearly $1 trillion.

The fear was specific: Anthropic was moving from models into the application layer, directly competing with enterprise SaaS companies. The market stabilized on February 24 after Dario Amodei joined Salesforce's Marc Benioff on stage to reframe AI as "human augmentation."

2. The Claude Code Security Selloff

On February 20, Anthropic launched Claude Code Security, a reasoning-based vulnerability scanner built on Opus 4.6. Anthropic claimed it identified over 500 previously unknown high-severity vulnerabilities in production open-source code.

On the first day, over $15 billion was wiped from cybersecurity stocks. Palo Alto Networks, Fortinet, Cloudflare, and Zscaler all declined sharply. By the end of the second day, CrowdStrike alone had fallen roughly 18%, wiping $20 billion in market cap.

3. The COBOL Modernization Shock

Anthropic published a blog post showing how Claude Code can automate COBOL modernization. IBM fell 13.2%, its biggest daily decline since 2000. This reaction was more questionable. IBM is a $200 billion company with massive consulting, cloud, and enterprise AI businesses. COBOL maintenance is a fraction of its revenue. Wiping out billions in market cap over a blog post about a language most developers have never touched is not a rational price signal.

Three selloffs in one month, each triggered by a single Anthropic announcement. Some of the reactions were rational. Some were pure panic. But the pattern is clear: the market is pricing in disruption before it fully materializes.

Anthropic and the Pentagon

This is the most consequential AI policy story of the month. Possibly the year.

Defense Secretary Hegseth gave Anthropic a deadline to allow Claude for "all lawful purposes," including autonomous weapons and mass surveillance. Anthropic refused. On February 27, Trump ordered every federal agency to cease all use of Anthropic's technology, and the Pentagon began designating Anthropic a "supply chain risk."

Meanwhile, xAI's Grok was approved for classified Pentagon systems after agreeing to the "all lawful purposes" standard. Officials acknowledged that Grok does not yet match Claude in performance.

Then came a twist that pulled in the opposite direction. On February 24, Anthropic dropped its core Responsible Scaling Policy. RSP Version 3.0 replaced binding safety commitments with transparency obligations and non-binding targets. The central pledge to never train a model unless safety measures were proven to work beforehand was removed. Bloomberg called it "one of the most dramatic policy shifts in the AI industry yet."

These events did not happen in isolation. On February 9, Mrinank Sharma, head of Anthropic's Safeguards Research Team, resigned and posted a letter that said "the world is in peril." He announced plans to move to the UK and pursue poetry. This was not unique to Anthropic. The same week, OpenAI researcher Zoe Hitzig resigned over ChatGPT ads, publishing a New York Times op-ed titled "OpenAI Is Making the Mistakes Facebook Made. I Quit." Different companies, different triggers, same pattern: the people closest to these systems are losing faith in the institutions building them.

Anthropic built its brand on being the safety-first AI company. In February, it weakened its own safety pledges, then refused the Pentagon's terms and lost government access days later. These two events pull in opposite directions, and together they reveal the impossible tension at the heart of the AI safety debate: be principled enough to matter, but competitive enough to survive.

The Agent Swarm Era

February was the month multi-agent systems went from research concept to shipping product. Three different companies shipped three different approaches in the same month.

Claude Code Agent Teams. Shipped with Opus 4.6 as an experimental feature. One session acts as team lead, coordinating work and assigning tasks. Teammates work independently, each in its own context window (up to 1M tokens, up to 16 agents). Unlike subagents, teammates can communicate directly with each other. Three display modes: in-process, split panes (tmux/iTerm2), and auto-detect. Community finding so far: small teams of 2 to 3 agents with a strict plan-first approach work best.

Kimi K2.5 Agent Swarm (Moonshot AI). Self-directs up to 100 sub-agents across up to 1,500 tool calls. Trained with Parallel-Agent Reinforcement Learning (PARL), meaning no predefined roles are needed. 4.5x faster than single-agent setups. Free consumer tier, API at $0.60 per million input tokens. 78.4% on BrowseComp in swarm mode, up from 60.6% single-agent.

Grok 4.20 Multi-Agent Debate. Four specialized agents: Grok (coordinator), Harper (researcher), Benjamin (logician), Lucas (creative contrarian). Lucas is literally trained to disagree with the other three. This reduces hallucinations from roughly 12% to roughly 4.2%. In the Alpha Arena trading competition, Grok 4.20 was the only model that finished in profit (+12.11%). Free users can watch the four agents debate in real time. SuperGrok ($30/month) scales to 16 agents.

The shift from "one agent, one task" to "agent teams, complex projects" is the new competitive frontier.

OpenAI's February

OpenAI shipped a coding app, an enterprise platform, and hardware plans in a single month while fighting a $134 billion lawsuit.

Codex App. A standalone macOS desktop application for AI-assisted coding. Different from the model: an orchestration layer for managing multiple AI coding agents. Run several agents in parallel, schedule background automations, use "Skills" to extend capabilities. Available to all ChatGPT tiers including free users temporarily.

Sam Altman called it "the most loved internal product we've ever had." A four-person team built the Sora for Android app in 28 days using Codex. Over a million developers have used Codex since mid-December 2025.

Frontier. An enterprise platform for building, deploying, and managing AI agents. Model agnostic: works with OpenAI, Google, Microsoft, and Anthropic agents. Agent identity management, IAM, governance, and Forward Deployed Engineers paired with customer teams. Early customers include HP, Intuit, Oracle, State Farm, Uber, and Thermo Fisher Scientific.

OpenAI followed up on February 23 with Frontier Alliances: multiyear deals with Accenture, BCG, Capgemini, and McKinsey.

Devices. A 200+ person team led by Jony Ive (acquired via the $6.5 billion io deal) is building hardware. First up: a smart speaker with camera at $200 to $300, shipping earliest February 2027. There is also a pen-shaped device codenamed "Gumdrop" and smart glasses potentially in 2028. Design philosophy: "more peaceful than a smartphone." Screenless to reduce addiction risk.

On the Legal Front. A federal judge dismissed xAI's trade secrets lawsuit against OpenAI on February 24, finding xAI failed to allege facts showing OpenAI directed, knew of, or used any trade secrets. Separately, Musk's deposition was released on February 27, in which he acknowledged his $100 million donation claim was wrong. The actual figure: $44.8 million. Jury selection for the main $134 billion lawsuit is scheduled for April 27.

Claude Goes Everywhere

Instead of asking users to come to Claude, Anthropic is sending Claude to the tools they already use.

Claude launched inside Microsoft Excel (January 24) and PowerPoint (February 20) as Microsoft 365 add-ins. PowerPoint support generates native, editable slides from natural language. It respects existing slide masters, fonts, and color schemes. It creates native charts, not static images. Excel support includes pivot tables, chart modifications, and conditional formatting.

Priced at $20/month (Claude Pro), undercutting Microsoft 365 Copilot's $30/month enterprise add-on. Partners include FactSet, S&P Global, LSEG, and Apollo for industry-specific plugins.

The enterprise distribution play is unmistakable. Undercutting Copilot on price while matching or exceeding its capabilities is an aggressive move.

Apple's Gemini-Powered Siri

Throughout February, details continued to emerge about Apple's revamped Siri, powered by Google's Gemini. Reports pointed to a late February or early March unveiling, with Apple's March 2 product event the most likely stage.

Google reportedly built a custom 1.2 trillion parameter Gemini model for Apple, running on Apple's Private Cloud Compute servers. Apple is paying roughly $1 billion per year for the partnership.

Phase 1 (iOS 26.4, beta expected spring 2026) brings personal context awareness, on-screen understanding, and in-app actions via voice. Phase 2, codenamed "Campos," is a full chatbot redesign launching with iOS 27 at WWDC in June. Craig Federighi told employees internally that the upgrade would be "bigger than what we originally envisioned." Apple still maintains ChatGPT integration alongside Gemini.

None of this has shipped yet, but the implications are already being priced in. Google's technology will power the default assistant on every iPhone. The partnership pushed Alphabet's market cap past $4 trillion briefly in January.

The Compute Race

The AI infrastructure arms race is reshaping the semiconductor industry.

Meta's $100B AMD Deal. Meta announced plans to buy up to $100 billion worth of AMD chips, specifically custom MI450 GPUs, covering 6GW of AI compute over five years. This comes weeks after Meta also expanded its Nvidia deal.

The deal's structure shows how high the stakes are: AMD issued Meta warrants for 160 million shares (roughly 10% of the company) at $0.01 per share, vesting as Meta hits purchase milestones. Part of Meta's "Meta Compute" initiative, 2026 capex guidance sits at $115 billion to $135 billion.

Other infrastructure deals.

OpenAI and Cerebras signed a $10 billion compute deal: 750MW of inference over three years, the largest high-speed AI inference deployment in the world. Cerebras re-filed for IPO.
A consortium including BlackRock, Microsoft, NVIDIA, MGX, and xAI acquired Aligned Data Centers for $40 billion: 50 campuses, 5GW capacity. The largest data center acquisition in history.
Micron broke ground on a $100 billion New York megafab: up to four fabs, nearly 50,000 jobs, producing advanced DRAM and HBM critical for AI.

The Nvidia monopoly is cracking as Meta, OpenAI, and others diversify to AMD and Cerebras. But the sheer scale of spending shows how far this race has escalated.

Perplexity Computer and Model Council

Perplexity's thesis is that AI models are specializing, not converging. The winner is whoever orchestrates them all.

Model Council runs queries across three frontier models (e.g., Opus 4.6, GPT 5.2, Gemini 3.0). A synthesizer resolves conflicts and shows where models agree and disagree.

Computer, launched February 25, goes further: it orchestrates 19 different AI models for complex workflows. Opus 4.6 for reasoning, Gemini for research, Veo 3.1 for video, Grok for speed tasks. Over 400 app integrations. Tasks can run for hours or months. Available on the $200/month Perplexity Max tier.

Perplexity engineers built Computer itself using Claude Code in about two months. From search engine to agent platform in one year.

The Distillation Wars

On February 23, Anthropic accused DeepSeek, Moonshot, and MiniMax of "industrial-scale" distillation attacks on Claude.

The numbers: approximately 24,000 fraudulent accounts generated over 16 million exchanges with Claude. MiniMax conducted over 13 million exchanges. Moonshot targeted agentic reasoning and tool use with 3.4 million. DeepSeek focused on reasoning tasks and censorship-safe rewrites with over 150,000.

DeepSeek's prompts specifically asked Claude to articulate its internal reasoning step by step, generating chain-of-thought training data. Anthropic calls the distributed infrastructure used to manage thousands of accounts simultaneously "hydra clusters."

Anthropic warned that distilled models lack safety protections and pose national security risks. OpenAI had previously made similar accusations against DeepSeek. This comes as the Trump administration allowed Nvidia to export advanced AI chips (H200) to China.

The accusation that Chinese labs used thousands of fake accounts to extract Claude's capabilities is a significant escalation in AI geopolitics. It frames model distillation not as a research technique but as industrial espionage. Whether or not the claims hold up legally, they're shaping the policy debate around export controls and AI access.

The Workforce Shift

Block Layoffs. Jack Dorsey cut over 4,000 employees, about 40% of Block's global workforce (from 10,000+ to under 6,000). He attributed it entirely to AI: "Intelligence tools have changed what it means to build and run a company." He predicted most companies will make similar changes within a year.

Block's stock surged roughly 24% in after-hours trading. Remaining employees must use generative AI daily, tracked and tied to performance evaluations. Critics questioned whether AI was the real reason. Block more than tripled its headcount from 3,900 to 12,500 during COVID, and Dorsey spent $68 million on a company event just five months before the cuts. Blaming AI is easier than admitting years of overhiring.

The broader layoff wave. Amazon cut 16,000 corporate jobs (second round in months, 30,000 total), with 2026 capex guided at $200 billion, most directed toward AI. Baker McKenzie cut 700 to 1,000 staff, becoming the first major BigLaw firm to explicitly blame AI for mass layoffs. Salesforce cut roughly 1,000, including its Agentforce AI team head.

Counter-trend: AI-era hiring. IBM is tripling entry-level US hiring. Its CHRO says AI handles tasks from 2 to 3 years ago, but cutting junior hiring risks a future shortage of mid-level managers. Dropbox expanded its intern and grad programs, citing Gen Z's AI fluency. Cognizant is quadrupling its early-career pipeline, calling Gen Z "AI-natives."

Enterprise AI tracking. Amazon tracks employee AI usage through "Clarity," factoring it into promotions. Accenture trained 550,000 of its roughly 780,000 employees in AI. Promotions now require "regular adoption" of AI tools. Meta, Microsoft, and KPMG all now tie AI usage to performance reviews.

The picture is contradictory. Some companies are cutting thousands and calling it AI. Others are hiring more juniors because they know AI cannot replace the pipeline that builds future leaders. The one thing both sides agree on: if you work in tech and you are not using AI daily, your employer has noticed.

Consolidation Wave

AI companies are no longer just building models. They're acquiring infrastructure, capabilities, distribution, and talent.

Mistral buys Koyeb. The French AI company's first acquisition. Koyeb is a Paris-based serverless cloud startup with 13 employees. The goal: make Mistral a full-stack AI player with its own inference infrastructure ("Mistral Compute"). Mistral had just announced $1.4 billion in Swedish data center investments and recently crossed $400 million annual recurring revenue.

Anthropic buys Vercept. A Seattle AI startup built by Allen Institute for AI alumni. Vercept built "Vy," a cloud-based computer-use agent operating a remote MacBook. The acquisition strengthens Claude's computer use capabilities (OSWorld improved from under 15% to 72.5% with Sonnet 4.6). Anthropic's second major acquisition in three months, after buying Bun in December.

Google DeepMind triple deal. Acquired Common Sense Machines (2D-to-3D AI models), licensed Hume AI (voice and emotional intelligence) and hired its team, and took a strategic stake in Sakana AI (Tokyo). The FTC is now scrutinizing the deals.

The pattern is consistent: AI companies are buying what they cannot build fast enough. Infrastructure, computer use, 3D models, emotional intelligence. Each acquisition fills a gap in the stack needed to ship complete products, not just models.

China's AI Push

China's AI push is not just models. It's robotics, distribution, and manufacturing.

Baidu adds OpenClaw. Baidu integrated OpenClaw into its main search app (700 million monthly active users), the first time OpenClaw is available outside messaging platforms. OpenClaw hit 150,000 GitHub stars in 10 weeks.

China's Robot Moment. The Spring Festival Gala (China's equivalent of the Super Bowl) featured two dozen humanoid robots doing martial arts, backflips, and parkour. Four startups powered the show: Unitree Robotics, Galbot, Noetix, and MagicLab. A major upgrade from last year's simple handkerchief-twirling.

China shipped nearly 90% of all humanoid robots sold globally in 2025. Morgan Stanley projects Chinese humanoid sales will more than double to 28,000 units in 2026. Robots sold out on JD.com during the live broadcast, with Galbot units priced at roughly $91,000 each.

The EU AI Act in Practice

The EU AI Act continued to dominate European regulatory headlines, though February was defined more by delays than enforcement.

Missed deadlines. The European Commission was required by February 2 to publish guidelines on how to classify AI systems as "high-risk" versus "not high-risk." It missed the deadline. Officials said they were still integrating stakeholder feedback, with a revised draft now expected sometime in March or April. This followed the fall 2025 delay in harmonized technical standards, which were the responsibility of standardization bodies CEN and CENELEC rather than the Commission itself.

The Digital Omnibus. The Commission's proposed Digital Omnibus on AI would delay high-risk AI obligations by up to 16 months past the August 2, 2026 start date. The Commission estimates it could save businesses up to EUR 5 billion in administrative costs by 2029. In practice, it is an acknowledgment that the original timeline was too aggressive.

Transparency Code of Practice. The first draft of the Code of Practice on AI-generated content labeling proposes watermarks, metadata, and visible labels. A second draft is expected in mid-March, with finalization in June.

AI Gigafactories. The EU Council adopted an amendment to the EuroHPC regulation in January, extending its mandate to include AI Gigafactories. The Commission received 76 expressions of interest across 16 Member States, totaling over EUR 230 billion in proposed investment. Each Gigafactory would house over 100,000 advanced AI processors for training next-generation models. A formal call for bids is expected in early 2026.

Meta and WhatsApp. On February 9, the Commission sent Meta a formal Statement of Objections, stating that Meta likely abused its dominant position by excluding third-party AI assistants from WhatsApp, leaving Meta AI as the sole assistant on the platform. The Commission is considering interim measures to force Meta to restore third-party access.

The EU's approach has been deliberate. Full enforcement for most provisions does not begin until August 2026. But the pattern is becoming familiar: ambitious regulation, missed deadlines, and a growing gap between the rules on paper and the industry they are trying to govern.

New Models

February saw a record number of model releases. More frontier-class models launched this month than in all of 2023.

Model	Provider	Type	Highlight
Opus 4.6	Anthropic	LLM	Top Finance Agent benchmark. 16 agents wrote a C compiler in Rust. 200K context (1M beta), 128K output.
Sonnet 4.6	Anthropic	LLM	$3/$15 per million tokens. Preferred over Opus 4.5 59% of the time.
GPT-5.3 Codex	OpenAI	LLM	25% faster than GPT-5.2. First rated "High" for cybersecurity capability.
GPT-5.3 Codex Spark	OpenAI	LLM	Runs on Cerebras chips (first OpenAI model off Nvidia). 1,000+ tokens/sec.
Gemini 3 Deep Think	Google	Reasoning	84.6% on ARC-AGI-2. Solved 18 previously unsolved research problems.
Gemini 3.1 Pro	Google	LLM	77.1% on ARC-AGI-2. 1M context. Multimodal across text, image, audio, video, code.
Lyria 3	Google	Music	Music generation in 8 languages. Available in Gemini app (750M+ MAU).
Nano Banana 2	Google	Image	Gemini 3.1 Flash Image. 2 to 3x faster than Nano Banana Pro. Up to 4K. 50% cheaper.
Grok 4.20	xAI	LLM	Multi-agent debate with four specialized agents. Only model to profit in Alpha Arena.
DeepSeek V4	DeepSeek	LLM	1T params, 32B active. Quantized version fits on dual RTX 4090s.
MiniMax M2.5	MiniMax	LLM	80.2% SWE-Bench (within 0.6% of Opus 4.6). 1/20th the cost. Open weight.
GLM-5	Zhipu	LLM	744B MoE. Trained entirely on Huawei Ascend chips. Zero Nvidia hardware. MIT license.
Qwen 3.5	Alibaba	LLM	397B MoE, 17B active. 8.6x faster than Qwen3 Max.
Seedance 2.0	ByteDance	Video	Simultaneous audio and video generation. Disney sent a cease-and-desist.
Mercury 2	Inception	LLM	First reasoning diffusion-based LLM. 1,009 tokens/sec on Blackwell (5x faster than conventional).
pplx-embed-v1	Perplexity	Embedding	0.6B and 4B sizes. Open-weight (MIT). Optimized for standalone queries and retrieval.
pplx-embed-context-v1	Perplexity	Embedding	0.6B and 4B sizes. New SOTA on ConTEB (81.96% nDCG@10 at 4B). Embeds chunks with document-level context.

Open-source (MiniMax, GLM-5, Qwen, DeepSeek V4, pplx-embed) is closing the gap with closed models fast. Chinese labs are building on non-Nvidia hardware. New architectures like Mercury 2's diffusion-based LLM and pplx-embed's diffusion-pretrained encoders hint that the transformer may not be the only game in town.

Looking Ahead

The pace of change is outrunning most people's ability to keep up with it. In one month, agent teams shipped, markets lost hundreds of billions, thousands of jobs were cut, and a major AI company lost all federal access for refusing to build autonomous weapons. Any one of these would have dominated headlines for weeks in 2022. In February 2026, they happened simultaneously and competed for attention.

The second-order effects are arriving faster than expected. Models ship and markets move the same day. Layoffs follow the same week. Policy decisions are forced within days. The distance between "AI can do this" and "your industry just changed" is shrinking.

This is only February. Ten months remain.