The AI Landscape: March 2026
The major labs are richer and faster than ever, but still must answer to political power.
February ended with high-stakes drama in the Pentagon, a sitting president blacklisting one of the country’s most valuable AI companies and a rival stepping into the vacancy the same night. It also delivered the two largest private funding rounds in AI history. The AI landscape entering March has never been more commercially powerful, or more politically uncertain.
Anthropic, the only AI model deployed on the Pentagon’s classified network, refused the Defense Department’s demand to remove two use restrictions from its contract: no mass domestic surveillance of Americans and no fully autonomous weapons. CEO Dario Amodei called the Pentagon’s final contract language essentially meaningless, with safeguards “paired with legalese that would allow those safeguards to be disregarded at will.” On Friday, President Trump ordered every federal agency to cease using Anthropic’s technology, and Defense Secretary Pete Hegseth designated the company a supply-chain risk to national security, a label previously reserved for foreign adversaries like Huawei. Anthropic has said it will challenge the designation in court.
The $200 million defense contract itself is not existential for a company generating $14 billion in annual revenue. The supply-chain designation could be. It requires any company doing business with the U.S. military to certify that it does not use Anthropic’s products, and for a company whose customer base is 80% enterprise, with clients that frequently hold government contracts, the effects could cascade well beyond the Pentagon. Multiple federal agencies, from HHS to NASA’s Jet Propulsion Laboratory, now face a six-month phaseout of a model that some had only recently integrated into core workflows.
Hours after the ban, OpenAI CEO Sam Altman announced a deal to deploy his company’s models on the Pentagon’s classified network, with stated principles mirroring Anthropic’s own: no mass surveillance, human responsibility for the use of force. Whether those principles are written into the contract the way Anthropic insisted, or exist as something softer, will become apparent soon enough. Altman told CNBC he shares Anthropic’s “red lines,” and hundreds of employees from Google and OpenAI signed a petition calling on their companies to mirror Anthropic’s stance. Elon Musk’s xAI, by contrast, had already signed its own classified-network agreement under the Pentagon’s “all lawful purposes” standard, the one Anthropic rejected. Defense officials acknowledged to Axios that Grok is not yet viewed as a capable substitute for Claude in classified settings, and that disentangling from Claude would be, in one official’s phrasing, a significant logistical challenge.
The Financial Picture
The Pentagon confrontation dominated the weekend’s headlines, but it landed in the middle of a month already reshaping the industry’s financial architecture. Anthropic closed a $30 billion Series G on February 12 at a $380 billion valuation, having grown revenue tenfold annually for three consecutive years. Claude Code’s annualized revenue crossed $2.5 billion and doubled since January 1, while business subscriptions quadrupled over the same period. OpenAI then closed $110 billion at a $730 billion pre-money valuation, the largest private round in history, with Amazon investing $50 billion, Nvidia $30 billion, and SoftBank $30 billion. Between them, two companies that did not exist five years ago raised $140 billion in a single month.
OpenAI generated $13.1 billion in 2025, beating its $10 billion target while burning $8 billion, and ChatGPT now supports over 900 million weekly active users. Anthropic’s $14 billion run rate is built on a different foundation: eight of the Fortune 10 are Claude customers, and the business is overwhelmingly enterprise rather than consumer. Ramp data suggests 79% of businesses paying for OpenAI also pay for Anthropic, which means this is not yet a zero-sum market; companies are buying everything available and sorting out the strategy later.
The products absorbing all that spending are also shifting. Anthropic’s Cowork, which extends Claude Code’s agentic capabilities to general knowledge work, triggered a roughly $2 trillion selloff in SaaS stocks when it debuted in January, and February brought a broader rollout with enterprise connectors for Google Drive, Gmail, DocuSign, and FactSet, plus domain-specific plugins for finance, HR, and design. Claude Sonnet 4.6 pushed computer-use capabilities to 72.5% on OSWorld, approaching human-level performance on tasks involving spreadsheets and multi-tab browser workflows, and Anthropic’s acquisition of Vercept, a perception-and-interaction startup, signaled that computer use is a core investment thesis rather than a research curiosity. OpenAI’s competing coding product, Codex, has passed 1.5 million weekly active users. As part of its Amazon deal, OpenAI is developing a “stateful runtime environment” on Bedrock with an additional $100 billion in compute commitments, binding the two companies’ infrastructure fortunes together at a scale that would be difficult to unwind.
Google shipped Gemini 3.1 Pro with what it describes as more than double the reasoning performance of its predecessor, a specialized Deep Think mode for scientific and engineering tasks, and a creative suite including Lyria 3 for music generation and improved image and video tools. Google’s position remains distinctive for its breadth: 650 million monthly Gemini users across mobile, web, TV, and Workspace apps, with an upcoming Siri integration via iOS 26.4 that could extend its reach further than any competitor can match through organic growth. With $185 billion in planned capital expenditures for the year, funded from its own cash flow rather than private rounds, Google can invest on a timeline that neither OpenAI nor Anthropic can match without continuous fundraising.
Meta’s trajectory is harder to read. After building its AI strategy around the open-weight Llama models, reports indicate the company is preparing closed-source successors codenamed Avocado and Mango, targeting the first half of this year. Yann LeCun’s departure and candid criticism of Meta’s tilt toward large language models at the expense of foundational research adds uncertainty about the company’s scientific direction. If Meta’s role shifts from open-source standard-bearer to infrastructure giant that happens to release some models openly, the thousands of companies that built on Llama assuming continuity will need to reassess.
The model release that kept investors anxious all month still has not arrived. DeepSeek V4, initially expected around mid-February to coincide with the Lunar New Year, missed its rumored window.