The Week in AI: February 23 – March 1, 2026 (Week 9)

0

This week’s news: February 23 – March 1, 2026 (Week 9)

The week was dominated by the dispute between Anthropic and the US Department of Defense, which ended with a presidential ban on Claude in federal agencies and a Pentagon deal with OpenAI. OpenAI also raised a record 110 billion dollars, and ChatGPT moved closer to a billion weekly users. Nvidia reported another record quarter, Meta signed a massive chip deal with AMD, and Google released Nano Banana 2. Meanwhile, markets remained on edge over how much work and software artificial intelligence will replace.

Anthropic vs. the Pentagon

Amodei refuses to remove Claude’s safeguards

Anthropic CEO Dario Amodei published a statement on February 26 saying the company would not comply with a demand from the US Department of Defense (which the administration now calls the Department of War). The department wanted two restrictions on the Claude model removed: the ban on its use for mass surveillance of American citizens and for fully autonomous weapons.

Amodei wrote that threats would not change the company’s position and that it “cannot in good conscience” accede to the demand. The department threatened to remove Anthropic from its systems and designate it a supply chain risk. (Anthropic, Fortune)

Trump orders agencies to stop using Claude

On the afternoon of Friday, February 27, President Donald Trump ordered all federal agencies on Truth Social to end their work with Anthropic. This came more than an hour before the deadline the Pentagon had set for the company. Agencies already using Claude have a six-month transition period.

Secretary of Defense Pete Hegseth announced the same day that Anthropic would be designated a supply chain risk. Companies working for the military would then have to prove that they do not use Claude in their work. Treasury Secretary Scott Bessent announced that his department was dropping Anthropic’s products. (CBS News, Washington Technology)

Anthropic said on February 28 that it would challenge any such designation in court. Legal experts pointed out that the law requires the Pentagon to exhaust less drastic options first. (Bloomberg, Fortune)

OpenAI strikes a deal with the Pentagon as rival employees back Anthropic

Just hours after the ban, Sam Altman announced on February 28 a deal under which OpenAI will deploy its models on the Pentagon’s classified network. Altman says the agreement rests on the same principles: a ban on domestic mass surveillance and human responsibility for the use of force. The models will run only in the cloud, not on devices in the field, and OpenAI will add its own technical safeguards. (Engadget, Bloomberg)

Already on February 27, hundreds of employees at Google DeepMind and OpenAI signed an open letter titled “We Will Not Be Divided.” In it, they urge their companies’ leadership to support Anthropic’s stance. The number of signatures grew rapidly over the weekend. (TechCrunch)

New models

Google releases Nano Banana 2

Google unveiled Nano Banana 2, officially Gemini 3.1 Flash Image, on February 26. According to the company, the image generation and editing model offers the quality of the more expensive Nano Banana Pro at the speed of the original model.

The new model launched simultaneously in the Gemini app, in Google Search (AI Mode and Lens), in AI Studio, in the Gemini API and in Vertex AI. In the Gemini app, it replaced the Pro version across all modes. (GSMArena, Google)

Alibaba releases mid-sized Qwen3.5 models

Alibaba’s Qwen team released the Qwen3.5 Medium series on February 24: the 27B, 35B-A3B and 122B-A10B models. These are open-weight models, meaning their weights can be freely downloaded under the Apache 2.0 license, and they can even run on a more powerful home computer.

The “A3B” label refers to a mixture of experts architecture: the model has 35 billion parameters but uses only about 3 billion for each response. Alibaba claims its performance comes close to Claude Sonnet 4.5; these are the manufacturer’s figures. (VentureBeat)

Products and agents

Claude Cowork heads into the enterprise

Anthropic introduced enterprise features for its desktop agent Claude Cowork on February 24. An agent is a program that carries out multi-step tasks on its own. The update adds ten new plug-ins for investment banking, HR, private equity and design, along with integrations with services such as Slack, DocuSign and FactSet.

Paradoxically, software stocks rose after the announcement, following weeks of selling driven by fears that AI would replace them. (Technobezz, Ground News)

Perplexity Computer as a “digital worker”

Perplexity launched Perplexity Computer on February 25, a cloud-based system that independently carries entire projects from research to final output. It coordinates 19 different AI models, each with its own specialty, such as programming, research or image creation.

For now, it is available only to subscribers of the Max plan at 200 dollars per month. (Business Today)

Other news

  • February 25 Samsung unveiled the Galaxy S26 lineup. Alongside the Gemini and Bixby assistants, it now also offers the Perplexity assistant, activated with “Hey, Plex.” (Engadget)
  • February 25 Anthropic acquired Seattle startup Vercept, which developed an agent that operates a computer the way a human does. Its team is set to improve Claude’s ability to work with the screen. (Anthropic)
  • February 27 OpenAI announced that ChatGPT has 900 million weekly active users and roughly 50 million paying subscribers. In October 2025, the figure was 800 million. (TechCrunch)

Business and investment

OpenAI raises 110 billion dollars

OpenAI announced on February 27 the largest private funding round in tech history: 110 billion dollars. Amazon pledged 50 billion (part of it conditional), while Nvidia and SoftBank committed 30 billion each.

The company was valued at 730 billion dollars before the investment, putting it at roughly 840 billion afterward. The deal also includes computing capacity agreements: among other things, OpenAI is set to use Nvidia’s Vera Rubin systems and expand its partnership with AWS. (Tom’s Hardware, Built In)

Another record for Nvidia

Nvidia reported its fourth-quarter results for fiscal year 2026 on February 25. Revenue reached 68.1 billion dollars, up 73% year over year, beating analyst expectations. The data center segment brought in 62.3 billion dollars, accounting for over 90% of revenue.

For the full fiscal year, the company took in 215.9 billion dollars. It expects around 78 billion for the next quarter, excluding sales of computing chips to China. (Fortune)

Meta to buy 6 gigawatts of chips from AMD

Meta signed a multi-year agreement with AMD on February 24 to purchase AI chips with a total power draw of up to 6 gigawatts. The deal covers Instinct MI450 GPU accelerators and EPYC processors, with the first gigawatt set to be deployed starting in the second half of 2026.

In return, AMD granted Meta warrants for up to 160 million of its shares, similar to its earlier deal with OpenAI. Sources differ on the value of the deal, with estimates in the tens of billions of dollars. (Engadget, AP / NY1)

Block lays off nearly half its staff because of AI

Jack Dorsey’s payments company Block announced on February 26 that it would cut its workforce from more than 10,000 to just under 6,000. For the first time, Dorsey explicitly cited artificial intelligence as the reason, even though the company is growing and profitable.

Block’s shares then rose by about a fifth. Some commentators, however, doubt that AI is really the main reason. (The Irish Times)

Other news

  • February 23 A fictional scenario titled “The 2028 Global Intelligence Crisis” by Citrini Research, describing mass layoffs caused by AI, triggered a sell-off in software, payments and delivery stocks. The scenario was criticized and rejected by, among others, Citadel Securities. (Invezz, Bloomberg)
  • February 25 Salesforce said that annual recurring revenue from its Agentforce agent platform reached 800 million dollars, up 169% from a year earlier. Its shares nonetheless fell after the results. (SiliconANGLE)

Security and incidents

Anthropic accuses Chinese labs of “distilling” Claude

Anthropic accused Chinese companies DeepSeek, Moonshot AI and MiniMax on February 23 of conducting more than 16 million conversations with Claude through roughly 24,000 fake accounts. According to Anthropic, they used the responses for distillation, that is, training their own smaller models on the outputs of someone else’s model.

The largest share of conversations (over 13 million) was attributed to MiniMax. Anthropic is strengthening its defenses and sharing technical indicators with other labs and authorities; the figures come from its own investigation. (Tom’s Hardware, The Register)

A new version of Anthropic’s safety policy

Since February 24, Responsible Scaling Policy 3.0 has been in effect: a revised set of rules that Anthropic uses to decide on the safe development of more powerful models. It now distinguishes between commitments made by the company itself and recommendations for the industry as a whole.

The company will also publish a plan of safety goals (Frontier Safety Roadmap) and a risk report every three to six months, subject to external review. Critics point out that some of the rules are no longer binding. (Anthropic)

Research

  • February 24 A study by researchers at Mount Sinai hospital tested ChatGPT Health on 60 clinical scenarios. In more than half of the serious cases, it failed to send the user to the emergency room, recommending instead a doctor’s visit within 24 to 48 hours. OpenAI objects that the test does not reflect how the service is actually used. (eWeek)

Sources

Categories

Leave a Reply